release: prepare for 4.2.rc5

This reverts commit 71d0d58f8c. Repair-based
node operations stil have a significant regression (See #7249).

.gitmodules

@@ -1,6 +1,6 @@
 [submodule "seastar"]
 	path = seastar
-	url = ../seastar
+	url = ../scylla-seastar
 	ignore = dirty
 [submodule "swagger-ui"]
 	path = swagger-ui

SCYLLA-VERSION-GEN

@@ -1,7 +1,7 @@
 #!/bin/sh
 
 PRODUCT=scylla
-VERSION=666.development
+VERSION=4.2.rc5
 
 if test -f version
 then

alternator/auth.cc

@@ -129,7 +129,7 @@ future<std::string> get_key_from_roles(cql3::query_processor& qp, std::string us
             auth::meta::roles_table::qualified_name(), auth::meta::roles_table::role_col_name);
 
     auto cl = auth::password_authenticator::consistency_for_user(username);
-    auto timeout = auth::internal_distributed_timeout_config();
+    auto& timeout = auth::internal_distributed_timeout_config();
     return qp.execute_internal(query, cl, timeout, {sstring(username)}, true).then_wrapped([username = std::move(username)] (future<::shared_ptr<cql3::untyped_result_set>> f) {
         auto res = f.get0();
         auto salted_hash = std::optional<sstring>();

alternator/conditions.cc

@@ -98,6 +98,11 @@ struct nonempty : public size_check {
 
 // Check that array has the expected number of elements
 static void verify_operand_count(const rjson::value* array, const size_check& expected, const rjson::value& op) {
+    if (!array && expected(0)) {
+        // If expected() allows an empty AttributeValueList, it is also fine
+        // that it is missing.
+        return;
+    }
     if (!array || !array->IsArray()) {
         throw api_error("ValidationException", "With ComparisonOperator, AttributeValueList must be given and an array");
     }

alternator/executor.cc

@@ -626,11 +626,8 @@ void rmw_operation::set_default_write_isolation(std::string_view value) {
     default_write_isolation = parse_write_isolation(value);
 }
 
-// FIXME: Updating tags currently relies on updating schema, which may be subject
-// to races during concurrent updates of the same table. Once Scylla schema updates
-// are fixed, this issue will automatically get fixed as well.
 enum class update_tags_action { add_tags, delete_tags };
-static future<> update_tags(service::migration_manager& mm, const rjson::value& tags, schema_ptr schema, std::map<sstring, sstring>&& tags_map, update_tags_action action) {
+static void update_tags_map(const rjson::value& tags, std::map<sstring, sstring>& tags_map, update_tags_action action) {
     if (action == update_tags_action::add_tags) {
         for (auto it = tags.Begin(); it != tags.End(); ++it) {
             const rjson::value& key = (*it)["Key"];
@@ -652,28 +649,20 @@ static future<> update_tags(service::migration_manager& mm, const rjson::value&
     }
 
     if (tags_map.size() > 50) {
-        return make_exception_future<>(api_error("ValidationException", "Number of Tags exceed the current limit for the provided ResourceArn"));
+        throw api_error("ValidationException", "Number of Tags exceed the current limit for the provided ResourceArn");
     }
     validate_tags(tags_map);
+}
 
+// FIXME: Updating tags currently relies on updating schema, which may be subject
+// to races during concurrent updates of the same table. Once Scylla schema updates
+// are fixed, this issue will automatically get fixed as well.
+static future<> update_tags(service::migration_manager& mm, schema_ptr schema, std::map<sstring, sstring>&& tags_map) {
     schema_builder builder(schema);
     builder.set_extensions(schema::extensions_map{{sstring(tags_extension::NAME), ::make_shared<tags_extension>(std::move(tags_map))}});
     return mm.announce_column_family_update(builder.build(), false, std::vector<view_ptr>(), false);
 }
 
-static future<> add_tags(service::migration_manager& mm, service::storage_proxy& proxy, schema_ptr schema, rjson::value& request_info) {
-    const rjson::value* tags = rjson::find(request_info, "Tags");
-    if (!tags || !tags->IsArray()) {
-        return make_exception_future<>(api_error("ValidationException", format("Cannot parse tags")));
-    }
-    if (tags->Size() < 1) {
-        return make_exception_future<>(api_error("ValidationException", "The number of tags must be at least 1"));
-    }
-
-    std::map<sstring, sstring> tags_map = get_tags_of_table(schema);
-    return update_tags(mm, rjson::copy(*tags), schema, std::move(tags_map), update_tags_action::add_tags);
-}
-
 future<executor::request_return_type> executor::tag_resource(client_state& client_state, service_permit permit, rjson::value request) {
     _stats.api_operations.tag_resource++;
 
@@ -683,7 +672,16 @@ future<executor::request_return_type> executor::tag_resource(client_state& clien
             return api_error("AccessDeniedException", "Incorrect resource identifier");
         }
         schema_ptr schema = get_table_from_arn(_proxy, rjson::to_string_view(*arn));
-        add_tags(_mm, _proxy, schema, request).get();
+        std::map<sstring, sstring> tags_map = get_tags_of_table(schema);
+        const rjson::value* tags = rjson::find(request, "Tags");
+        if (!tags || !tags->IsArray()) {
+            return api_error("ValidationException", format("Cannot parse tags"));
+        }
+        if (tags->Size() < 1) {
+            return api_error("ValidationException", "The number of tags must be at least 1") ;
+        }
+        update_tags_map(*tags, tags_map,  update_tags_action::add_tags);
+        update_tags(_mm, schema, std::move(tags_map)).get();
         return json_string("");
     });
 }
@@ -704,7 +702,8 @@ future<executor::request_return_type> executor::untag_resource(client_state& cli
         schema_ptr schema = get_table_from_arn(_proxy, rjson::to_string_view(*arn));
 
         std::map<sstring, sstring> tags_map = get_tags_of_table(schema);
-        update_tags(_mm, *tags, schema, std::move(tags_map), update_tags_action::delete_tags).get();
+        update_tags_map(*tags, tags_map, update_tags_action::delete_tags);
+        update_tags(_mm, schema, std::move(tags_map)).get();
         return json_string("");
     });
 }
@@ -891,9 +890,22 @@ future<executor::request_return_type> executor::create_table(client_state& clien
             view_builders.emplace_back(std::move(view_builder));
         }
     }
-    if (rjson::find(request, "SSESpecification")) {
-        return make_ready_future<request_return_type>(api_error("ValidationException", "SSESpecification: configuring encryption-at-rest is not yet supported."));
+
+    // We don't yet support configuring server-side encryption (SSE) via the
+    // SSESpecifiction attribute, but an SSESpecification with Enabled=false
+    // is simply the default, and should be accepted:
+    rjson::value* sse_specification = rjson::find(request, "SSESpecification");
+    if (sse_specification && sse_specification->IsObject()) {
+        rjson::value* enabled = rjson::find(*sse_specification, "Enabled");
+        if (!enabled || !enabled->IsBool()) {
+            return make_ready_future<request_return_type>(api_error("ValidationException", "SSESpecification needs boolean Enabled"));
+        }
+        if (enabled->GetBool()) {
+            // TODO: full support for SSESpecification
+            return make_ready_future<request_return_type>(api_error("ValidationException", "SSESpecification: configuring encryption-at-rest is not yet supported."));
+        }
     }
+
     // We don't yet support streams (CDC), but a StreamSpecification asking
     // *not* to use streams should be accepted:
     rjson::value* stream_specification = rjson::find(request, "StreamSpecification");
@@ -908,6 +920,14 @@ future<executor::request_return_type> executor::create_table(client_state& clien
         }
     }
 
+    // Parse the "Tags" parameter early, so we can avoid creating the table
+    // at all if this parsing failed.
+    const rjson::value* tags = rjson::find(request, "Tags");
+    std::map<sstring, sstring> tags_map;
+    if (tags && tags->IsArray()) {
+        update_tags_map(*tags, tags_map, update_tags_action::add_tags);
+    }
+
     builder.set_extensions(schema::extensions_map{{sstring(tags_extension::NAME), ::make_shared<tags_extension>()}});
     schema_ptr schema = builder.build();
     auto where_clause_it = where_clauses.begin();
@@ -928,14 +948,14 @@ future<executor::request_return_type> executor::create_table(client_state& clien
 
     return create_keyspace(keyspace_name).handle_exception_type([] (exceptions::already_exists_exception&) {
             // Ignore the fact that the keyspace may already exist. See discussion in #6340
-        }).then([this, table_name, request = std::move(request), schema, view_builders = std::move(view_builders)] () mutable {
-        return futurize_invoke([&] { return _mm.announce_new_column_family(schema, false); }).then([this, table_info = std::move(request), schema, view_builders = std::move(view_builders)] () mutable {
+        }).then([this, table_name, request = std::move(request), schema, view_builders = std::move(view_builders), tags_map = std::move(tags_map)] () mutable {
+        return futurize_invoke([&] { return _mm.announce_new_column_family(schema, false); }).then([this, table_info = std::move(request), schema, view_builders = std::move(view_builders), tags_map = std::move(tags_map)] () mutable {
             return parallel_for_each(std::move(view_builders), [this, schema] (schema_builder builder) {
                 return _mm.announce_new_view(view_ptr(builder.build()));
-            }).then([this, table_info = std::move(table_info), schema] () mutable {
+            }).then([this, table_info = std::move(table_info), schema, tags_map = std::move(tags_map)] () mutable {
                 future<> f = make_ready_future<>();
-                if (rjson::find(table_info, "Tags")) {
-                    f = add_tags(_mm, _proxy, schema, table_info);
+                if (!tags_map.empty()) {
+                    f = update_tags(_mm, schema, std::move(tags_map));
                 }
                 return f.then([this] {
                     return wait_for_schema_agreement(_mm, db::timeout_clock::now() + 10s);
@@ -963,15 +983,24 @@ class attribute_collector {
     void add(bytes&& name, atomic_cell&& cell) {
         collected.emplace(std::move(name), std::move(cell));
     }
+    void add(const bytes& name, atomic_cell&& cell) {
+        collected.emplace(name, std::move(cell));
+    }
 public:
     attribute_collector() : collected(attrs_type()->get_keys_type()->as_less_comparator()) { }
-    void put(bytes&& name, bytes&& val, api::timestamp_type ts) {
-        add(std::move(name), atomic_cell::make_live(*bytes_type, ts, std::move(val), atomic_cell::collection_member::yes));
+    void put(bytes&& name, const bytes& val, api::timestamp_type ts) {
+        add(std::move(name), atomic_cell::make_live(*bytes_type, ts, val, atomic_cell::collection_member::yes));
 
     }
+    void put(const bytes& name, const bytes& val, api::timestamp_type ts) {
+        add(name, atomic_cell::make_live(*bytes_type, ts, val, atomic_cell::collection_member::yes));
+    }
     void del(bytes&& name, api::timestamp_type ts) {
         add(std::move(name), atomic_cell::make_dead(ts, gc_clock::now()));
     }
+    void del(const bytes& name, api::timestamp_type ts) {
+        add(name, atomic_cell::make_dead(ts, gc_clock::now()));
+    }
     collection_mutation_description to_mut() {
         collection_mutation_description ret;
         for (auto&& e : collected) {
@@ -1048,7 +1077,7 @@ public:
     put_or_delete_item(const rjson::value& item, schema_ptr schema, put_item);
     // put_or_delete_item doesn't keep a reference to schema (so it can be
     // moved between shards for LWT) so it needs to be given again to build():
-    mutation build(schema_ptr schema, api::timestamp_type ts);
+    mutation build(schema_ptr schema, api::timestamp_type ts) const;
     const partition_key& pk() const { return _pk; }
     const clustering_key& ck() const { return _ck; }
 };
@@ -1077,7 +1106,7 @@ put_or_delete_item::put_or_delete_item(const rjson::value& item, schema_ptr sche
     }
 }
 
-mutation put_or_delete_item::build(schema_ptr schema, api::timestamp_type ts) {
+mutation put_or_delete_item::build(schema_ptr schema, api::timestamp_type ts) const {
     mutation m(schema, _pk);
     // If there's no clustering key, a tombstone should be created directly
     // on a partition, not on a clustering row - otherwise it will look like
@@ -1099,7 +1128,7 @@ mutation put_or_delete_item::build(schema_ptr schema, api::timestamp_type ts) {
     for (auto& c : *_cells) {
         const column_definition* cdef = schema->get_column_definition(c.column_name);
         if (!cdef) {
-            attrs_collector.put(std::move(c.column_name), std::move(c.value), ts);
+            attrs_collector.put(c.column_name, c.value, ts);
         } else {
             row.cells().apply(*cdef, atomic_cell::make_live(*cdef->type, ts, std::move(c.value)));
         }
@@ -1410,7 +1439,7 @@ public:
                check_needs_read_before_write(_condition_expression) ||
                _returnvalues == returnvalues::ALL_OLD;
     }
-    virtual std::optional<mutation> apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) override {
+    virtual std::optional<mutation> apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) const override {
         if (!verify_expected(_request, previous_item.get()) ||
             !verify_condition_expression(_condition_expression, previous_item.get())) {
             // If the update is to be cancelled because of an unfulfilled Expected
@@ -1420,6 +1449,8 @@ public:
         }
         if (_returnvalues == returnvalues::ALL_OLD && previous_item) {
             _return_attributes = std::move(*previous_item);
+        } else {
+            _return_attributes = {};
         }
         return _mutation_builder.build(_schema, ts);
     }
@@ -1493,7 +1524,7 @@ public:
                 check_needs_read_before_write(_condition_expression) ||
                 _returnvalues == returnvalues::ALL_OLD;
     }
-    virtual std::optional<mutation> apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) override {
+    virtual std::optional<mutation> apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) const override {
         if (!verify_expected(_request, previous_item.get()) ||
             !verify_condition_expression(_condition_expression, previous_item.get())) {
             // If the update is to be cancelled because of an unfulfilled Expected
@@ -1503,6 +1534,8 @@ public:
         }
         if (_returnvalues == returnvalues::ALL_OLD && previous_item) {
             _return_attributes = std::move(*previous_item);
+        } else {
+            _return_attributes = {};
         }
         return _mutation_builder.build(_schema, ts);
     }
@@ -1577,7 +1610,7 @@ public:
     virtual ~put_or_delete_item_cas_request() = default;
     virtual std::optional<mutation> apply(foreign_ptr<lw_shared_ptr<query::result>> qr, const query::partition_slice& slice, api::timestamp_type ts) override {
         std::optional<mutation> ret;
-        for (put_or_delete_item& mutation_builder : _mutation_builders) {
+        for (const put_or_delete_item& mutation_builder : _mutation_builders) {
             // We assume all these builders have the same partition.
             if (ret) {
                 ret->apply(mutation_builder.build(schema, ts));
@@ -1906,7 +1939,7 @@ public:
 
     update_item_operation(service::storage_proxy& proxy, rjson::value&& request);
     virtual ~update_item_operation() = default;
-    virtual std::optional<mutation> apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) override;
+    virtual std::optional<mutation> apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) const override;
     bool needs_read_before_write() const;
 };
 
@@ -1984,7 +2017,7 @@ update_item_operation::needs_read_before_write() const {
 }
 
 std::optional<mutation>
-update_item_operation::apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) {
+update_item_operation::apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) const {
     if (!verify_expected(_request, previous_item.get()) ||
         !verify_condition_expression(_condition_expression, previous_item.get())) {
         // If the update is to be cancelled because of an unfulfilled

alternator/rmw_operation.hh

@@ -87,7 +87,11 @@ protected:
     // When _returnvalues != NONE, apply() should store here, in JSON form,
     // the values which are to be returned in the "Attributes" field.
     // The default null JSON means do not return an Attributes field at all.
-    rjson::value _return_attributes;
+    // This field is marked "mutable" so that the const apply() can modify
+    // it (see explanation below), but note that because apply() may be
+    // called more than once, if apply() will sometimes set this field it
+    // must set it (even if just to the default empty value) every time.
+    mutable rjson::value _return_attributes;
 public:
     // The constructor of a rmw_operation subclass should parse the request
     // and try to discover as many input errors as it can before really
@@ -100,7 +104,12 @@ public:
     // conditional expression, apply() should return an empty optional.
     // apply() may throw if it encounters input errors not discovered during
     // the constructor.
-    virtual std::optional<mutation> apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) = 0;
+    // apply() may be called more than once in case of contention, so it must
+    // not change the state saved in the object (issue #7218 was caused by
+    // violating this). We mark apply() "const" to let the compiler validate
+    // this for us. The output-only field _return_attributes is marked
+    // "mutable" above so that apply() can still write to it.
+    virtual std::optional<mutation> apply(std::unique_ptr<rjson::value> previous_item, api::timestamp_type ts) const = 0;
     // Convert the above apply() into the signature needed by cas_request:
     virtual std::optional<mutation> apply(foreign_ptr<lw_shared_ptr<query::result>> qr, const query::partition_slice& slice, api::timestamp_type ts) override;
     virtual ~rmw_operation() = default;

api/storage_service.cc

@@ -322,8 +322,8 @@ void set_storage_service(http_context& ctx, routes& r) {
                 for (auto cf : column_families) {
                     column_families_vec.push_back(&db.find_column_family(keyspace, cf));
                 }
-                return parallel_for_each(column_families_vec, [&cm] (column_family* cf) {
-                    return cm.perform_cleanup(cf);
+                return parallel_for_each(column_families_vec, [&cm, &db] (column_family* cf) {
+                    return cm.perform_cleanup(db, cf);
                 });
             }).then([]{
                 return make_ready_future<json::json_return_type>(0);

configure.py

@@ -386,6 +386,7 @@ scylla_tests = set([
     'test/boost/view_schema_ckey_test',
     'test/boost/vint_serialization_test',
     'test/boost/virtual_reader_test',
+    'test/boost/stall_free_test',
     'test/manual/ec2_snitch_test',
     'test/manual/gce_snitch_test',
     'test/manual/gossip',

cql3/functions/aggregate_fcts.cc

@@ -267,10 +267,13 @@ public:
     }
 };
 
-/// The same as `impl_max_function_for' but without knowledge of `Type'.
+/// The same as `impl_max_function_for' but without compile-time dependency on `Type'.
 class impl_max_dynamic_function final : public aggregate_function::aggregate {
+    data_type _io_type;
     opt_bytes _max;
 public:
+    impl_max_dynamic_function(data_type io_type) : _io_type(std::move(io_type)) {}
+
     virtual void reset() override {
         _max = {};
     }
@@ -278,12 +281,11 @@ public:
         return _max.value_or(bytes{});
     }
     virtual void add_input(cql_serialization_format sf, const std::vector<opt_bytes>& values) override {
-        if (!values[0]) {
+        if (values.empty() || !values[0]) {
             return;
         }
-        const auto val = *values[0];
-        if (!_max || *_max < val) {
-            _max = val;
+        if (!_max || _io_type->less(*_max, *values[0])) {
+            _max = values[0];
         }
     }
 };
@@ -298,10 +300,13 @@ public:
 };
 
 class max_dynamic_function final : public native_aggregate_function {
+    data_type _io_type;
 public:
-    max_dynamic_function(data_type io_type) : native_aggregate_function("max", io_type, { io_type }) {}
+    max_dynamic_function(data_type io_type)
+            : native_aggregate_function("max", io_type, { io_type })
+            , _io_type(std::move(io_type)) {}
     virtual std::unique_ptr<aggregate> new_aggregate() override {
-        return std::make_unique<impl_max_dynamic_function>();
+        return std::make_unique<impl_max_dynamic_function>(_io_type);
     }
 };
 
@@ -358,10 +363,13 @@ public:
     }
 };
 
-/// The same as `impl_min_function_for' but without knowledge of `Type'.
+/// The same as `impl_min_function_for' but without compile-time dependency on `Type'.
 class impl_min_dynamic_function final : public aggregate_function::aggregate {
+    data_type _io_type;
     opt_bytes _min;
 public:
+    impl_min_dynamic_function(data_type io_type) : _io_type(std::move(io_type)) {}
+
     virtual void reset() override {
         _min = {};
     }
@@ -369,12 +377,11 @@ public:
         return _min.value_or(bytes{});
     }
     virtual void add_input(cql_serialization_format sf, const std::vector<opt_bytes>& values) override {
-        if (!values[0]) {
+        if (values.empty() || !values[0]) {
             return;
         }
-        const auto val = *values[0];
-        if (!_min || val < *_min) {
-            _min = val;
+        if (!_min || _io_type->less(*values[0], *_min)) {
+            _min = values[0];
         }
     }
 };
@@ -389,10 +396,13 @@ public:
 };
 
 class min_dynamic_function final : public native_aggregate_function {
+    data_type _io_type;
 public:
-    min_dynamic_function(data_type io_type) : native_aggregate_function("min", io_type, { io_type }) {}
+    min_dynamic_function(data_type io_type)
+            : native_aggregate_function("min", io_type, { io_type })
+            , _io_type(std::move(io_type)) {}
     virtual std::unique_ptr<aggregate> new_aggregate() override {
-        return std::make_unique<impl_min_dynamic_function>();
+        return std::make_unique<impl_min_dynamic_function>(_io_type);
     }
 };
 

cql3/functions/castas_fcts.cc

@@ -88,16 +88,13 @@ static data_value castas_fctn_simple(data_value from) {
 template<typename ToType>
 static data_value castas_fctn_from_decimal_to_float(data_value from) {
     auto val_from = value_cast<big_decimal>(from);
-    boost::multiprecision::cpp_int ten(10);
-    boost::multiprecision::cpp_rational r = val_from.unscaled_value();
-    r /= boost::multiprecision::pow(ten, val_from.scale());
-    return static_cast<ToType>(r);
+    return static_cast<ToType>(val_from.as_rational());
 }
 
 static utils::multiprecision_int from_decimal_to_cppint(const data_value& from) {
     const auto& val_from = value_cast<big_decimal>(from);
-    boost::multiprecision::cpp_int ten(10);
-    return boost::multiprecision::cpp_int(val_from.unscaled_value() / boost::multiprecision::pow(ten, val_from.scale()));
+    auto r = val_from.as_rational();
+    return utils::multiprecision_int(numerator(r)/denominator(r));
 }
 
 template<typename ToType>

cql3/lists.cc

@@ -357,7 +357,12 @@ lists::setter_by_uuid::execute(mutation& m, const clustering_key_prefix& prefix,
 
     collection_mutation_description mut;
     mut.cells.reserve(1);
-    mut.cells.emplace_back(to_bytes(*index), params.make_cell(*ltype->value_comparator(), *value, atomic_cell::collection_member::yes));
+
+    if (!value) {
+        mut.cells.emplace_back(to_bytes(*index), params.make_dead_cell());
+    } else {
+        mut.cells.emplace_back(to_bytes(*index), params.make_cell(*ltype->value_comparator(), *value, atomic_cell::collection_member::yes));
+    }
 
     m.set_cell(prefix, column, mut.serialize(*ltype));
 }

cql3/restrictions/statement_restrictions.cc

@@ -417,7 +417,7 @@ std::vector<const column_definition*> statement_restrictions::get_column_defs_fo
                     _clustering_columns_restrictions->num_prefix_columns_that_need_not_be_filtered();
             for (auto&& cdef : _clustering_columns_restrictions->get_column_defs()) {
                 ::shared_ptr<single_column_restriction> restr;
-                if (single_pk_restrs) {
+                if (single_ck_restrs) {
                     auto it = single_ck_restrs->restrictions().find(cdef);
                     if (it != single_ck_restrs->restrictions().end()) {
                         restr = dynamic_pointer_cast<single_column_restriction>(it->second);
@@ -688,6 +688,11 @@ static query::range<bytes_view> to_range(const term_slice& slice, const query_op
         extract_bound(statements::bound::END));
 }
 
+static bool contains_without_wraparound(
+        const query::range<bytes_view>& range, bytes_view value, const serialized_tri_compare& cmp) {
+    return !range.is_wrap_around(cmp) && range.contains(value, cmp);
+}
+
 bool single_column_restriction::slice::is_satisfied_by(const schema& schema,
         const partition_key& key,
         const clustering_key_prefix& ckey,
@@ -702,13 +707,13 @@ bool single_column_restriction::slice::is_satisfied_by(const schema& schema,
         return false;
     }
     return cell_value->with_linearized([&] (bytes_view cell_value_bv) {
-        return to_range(_slice, options, _column_def.name_as_text()).contains(
+        return contains_without_wraparound(to_range(_slice, options, _column_def.name_as_text()),
                 cell_value_bv, _column_def.type->as_tri_comparator());
     });
 }
 
 bool single_column_restriction::slice::is_satisfied_by(bytes_view data, const query_options& options) const {
-    return to_range(_slice, options, _column_def.name_as_text()).contains(
+    return contains_without_wraparound(to_range(_slice, options, _column_def.name_as_text()),
             data, _column_def.type->underlying_type()->as_tri_comparator());
 }
 

cql3/statements/alter_table_statement.cc

@@ -207,6 +207,9 @@ void alter_table_statement::add_column(const schema& schema, const table& cf, sc
                 "because a collection with the same name and a different type has already been used in the past", column_name));
         }
     }
+    if (type->is_counter() && !schema.is_counter()) {
+        throw exceptions::configuration_exception(format("Cannot add a counter column ({}) in a non counter column family", column_name));
+    }
 
     cfm.with_column(column_name.name(), type, is_static ? column_kind::static_column : column_kind::regular_column);
 
@@ -222,7 +225,7 @@ void alter_table_statement::add_column(const schema& schema, const table& cf, sc
             schema_builder builder(view);
             if (view->view_info()->include_all_columns()) {
                 builder.with_column(column_name.name(), type);
-            } else if (view->view_info()->base_non_pk_columns_in_view_pk().empty()) {
+            } else if (!view->view_info()->has_base_non_pk_columns_in_view_pk()) {
                 db::view::create_virtual_column(builder, column_name.name(), type);
             }
             view_updates.push_back(view_ptr(builder.build()));

database.cc

@@ -1851,7 +1851,11 @@ future<> database::truncate(const keyspace& ks, column_family& cf, timestamp_fun
                             // TODO: indexes.
                             // Note: since discard_sstables was changed to only count tables owned by this shard,
                             // we can get zero rp back. Changed assert, and ensure we save at least low_mark.
-                            assert(low_mark <= rp || rp == db::replay_position());
+                            // #6995 - the assert below was broken in c2c6c71 and remained so for many years. 
+                            // We nowadays do not flush tables with sstables but autosnapshot=false. This means
+                            // the low_mark assertion does not hold, because we maybe/probably never got around to 
+                            // creating the sstables that would create them.
+                            assert(!should_flush || low_mark <= rp || rp == db::replay_position());
                             rp = std::max(low_mark, rp);
                             return truncate_views(cf, truncated_at, should_flush).then([&cf, truncated_at, rp] {
                                 // save_truncation_record() may actually fail after we cached the truncation time

database.hh

@@ -55,6 +55,7 @@
 #include <limits>
 #include <cstddef>
 #include "schema_fwd.hh"
+#include "db/view/view.hh"
 #include "db/schema_features.hh"
 #include "gms/feature.hh"
 #include "timestamp.hh"
@@ -901,7 +902,7 @@ public:
     lw_shared_ptr<const sstable_list> get_sstables_including_compacted_undeleted() const;
     const std::vector<sstables::shared_sstable>& compacted_undeleted_sstables() const;
     std::vector<sstables::shared_sstable> select_sstables(const dht::partition_range& range) const;
-    std::vector<sstables::shared_sstable> candidates_for_compaction() const;
+    std::vector<sstables::shared_sstable> non_staging_sstables() const;
     std::vector<sstables::shared_sstable> sstables_need_rewrite() const;
     size_t sstables_count() const;
     std::vector<uint64_t> sstable_count_per_level() const;
@@ -1008,8 +1009,9 @@ public:
         return *_config.sstables_manager;
     }
 
+    // Reader's schema must be the same as the base schema of each of the views.
     future<> populate_views(
-            std::vector<view_ptr>,
+            std::vector<db::view::view_and_base>,
             dht::token base_token,
             flat_mutation_reader&&,
             gc_clock::time_point);
@@ -1027,7 +1029,7 @@ private:
             tracing::trace_state_ptr tr_state, reader_concurrency_semaphore& sem, const io_priority_class& io_priority, query::partition_slice::option_set custom_opts) const;
     std::vector<view_ptr> affected_views(const schema_ptr& base, const mutation& update, gc_clock::time_point now) const;
     future<> generate_and_propagate_view_updates(const schema_ptr& base,
-            std::vector<view_ptr>&& views,
+            std::vector<db::view::view_and_base>&& views,
             mutation&& m,
             flat_mutation_reader_opt existings,
             tracing::trace_state_ptr tr_state,
@@ -1099,6 +1101,10 @@ flat_mutation_reader make_local_shard_sstable_reader(schema_ptr s,
         mutation_reader::forwarding fwd_mr,
         sstables::read_monitor_generator& monitor_generator = sstables::default_read_monitor_generator());
 
+/// Read a range from the passed-in sstables.
+///
+/// The reader is unrestricted, but will account its resource usage on the
+/// semaphore belonging to the passed-in permit.
 flat_mutation_reader make_range_sstable_reader(schema_ptr s,
         reader_permit permit,
         lw_shared_ptr<sstables::sstable_set> sstables,
@@ -1110,6 +1116,21 @@ flat_mutation_reader make_range_sstable_reader(schema_ptr s,
         mutation_reader::forwarding fwd_mr,
         sstables::read_monitor_generator& monitor_generator = sstables::default_read_monitor_generator());
 
+/// Read a range from the passed-in sstables.
+///
+/// The reader is restricted, that is it will wait for admission on the semaphore
+/// belonging to the passed-in permit, before starting to read.
+flat_mutation_reader make_restricted_range_sstable_reader(schema_ptr s,
+        reader_permit permit,
+        lw_shared_ptr<sstables::sstable_set> sstables,
+        const dht::partition_range& pr,
+        const query::partition_slice& slice,
+        const io_priority_class& pc,
+        tracing::trace_state_ptr trace_state,
+        streamed_mutation::forwarding fwd,
+        mutation_reader::forwarding fwd_mr,
+        sstables::read_monitor_generator& monitor_generator = sstables::default_read_monitor_generator());
+
 class user_types_metadata;
 
 class keyspace_metadata final {

db/batchlog_manager.cc

@@ -290,7 +290,7 @@ future<> db::batchlog_manager::replay_all_failed_batches() {
             mutation m(schema, key);
             auto now = service::client_state(service::client_state::internal_tag()).get_timestamp();
             m.partition().apply_delete(*schema, clustering_key_prefix::make_empty(), tombstone(now, gc_clock::now()));
-            return _qp.proxy().mutate_locally(m, tracing::trace_state_ptr());
+            return _qp.proxy().mutate_locally(m, tracing::trace_state_ptr(), db::commitlog::force_sync::no);
         });
     };
 

db/commitlog/commitlog.cc

@@ -521,7 +521,7 @@ public:
             _segment_manager->totals.total_size_on_disk -= size_on_disk();
             _segment_manager->totals.total_size -= (size_on_disk() + _buffer.size_bytes());
             _segment_manager->add_file_to_delete(_file_name, _desc);
-        } else {
+        } else if (_segment_manager->cfg.warn_about_segments_left_on_disk_after_shutdown) {
             clogger.warn("Segment {} is dirty and is left on disk.", *this);
         }
     }

db/commitlog/commitlog.hh

@@ -137,6 +137,7 @@ public:
 
         bool reuse_segments = true;
         bool use_o_dsync = false;
+        bool warn_about_segments_left_on_disk_after_shutdown = true;
 
         const db::extensions * extensions = nullptr;
     };

db/commitlog/commitlog_replayer.cc

@@ -304,7 +304,7 @@ future<> db::commitlog_replayer::impl::process(stats* s, commitlog::buffer_and_r
                 mutation m(cf.schema(), fm.decorated_key(*cf.schema()));
                 converting_mutation_partition_applier v(cm, *cf.schema(), m.partition());
                 fm.partition().accept(cm, v);
-                return do_with(std::move(m), [&db, &cf] (mutation m) {
+                return do_with(std::move(m), [&db, &cf] (const mutation& m) {
                     return db.apply_in_memory(m, cf, db::rp_handle(), db::no_timeout);
                 });
             } else {

db/config.cc

@@ -681,7 +681,7 @@ db::config::config(std::shared_ptr<db::extensions> exts)
     , replace_address(this, "replace_address", value_status::Used, "", "The listen_address or broadcast_address of the dead node to replace. Same as -Dcassandra.replace_address.")
     , replace_address_first_boot(this, "replace_address_first_boot", value_status::Used, "", "Like replace_address option, but if the node has been bootstrapped successfully it will be ignored. Same as -Dcassandra.replace_address_first_boot.")
     , override_decommission(this, "override_decommission", value_status::Used, false, "Set true to force a decommissioned node to join the cluster")
-    , enable_repair_based_node_ops(this, "enable_repair_based_node_ops", liveness::LiveUpdate, value_status::Used, true, "Set true to use enable repair based node operations instead of streaming based")
+    , enable_repair_based_node_ops(this, "enable_repair_based_node_ops", liveness::LiveUpdate, value_status::Used, false, "Set true to use enable repair based node operations instead of streaming based")
     , ring_delay_ms(this, "ring_delay_ms", value_status::Used, 30 * 1000, "Time a node waits to hear from other nodes before joining the ring in milliseconds. Same as -Dcassandra.ring_delay_ms in cassandra.")
     , shadow_round_ms(this, "shadow_round_ms", value_status::Used, 300 * 1000, "The maximum gossip shadow round time. Can be used to reduce the gossip feature check time during node boot up.")
     , fd_max_interval_ms(this, "fd_max_interval_ms", value_status::Used, 2 * 1000, "The maximum failure_detector interval time in milliseconds. Interval larger than the maximum will be ignored. Larger cluster may need to increase the default.")

db/hints/manager.cc

@@ -224,7 +224,9 @@ future<> manager::end_point_hints_manager::stop(drain should_drain) noexcept {
         with_lock(file_update_mutex(), [this] {
             if (_hints_store_anchor) {
                 hints_store_ptr tmp = std::exchange(_hints_store_anchor, nullptr);
-                return tmp->shutdown().finally([tmp] {});
+                return tmp->shutdown().finally([tmp] {
+                    return tmp->release();
+                }).finally([tmp] {});
             }
             return make_ready_future<>();
         }).handle_exception([&eptr] (auto e) { eptr = std::move(e); }).get();
@@ -290,7 +292,7 @@ inline bool manager::have_ep_manager(ep_key_type ep) const noexcept {
 }
 
 bool manager::store_hint(ep_key_type ep, schema_ptr s, lw_shared_ptr<const frozen_mutation> fm, tracing::trace_state_ptr tr_state) noexcept {
-    if (stopping() || !started() || !can_hint_for(ep)) {
+    if (stopping() || draining_all() || !started() || !can_hint_for(ep)) {
         manager_logger.trace("Can't store a hint to {}", ep);
         ++_stats.dropped;
         return false;
@@ -326,6 +328,10 @@ future<db::commitlog> manager::end_point_hints_manager::add_store() noexcept {
             // HH doesn't utilize the flow that benefits from reusing segments.
             // Therefore let's simply disable it to avoid any possible confusion.
             cfg.reuse_segments = false;
+            // HH leaves segments on disk after commitlog shutdown, and later reads
+            // them when commitlog is re-created. This is expected to happen regularly
+            // during standard HH workload, so no need to print a warning about it.
+            cfg.warn_about_segments_left_on_disk_after_shutdown = false;
 
             return commitlog::create_commitlog(std::move(cfg)).then([this] (commitlog l) {
                 // add_store() is triggered every time hint files are forcefully flushed to I/O (every hints_flush_period).
@@ -352,7 +358,9 @@ future<> manager::end_point_hints_manager::flush_current_hints() noexcept {
         return futurize_invoke([this] {
             return with_lock(file_update_mutex(), [this]() -> future<> {
                 return get_or_load().then([] (hints_store_ptr cptr) {
-                    return cptr->shutdown();
+                    return cptr->shutdown().finally([cptr] {
+                        return cptr->release();
+                    }).finally([cptr] {});
                 }).then([this] {
                     // Un-hold the commitlog object. Since we are under the exclusive _file_update_mutex lock there are no
                     // other hints_store_ptr copies and this would destroy the commitlog shared value.
@@ -529,7 +537,7 @@ bool manager::check_dc_for(ep_key_type ep) const noexcept {
 }
 
 void manager::drain_for(gms::inet_address endpoint) {
-    if (stopping()) {
+    if (stopping() || draining_all()) {
         return;
     }
 
@@ -540,6 +548,7 @@ void manager::drain_for(gms::inet_address endpoint) {
         return with_semaphore(drain_lock(), 1, [this, endpoint] {
             return futurize_invoke([this, endpoint] () {
                 if (utils::fb_utilities::is_me(endpoint)) {
+                    set_draining_all();
                     return parallel_for_each(_ep_managers, [] (auto& pair) {
                         return pair.second.stop(drain::yes).finally([&pair] {
                             return with_file_update_mutex(pair.second, [&pair] {

db/hints/manager.hh

@@ -424,12 +424,14 @@ public:
     enum class state {
         started,                // hinting is currently allowed (start() call is complete)
         replay_allowed,         // replaying (hints sending) is allowed
+        draining_all,           // hinting is not allowed - all ep managers are being stopped because this node is leaving the cluster
         stopping                // hinting is not allowed - stopping is in progress (stop() method has been called)
     };
 
     using state_set = enum_set<super_enum<state,
         state::started,
         state::replay_allowed,
+        state::draining_all,
         state::stopping>>;
 
 private:
@@ -690,6 +692,14 @@ private:
         return _state.contains(state::replay_allowed);
     }
 
+    void set_draining_all() noexcept {
+        _state.set(state::draining_all);
+    }
+
+    bool draining_all() noexcept {
+        return _state.contains(state::draining_all);
+    }
+
 public:
     ep_managers_map_type::iterator find_ep_manager(ep_key_type ep_key) noexcept {
         return _ep_managers.find(ep_key);

db/view/view.cc

@@ -58,6 +58,7 @@
 #include "cql3/util.hh"
 #include "db/view/view.hh"
 #include "db/view/view_builder.hh"
+#include "db/view/view_updating_consumer.hh"
 #include "db/system_keyspace_view_types.hh"
 #include "db/system_keyspace.hh"
 #include "frozen_mutation.hh"
@@ -136,17 +137,26 @@ const column_definition* view_info::view_column(const column_definition& base_de
     return _schema.get_column_definition(base_def.name());
 }
 
-const std::vector<column_id>& view_info::base_non_pk_columns_in_view_pk() const {
-    return _base_non_pk_columns_in_view_pk;
+void view_info::set_base_info(db::view::base_info_ptr base_info) {
+    _base_info = std::move(base_info);
 }
 
-void view_info::initialize_base_dependent_fields(const schema& base) {
+db::view::base_info_ptr view_info::make_base_dependent_view_info(const schema& base) const {
+    std::vector<column_id> base_non_pk_columns_in_view_pk;
     for (auto&& view_col : boost::range::join(_schema.partition_key_columns(), _schema.clustering_key_columns())) {
         auto* base_col = base.get_column_definition(view_col.name());
         if (base_col && !base_col->is_primary_key()) {
-            _base_non_pk_columns_in_view_pk.push_back(base_col->id);
+            base_non_pk_columns_in_view_pk.push_back(base_col->id);
         }
     }
+    return make_lw_shared<db::view::base_dependent_view_info>({
+        .base_schema = base.shared_from_this(),
+        .base_non_pk_columns_in_view_pk = std::move(base_non_pk_columns_in_view_pk)
+    });
+}
+
+bool view_info::has_base_non_pk_columns_in_view_pk() const {
+    return !_base_info->base_non_pk_columns_in_view_pk.empty();
 }
 
 namespace db {
@@ -194,12 +204,12 @@ bool may_be_affected_by(const schema& base, const view_info& view, const dht::de
 }
 
 static bool update_requires_read_before_write(const schema& base,
-        const std::vector<view_ptr>& views,
+        const std::vector<view_and_base>& views,
         const dht::decorated_key& key,
         const rows_entry& update,
         gc_clock::time_point now) {
     for (auto&& v : views) {
-        view_info& vf = *v->view_info();
+        view_info& vf = *v.view->view_info();
         if (may_be_affected_by(base, vf, key, update, now)) {
             return true;
         }
@@ -246,12 +256,14 @@ class view_updates final {
     view_ptr _view;
     const view_info& _view_info;
     schema_ptr _base;
+    base_info_ptr _base_info;
     std::unordered_map<partition_key, mutation_partition, partition_key::hashing, partition_key::equality> _updates;
 public:
-    explicit view_updates(view_ptr view, schema_ptr base)
-            : _view(std::move(view))
+    explicit view_updates(view_and_base vab)
+            : _view(std::move(vab.view))
             , _view_info(*_view->view_info())
-            , _base(std::move(base))
+            , _base(vab.base->base_schema)
+            , _base_info(vab.base)
             , _updates(8, partition_key::hashing(*_view), partition_key::equality(*_view)) {
     }
 
@@ -313,7 +325,7 @@ row_marker view_updates::compute_row_marker(const clustering_row& base_row) cons
     // they share liveness information. It's true especially in the only case currently allowed by CQL,
     // which assumes there's up to one non-pk column in the view key. It's also true in alternator,
     // which does not carry TTL information.
-    const auto& col_ids = _view_info.base_non_pk_columns_in_view_pk();
+    const auto& col_ids = _base_info->base_non_pk_columns_in_view_pk;
     if (!col_ids.empty()) {
         auto& def = _base->regular_column_at(col_ids[0]);
         // Note: multi-cell columns can't be part of the primary key.
@@ -544,7 +556,7 @@ void view_updates::delete_old_entry(const partition_key& base_key, const cluster
 
 void view_updates::do_delete_old_entry(const partition_key& base_key, const clustering_row& existing, const clustering_row& update, gc_clock::time_point now) {
     auto& r = get_view_row(base_key, existing);
-    const auto& col_ids = _view_info.base_non_pk_columns_in_view_pk();
+    const auto& col_ids = _base_info->base_non_pk_columns_in_view_pk;
     if (!col_ids.empty()) {
         // We delete the old row using a shadowable row tombstone, making sure that
         // the tombstone deletes everything in the row (or it might still show up).
@@ -685,7 +697,7 @@ void view_updates::generate_update(
         return;
     }
 
-    const auto& col_ids = _view_info.base_non_pk_columns_in_view_pk();
+    const auto& col_ids = _base_info->base_non_pk_columns_in_view_pk;
     if (col_ids.empty()) {
         // The view key is necessarily the same pre and post update.
         if (existing && existing->is_live(*_base)) {
@@ -940,12 +952,17 @@ future<stop_iteration> view_update_builder::on_results() {
 
 future<std::vector<frozen_mutation_and_schema>> generate_view_updates(
         const schema_ptr& base,
-        std::vector<view_ptr>&& views_to_update,
+        std::vector<view_and_base>&& views_to_update,
         flat_mutation_reader&& updates,
         flat_mutation_reader_opt&& existings,
         gc_clock::time_point now) {
-    auto vs = boost::copy_range<std::vector<view_updates>>(views_to_update | boost::adaptors::transformed([&] (auto&& v) {
-        return view_updates(std::move(v), base);
+    auto vs = boost::copy_range<std::vector<view_updates>>(views_to_update | boost::adaptors::transformed([&] (view_and_base v) {
+        if (base->version() != v.base->base_schema->version()) {
+            on_internal_error(vlogger, format("Schema version used for view updates ({}) does not match the current"
+                                              " base schema version of the view ({}) for view {}.{} of {}.{}",
+                base->version(), v.base->base_schema->version(), v.view->ks_name(), v.view->cf_name(), base->ks_name(), base->cf_name()));
+        }
+        return view_updates(std::move(v));
     }));
     auto builder = std::make_unique<view_update_builder>(base, std::move(vs), std::move(updates), std::move(existings), now);
     auto f = builder->build();
@@ -955,7 +972,7 @@ future<std::vector<frozen_mutation_and_schema>> generate_view_updates(
 query::clustering_row_ranges calculate_affected_clustering_ranges(const schema& base,
         const dht::decorated_key& key,
         const mutation_partition& mp,
-        const std::vector<view_ptr>& views,
+        const std::vector<view_and_base>& views,
         gc_clock::time_point now) {
     std::vector<nonwrapping_range<clustering_key_prefix_view>> row_ranges;
     std::vector<nonwrapping_range<clustering_key_prefix_view>> view_row_ranges;
@@ -963,11 +980,11 @@ query::clustering_row_ranges calculate_affected_clustering_ranges(const schema&
     if (mp.partition_tombstone() || !mp.row_tombstones().empty()) {
         for (auto&& v : views) {
             // FIXME: #2371
-            if (v->view_info()->select_statement().get_restrictions()->has_unrestricted_clustering_columns()) {
+            if (v.view->view_info()->select_statement().get_restrictions()->has_unrestricted_clustering_columns()) {
                 view_row_ranges.push_back(nonwrapping_range<clustering_key_prefix_view>::make_open_ended_both_sides());
                 break;
             }
-            for (auto&& r : v->view_info()->partition_slice().default_row_ranges()) {
+            for (auto&& r : v.view->view_info()->partition_slice().default_row_ranges()) {
                 view_row_ranges.push_back(r.transform(std::mem_fn(&clustering_key_prefix::view)));
             }
         }
@@ -1732,7 +1749,7 @@ public:
             return stop_iteration::yes;
         }
 
-        _fragments_memory_usage += cr.memory_usage(*_step.base->schema());
+        _fragments_memory_usage += cr.memory_usage(*_step.reader.schema());
         _fragments.push_back(std::move(cr));
         if (_fragments_memory_usage > batch_memory_max) {
             // Although we have not yet completed the batch of base rows that
@@ -1754,10 +1771,14 @@ public:
         _builder._as.check();
         if (!_fragments.empty()) {
             _fragments.push_front(partition_start(_step.current_key, tombstone()));
+            auto base_schema = _step.base->schema();
+            auto views = with_base_info_snapshot(_views_to_build);
+            auto reader = make_flat_mutation_reader_from_fragments(_step.reader.schema(), std::move(_fragments));
+            reader.upgrade_schema(base_schema);
             _step.base->populate_views(
-                    _views_to_build,
+                    std::move(views),
                     _step.current_token(),
-                    make_flat_mutation_reader_from_fragments(_step.base->schema(), std::move(_fragments)),
+                    std::move(reader),
                     _now).get();
             _fragments.clear();
             _fragments_memory_usage = 0;
@@ -1909,5 +1930,54 @@ future<bool> check_needs_view_update_path(db::system_distributed_keyspace& sys_d
     });
 }
 
+const size_t view_updating_consumer::buffer_size_soft_limit{1 * 1024 * 1024};
+const size_t view_updating_consumer::buffer_size_hard_limit{2 * 1024 * 1024};
+
+void view_updating_consumer::do_flush_buffer() {
+    _staging_reader_handle.pause();
+
+    if (_buffer.front().partition().empty()) {
+        // If we flushed mid-partition we can have an empty mutation if we
+        // flushed right before getting the end-of-partition fragment.
+        _buffer.pop_front();
+    }
+
+    while (!_buffer.empty()) {
+        try {
+            auto lock_holder = _view_update_pusher(std::move(_buffer.front())).get();
+        } catch (...) {
+            vlogger.warn("Failed to push replica updates for table {}.{}: {}", _schema->ks_name(), _schema->cf_name(), std::current_exception());
+        }
+        _buffer.pop_front();
+    }
+
+    _buffer_size = 0;
+    _m = nullptr;
+}
+
+void view_updating_consumer::maybe_flush_buffer_mid_partition() {
+    if (_buffer_size >= buffer_size_hard_limit) {
+        auto m = mutation(_schema, _m->decorated_key(), mutation_partition(_schema));
+        do_flush_buffer();
+        _buffer.emplace_back(std::move(m));
+        _m = &_buffer.back();
+    }
+}
+
+view_updating_consumer::view_updating_consumer(schema_ptr schema, table& table, std::vector<sstables::shared_sstable> excluded_sstables, const seastar::abort_source& as,
+        evictable_reader_handle& staging_reader_handle)
+    : view_updating_consumer(std::move(schema), as, staging_reader_handle,
+            [table = table.shared_from_this(), excluded_sstables = std::move(excluded_sstables)] (mutation m) mutable {
+        auto s = m.schema();
+        return table->stream_view_replica_updates(std::move(s), std::move(m), db::no_timeout, excluded_sstables);
+    })
+{ }
+
+std::vector<db::view::view_and_base> with_base_info_snapshot(std::vector<view_ptr> vs) {
+    return boost::copy_range<std::vector<db::view::view_and_base>>(vs | boost::adaptors::transformed([] (const view_ptr& v) {
+        return db::view::view_and_base{v, v->view_info()->base_info()};
+    }));
+}
+
 } // namespace view
 } // namespace db

db/view/view.hh

@@ -43,6 +43,27 @@ namespace db {
 
 namespace view {
 
+// Part of the view description which depends on the base schema version.
+//
+// This structure may change even though the view schema doesn't change, so
+// it needs to live outside view_ptr.
+struct base_dependent_view_info {
+    schema_ptr base_schema;
+
+    // Id of a regular base table column included in the view's PK, if any.
+    // Scylla views only allow one such column, alternator can have up to two.
+    std::vector<column_id> base_non_pk_columns_in_view_pk;
+};
+
+// Immutable snapshot of view's base-schema-dependent part.
+using base_info_ptr = lw_shared_ptr<const base_dependent_view_info>;
+
+// Snapshot of the view schema and its base-schema-dependent part.
+struct view_and_base {
+    view_ptr view;
+    base_info_ptr base;
+};
+
 /**
  * Whether the view filter considers the specified partition key.
  *
@@ -94,7 +115,7 @@ bool clustering_prefix_matches(const schema& base, const partition_key& key, con
 
 future<std::vector<frozen_mutation_and_schema>> generate_view_updates(
         const schema_ptr& base,
-        std::vector<view_ptr>&& views_to_update,
+        std::vector<view_and_base>&& views_to_update,
         flat_mutation_reader&& updates,
         flat_mutation_reader_opt&& existings,
         gc_clock::time_point now);
@@ -103,7 +124,7 @@ query::clustering_row_ranges calculate_affected_clustering_ranges(
         const schema& base,
         const dht::decorated_key& key,
         const mutation_partition& mp,
-        const std::vector<view_ptr>& views,
+        const std::vector<view_and_base>& views,
         gc_clock::time_point now);
 
 struct wait_for_all_updates_tag {};
@@ -133,6 +154,13 @@ future<> mutate_MV(
  */
  void create_virtual_column(schema_builder& builder, const bytes& name, const data_type& type);
 
+/**
+ * Converts a collection of view schema snapshots into a collection of
+ * view_and_base objects, which are snapshots of both the view schema
+ * and the base-schema-dependent part of view description.
+ */
+std::vector<view_and_base> with_base_info_snapshot(std::vector<view_ptr>);
+
 }
 
 }

db/view/view_update_generator.cc

@@ -42,35 +42,52 @@ future<> view_update_generator::start() {
                 _pending_sstables.wait().get();
             }
 
+            // To ensure we don't race with updates, move the entire content
+            // into a local variable.
+            auto sstables_with_tables = std::exchange(_sstables_with_tables, {});
+
             // If we got here, we will process all tables we know about so far eventually so there
             // is no starvation
-            for (auto& t : _sstables_with_tables | boost::adaptors::map_keys) {
+            for (auto table_it = sstables_with_tables.begin(); table_it != sstables_with_tables.end(); table_it = sstables_with_tables.erase(table_it)) {
+                auto& [t, sstables] = *table_it;
                 schema_ptr s = t->schema();
 
-                // Copy what we have so far so we don't miss new updates
-                auto sstables = std::exchange(_sstables_with_tables[t], {});
+                vug_logger.trace("Processing {}.{}: {} sstables", s->ks_name(), s->cf_name(), sstables.size());
+
+                const auto num_sstables = sstables.size();
 
                 try {
-                    // temporary: need an sstable set for the flat mutation reader, but the
-                    // compaction_descriptor takes a vector. Soon this will become a compaction
-                    // so the transformation to the SSTable set will not be needed.
-                    auto ssts = make_lw_shared(t->get_compaction_strategy().make_sstable_set(s));
+                    // Exploit the fact that sstables in the staging directory
+                    // are usually non-overlapping and use a partitioned set for
+                    // the read.
+                    auto ssts = make_lw_shared(sstables::make_partitioned_sstable_set(s, make_lw_shared<sstable_list>(sstable_list{}), false));
                     for (auto& sst : sstables) {
                         ssts->insert(sst);
                     }
 
-                    flat_mutation_reader staging_sstable_reader = ::make_range_sstable_reader(s,
+                    auto ms = mutation_source([this, ssts] (
+                                schema_ptr s,
+                                reader_permit permit,
+                                const dht::partition_range& pr,
+                                const query::partition_slice& ps,
+                                const io_priority_class& pc,
+                                tracing::trace_state_ptr ts,
+                                streamed_mutation::forwarding fwd_ms,
+                                mutation_reader::forwarding fwd_mr) {
+                        return ::make_restricted_range_sstable_reader(s, std::move(permit), std::move(ssts), pr, ps, pc, std::move(ts), fwd_ms, fwd_mr);
+                    });
+                    auto [staging_sstable_reader, staging_sstable_reader_handle] = make_manually_paused_evictable_reader(
+                            std::move(ms),
+                            s,
                             _db.make_query_class_config().semaphore.make_permit(),
-                            std::move(ssts),
                             query::full_partition_range,
                             s->full_slice(),
                             service::get_local_streaming_priority(),
                             nullptr,
-                            ::streamed_mutation::forwarding::no,
                             ::mutation_reader::forwarding::no);
 
                     inject_failure("view_update_generator_consume_staging_sstable");
-                    auto result = staging_sstable_reader.consume_in_thread(view_updating_consumer(s, *t, sstables, _as), db::no_timeout);
+                    auto result = staging_sstable_reader.consume_in_thread(view_updating_consumer(s, *t, sstables, _as, staging_sstable_reader_handle), db::no_timeout);
                     if (result == stop_iteration::yes) {
                         break;
                     }
@@ -89,7 +106,7 @@ future<> view_update_generator::start() {
                     // Move from staging will be retried upon restart.
                     vug_logger.warn("Moving {} from staging failed: {}:{}. Ignoring...", s->ks_name(), s->cf_name(), std::current_exception());
                 }
-                _registration_sem.signal();
+                _registration_sem.signal(num_sstables);
             }
             // For each table, move the processed staging sstables into the table's base dir.
             for (auto it = _sstables_to_move.begin(); it != _sstables_to_move.end(); ) {

db/view/view_update_generator.hh

@@ -32,7 +32,10 @@
 namespace db::view {
 
 class view_update_generator {
+public:
     static constexpr size_t registration_queue_size = 5;
+
+private:
     database& _db;
     seastar::abort_source _as;
     future<> _started = make_ready_future<>();
@@ -51,6 +54,8 @@ public:
     future<> start();
     future<> stop();
     future<> register_staging_sstable(sstables::shared_sstable sst, lw_shared_ptr<table> table);
+
+    ssize_t available_register_units() const { return _registration_sem.available_units(); }
 private:
     bool should_throttle() const;
 };

db/view/view_updating_consumer.hh

@@ -27,6 +27,8 @@
 #include "sstables/shared_sstable.hh"
 #include "database.hh"
 
+class evictable_reader_handle;
+
 namespace db::view {
 
 /*
@@ -34,22 +36,46 @@ namespace db::view {
  * It is expected to be run in seastar::async threaded context through consume_in_thread()
  */
 class view_updating_consumer {
-    schema_ptr _schema;
-    lw_shared_ptr<table> _table;
-    std::vector<sstables::shared_sstable> _excluded_sstables;
-    const seastar::abort_source* _as;
-    std::optional<mutation> _m;
 public:
-    view_updating_consumer(schema_ptr schema, table& table, std::vector<sstables::shared_sstable> excluded_sstables, const seastar::abort_source& as)
+    // We prefer flushing on partition boundaries, so at the end of a partition,
+    // we flush on reaching the soft limit. Otherwise we continue accumulating
+    // data. We flush mid-partition if we reach the hard limit.
+    static const size_t buffer_size_soft_limit;
+    static const size_t buffer_size_hard_limit;
+
+private:
+    schema_ptr _schema;
+    const seastar::abort_source* _as;
+    evictable_reader_handle& _staging_reader_handle;
+    circular_buffer<mutation> _buffer;
+    mutation* _m{nullptr};
+    size_t _buffer_size{0};
+    noncopyable_function<future<row_locker::lock_holder>(mutation)> _view_update_pusher;
+
+private:
+    void do_flush_buffer();
+    void maybe_flush_buffer_mid_partition();
+
+public:
+    // Push updates with a custom pusher. Mainly for tests.
+    view_updating_consumer(schema_ptr schema, const seastar::abort_source& as, evictable_reader_handle& staging_reader_handle,
+            noncopyable_function<future<row_locker::lock_holder>(mutation)> view_update_pusher)
             : _schema(std::move(schema))
-            , _table(table.shared_from_this())
-            , _excluded_sstables(std::move(excluded_sstables))
             , _as(&as)
-            , _m()
+            , _staging_reader_handle(staging_reader_handle)
+            , _view_update_pusher(std::move(view_update_pusher))
     { }
 
+    view_updating_consumer(schema_ptr schema, table& table, std::vector<sstables::shared_sstable> excluded_sstables, const seastar::abort_source& as,
+            evictable_reader_handle& staging_reader_handle);
+
+    view_updating_consumer(view_updating_consumer&&) = default;
+
+    view_updating_consumer& operator=(view_updating_consumer&&) = delete;
+
     void consume_new_partition(const dht::decorated_key& dk) {
-        _m = mutation(_schema, dk, mutation_partition(_schema));
+        _buffer.emplace_back(_schema, dk, mutation_partition(_schema));
+        _m = &_buffer.back();
     }
 
     void consume(tombstone t) {
@@ -60,7 +86,9 @@ public:
         if (_as->abort_requested()) {
             return stop_iteration::yes;
         }
+        _buffer_size += sr.memory_usage(*_schema);
         _m->partition().apply(*_schema, std::move(sr));
+        maybe_flush_buffer_mid_partition();
         return stop_iteration::no;
     }
 
@@ -68,7 +96,9 @@ public:
         if (_as->abort_requested()) {
             return stop_iteration::yes;
         }
+        _buffer_size += cr.memory_usage(*_schema);
         _m->partition().apply(*_schema, std::move(cr));
+        maybe_flush_buffer_mid_partition();
         return stop_iteration::no;
     }
 
@@ -76,14 +106,27 @@ public:
         if (_as->abort_requested()) {
             return stop_iteration::yes;
         }
+        _buffer_size += rt.memory_usage(*_schema);
         _m->partition().apply(*_schema, std::move(rt));
+        maybe_flush_buffer_mid_partition();
         return stop_iteration::no;
     }
 
     // Expected to be run in seastar::async threaded context (consume_in_thread())
-    stop_iteration consume_end_of_partition();
+    stop_iteration consume_end_of_partition() {
+        if (_as->abort_requested()) {
+            return stop_iteration::yes;
+        }
+        if (_buffer_size >= buffer_size_soft_limit) {
+            do_flush_buffer();
+        }
+        return stop_iteration::no;
+    }
 
     stop_iteration consume_end_of_stream() {
+        if (!_buffer.empty()) {
+            do_flush_buffer();
+        }
         return stop_iteration(_as->abort_requested());
     }
 };

dht/boot_strapper.cc

@@ -59,7 +59,12 @@ future<> boot_strapper::bootstrap(streaming::stream_reason reason) {
         return make_exception_future<>(std::runtime_error("Wrong stream_reason provided: it can only be replace or bootstrap"));
     }
     auto streamer = make_lw_shared<range_streamer>(_db, _token_metadata, _abort_source, _tokens, _address, description, reason);
-    streamer->add_source_filter(std::make_unique<range_streamer::failure_detector_source_filter>(gms::get_local_gossiper().get_unreachable_members()));
+    auto nodes_to_filter = gms::get_local_gossiper().get_unreachable_members();
+    if (reason == streaming::stream_reason::replace && _db.local().get_replace_address()) {
+        nodes_to_filter.insert(_db.local().get_replace_address().value());
+    }
+    blogger.debug("nodes_to_filter={}", nodes_to_filter);
+    streamer->add_source_filter(std::make_unique<range_streamer::failure_detector_source_filter>(nodes_to_filter));
     auto keyspaces = make_lw_shared<std::vector<sstring>>(_db.local().get_non_system_keyspaces());
     return do_for_each(*keyspaces, [this, keyspaces, streamer] (sstring& keyspace_name) {
         auto& ks = _db.local().find_keyspace(keyspace_name);

dist/common/scripts/scylla-housekeeping

@@ -61,7 +61,15 @@ def sh_command(*args):
     return out
 
 def get_url(path):
-    return urllib.request.urlopen(path).read().decode('utf-8')
+    # If server returns any error, like 403, or 500 urllib.request throws exception, which is not serializable.
+    # When multiprocessing routines fail to serialize it, it throws ambiguous serialization exception
+    #   from get_json_from_url.
+    # In order to see legit error we catch it from the inside of process, covert to string and
+    #   pass it as part of return value
+    try:
+        return 0, urllib.request.urlopen(path).read().decode('utf-8')
+    except Exception as exc:
+        return 1, str(exc)
 
 def get_json_from_url(path):
     pool = mp.Pool(processes=1)
@@ -71,13 +79,16 @@ def get_json_from_url(path):
     # to enforce a wallclock timeout.
     result = pool.apply_async(get_url, args=(path,))
     try:
-        retval = result.get(timeout=5)
+        status, retval = result.get(timeout=5)
     except mp.TimeoutError as err:
         pool.terminate()
         pool.join()
         raise
+    if status == 1:
+        raise RuntimeError(f'Failed to get "{path}" due to the following error: {retval}')
     return json.loads(retval)
 
+
 def get_api(path):
     return get_json_from_url("http://" + api_address + path)
 

dist/common/scripts/scylla_setup

@@ -27,6 +27,7 @@ import glob
 import shutil
 import io
 import stat
+import distro
 from scylla_util import *
 
 interactive = False
@@ -385,6 +386,9 @@ if __name__ == '__main__':
                 if not stat.S_ISBLK(os.stat(dsk).st_mode):
                     print('{} is not block device'.format(dsk))
                     continue
+                if dsk in selected:
+                    print(f'{dsk} is already added')
+                    continue
                 selected.append(dsk)
                 devices.remove(dsk)
             disks = ','.join(selected)
@@ -468,5 +472,10 @@ if __name__ == '__main__':
             print('Please restart your machine before using ScyllaDB, as you have disabled')
             print(' SELinux.')
 
-        if dist_name() == 'Ubuntu':
-            run('apt-get install -y hugepages')
+        if distro.id() == 'ubuntu':
+            # Ubuntu version is 20.04 or later
+            if int(distro.major_version()) >= 20:
+                hugepkg = 'libhugetlbfs-bin'
+            else:
+                hugepkg = 'hugepages'
+            run(f'apt-get install -y {hugepkg}')

dist/common/scripts/scylla_swap_setup

@@ -40,6 +40,10 @@ if __name__ == '__main__':
         sys.exit(1)
 
     memtotal = get_memtotal_gb()
+    if memtotal == 0:
+        print('memory too small: {} KB'.format(get_memtotal()))
+        sys.exit(1)
+
     # Scylla document says 'swap size should be set to either total_mem/3 or
     # 16GB - lower of the two', so we need to compare 16g vs memtotal/3 and
     # choose lower one

dist/common/scripts/scylla_util.py

@@ -184,7 +184,7 @@ class aws_instance:
         instance_size = self.instance_size()
         if instance_class in ['c3', 'c4', 'd2', 'i2', 'r3']:
             return 'ixgbevf'
-        if instance_class in ['a1', 'c5', 'c5d', 'f1', 'g3', 'g4', 'h1', 'i3', 'i3en', 'inf1', 'm5', 'm5a', 'm5ad', 'm5d', 'm5dn', 'm5n', 'm6g', 'p2', 'p3', 'r4', 'r5', 'r5a', 'r5ad', 'r5d', 'r5dn', 'r5n', 't3', 't3a', 'u-6tb1', 'u-9tb1', 'u-12tb1', 'u-18tn1', 'u-24tb1', 'x1', 'x1e', 'z1d']:
+        if instance_class in ['a1', 'c5', 'c5a', 'c5d', 'c5n', 'c6g', 'c6gd', 'f1', 'g3', 'g4', 'h1', 'i3', 'i3en', 'inf1', 'm5', 'm5a', 'm5ad', 'm5d', 'm5dn', 'm5n', 'm6g', 'm6gd', 'p2', 'p3', 'r4', 'r5', 'r5a', 'r5ad', 'r5d', 'r5dn', 'r5n', 't3', 't3a', 'u-6tb1', 'u-9tb1', 'u-12tb1', 'u-18tn1', 'u-24tb1', 'x1', 'x1e', 'z1d']:
             return 'ena'
         if instance_class == 'm4':
             if instance_size == '16xlarge':
@@ -331,7 +331,7 @@ class scylla_cpuinfo:
 
 # When a CLI tool is not installed, use relocatable CLI tool provided by Scylla
 scylla_env = os.environ.copy()
-scylla_env['PATH'] =  '{}:{}'.format(scylla_env['PATH'], scyllabindir())
+scylla_env['PATH'] =  '{}:{}'.format(scyllabindir(), scylla_env['PATH'])
 
 def run(cmd, shell=False, silent=False, exception=True):
     stdout = subprocess.DEVNULL if silent else None
@@ -446,6 +446,19 @@ def dist_ver():
     return distro.version()
 
 
+SYSTEM_PARTITION_UUIDS = [
+        '21686148-6449-6e6f-744e-656564454649', # BIOS boot partition
+        'c12a7328-f81f-11d2-ba4b-00a0c93ec93b', # EFI system partition
+        '024dee41-33e7-11d3-9d69-0008c781f39f'  # MBR partition scheme
+]
+
+def get_partition_uuid(dev):
+    return out(f'lsblk -n -oPARTTYPE {dev}')
+
+def is_system_partition(dev):
+    uuid = get_partition_uuid(dev)
+    return (uuid in SYSTEM_PARTITION_UUIDS)
+
 def is_unused_disk(dev):
     # dev is not in /sys/class/block/, like /dev/nvme[0-9]+
     if not os.path.isdir('/sys/class/block/{dev}'.format(dev=dev.replace('/dev/', ''))):
@@ -453,7 +466,8 @@ def is_unused_disk(dev):
     try:
         fd = os.open(dev, os.O_EXCL)
         os.close(fd)
-        return True
+        # dev is not reserved for system
+        return not is_system_partition(dev)
     except OSError:
         return False
 

dist/debian/debian/rules

@@ -39,6 +39,7 @@ override_dh_strip:
 	# The binaries (ethtool...patchelf) don't pass dh_strip after going through patchelf. Since they are
 	# already stripped, nothing is lost if we exclude them, so that's what we do.
 	dh_strip -Xlibprotobuf.so.15 -Xld.so -Xethtool -Xgawk -Xgzip -Xhwloc-calc -Xhwloc-distrib -Xifconfig -Xlscpu -Xnetstat -Xpatchelf --dbg-package=$(product)-server-dbg
+	find $(CURDIR)/debian/$(product)-server-dbg/usr/lib/debug/.build-id/ -name "*.debug" -exec objcopy --decompress-debug-sections {} \;
 
 override_dh_makeshlibs:
 

dist/docker/redhat/Dockerfile

@@ -5,8 +5,8 @@ MAINTAINER Avi Kivity <avi@cloudius-systems.com>
 ENV container docker
 
 # The SCYLLA_REPO_URL argument specifies the URL to the RPM repository this Docker image uses to install Scylla. The default value is the Scylla's unstable RPM repository, which contains the daily build.
-ARG SCYLLA_REPO_URL=http://downloads.scylladb.com/rpm/unstable/centos/master/latest/scylla.repo
-ARG VERSION=666.development
+ARG SCYLLA_REPO_URL=http://downloads.scylladb.com/rpm/unstable/centos/scylla-4.2/latest/scylla.repo
+ARG VERSION=4.2
 
 ADD scylla_bashrc /scylla_bashrc
 

flat_mutation_reader.hh

@@ -468,6 +468,9 @@ public:
     size_t buffer_size() const {
         return _impl->buffer_size();
     }
+    const circular_buffer<mutation_fragment>& buffer() const {
+        return _impl->buffer();
+    }
     // Detach the internal buffer of the reader.
     // Roughly equivalent to depleting it by calling pop_mutation_fragment()
     // until is_buffer_empty() returns true.

gms/gossiper.cc

@@ -428,6 +428,7 @@ future<> gossiper::handle_shutdown_msg(inet_address from) {
         return make_ready_future<>();
     }
     return seastar::async([this, from] {
+        auto permit = this->lock_endpoint(from).get0();
         this->mark_as_shutdown(from);
     });
 }

install-dependencies.sh

@@ -98,6 +98,7 @@ fedora_packages=(
     debhelper
     fakeroot
     file
+    dpkg-dev
 )
 
 centos_packages=(

locator/abstract_replication_strategy.cc

@@ -168,15 +168,33 @@ insert_token_range_to_sorted_container_while_unwrapping(
 
 dht::token_range_vector
 abstract_replication_strategy::get_ranges(inet_address ep) const {
-    return get_ranges(ep, _token_metadata);
+    return do_get_ranges(ep, _token_metadata, false);
+}
+
+dht::token_range_vector
+abstract_replication_strategy::get_ranges_in_thread(inet_address ep) const {
+    return do_get_ranges(ep, _token_metadata, true);
 }
 
 dht::token_range_vector
 abstract_replication_strategy::get_ranges(inet_address ep, token_metadata& tm) const {
+    return do_get_ranges(ep, tm, false);
+}
+
+dht::token_range_vector
+abstract_replication_strategy::get_ranges_in_thread(inet_address ep, token_metadata& tm) const {
+    return do_get_ranges(ep, tm, true);
+}
+
+dht::token_range_vector
+abstract_replication_strategy::do_get_ranges(inet_address ep, token_metadata& tm, bool can_yield) const {
     dht::token_range_vector ret;
     auto prev_tok = tm.sorted_tokens().back();
     for (auto tok : tm.sorted_tokens()) {
         for (inet_address a : calculate_natural_endpoints(tok, tm)) {
+            if (can_yield) {
+                seastar::thread::maybe_yield();
+            }
             if (a == ep) {
                 insert_token_range_to_sorted_container_while_unwrapping(prev_tok, tok, ret);
                 break;

locator/abstract_replication_strategy.hh

@@ -113,10 +113,15 @@ public:
     // It the analogue of Origin's getAddressRanges().get(endpoint).
     // This function is not efficient, and not meant for the fast path.
     dht::token_range_vector get_ranges(inet_address ep) const;
+    dht::token_range_vector get_ranges_in_thread(inet_address ep) const;
 
     // Use the token_metadata provided by the caller instead of _token_metadata
     dht::token_range_vector get_ranges(inet_address ep, token_metadata& tm) const;
+    dht::token_range_vector get_ranges_in_thread(inet_address ep, token_metadata& tm) const;
+private:
+    dht::token_range_vector do_get_ranges(inet_address ep, token_metadata& tm, bool can_yield) const;
 
+public:
     // get_primary_ranges() returns the list of "primary ranges" for the given
     // endpoint. "Primary ranges" are the ranges that the node is responsible
     // for storing replica primarily, which means this is the first node

lua.cc

@@ -262,14 +262,12 @@ static auto visit_lua_raw_value(lua_State* l, int index, Func&& f) {
 
 template <typename Func>
 static auto visit_decimal(const big_decimal &v, Func&& f) {
-    boost::multiprecision::cpp_int ten(10);
-    const auto& dividend = v.unscaled_value();
-    auto divisor = boost::multiprecision::pow(ten, v.scale());
+    boost::multiprecision::cpp_rational r = v.as_rational();
+    const boost::multiprecision::cpp_int& dividend = numerator(r);
+    const boost::multiprecision::cpp_int& divisor = denominator(r);
     if (dividend % divisor == 0) {
-        return f(utils::multiprecision_int(boost::multiprecision::cpp_int(dividend/divisor)));
+        return f(utils::multiprecision_int(dividend/divisor));
     }
-    boost::multiprecision::cpp_rational r = dividend;
-    r /= divisor;
     return f(r.convert_to<double>());
 }
 

main.cc

@@ -830,6 +830,7 @@ int main(int ac, char** av) {
             storage_proxy_smp_service_group_config.max_nonlocal_requests = 5000;
             spcfg.read_smp_service_group = create_smp_service_group(storage_proxy_smp_service_group_config).get0();
             spcfg.write_smp_service_group = create_smp_service_group(storage_proxy_smp_service_group_config).get0();
+            spcfg.hints_write_smp_service_group = create_smp_service_group(storage_proxy_smp_service_group_config).get0();
             spcfg.write_ack_smp_service_group = create_smp_service_group(storage_proxy_smp_service_group_config).get0();
             static db::view::node_update_backlog node_backlog(smp::count, 10ms);
             scheduling_group_key_config storage_proxy_stats_cfg =
@@ -967,12 +968,16 @@ int main(int ac, char** av) {
                 mm.init_messaging_service();
             }).get();
             supervisor::notify("initializing storage proxy RPC verbs");
-            proxy.invoke_on_all([] (service::storage_proxy& p) {
-                p.init_messaging_service();
-            }).get();
+            proxy.invoke_on_all(&service::storage_proxy::init_messaging_service).get();
+            auto stop_proxy_handlers = defer_verbose_shutdown("storage proxy RPC verbs", [&proxy] {
+                proxy.invoke_on_all(&service::storage_proxy::uninit_messaging_service).get();
+            });
 
             supervisor::notify("starting streaming service");
             streaming::stream_session::init_streaming_service(db, sys_dist_ks, view_update_generator).get();
+            auto stop_streaming_service = defer_verbose_shutdown("streaming service", [] {
+                streaming::stream_session::uninit_streaming_service().get();
+            });
             api::set_server_stream_manager(ctx).get();
 
             supervisor::notify("starting hinted handoff manager");
@@ -1005,6 +1010,9 @@ int main(int ac, char** av) {
                 rs.stop().get();
             });
             repair_init_messaging_service_handler(rs, sys_dist_ks, view_update_generator).get();
+            auto stop_repair_messages = defer_verbose_shutdown("repair message handlers", [] {
+                repair_uninit_messaging_service_handler().get();
+            });
             supervisor::notify("starting storage service", true);
             auto& ss = service::get_local_storage_service();
             ss.init_messaging_service_part().get();

message/messaging_service.cc

@@ -572,7 +572,12 @@ messaging_service::initial_scheduling_info() const {
 
 scheduling_group
 messaging_service::scheduling_group_for_verb(messaging_verb verb) const {
-    return _scheduling_info_for_connection_index[get_rpc_client_idx(verb)].sched_group;
+    // We are not using get_rpc_client_idx() because it figures out the client
+    // index based on the current scheduling group, which is relevant when
+    // selecting the right client for sending a message, but is not relevant
+    // when registering handlers.
+    const auto idx = s_rpc_client_idx_table[static_cast<size_t>(verb)];
+    return _scheduling_info_for_connection_index[idx].sched_group;
 }
 
 scheduling_group
@@ -791,6 +796,10 @@ void messaging_service::register_stream_mutation_fragments(std::function<future<
     register_handler(this, messaging_verb::STREAM_MUTATION_FRAGMENTS, std::move(func));
 }
 
+future<> messaging_service::unregister_stream_mutation_fragments() {
+    return unregister_handler(messaging_verb::STREAM_MUTATION_FRAGMENTS);
+}
+
 template<class SinkType, class SourceType>
 future<rpc::sink<SinkType>, rpc::source<SourceType>>
 do_make_sink_source(messaging_verb verb, uint32_t repair_meta_id, shared_ptr<messaging_service::rpc_protocol_client_wrapper> rpc_client, std::unique_ptr<messaging_service::rpc_protocol_wrapper>& rpc) {
@@ -822,6 +831,9 @@ rpc::sink<repair_row_on_wire_with_cmd> messaging_service::make_sink_for_repair_g
 void messaging_service::register_repair_get_row_diff_with_rpc_stream(std::function<future<rpc::sink<repair_row_on_wire_with_cmd>> (const rpc::client_info& cinfo, uint32_t repair_meta_id, rpc::source<repair_hash_with_cmd> source)>&& func) {
     register_handler(this, messaging_verb::REPAIR_GET_ROW_DIFF_WITH_RPC_STREAM, std::move(func));
 }
+future<> messaging_service::unregister_repair_get_row_diff_with_rpc_stream() {
+    return unregister_handler(messaging_verb::REPAIR_GET_ROW_DIFF_WITH_RPC_STREAM);
+}
 
 // Wrapper for REPAIR_PUT_ROW_DIFF_WITH_RPC_STREAM
 future<rpc::sink<repair_row_on_wire_with_cmd>, rpc::source<repair_stream_cmd>>
@@ -841,6 +853,9 @@ rpc::sink<repair_stream_cmd> messaging_service::make_sink_for_repair_put_row_dif
 void messaging_service::register_repair_put_row_diff_with_rpc_stream(std::function<future<rpc::sink<repair_stream_cmd>> (const rpc::client_info& cinfo, uint32_t repair_meta_id, rpc::source<repair_row_on_wire_with_cmd> source)>&& func) {
     register_handler(this, messaging_verb::REPAIR_PUT_ROW_DIFF_WITH_RPC_STREAM, std::move(func));
 }
+future<> messaging_service::unregister_repair_put_row_diff_with_rpc_stream() {
+    return unregister_handler(messaging_verb::REPAIR_PUT_ROW_DIFF_WITH_RPC_STREAM);
+}
 
 // Wrapper for REPAIR_GET_FULL_ROW_HASHES_WITH_RPC_STREAM
 future<rpc::sink<repair_stream_cmd>, rpc::source<repair_hash_with_cmd>>
@@ -860,6 +875,9 @@ rpc::sink<repair_hash_with_cmd> messaging_service::make_sink_for_repair_get_full
 void messaging_service::register_repair_get_full_row_hashes_with_rpc_stream(std::function<future<rpc::sink<repair_hash_with_cmd>> (const rpc::client_info& cinfo, uint32_t repair_meta_id, rpc::source<repair_stream_cmd> source)>&& func) {
     register_handler(this, messaging_verb::REPAIR_GET_FULL_ROW_HASHES_WITH_RPC_STREAM, std::move(func));
 }
+future<> messaging_service::unregister_repair_get_full_row_hashes_with_rpc_stream() {
+    return unregister_handler(messaging_verb::REPAIR_GET_FULL_ROW_HASHES_WITH_RPC_STREAM);
+}
 
 // Send a message for verb
 template <typename MsgIn, typename... MsgOut>
@@ -943,6 +961,9 @@ future<streaming::prepare_message> messaging_service::send_prepare_message(msg_a
     return send_message<streaming::prepare_message>(this, messaging_verb::PREPARE_MESSAGE, id,
         std::move(msg), plan_id, std::move(description), reason);
 }
+future<> messaging_service::unregister_prepare_message() {
+    return unregister_handler(messaging_verb::PREPARE_MESSAGE);
+}
 
 // PREPARE_DONE_MESSAGE
 void messaging_service::register_prepare_done_message(std::function<future<> (const rpc::client_info& cinfo, UUID plan_id, unsigned dst_cpu_id)>&& func) {
@@ -952,6 +973,9 @@ future<> messaging_service::send_prepare_done_message(msg_addr id, UUID plan_id,
     return send_message<void>(this, messaging_verb::PREPARE_DONE_MESSAGE, id,
         plan_id, dst_cpu_id);
 }
+future<> messaging_service::unregister_prepare_done_message() {
+    return unregister_handler(messaging_verb::PREPARE_DONE_MESSAGE);
+}
 
 // STREAM_MUTATION
 void messaging_service::register_stream_mutation(std::function<future<> (const rpc::client_info& cinfo, UUID plan_id, frozen_mutation fm, unsigned dst_cpu_id, rpc::optional<bool> fragmented, rpc::optional<streaming::stream_reason> reason)>&& func) {
@@ -976,6 +1000,9 @@ future<> messaging_service::send_stream_mutation_done(msg_addr id, UUID plan_id,
     return send_message<void>(this, messaging_verb::STREAM_MUTATION_DONE, id,
         plan_id, std::move(ranges), cf_id, dst_cpu_id);
 }
+future<> messaging_service::unregister_stream_mutation_done() {
+    return unregister_handler(messaging_verb::STREAM_MUTATION_DONE);
+}
 
 // COMPLETE_MESSAGE
 void messaging_service::register_complete_message(std::function<future<> (const rpc::client_info& cinfo, UUID plan_id, unsigned dst_cpu_id, rpc::optional<bool> failed)>&& func) {
@@ -985,6 +1012,9 @@ future<> messaging_service::send_complete_message(msg_addr id, UUID plan_id, uns
     return send_message<void>(this, messaging_verb::COMPLETE_MESSAGE, id,
         plan_id, dst_cpu_id, failed);
 }
+future<> messaging_service::unregister_complete_message() {
+    return unregister_handler(messaging_verb::COMPLETE_MESSAGE);
+}
 
 void messaging_service::register_gossip_echo(std::function<future<> ()>&& func) {
     register_handler(this, messaging_verb::GOSSIP_ECHO, std::move(func));
@@ -1199,14 +1229,14 @@ future<partition_checksum> messaging_service::send_repair_checksum_range(
 }
 
 // Wrapper for REPAIR_GET_FULL_ROW_HASHES
-void messaging_service::register_repair_get_full_row_hashes(std::function<future<std::unordered_set<repair_hash>> (const rpc::client_info& cinfo, uint32_t repair_meta_id)>&& func) {
+void messaging_service::register_repair_get_full_row_hashes(std::function<future<repair_hash_set> (const rpc::client_info& cinfo, uint32_t repair_meta_id)>&& func) {
     register_handler(this, messaging_verb::REPAIR_GET_FULL_ROW_HASHES, std::move(func));
 }
 future<> messaging_service::unregister_repair_get_full_row_hashes() {
     return unregister_handler(messaging_verb::REPAIR_GET_FULL_ROW_HASHES);
 }
-future<std::unordered_set<repair_hash>> messaging_service::send_repair_get_full_row_hashes(msg_addr id, uint32_t repair_meta_id) {
-    return send_message<future<std::unordered_set<repair_hash>>>(this, messaging_verb::REPAIR_GET_FULL_ROW_HASHES, std::move(id), repair_meta_id);
+future<repair_hash_set> messaging_service::send_repair_get_full_row_hashes(msg_addr id, uint32_t repair_meta_id) {
+    return send_message<future<repair_hash_set>>(this, messaging_verb::REPAIR_GET_FULL_ROW_HASHES, std::move(id), repair_meta_id);
 }
 
 // Wrapper for REPAIR_GET_COMBINED_ROW_HASH
@@ -1231,13 +1261,13 @@ future<get_sync_boundary_response> messaging_service::send_repair_get_sync_bound
 }
 
 // Wrapper for REPAIR_GET_ROW_DIFF
-void messaging_service::register_repair_get_row_diff(std::function<future<repair_rows_on_wire> (const rpc::client_info& cinfo, uint32_t repair_meta_id, std::unordered_set<repair_hash> set_diff, bool needs_all_rows)>&& func) {
+void messaging_service::register_repair_get_row_diff(std::function<future<repair_rows_on_wire> (const rpc::client_info& cinfo, uint32_t repair_meta_id, repair_hash_set set_diff, bool needs_all_rows)>&& func) {
     register_handler(this, messaging_verb::REPAIR_GET_ROW_DIFF, std::move(func));
 }
 future<> messaging_service::unregister_repair_get_row_diff() {
     return unregister_handler(messaging_verb::REPAIR_GET_ROW_DIFF);
 }
-future<repair_rows_on_wire> messaging_service::send_repair_get_row_diff(msg_addr id, uint32_t repair_meta_id, std::unordered_set<repair_hash> set_diff, bool needs_all_rows) {
+future<repair_rows_on_wire> messaging_service::send_repair_get_row_diff(msg_addr id, uint32_t repair_meta_id, repair_hash_set set_diff, bool needs_all_rows) {
     return send_message<future<repair_rows_on_wire>>(this, messaging_verb::REPAIR_GET_ROW_DIFF, std::move(id), repair_meta_id, std::move(set_diff), needs_all_rows);
 }
 

message/messaging_service.hh

@@ -297,10 +297,12 @@ public:
             streaming::prepare_message msg, UUID plan_id, sstring description, rpc::optional<streaming::stream_reason> reason)>&& func);
     future<streaming::prepare_message> send_prepare_message(msg_addr id, streaming::prepare_message msg, UUID plan_id,
             sstring description, streaming::stream_reason);
+    future<> unregister_prepare_message();
 
     // Wrapper for PREPARE_DONE_MESSAGE verb
     void register_prepare_done_message(std::function<future<> (const rpc::client_info& cinfo, UUID plan_id, unsigned dst_cpu_id)>&& func);
     future<> send_prepare_done_message(msg_addr id, UUID plan_id, unsigned dst_cpu_id);
+    future<> unregister_prepare_done_message();
 
     // Wrapper for STREAM_MUTATION verb
     void register_stream_mutation(std::function<future<> (const rpc::client_info& cinfo, UUID plan_id, frozen_mutation fm, unsigned dst_cpu_id, rpc::optional<bool>, rpc::optional<streaming::stream_reason>)>&& func);
@@ -309,6 +311,7 @@ public:
     // Wrapper for STREAM_MUTATION_FRAGMENTS
     // The receiver of STREAM_MUTATION_FRAGMENTS sends status code to the sender to notify any error on the receiver side. The status code is of type int32_t. 0 means successful, -1 means error, other status code value are reserved for future use.
     void register_stream_mutation_fragments(std::function<future<rpc::sink<int32_t>> (const rpc::client_info& cinfo, UUID plan_id, UUID schema_id, UUID cf_id, uint64_t estimated_partitions, rpc::optional<streaming::stream_reason> reason_opt, rpc::source<frozen_mutation_fragment, rpc::optional<streaming::stream_mutation_fragments_cmd>> source)>&& func);
+    future<> unregister_stream_mutation_fragments();
     rpc::sink<int32_t> make_sink_for_stream_mutation_fragments(rpc::source<frozen_mutation_fragment, rpc::optional<streaming::stream_mutation_fragments_cmd>>& source);
     future<rpc::sink<frozen_mutation_fragment, streaming::stream_mutation_fragments_cmd>, rpc::source<int32_t>> make_sink_and_source_for_stream_mutation_fragments(utils::UUID schema_id, utils::UUID plan_id, utils::UUID cf_id, uint64_t estimated_partitions, streaming::stream_reason reason, msg_addr id);
 
@@ -316,22 +319,27 @@ public:
     future<rpc::sink<repair_hash_with_cmd>, rpc::source<repair_row_on_wire_with_cmd>> make_sink_and_source_for_repair_get_row_diff_with_rpc_stream(uint32_t repair_meta_id, msg_addr id);
     rpc::sink<repair_row_on_wire_with_cmd> make_sink_for_repair_get_row_diff_with_rpc_stream(rpc::source<repair_hash_with_cmd>& source);
     void register_repair_get_row_diff_with_rpc_stream(std::function<future<rpc::sink<repair_row_on_wire_with_cmd>> (const rpc::client_info& cinfo, uint32_t repair_meta_id, rpc::source<repair_hash_with_cmd> source)>&& func);
+    future<> unregister_repair_get_row_diff_with_rpc_stream();
 
     // Wrapper for REPAIR_PUT_ROW_DIFF_WITH_RPC_STREAM
     future<rpc::sink<repair_row_on_wire_with_cmd>, rpc::source<repair_stream_cmd>> make_sink_and_source_for_repair_put_row_diff_with_rpc_stream(uint32_t repair_meta_id, msg_addr id);
     rpc::sink<repair_stream_cmd> make_sink_for_repair_put_row_diff_with_rpc_stream(rpc::source<repair_row_on_wire_with_cmd>& source);
     void register_repair_put_row_diff_with_rpc_stream(std::function<future<rpc::sink<repair_stream_cmd>> (const rpc::client_info& cinfo, uint32_t repair_meta_id, rpc::source<repair_row_on_wire_with_cmd> source)>&& func);
+    future<> unregister_repair_put_row_diff_with_rpc_stream();
 
     // Wrapper for REPAIR_GET_FULL_ROW_HASHES_WITH_RPC_STREAM
     future<rpc::sink<repair_stream_cmd>, rpc::source<repair_hash_with_cmd>> make_sink_and_source_for_repair_get_full_row_hashes_with_rpc_stream(uint32_t repair_meta_id, msg_addr id);
     rpc::sink<repair_hash_with_cmd> make_sink_for_repair_get_full_row_hashes_with_rpc_stream(rpc::source<repair_stream_cmd>& source);
     void register_repair_get_full_row_hashes_with_rpc_stream(std::function<future<rpc::sink<repair_hash_with_cmd>> (const rpc::client_info& cinfo, uint32_t repair_meta_id, rpc::source<repair_stream_cmd> source)>&& func);
+    future<> unregister_repair_get_full_row_hashes_with_rpc_stream();
 
     void register_stream_mutation_done(std::function<future<> (const rpc::client_info& cinfo, UUID plan_id, dht::token_range_vector ranges, UUID cf_id, unsigned dst_cpu_id)>&& func);
     future<> send_stream_mutation_done(msg_addr id, UUID plan_id, dht::token_range_vector ranges, UUID cf_id, unsigned dst_cpu_id);
+    future<> unregister_stream_mutation_done();
 
     void register_complete_message(std::function<future<> (const rpc::client_info& cinfo, UUID plan_id, unsigned dst_cpu_id, rpc::optional<bool> failed)>&& func);
     future<> send_complete_message(msg_addr id, UUID plan_id, unsigned dst_cpu_id, bool failed = false);
+    future<> unregister_complete_message();
 
     // Wrapper for REPAIR_CHECKSUM_RANGE verb
     void register_repair_checksum_range(std::function<future<partition_checksum> (sstring keyspace, sstring cf, dht::token_range range, rpc::optional<repair_checksum> hash_version)>&& func);
@@ -339,9 +347,9 @@ public:
     future<partition_checksum> send_repair_checksum_range(msg_addr id, sstring keyspace, sstring cf, dht::token_range range, repair_checksum hash_version);
 
     // Wrapper for REPAIR_GET_FULL_ROW_HASHES
-    void register_repair_get_full_row_hashes(std::function<future<std::unordered_set<repair_hash>> (const rpc::client_info& cinfo, uint32_t repair_meta_id)>&& func);
+    void register_repair_get_full_row_hashes(std::function<future<repair_hash_set> (const rpc::client_info& cinfo, uint32_t repair_meta_id)>&& func);
     future<> unregister_repair_get_full_row_hashes();
-    future<std::unordered_set<repair_hash>> send_repair_get_full_row_hashes(msg_addr id, uint32_t repair_meta_id);
+    future<repair_hash_set> send_repair_get_full_row_hashes(msg_addr id, uint32_t repair_meta_id);
 
     // Wrapper for REPAIR_GET_COMBINED_ROW_HASH
     void register_repair_get_combined_row_hash(std::function<future<get_combined_row_hash_response> (const rpc::client_info& cinfo, uint32_t repair_meta_id, std::optional<repair_sync_boundary> common_sync_boundary)>&& func);
@@ -354,9 +362,9 @@ public:
     future<get_sync_boundary_response> send_repair_get_sync_boundary(msg_addr id, uint32_t repair_meta_id, std::optional<repair_sync_boundary> skipped_sync_boundary);
 
     // Wrapper for REPAIR_GET_ROW_DIFF
-    void register_repair_get_row_diff(std::function<future<repair_rows_on_wire> (const rpc::client_info& cinfo, uint32_t repair_meta_id, std::unordered_set<repair_hash> set_diff, bool needs_all_rows)>&& func);
+    void register_repair_get_row_diff(std::function<future<repair_rows_on_wire> (const rpc::client_info& cinfo, uint32_t repair_meta_id, repair_hash_set set_diff, bool needs_all_rows)>&& func);
     future<> unregister_repair_get_row_diff();
-    future<repair_rows_on_wire> send_repair_get_row_diff(msg_addr id, uint32_t repair_meta_id, std::unordered_set<repair_hash> set_diff, bool needs_all_rows);
+    future<repair_rows_on_wire> send_repair_get_row_diff(msg_addr id, uint32_t repair_meta_id, repair_hash_set set_diff, bool needs_all_rows);
 
     // Wrapper for REPAIR_PUT_ROW_DIFF
     void register_repair_put_row_diff(std::function<future<> (const rpc::client_info& cinfo, uint32_t repair_meta_id, repair_rows_on_wire row_diff)>&& func);

multishard_mutation_query.cc

@@ -300,10 +300,9 @@ flat_mutation_reader read_context::create_reader(
     }
 
     auto& table = _db.local().find_column_family(schema);
-    auto class_config = _db.local().make_query_class_config();
 
     if (!rm.rparts) {
-        rm.rparts = make_foreign(std::make_unique<reader_meta::remote_parts>(class_config.semaphore));
+        rm.rparts = make_foreign(std::make_unique<reader_meta::remote_parts>(semaphore()));
     }
 
     rm.rparts->range = std::make_unique<const dht::partition_range>(pr);
@@ -513,18 +512,28 @@ future<> read_context::lookup_readers() {
     }
 
     return parallel_for_each(boost::irange(0u, smp::count), [this] (shard_id shard) {
-        return _db.invoke_on(shard, [shard, cmd = &_cmd, ranges = &_ranges, gs = global_schema_ptr(_schema),
+        return _db.invoke_on(shard, [this, shard, cmd = &_cmd, ranges = &_ranges, gs = global_schema_ptr(_schema),
                 gts = tracing::global_trace_state_ptr(_trace_state)] (database& db) mutable {
             auto schema = gs.get();
             auto querier_opt = db.get_querier_cache().lookup_shard_mutation_querier(cmd->query_uuid, *schema, *ranges, cmd->slice, gts.get());
             auto& table = db.find_column_family(schema);
-            auto& semaphore = db.make_query_class_config().semaphore;
+            auto& semaphore = this->semaphore();
 
             if (!querier_opt) {
                 return reader_meta(reader_state::inexistent, reader_meta::remote_parts(semaphore));
             }
 
             auto& q = *querier_opt;
+
+            if (&q.permit().semaphore() != &semaphore) {
+                on_internal_error(mmq_log, format("looked-up reader belongs to different semaphore than the one appropriate for this query class: "
+                        "looked-up reader belongs to {} (0x{:x}) the query class appropriate is {} (0x{:x})",
+                        q.permit().semaphore().name(),
+                        reinterpret_cast<uintptr_t>(&q.permit().semaphore()),
+                        semaphore.name(),
+                        reinterpret_cast<uintptr_t>(&semaphore)));
+            }
+
             auto handle = pause(semaphore, std::move(q).reader());
             return reader_meta(
                     reader_state::successful_lookup,

mutation_partition.cc

@@ -1721,7 +1721,7 @@ void row::apply_monotonically(const schema& s, column_kind kind, row&& other) {
 // we erase the live cells according to the shadowable_tombstone rules.
 static bool dead_marker_shadows_row(const schema& s, column_kind kind, const row_marker& marker) {
     return s.is_view()
-            && !s.view_info()->base_non_pk_columns_in_view_pk().empty()
+            && s.view_info()->has_base_non_pk_columns_in_view_pk()
             && !marker.is_live()
             && kind == column_kind::regular_column; // not applicable to static rows
 }

mutation_query.hh

@@ -114,9 +114,6 @@ class reconcilable_result_builder {
     const schema& _schema;
     const query::partition_slice& _slice;
 
-    utils::chunked_vector<partition> _result;
-    uint32_t _live_rows{};
-
     bool _return_static_content_on_partition_with_no_rows{};
     bool _static_row_is_alive{};
     uint32_t _total_live_rows = 0;
@@ -124,6 +121,10 @@ class reconcilable_result_builder {
     stop_iteration _stop;
     bool _short_read_allowed;
     std::optional<streamed_mutation_freezer> _mutation_consumer;
+
+    uint32_t _live_rows{};
+    // make this the last member so it is destroyed first. #7240
+    utils::chunked_vector<partition> _result;
 public:
     reconcilable_result_builder(const schema& s, const query::partition_slice& slice,
                                 query::result_memory_accounter&& accounter)

mutation_reader.cc

@@ -30,6 +30,7 @@
 #include "schema_registry.hh"
 #include "mutation_compactor.hh"
 
+logging::logger mrlog("mutation_reader");
 
 static constexpr size_t merger_small_vector_size = 4;
 
@@ -659,6 +660,8 @@ flat_mutation_reader make_combined_reader(schema_ptr schema,
     return make_combined_reader(std::move(schema), std::move(v), fwd_sm, fwd_mr);
 }
 
+const ssize_t new_reader_base_cost{16 * 1024};
+
 class restricting_mutation_reader : public flat_mutation_reader::impl {
     struct mutation_source_and_params {
         mutation_source _ms;
@@ -685,8 +688,6 @@ class restricting_mutation_reader : public flat_mutation_reader::impl {
     };
     std::variant<pending_state, admitted_state> _state;
 
-    static const ssize_t new_reader_base_cost{16 * 1024};
-
     template<typename Function>
     requires std::is_move_constructible<Function>::value
         && requires(Function fn, flat_mutation_reader& reader) {
@@ -1026,6 +1027,13 @@ private:
     bool _reader_created = false;
     bool _drop_partition_start = false;
     bool _drop_static_row = false;
+    // Trim range tombstones on the start of the buffer to the start of the read
+    // range (_next_position_in_partition). Set after reader recreation.
+    // Also validate the first not-trimmed mutation fragment's position.
+    bool _trim_range_tombstones = false;
+    // Validate the partition key of the first emitted partition, set after the
+    // reader was recreated.
+    bool _validate_partition_key = false;
     position_in_partition::tri_compare _tri_cmp;
 
     std::optional<dht::decorated_key> _last_pkey;
@@ -1047,7 +1055,10 @@ private:
     void adjust_partition_slice();
     flat_mutation_reader recreate_reader();
     flat_mutation_reader resume_or_create_reader();
+    void maybe_validate_partition_start(const circular_buffer<mutation_fragment>& buffer);
+    void validate_position_in_partition(position_in_partition_view pos) const;
     bool should_drop_fragment(const mutation_fragment& mf);
+    bool maybe_trim_range_tombstone(mutation_fragment& mf) const;
     future<> do_fill_buffer(flat_mutation_reader& reader, db::timeout_clock::time_point timeout);
     future<> fill_buffer(flat_mutation_reader& reader, db::timeout_clock::time_point timeout);
 
@@ -1120,16 +1131,11 @@ void evictable_reader::update_next_position(flat_mutation_reader& reader) {
             _next_position_in_partition = position_in_partition::before_all_clustered_rows();
             break;
         case partition_region::clustered:
-            if (reader.is_buffer_empty()) {
-                _next_position_in_partition = position_in_partition::after_key(last_pos);
-            } else {
-               const auto& next_frag = reader.peek_buffer();
-               if (next_frag.is_end_of_partition()) {
+            if (!reader.is_buffer_empty() && reader.peek_buffer().is_end_of_partition()) {
                    push_mutation_fragment(reader.pop_mutation_fragment());
                    _next_position_in_partition = position_in_partition::for_partition_start();
-               } else {
-                   _next_position_in_partition = position_in_partition(next_frag.position());
-               }
+            } else {
+                _next_position_in_partition = position_in_partition::after_key(last_pos);
             }
             break;
         case partition_region::partition_end:
@@ -1154,6 +1160,9 @@ flat_mutation_reader evictable_reader::recreate_reader() {
     const dht::partition_range* range = _pr;
     const query::partition_slice* slice = &_ps;
 
+    _range_override.reset();
+    _slice_override.reset();
+
     if (_last_pkey) {
         bool partition_range_is_inclusive = true;
 
@@ -1190,6 +1199,9 @@ flat_mutation_reader evictable_reader::recreate_reader() {
         range = &*_range_override;
     }
 
+    _trim_range_tombstones = true;
+    _validate_partition_key = true;
+
     return _ms.make_reader(
             _schema,
             _permit,
@@ -1216,6 +1228,78 @@ flat_mutation_reader evictable_reader::resume_or_create_reader() {
     return recreate_reader();
 }
 
+template <typename... Arg>
+static void require(bool condition, const char* msg, const Arg&... arg) {
+    if (!condition) {
+        on_internal_error(mrlog, format(msg, arg...));
+    }
+}
+
+void evictable_reader::maybe_validate_partition_start(const circular_buffer<mutation_fragment>& buffer) {
+    if (!_validate_partition_key || buffer.empty()) {
+        return;
+    }
+
+    // If this is set we can assume the first fragment is a partition-start.
+    const auto& ps = buffer.front().as_partition_start();
+    const auto tri_cmp = dht::ring_position_comparator(*_schema);
+    // If we recreated the reader after fast-forwarding it we won't have
+    // _last_pkey set. In this case it is enough to check if the partition
+    // is in range.
+    if (_last_pkey) {
+        const auto cmp_res = tri_cmp(*_last_pkey, ps.key());
+        if (_drop_partition_start) { // should be the same partition
+            require(
+                    cmp_res == 0,
+                    "{}(): validation failed, expected partition with key equal to _last_pkey {} due to _drop_partition_start being set, but got {}",
+                    __FUNCTION__,
+                    *_last_pkey,
+                    ps.key());
+        } else { // should be a larger partition
+            require(
+                    cmp_res < 0,
+                    "{}(): validation failed, expected partition with key larger than _last_pkey {} due to _drop_partition_start being unset, but got {}",
+                    __FUNCTION__,
+                    *_last_pkey,
+                    ps.key());
+        }
+    }
+    const auto& prange = _range_override ? *_range_override : *_pr;
+    require(
+            // TODO: somehow avoid this copy
+            prange.contains(ps.key(), tri_cmp),
+            "{}(): validation failed, expected partition with key that falls into current range {}, but got {}",
+            __FUNCTION__,
+            prange,
+            ps.key());
+
+    _validate_partition_key = false;
+}
+
+void evictable_reader::validate_position_in_partition(position_in_partition_view pos) const {
+    require(
+            _tri_cmp(_next_position_in_partition, pos) <= 0,
+            "{}(): validation failed, expected position in partition that is larger-than-equal than _next_position_in_partition {}, but got {}",
+            __FUNCTION__,
+            _next_position_in_partition,
+            pos);
+
+    if (_slice_override && pos.region() == partition_region::clustered) {
+        const auto ranges = _slice_override->row_ranges(*_schema, _last_pkey->key());
+        const bool any_contains = std::any_of(ranges.begin(), ranges.end(), [this, &pos] (const query::clustering_range& cr) {
+            // TODO: somehow avoid this copy
+            auto range = position_range(cr);
+            return range.contains(*_schema, pos);
+        });
+        require(
+                any_contains,
+                "{}(): validation failed, expected clustering fragment that is included in the slice {}, but got {}",
+                __FUNCTION__,
+                *_slice_override,
+                pos);
+    }
+}
+
 bool evictable_reader::should_drop_fragment(const mutation_fragment& mf) {
     if (_drop_partition_start && mf.is_partition_start()) {
         _drop_partition_start = false;
@@ -1228,12 +1312,50 @@ bool evictable_reader::should_drop_fragment(const mutation_fragment& mf) {
     return false;
 }
 
+bool evictable_reader::maybe_trim_range_tombstone(mutation_fragment& mf) const {
+    // We either didn't read a partition yet (evicted after fast-forwarding) or
+    // didn't stop in a clustering region. We don't need to trim range
+    // tombstones in either case.
+    if (!_last_pkey || _next_position_in_partition.region() != partition_region::clustered) {
+        return false;
+    }
+    if (!mf.is_range_tombstone()) {
+        validate_position_in_partition(mf.position());
+        return false;
+    }
+
+    if (_tri_cmp(mf.position(), _next_position_in_partition) >= 0) {
+        validate_position_in_partition(mf.position());
+        return false; // rt in range, no need to trim
+    }
+
+    auto& rt = mf.as_mutable_range_tombstone();
+
+    require(
+            _tri_cmp(_next_position_in_partition, rt.end_position()) <= 0,
+            "{}(): validation failed, expected range tombstone with end pos larger than _next_position_in_partition {}, but got {}",
+            __FUNCTION__,
+            _next_position_in_partition,
+            rt.end_position());
+
+    rt.set_start(*_schema, position_in_partition_view::before_key(_next_position_in_partition));
+
+    return true;
+}
+
 future<> evictable_reader::do_fill_buffer(flat_mutation_reader& reader, db::timeout_clock::time_point timeout) {
     if (!_drop_partition_start && !_drop_static_row) {
-        return reader.fill_buffer(timeout);
+        auto fill_buf_fut = reader.fill_buffer(timeout);
+        if (_validate_partition_key) {
+            fill_buf_fut = fill_buf_fut.then([this, &reader] {
+                maybe_validate_partition_start(reader.buffer());
+            });
+        }
+        return fill_buf_fut;
     }
     return repeat([this, &reader, timeout] {
         return reader.fill_buffer(timeout).then([this, &reader] {
+            maybe_validate_partition_start(reader.buffer());
             while (!reader.is_buffer_empty() && should_drop_fragment(reader.peek_buffer())) {
                 reader.pop_mutation_fragment();
             }
@@ -1247,6 +1369,11 @@ future<> evictable_reader::fill_buffer(flat_mutation_reader& reader, db::timeout
         if (reader.is_buffer_empty()) {
             return make_ready_future<>();
         }
+        while (_trim_range_tombstones && !reader.is_buffer_empty()) {
+            auto mf = reader.pop_mutation_fragment();
+            _trim_range_tombstones = maybe_trim_range_tombstone(mf);
+            push_mutation_fragment(std::move(mf));
+        }
         reader.move_buffer_content_to(*this);
         auto stop = [this, &reader] {
             // The only problematic fragment kind is the range tombstone.
@@ -1287,7 +1414,13 @@ future<> evictable_reader::fill_buffer(flat_mutation_reader& reader, db::timeout
             if (reader.is_buffer_empty()) {
                 return do_fill_buffer(reader, timeout);
             }
-            push_mutation_fragment(reader.pop_mutation_fragment());
+            if (_trim_range_tombstones) {
+                auto mf = reader.pop_mutation_fragment();
+                _trim_range_tombstones = maybe_trim_range_tombstone(mf);
+                push_mutation_fragment(std::move(mf));
+            } else {
+                push_mutation_fragment(reader.pop_mutation_fragment());
+            }
             return make_ready_future<>();
         });
     }).then([this, &reader] {

mutation_reader.hh

@@ -304,6 +304,8 @@ public:
 mutation_source make_empty_mutation_source();
 snapshot_source make_empty_snapshot_source();
 
+extern const ssize_t new_reader_base_cost;
+
 // Creates a restricted reader whose resource usages will be tracked
 // during it's lifetime. If there are not enough resources (dues to
 // existing readers) to create the new reader, it's construction will

position_in_partition.hh

@@ -163,6 +163,11 @@ public:
         return {partition_region::clustered, bound_weight::before_all_prefixed, &ck};
     }
 
+    // Returns a view to before_key(pos._ck) if pos.is_clustering_row() else returns pos as-is.
+    static position_in_partition_view before_key(position_in_partition_view pos) {
+        return {partition_region::clustered, pos._bound_weight == bound_weight::equal ? bound_weight::before_all_prefixed : pos._bound_weight, pos._ck};
+    }
+
     partition_region region() const { return _type; }
     bound_weight get_bound_weight() const { return _bound_weight; }
     bool is_partition_start() const { return _type == partition_region::partition_start; }

reader_concurrency_semaphore.cc

@@ -104,7 +104,7 @@ reader_concurrency_semaphore::inactive_read_handle reader_concurrency_semaphore:
         const auto [it, _] = _inactive_reads.emplace(_next_id++, std::move(ir));
         (void)_;
         ++_inactive_read_stats.population;
-        return inactive_read_handle(it->first);
+        return inactive_read_handle(*this, it->first);
     }
 
     // The evicted reader will release its permit, hopefully allowing us to
@@ -115,6 +115,17 @@ reader_concurrency_semaphore::inactive_read_handle reader_concurrency_semaphore:
 }
 
 std::unique_ptr<reader_concurrency_semaphore::inactive_read> reader_concurrency_semaphore::unregister_inactive_read(inactive_read_handle irh) {
+    if (irh && irh._sem != this) {
+        throw std::runtime_error(fmt::format(
+                    "reader_concurrency_semaphore::unregister_inactive_read(): "
+                    "attempted to unregister an inactive read with a handle belonging to another semaphore: "
+                    "this is {} (0x{:x}) but the handle belongs to {} (0x{:x})",
+                    name(),
+                    reinterpret_cast<uintptr_t>(this),
+                    irh._sem->name(),
+                    reinterpret_cast<uintptr_t>(irh._sem)));
+    }
+
     if (auto it = _inactive_reads.find(irh._id); it != _inactive_reads.end()) {
         auto ir = std::move(it->second);
         _inactive_reads.erase(it);

reader_concurrency_semaphore.hh

@@ -60,18 +60,20 @@ public:
     };
 
     class inactive_read_handle {
+        reader_concurrency_semaphore* _sem = nullptr;
         uint64_t _id = 0;
 
         friend class reader_concurrency_semaphore;
 
-        explicit inactive_read_handle(uint64_t id)
-            : _id(id) {
+        explicit inactive_read_handle(reader_concurrency_semaphore& sem, uint64_t id)
+            : _sem(&sem), _id(id) {
         }
     public:
         inactive_read_handle() = default;
-        inactive_read_handle(inactive_read_handle&& o) : _id(std::exchange(o._id, 0)) {
+        inactive_read_handle(inactive_read_handle&& o) : _sem(std::exchange(o._sem, nullptr)), _id(std::exchange(o._id, 0)) {
         }
         inactive_read_handle& operator=(inactive_read_handle&& o) {
+            _sem = std::exchange(o._sem, nullptr);
             _id = std::exchange(o._id, 0);
             return *this;
         }
@@ -105,6 +107,7 @@ private:
     };
 
 private:
+    const resources _initial_resources;
     resources _resources;
 
     expiring_fifo<entry, expiry_handler, db::timeout_clock> _wait_list;
@@ -135,7 +138,8 @@ public:
             sstring name,
             size_t max_queue_length = std::numeric_limits<size_t>::max(),
             std::function<void()> prethrow_action = nullptr)
-        : _resources(count, memory)
+        : _initial_resources(count, memory)
+        , _resources(count, memory)
         , _wait_list(expiry_handler(name))
         , _name(std::move(name))
         , _max_queue_length(max_queue_length)
@@ -144,11 +148,11 @@ public:
     /// Create a semaphore with practically unlimited count and memory.
     ///
     /// And conversely, no queue limit either.
-    explicit reader_concurrency_semaphore(no_limits)
+    explicit reader_concurrency_semaphore(no_limits, sstring name = "unlimited reader_concurrency_semaphore")
         : reader_concurrency_semaphore(
                 std::numeric_limits<int>::max(),
                 std::numeric_limits<ssize_t>::max(),
-                "unlimited reader_concurrency_semaphore") {}
+                std::move(name)) {}
 
     ~reader_concurrency_semaphore();
 
@@ -158,6 +162,13 @@ public:
     reader_concurrency_semaphore(reader_concurrency_semaphore&&) = delete;
     reader_concurrency_semaphore& operator=(reader_concurrency_semaphore&&) = delete;
 
+    /// Returns the name of the semaphore
+    ///
+    /// If the semaphore has no name, "unnamed reader concurrency semaphore" is returned.
+    std::string_view name() const {
+        return _name.empty() ? "unnamed reader concurrency semaphore" : std::string_view(_name);
+    }
+
     /// Register an inactive read.
     ///
     /// The semaphore will evict this read when there is a shortage of
@@ -193,6 +204,10 @@ public:
 
     reader_permit make_permit();
 
+    const resources initial_resources() const {
+        return _initial_resources;
+    }
+
     const resources available_resources() const {
         return _resources;
     }

reader_permit.hh

@@ -42,12 +42,20 @@ struct reader_resources {
         return count >= other.count && memory >= other.memory;
     }
 
+    reader_resources operator-(const reader_resources& other) const {
+        return reader_resources{count - other.count, memory - other.memory};
+    }
+
     reader_resources& operator-=(const reader_resources& other) {
         count -= other.count;
         memory -= other.memory;
         return *this;
     }
 
+    reader_resources operator+(const reader_resources& other) const {
+        return reader_resources{count + other.count, memory + other.memory};
+    }
+
     reader_resources& operator+=(const reader_resources& other) {
         count += other.count;
         memory += other.memory;

redis/commands.cc

@@ -62,7 +62,7 @@ shared_ptr<abstract_command> exists::prepare(service::storage_proxy& proxy, requ
 }
 
 future<redis_message> exists::execute(service::storage_proxy& proxy, redis::redis_options& options, service_permit permit) {
-    return seastar::do_for_each(_keys, [&proxy, &options, &permit, this] (bytes key) {
+    return seastar::do_for_each(_keys, [&proxy, &options, permit, this] (bytes& key) {
         return redis::read_strings(proxy, options, key, permit).then([this] (lw_shared_ptr<strings_result> result) {
             if (result->has_result()) {
                 _count++;

reloc/build_deb.sh

@@ -44,15 +44,15 @@ mkdir -p $BUILDDIR/scylla-package
 tar -C $BUILDDIR/scylla-package -xpf $RELOC_PKG
 cd $BUILDDIR/scylla-package
 
-PRODUCT=$(cat scylla/SCYLLA-PRODUCT-FILE)
-SCYLLA_VERSION=$(cat scylla/SCYLLA-VERSION-FILE)
-SCYLLA_RELEASE=$(cat scylla/SCYLLA-RELEASE-FILE)
-
-ln -fv $RELOC_PKG ../$PRODUCT-server_$SCYLLA_VERSION-$SCYLLA_RELEASE.orig.tar.gz
-
 if $DIST; then
     export DEB_BUILD_OPTIONS="housekeeping"
 fi
 
 mv scylla/debian debian
+
+PKG_NAME=$(dpkg-parsechangelog --show-field Source)
+# XXX: Drop revision number from version string.
+#      Since it always '1', this should be okay for now.
+PKG_VERSION=$(dpkg-parsechangelog --show-field Version |sed -e 's/-1$//')
+ln -fv $RELOC_PKG ../"$PKG_NAME"_"$PKG_VERSION".orig.tar.gz
 debuild -rfakeroot -us -uc

repair/repair.cc

@@ -1633,6 +1633,7 @@ future<> bootstrap_with_repair(seastar::sharded<database>& db, locator::token_me
             auto& ks = db.local().find_keyspace(keyspace_name);
             auto& strat = ks.get_replication_strategy();
             dht::token_range_vector desired_ranges = strat.get_pending_address_ranges(tm, tokens, myip);
+            bool find_node_in_local_dc_only = strat.get_type() == locator::replication_strategy_type::network_topology;
 
             //Active ranges
             auto metadata_clone = tm.clone_only_token_map();
@@ -1719,6 +1720,9 @@ future<> bootstrap_with_repair(seastar::sharded<database>& db, locator::token_me
                             mandatory_neighbors = get_node_losing_the_ranges(old_endpoints, new_endpoints);
                             neighbors = mandatory_neighbors;
                         } else if (old_endpoints.size() < strat.get_replication_factor()) {
+                          if (!find_node_in_local_dc_only) {
+                            neighbors = old_endpoints;
+                          } else {
                             if (old_endpoints_in_local_dc.size() == rf_in_local_dc) {
                                 // Local DC has enough replica nodes.
                                 mandatory_neighbors = get_node_losing_the_ranges(old_endpoints_in_local_dc, new_endpoints);
@@ -1746,6 +1750,7 @@ future<> bootstrap_with_repair(seastar::sharded<database>& db, locator::token_me
                                 throw std::runtime_error(format("bootstrap_with_repair: keyspace={}, range={}, wrong number of old_endpoints_in_local_dc={}, rf_in_local_dc={}",
                                         keyspace_name, desired_range, old_endpoints_in_local_dc.size(), rf_in_local_dc));
                             }
+                          }
                         } else {
                             throw std::runtime_error(format("bootstrap_with_repair: keyspace={}, range={}, wrong number of old_endpoints={}, rf={}",
                                         keyspace_name, desired_range, old_endpoints, strat.get_replication_factor()));

repair/repair.hh

@@ -23,6 +23,7 @@
 
 #include <unordered_map>
 #include <exception>
+#include <absl/container/btree_set.h>
 
 #include <seastar/core/sstring.hh>
 #include <seastar/core/sharded.hh>
@@ -339,6 +340,8 @@ public:
     }
 };
 
+using repair_hash_set = absl::btree_set<repair_hash>;
+
 enum class repair_row_level_start_status: uint8_t {
     ok,
     no_such_column_family,

repair/row_level.cc

@@ -47,6 +47,7 @@
 #include "gms/gossiper.hh"
 #include "repair/row_level.hh"
 #include "mutation_source_metadata.hh"
+#include "utils/stall_free.hh"
 
 extern logging::logger rlogger;
 
@@ -529,7 +530,7 @@ public:
                     sstables::shared_sstable sst = use_view_update_path ? t->make_streaming_staging_sstable() : t->make_streaming_sstable_for_write();
                     schema_ptr s = reader.schema();
                     auto& pc = service::get_local_streaming_priority();
-                    return sst->write_components(std::move(reader), std::max(1ul, adjusted_estimated_partitions), s,
+                    return sst->write_components(std::move(reader), adjusted_estimated_partitions, s,
                                                  t->get_sstables_manager().configure_writer(),
                                                  encoding_stats{}, pc).then([sst] {
                         return sst->open_data();
@@ -666,7 +667,7 @@ private:
     // Tracks current sync boundary
     std::optional<repair_sync_boundary> _current_sync_boundary;
     // Contains the hashes of rows in the _working_row_buffor for all peer nodes
-    std::vector<std::unordered_set<repair_hash>> _peer_row_hash_sets;
+    std::vector<repair_hash_set> _peer_row_hash_sets;
     // Gate used to make sure pending operation of meta data is done
     seastar::gate _gate;
     sink_source_for_get_full_row_hashes _sink_source_for_get_full_row_hashes;
@@ -754,11 +755,12 @@ public:
 public:
     future<> stop() {
         auto gate_future = _gate.close();
-        auto writer_future = _repair_writer.wait_for_writer_done();
         auto f1 = _sink_source_for_get_full_row_hashes.close();
         auto f2 = _sink_source_for_get_row_diff.close();
         auto f3 = _sink_source_for_put_row_diff.close();
-        return when_all_succeed(std::move(gate_future), std::move(writer_future), std::move(f1), std::move(f2), std::move(f3)).discard_result();
+        return when_all_succeed(std::move(gate_future), std::move(f1), std::move(f2), std::move(f3)).discard_result().finally([this] {
+            return _repair_writer.wait_for_writer_done();
+        });
     }
 
     static std::unordered_map<node_repair_meta_id, lw_shared_ptr<repair_meta>>& repair_meta_map() {
@@ -886,9 +888,9 @@ public:
     }
 
     // Must run inside a seastar thread
-    static std::unordered_set<repair_hash>
-    get_set_diff(const std::unordered_set<repair_hash>& x, const std::unordered_set<repair_hash>& y) {
-        std::unordered_set<repair_hash> set_diff;
+    static repair_hash_set
+    get_set_diff(const repair_hash_set& x, const repair_hash_set& y) {
+        repair_hash_set set_diff;
         // Note std::set_difference needs x and y are sorted.
         std::copy_if(x.begin(), x.end(), std::inserter(set_diff, set_diff.end()),
                 [&y] (auto& item) { thread::maybe_yield(); return y.find(item) == y.end(); });
@@ -906,14 +908,14 @@ public:
 
     }
 
-    std::unordered_set<repair_hash>& peer_row_hash_sets(unsigned node_idx) {
+    repair_hash_set& peer_row_hash_sets(unsigned node_idx) {
         return _peer_row_hash_sets[node_idx];
     }
 
     // Get a list of row hashes in _working_row_buf
-    future<std::unordered_set<repair_hash>>
+    future<repair_hash_set>
     working_row_hashes() {
-        return do_with(std::unordered_set<repair_hash>(), [this] (std::unordered_set<repair_hash>& hashes) {
+        return do_with(repair_hash_set(), [this] (repair_hash_set& hashes) {
             return do_for_each(_working_row_buf, [&hashes] (repair_row& r) {
                 hashes.emplace(r.hash());
             }).then([&hashes] {
@@ -1090,24 +1092,32 @@ private:
         });
     }
 
+    future<> clear_row_buf() {
+        return utils::clear_gently(_row_buf);
+    }
+
+    future<> clear_working_row_buf() {
+        return utils::clear_gently(_working_row_buf).then([this] {
+            _working_row_buf_combined_hash.clear();
+        });
+    }
+
     // Read rows from disk until _max_row_buf_size of rows are filled into _row_buf.
     // Calculate the combined checksum of the rows
     // Calculate the total size of the rows in _row_buf
     future<get_sync_boundary_response>
     get_sync_boundary(std::optional<repair_sync_boundary> skipped_sync_boundary) {
+        auto f = make_ready_future<>();
         if (skipped_sync_boundary) {
             _current_sync_boundary = skipped_sync_boundary;
-            _row_buf.clear();
-            _working_row_buf.clear();
-            _working_row_buf_combined_hash.clear();
-        } else {
-            _working_row_buf.clear();
-            _working_row_buf_combined_hash.clear();
+            f = clear_row_buf();
         }
         // Here is the place we update _last_sync_boundary
         rlogger.trace("SET _last_sync_boundary from {} to {}", _last_sync_boundary, _current_sync_boundary);
         _last_sync_boundary = _current_sync_boundary;
-        return row_buf_size().then([this, sb = std::move(skipped_sync_boundary)] (size_t cur_size) {
+      return f.then([this, sb = std::move(skipped_sync_boundary)] () mutable {
+       return clear_working_row_buf().then([this, sb = sb] () mutable {
+        return row_buf_size().then([this, sb = std::move(sb)] (size_t cur_size) {
             return read_rows_from_disk(cur_size).then([this, sb = std::move(sb)] (std::list<repair_row> new_rows, size_t new_rows_size) mutable {
                 size_t new_rows_nr = new_rows.size();
                 _row_buf.splice(_row_buf.end(), new_rows);
@@ -1124,6 +1134,8 @@ private:
                 });
             });
         });
+       });
+      });
     }
 
     future<> move_row_buf_to_working_row_buf() {
@@ -1199,9 +1211,9 @@ private:
     }
 
     future<std::list<repair_row>>
-    copy_rows_from_working_row_buf_within_set_diff(std::unordered_set<repair_hash> set_diff) {
+    copy_rows_from_working_row_buf_within_set_diff(repair_hash_set set_diff) {
         return do_with(std::list<repair_row>(), std::move(set_diff),
-                [this] (std::list<repair_row>& rows, std::unordered_set<repair_hash>& set_diff) {
+                [this] (std::list<repair_row>& rows, repair_hash_set& set_diff) {
             return do_for_each(_working_row_buf, [this, &set_diff, &rows] (const repair_row& r) {
                 if (set_diff.count(r.hash()) > 0) {
                     rows.push_back(r);
@@ -1216,7 +1228,7 @@ private:
     // Give a set of row hashes, return the corresponding rows
     // If needs_all_rows is set, return all the rows in _working_row_buf, ignore the set_diff
     future<std::list<repair_row>>
-    get_row_diff(std::unordered_set<repair_hash> set_diff, needs_all_rows_t needs_all_rows = needs_all_rows_t::no) {
+    get_row_diff(repair_hash_set set_diff, needs_all_rows_t needs_all_rows = needs_all_rows_t::no) {
         if (needs_all_rows) {
             if (!_repair_master || _nr_peer_nodes == 1) {
                 return make_ready_future<std::list<repair_row>>(std::move(_working_row_buf));
@@ -1227,19 +1239,28 @@ private:
         }
     }
 
-    future<> do_apply_rows(std::list<repair_row>& row_diff, unsigned node_idx, update_working_row_buf update_buf) {
-        return with_semaphore(_repair_writer.sem(), 1, [this, node_idx, update_buf, &row_diff] {
-            _repair_writer.create_writer(_db, node_idx);
-            return do_for_each(row_diff, [this, node_idx, update_buf] (repair_row& r) {
-                if (update_buf) {
-                    _working_row_buf_combined_hash.add(r.hash());
-                }
-                // The repair_row here is supposed to have
-                // mutation_fragment attached because we have stored it in
-                // to_repair_rows_list above where the repair_row is created.
-                mutation_fragment mf = std::move(r.get_mutation_fragment());
-                auto dk_with_hash = r.get_dk_with_hash();
-                return _repair_writer.do_write(node_idx, std::move(dk_with_hash), std::move(mf));
+    future<> do_apply_rows(std::list<repair_row>&& row_diff, unsigned node_idx, update_working_row_buf update_buf) {
+        return do_with(std::move(row_diff), [this, node_idx, update_buf] (std::list<repair_row>& row_diff) {
+            return with_semaphore(_repair_writer.sem(), 1, [this, node_idx, update_buf, &row_diff] {
+                _repair_writer.create_writer(_db, node_idx);
+                return repeat([this, node_idx, update_buf, &row_diff] () mutable {
+                    if (row_diff.empty()) {
+                        return make_ready_future<stop_iteration>(stop_iteration::yes);
+                    }
+                    repair_row& r = row_diff.front();
+                    if (update_buf) {
+                        _working_row_buf_combined_hash.add(r.hash());
+                    }
+                    // The repair_row here is supposed to have
+                    // mutation_fragment attached because we have stored it in
+                    // to_repair_rows_list above where the repair_row is created.
+                    mutation_fragment mf = std::move(r.get_mutation_fragment());
+                    auto dk_with_hash = r.get_dk_with_hash();
+                    return _repair_writer.do_write(node_idx, std::move(dk_with_hash), std::move(mf)).then([&row_diff] {
+                        row_diff.pop_front();
+                        return make_ready_future<stop_iteration>(stop_iteration::no);
+                    });
+                });
             });
         });
     }
@@ -1257,19 +1278,17 @@ private:
         stats().rx_row_nr += row_diff.size();
         stats().rx_row_nr_peer[from] += row_diff.size();
         if (update_buf) {
-            std::list<repair_row> tmp;
-            tmp.swap(_working_row_buf);
             // Both row_diff and _working_row_buf and are ordered, merging
             // two sored list to make sure the combination of row_diff
             // and _working_row_buf are ordered.
-            std::merge(tmp.begin(), tmp.end(), row_diff.begin(), row_diff.end(), std::back_inserter(_working_row_buf),
-                [this] (const repair_row& x, const repair_row& y) { thread::maybe_yield(); return _cmp(x.boundary(), y.boundary()) < 0; });
+            utils::merge_to_gently(_working_row_buf, row_diff,
+                 [this] (const repair_row& x, const repair_row& y) { return _cmp(x.boundary(), y.boundary()) < 0; });
         }
         if (update_hash_set) {
-            _peer_row_hash_sets[node_idx] = boost::copy_range<std::unordered_set<repair_hash>>(row_diff |
+            _peer_row_hash_sets[node_idx] = boost::copy_range<repair_hash_set>(row_diff |
                     boost::adaptors::transformed([] (repair_row& r) { thread::maybe_yield(); return r.hash(); }));
         }
-        do_apply_rows(row_diff, node_idx, update_buf).get();
+        do_apply_rows(std::move(row_diff), node_idx, update_buf).get();
     }
 
     future<>
@@ -1277,11 +1296,9 @@ private:
         if (rows.empty()) {
             return make_ready_future<>();
         }
-        return to_repair_rows_list(rows).then([this] (std::list<repair_row> row_diff) {
-            return do_with(std::move(row_diff), [this] (std::list<repair_row>& row_diff) {
-                unsigned node_idx = 0;
-                return do_apply_rows(row_diff, node_idx, update_working_row_buf::no);
-            });
+        return to_repair_rows_list(std::move(rows)).then([this] (std::list<repair_row> row_diff) {
+            unsigned node_idx = 0;
+            return do_apply_rows(std::move(row_diff), node_idx, update_working_row_buf::no);
         });
     }
 
@@ -1360,13 +1377,13 @@ private:
 public:
     // RPC API
     // Return the hashes of the rows in _working_row_buf
-    future<std::unordered_set<repair_hash>>
+    future<repair_hash_set>
     get_full_row_hashes(gms::inet_address remote_node) {
         if (remote_node == _myip) {
             return get_full_row_hashes_handler();
         }
         return netw::get_local_messaging_service().send_repair_get_full_row_hashes(msg_addr(remote_node),
-                _repair_meta_id).then([this, remote_node] (std::unordered_set<repair_hash> hashes) {
+                _repair_meta_id).then([this, remote_node] (repair_hash_set hashes) {
             rlogger.debug("Got full hashes from peer={}, nr_hashes={}", remote_node, hashes.size());
             _metrics.rx_hashes_nr += hashes.size();
             stats().rx_hashes_nr += hashes.size();
@@ -1377,7 +1394,7 @@ public:
 
 private:
     future<> get_full_row_hashes_source_op(
-            lw_shared_ptr<std::unordered_set<repair_hash>> current_hashes,
+            lw_shared_ptr<repair_hash_set> current_hashes,
             gms::inet_address remote_node,
             unsigned node_idx,
             rpc::source<repair_hash_with_cmd>& source) {
@@ -1415,12 +1432,12 @@ private:
     }
 
 public:
-    future<std::unordered_set<repair_hash>>
+    future<repair_hash_set>
     get_full_row_hashes_with_rpc_stream(gms::inet_address remote_node, unsigned node_idx) {
         if (remote_node == _myip) {
             return get_full_row_hashes_handler();
         }
-        auto current_hashes = make_lw_shared<std::unordered_set<repair_hash>>();
+        auto current_hashes = make_lw_shared<repair_hash_set>();
         return _sink_source_for_get_full_row_hashes.get_sink_source(remote_node, node_idx).then(
                 [this, current_hashes, remote_node, node_idx]
                 (rpc::sink<repair_stream_cmd>& sink, rpc::source<repair_hash_with_cmd>& source) mutable {
@@ -1435,7 +1452,7 @@ public:
     }
 
     // RPC handler
-    future<std::unordered_set<repair_hash>>
+    future<repair_hash_set>
     get_full_row_hashes_handler() {
         return with_gate(_gate, [this] {
             return working_row_hashes();
@@ -1585,7 +1602,7 @@ public:
     // RPC API
     // Return rows in the _working_row_buf with hash within the given sef_diff
     // Must run inside a seastar thread
-    void get_row_diff(std::unordered_set<repair_hash> set_diff, needs_all_rows_t needs_all_rows, gms::inet_address remote_node, unsigned node_idx) {
+    void get_row_diff(repair_hash_set set_diff, needs_all_rows_t needs_all_rows, gms::inet_address remote_node, unsigned node_idx) {
         if (needs_all_rows || !set_diff.empty()) {
             if (remote_node == _myip) {
                 return;
@@ -1654,11 +1671,11 @@ private:
     }
 
     future<> get_row_diff_sink_op(
-            std::unordered_set<repair_hash> set_diff,
+            repair_hash_set set_diff,
             needs_all_rows_t needs_all_rows,
             rpc::sink<repair_hash_with_cmd>& sink,
             gms::inet_address remote_node) {
-        return do_with(std::move(set_diff), [needs_all_rows, remote_node, &sink] (std::unordered_set<repair_hash>& set_diff) mutable {
+        return do_with(std::move(set_diff), [needs_all_rows, remote_node, &sink] (repair_hash_set& set_diff) mutable {
             if (inject_rpc_stream_error) {
                 return make_exception_future<>(std::runtime_error("get_row_diff: Inject sender error in sink loop"));
             }
@@ -1685,7 +1702,7 @@ private:
 public:
     // Must run inside a seastar thread
     void get_row_diff_with_rpc_stream(
-            std::unordered_set<repair_hash> set_diff,
+            repair_hash_set set_diff,
             needs_all_rows_t needs_all_rows,
             update_peer_row_hash_sets update_hash_set,
             gms::inet_address remote_node,
@@ -1711,7 +1728,7 @@ public:
     }
 
     // RPC handler
-    future<repair_rows_on_wire> get_row_diff_handler(std::unordered_set<repair_hash> set_diff, needs_all_rows_t needs_all_rows) {
+    future<repair_rows_on_wire> get_row_diff_handler(repair_hash_set set_diff, needs_all_rows_t needs_all_rows) {
         return with_gate(_gate, [this, set_diff = std::move(set_diff), needs_all_rows] () mutable {
             return get_row_diff(std::move(set_diff), needs_all_rows).then([this] (std::list<repair_row> row_diff) {
                 return to_repair_rows_on_wire(std::move(row_diff));
@@ -1721,15 +1738,16 @@ public:
 
     // RPC API
     // Send rows in the _working_row_buf with hash within the given sef_diff
-    future<> put_row_diff(std::unordered_set<repair_hash> set_diff, needs_all_rows_t needs_all_rows, gms::inet_address remote_node) {
+    future<> put_row_diff(repair_hash_set set_diff, needs_all_rows_t needs_all_rows, gms::inet_address remote_node) {
         if (!set_diff.empty()) {
             if (remote_node == _myip) {
                 return make_ready_future<>();
             }
-            auto sz = set_diff.size();
+            size_t sz = set_diff.size();
             return get_row_diff(std::move(set_diff), needs_all_rows).then([this, remote_node, sz] (std::list<repair_row> row_diff) {
                 if (row_diff.size() != sz) {
-                    throw std::runtime_error("row_diff.size() != set_diff.size()");
+                    rlogger.warn("Hash conflict detected, keyspace={}, table={}, range={}, row_diff.size={}, set_diff.size={}. It is recommended to compact the table and rerun repair for the range.",
+                            _schema->ks_name(), _schema->cf_name(), _range, row_diff.size(), sz);
                 }
                 return do_with(std::move(row_diff), [this, remote_node] (std::list<repair_row>& row_diff) {
                     return get_repair_rows_size(row_diff).then([this, remote_node, &row_diff] (size_t row_bytes) mutable {
@@ -1796,17 +1814,18 @@ private:
 
 public:
     future<> put_row_diff_with_rpc_stream(
-            std::unordered_set<repair_hash> set_diff,
+            repair_hash_set set_diff,
             needs_all_rows_t needs_all_rows,
             gms::inet_address remote_node, unsigned node_idx) {
         if (!set_diff.empty()) {
             if (remote_node == _myip) {
                 return make_ready_future<>();
             }
-            auto sz = set_diff.size();
+            size_t sz = set_diff.size();
             return get_row_diff(std::move(set_diff), needs_all_rows).then([this, remote_node, node_idx, sz] (std::list<repair_row> row_diff) {
                 if (row_diff.size() != sz) {
-                    throw std::runtime_error("row_diff.size() != set_diff.size()");
+                    rlogger.warn("Hash conflict detected, keyspace={}, table={}, range={}, row_diff.size={}, set_diff.size={}. It is recommended to compact the table and rerun repair for the range.",
+                            _schema->ks_name(), _schema->cf_name(), _range, row_diff.size(), sz);
                 }
                 return do_with(std::move(row_diff), [this, remote_node, node_idx] (std::list<repair_row>& row_diff) {
                     return get_repair_rows_size(row_diff).then([this, remote_node, node_idx, &row_diff] (size_t row_bytes) mutable {
@@ -1845,7 +1864,7 @@ static future<stop_iteration> repair_get_row_diff_with_rpc_stream_process_op(
         rpc::sink<repair_row_on_wire_with_cmd> sink,
         rpc::source<repair_hash_with_cmd> source,
         bool &error,
-        std::unordered_set<repair_hash>& current_set_diff,
+        repair_hash_set& current_set_diff,
         std::optional<std::tuple<repair_hash_with_cmd>> hash_cmd_opt) {
     repair_hash_with_cmd hash_cmd = std::get<0>(hash_cmd_opt.value());
     rlogger.trace("Got repair_hash_with_cmd from peer={}, hash={}, cmd={}", from, hash_cmd.hash, int(hash_cmd.cmd));
@@ -1858,7 +1877,7 @@ static future<stop_iteration> repair_get_row_diff_with_rpc_stream_process_op(
         }
         bool needs_all_rows = hash_cmd.cmd == repair_stream_cmd::needs_all_rows;
         _metrics.rx_hashes_nr += current_set_diff.size();
-        auto fp = make_foreign(std::make_unique<std::unordered_set<repair_hash>>(std::move(current_set_diff)));
+        auto fp = make_foreign(std::make_unique<repair_hash_set>(std::move(current_set_diff)));
         return smp::submit_to(src_cpu_id % smp::count, [from, repair_meta_id, needs_all_rows, fp = std::move(fp)] {
             auto rm = repair_meta::get_repair_meta(from, repair_meta_id);
             if (fp.get_owner_shard() == this_shard_id()) {
@@ -1936,12 +1955,12 @@ static future<stop_iteration> repair_get_full_row_hashes_with_rpc_stream_process
     if (status == repair_stream_cmd::get_full_row_hashes) {
         return smp::submit_to(src_cpu_id % smp::count, [from, repair_meta_id] {
             auto rm = repair_meta::get_repair_meta(from, repair_meta_id);
-            return rm->get_full_row_hashes_handler().then([] (std::unordered_set<repair_hash> hashes) {
+            return rm->get_full_row_hashes_handler().then([] (repair_hash_set hashes) {
                 _metrics.tx_hashes_nr += hashes.size();
                 return hashes;
             });
-        }).then([sink] (std::unordered_set<repair_hash> hashes) mutable {
-            return do_with(std::move(hashes), [sink] (std::unordered_set<repair_hash>& hashes) mutable {
+        }).then([sink] (repair_hash_set hashes) mutable {
+            return do_with(std::move(hashes), [sink] (repair_hash_set& hashes) mutable {
                 return do_for_each(hashes, [sink] (const repair_hash& hash) mutable {
                     return sink(repair_hash_with_cmd{repair_stream_cmd::hash_data, hash});
                 }).then([sink] () mutable {
@@ -1964,7 +1983,7 @@ static future<> repair_get_row_diff_with_rpc_stream_handler(
         uint32_t repair_meta_id,
         rpc::sink<repair_row_on_wire_with_cmd> sink,
         rpc::source<repair_hash_with_cmd> source) {
-    return do_with(false, std::unordered_set<repair_hash>(), [from, src_cpu_id, repair_meta_id, sink, source] (bool& error, std::unordered_set<repair_hash>& current_set_diff) mutable {
+    return do_with(false, repair_hash_set(), [from, src_cpu_id, repair_meta_id, sink, source] (bool& error, repair_hash_set& current_set_diff) mutable {
         return repeat([from, src_cpu_id, repair_meta_id, sink, source, &error, &current_set_diff] () mutable {
             return source().then([from, src_cpu_id, repair_meta_id, sink, source, &error, &current_set_diff] (std::optional<std::tuple<repair_hash_with_cmd>> hash_cmd_opt) mutable {
                 if (hash_cmd_opt) {
@@ -2107,7 +2126,7 @@ future<> repair_init_messaging_service_handler(repair_service& rs, distributed<d
             auto from = cinfo.retrieve_auxiliary<gms::inet_address>("baddr");
             return smp::submit_to(src_cpu_id % smp::count, [from, repair_meta_id] {
                 auto rm = repair_meta::get_repair_meta(from, repair_meta_id);
-                return rm->get_full_row_hashes_handler().then([] (std::unordered_set<repair_hash> hashes) {
+                return rm->get_full_row_hashes_handler().then([] (repair_hash_set hashes) {
                     _metrics.tx_hashes_nr += hashes.size();
                     return hashes;
                 });
@@ -2135,11 +2154,11 @@ future<> repair_init_messaging_service_handler(repair_service& rs, distributed<d
             });
         });
         ms.register_repair_get_row_diff([] (const rpc::client_info& cinfo, uint32_t repair_meta_id,
-                std::unordered_set<repair_hash> set_diff, bool needs_all_rows) {
+                repair_hash_set set_diff, bool needs_all_rows) {
             auto src_cpu_id = cinfo.retrieve_auxiliary<uint32_t>("src_cpu_id");
             auto from = cinfo.retrieve_auxiliary<gms::inet_address>("baddr");
             _metrics.rx_hashes_nr += set_diff.size();
-            auto fp = make_foreign(std::make_unique<std::unordered_set<repair_hash>>(std::move(set_diff)));
+            auto fp = make_foreign(std::make_unique<repair_hash_set>(std::move(set_diff)));
             return smp::submit_to(src_cpu_id % smp::count, [from, repair_meta_id, fp = std::move(fp), needs_all_rows] () mutable {
                 auto rm = repair_meta::get_repair_meta(from, repair_meta_id);
                 if (fp.get_owner_shard() == this_shard_id()) {
@@ -2207,6 +2226,25 @@ future<> repair_init_messaging_service_handler(repair_service& rs, distributed<d
     });
 }
 
+future<> repair_uninit_messaging_service_handler() {
+    return netw::get_messaging_service().invoke_on_all([] (auto& ms) {
+        return when_all_succeed(
+            ms.unregister_repair_get_row_diff_with_rpc_stream(),
+            ms.unregister_repair_put_row_diff_with_rpc_stream(),
+            ms.unregister_repair_get_full_row_hashes_with_rpc_stream(),
+            ms.unregister_repair_get_full_row_hashes(),
+            ms.unregister_repair_get_combined_row_hash(),
+            ms.unregister_repair_get_sync_boundary(),
+            ms.unregister_repair_get_row_diff(),
+            ms.unregister_repair_put_row_diff(),
+            ms.unregister_repair_row_level_start(),
+            ms.unregister_repair_row_level_stop(),
+            ms.unregister_repair_get_estimated_partitions(),
+            ms.unregister_repair_set_estimated_partitions(),
+            ms.unregister_repair_get_diff_algorithms()).discard_result();
+    });
+}
+
 class row_level_repair {
     repair_info& _ri;
     sstring _cf_name;
@@ -2439,7 +2477,7 @@ private:
             // sequentially because the rows from repair follower 1 to
             // repair master might reduce the amount of missing data
             // between repair master and repair follower 2.
-            std::unordered_set<repair_hash> set_diff = repair_meta::get_set_diff(master.peer_row_hash_sets(node_idx), master.working_row_hashes().get0());
+            repair_hash_set set_diff = repair_meta::get_set_diff(master.peer_row_hash_sets(node_idx), master.working_row_hashes().get0());
             // Request missing sets from peer node
             rlogger.debug("Before get_row_diff to node {}, local={}, peer={}, set_diff={}",
                     node, master.working_row_hashes().get0().size(), master.peer_row_hash_sets(node_idx).size(), set_diff.size());
@@ -2462,9 +2500,9 @@ private:
         // So we can figure out which rows peer node are missing and send the missing rows to them
         check_in_shutdown();
         _ri.check_in_abort();
-        std::unordered_set<repair_hash> local_row_hash_sets = master.working_row_hashes().get0();
+        repair_hash_set local_row_hash_sets = master.working_row_hashes().get0();
         auto sz = _all_live_peer_nodes.size();
-        std::vector<std::unordered_set<repair_hash>> set_diffs(sz);
+        std::vector<repair_hash_set> set_diffs(sz);
         for (size_t idx : boost::irange(size_t(0), sz)) {
             set_diffs[idx] = repair_meta::get_set_diff(local_row_hash_sets, master.peer_row_hash_sets(idx));
         }

repair/row_level.hh

@@ -45,6 +45,7 @@ private:
 };
 
 future<> repair_init_messaging_service_handler(repair_service& rs, distributed<db::system_distributed_keyspace>& sys_dist_ks, distributed<db::view::view_update_generator>& view_update_generator);
+future<> repair_uninit_messaging_service_handler();
 
 class repair_info;
 

schema.cc

@@ -19,6 +19,7 @@
  * along with Scylla.  If not, see <http://www.gnu.org/licenses/>.
  */
 
+#include <seastar/core/on_internal_error.hh>
 #include <map>
 #include "utils/UUID_gen.hh"
 #include "cql3/column_identifier.hh"
@@ -43,6 +44,8 @@
 
 constexpr int32_t schema::NAME_LENGTH;
 
+extern logging::logger dblog;
+
 sstring to_sstring(column_kind k) {
     switch (k) {
     case column_kind::partition_key:  return "PARTITION_KEY";
@@ -592,11 +595,15 @@ schema::get_column_definition(const bytes& name) const {
 
 const column_definition&
 schema::column_at(column_kind kind, column_id id) const {
-    return _raw._columns.at(column_offset(kind) + id);
+    return column_at(static_cast<ordinal_column_id>(column_offset(kind) + id));
 }
 
 const column_definition&
 schema::column_at(ordinal_column_id ordinal_id) const {
+    if (size_t(ordinal_id) >= _raw._columns.size()) [[unlikely]] {
+        on_internal_error(dblog, format("{}.{}@{}: column id {:d} >= {:d}",
+            ks_name(), cf_name(), version(), size_t(ordinal_id), _raw._columns.size()));
+    }
     return _raw._columns.at(static_cast<column_count_type>(ordinal_id));
 }
 

scripts/create-relocatable-package.py

@@ -92,7 +92,8 @@ executables = ['build/{}/scylla'.format(args.mode),
                '/usr/sbin/ethtool',
                '/usr/bin/netstat',
                '/usr/bin/hwloc-distrib',
-               '/usr/bin/hwloc-calc']
+               '/usr/bin/hwloc-calc',
+               '/usr/bin/lsblk']
 
 output = args.dest
 

scylla-gdb.py

@@ -597,7 +597,7 @@ def current_shard():
 
 
 def find_db(shard=None):
-    if not shard:
+    if shard is None:
         shard = current_shard()
     return gdb.parse_and_eval('::debug::db')['_instances']['_M_impl']['_M_start'][shard]['service']['_p']
 

scylla_post_install.sh

@@ -63,6 +63,17 @@ MemoryHigh=1200M
 MemoryMax=1400M
 MemoryLimit=1400M
 EOS
+
+# On CentOS7, systemd does not support percentage-based parameter.
+# To apply memory parameter on CentOS7, we need to override the parameter
+# in bytes, instead of percentage.
+elif [ "$RHEL" -a "$VERSION_ID" = "7" ]; then
+    MEMORY_LIMIT=$((MEMTOTAL_BYTES / 100 * 5))
+    mkdir -p /etc/systemd/system/scylla-helper.slice.d/
+    cat << EOS > /etc/systemd/system/scylla-helper.slice.d/memory.conf
+[Slice]
+MemoryLimit=$MEMORY_LIMIT
+EOS
 fi
 
 systemctl --system daemon-reload >/dev/null || true

serializer_impl.hh

@@ -25,6 +25,7 @@
 #include <seastar/util/bool_class.hh>
 #include <boost/range/algorithm/for_each.hpp>
 #include "utils/small_vector.hh"
+#include <absl/container/btree_set.h>
 
 namespace ser {
 
@@ -81,6 +82,17 @@ static inline void serialize_array(Output& out, const Container& v) {
 template<typename Container>
 struct container_traits;
 
+template<typename T>
+struct container_traits<absl::btree_set<T>> {
+    struct back_emplacer {
+        absl::btree_set<T>& c;
+        back_emplacer(absl::btree_set<T>& c_) : c(c_) {}
+        void operator()(T&& v) {
+            c.emplace(std::move(v));
+        }
+    };
+};
+
 template<typename T>
 struct container_traits<std::unordered_set<T>> {
     struct back_emplacer {
@@ -253,6 +265,27 @@ struct serializer<std::list<T>> {
     }
 };
 
+template<typename T>
+struct serializer<absl::btree_set<T>> {
+    template<typename Input>
+    static absl::btree_set<T> read(Input& in) {
+        auto sz = deserialize(in, boost::type<uint32_t>());
+        absl::btree_set<T> v;
+        deserialize_array_helper<false, T>::doit(in, v, sz);
+        return v;
+    }
+    template<typename Output>
+    static void write(Output& out, const absl::btree_set<T>& v) {
+        safe_serialize_as_uint32(out, v.size());
+        serialize_array_helper<false, T>::doit(out, v);
+    }
+    template<typename Input>
+    static void skip(Input& in) {
+        auto sz = deserialize(in, boost::type<uint32_t>());
+        skip_array<T>(in, sz);
+    }
+};
+
 template<typename T>
 struct serializer<std::unordered_set<T>> {
     template<typename Input>

service/storage_proxy.cc

@@ -1760,6 +1760,7 @@ storage_proxy::storage_proxy(distributed<database>& db, storage_proxy::config cf
     , _token_metadata(tm)
     , _read_smp_service_group(cfg.read_smp_service_group)
     , _write_smp_service_group(cfg.write_smp_service_group)
+    , _hints_write_smp_service_group(cfg.hints_write_smp_service_group)
     , _write_ack_smp_service_group(cfg.write_ack_smp_service_group)
     , _next_response_id(std::chrono::system_clock::now().time_since_epoch()/1ms)
     , _hints_resource_manager(cfg.available_memory / 10)
@@ -1803,39 +1804,48 @@ storage_proxy::response_id_type storage_proxy::unique_response_handler::release(
 }
 
 future<>
-storage_proxy::mutate_locally(const mutation& m, tracing::trace_state_ptr tr_state, clock_type::time_point timeout) {
+storage_proxy::mutate_locally(const mutation& m, tracing::trace_state_ptr tr_state, db::commitlog::force_sync sync, clock_type::time_point timeout, smp_service_group smp_grp) {
     auto shard = _db.local().shard_of(m);
     get_stats().replica_cross_shard_ops += shard != this_shard_id();
-    return _db.invoke_on(shard, {_write_smp_service_group, timeout},
-            [s = global_schema_ptr(m.schema()), m = freeze(m), gtr = tracing::global_trace_state_ptr(std::move(tr_state)), timeout] (database& db) mutable -> future<> {
-        return db.apply(s, m, gtr.get(), db::commitlog::force_sync::no, timeout);
+    return _db.invoke_on(shard, {smp_grp, timeout},
+            [s = global_schema_ptr(m.schema()),
+             m = freeze(m),
+             gtr = tracing::global_trace_state_ptr(std::move(tr_state)),
+             timeout,
+             sync] (database& db) mutable -> future<> {
+        return db.apply(s, m, gtr.get(), sync, timeout);
     });
 }
 
 future<>
-storage_proxy::mutate_locally(const schema_ptr& s, const frozen_mutation& m, tracing::trace_state_ptr tr_state, db::commitlog::force_sync sync, clock_type::time_point timeout) {
+storage_proxy::mutate_locally(const schema_ptr& s, const frozen_mutation& m, tracing::trace_state_ptr tr_state, db::commitlog::force_sync sync, clock_type::time_point timeout,
+        smp_service_group smp_grp) {
     auto shard = _db.local().shard_of(m);
     get_stats().replica_cross_shard_ops += shard != this_shard_id();
-    return _db.invoke_on(shard, {_write_smp_service_group, timeout},
+    return _db.invoke_on(shard, {smp_grp, timeout},
             [&m, gs = global_schema_ptr(s), gtr = tracing::global_trace_state_ptr(std::move(tr_state)), timeout, sync] (database& db) mutable -> future<> {
         return db.apply(gs, m, gtr.get(), sync, timeout);
     });
 }
 
 future<>
-storage_proxy::mutate_locally(std::vector<mutation> mutations, tracing::trace_state_ptr tr_state, clock_type::time_point timeout) {
-    return do_with(std::move(mutations), [this, timeout, tr_state = std::move(tr_state)] (std::vector<mutation>& pmut) mutable {
-        return parallel_for_each(pmut.begin(), pmut.end(), [this, tr_state = std::move(tr_state), timeout] (const mutation& m) mutable {
-            return mutate_locally(m, tr_state, timeout);
+storage_proxy::mutate_locally(std::vector<mutation> mutations, tracing::trace_state_ptr tr_state, clock_type::time_point timeout, smp_service_group smp_grp) {
+    return do_with(std::move(mutations), [this, timeout, tr_state = std::move(tr_state), smp_grp] (std::vector<mutation>& pmut) mutable {
+        return parallel_for_each(pmut.begin(), pmut.end(), [this, tr_state = std::move(tr_state), timeout, smp_grp] (const mutation& m) mutable {
+            return mutate_locally(m, tr_state, db::commitlog::force_sync::no, timeout, smp_grp);
         });
     });
 }
 
+future<> 
+storage_proxy::mutate_locally(std::vector<mutation> mutation, tracing::trace_state_ptr tr_state, clock_type::time_point timeout) {
+        return mutate_locally(std::move(mutation), tr_state, timeout, _write_smp_service_group);
+}
 future<>
 storage_proxy::mutate_hint(const schema_ptr& s, const frozen_mutation& m, tracing::trace_state_ptr tr_state, clock_type::time_point timeout) {
     auto shard = _db.local().shard_of(m);
     get_stats().replica_cross_shard_ops += shard != this_shard_id();
-    return _db.invoke_on(shard, {_write_smp_service_group, timeout}, [&m, gs = global_schema_ptr(s), tr_state = std::move(tr_state), timeout] (database& db) mutable -> future<> {
+    return _db.invoke_on(shard, {_hints_write_smp_service_group, timeout}, [&m, gs = global_schema_ptr(s), tr_state = std::move(tr_state), timeout] (database& db) mutable -> future<> {
         return db.apply_hint(gs, m, std::move(tr_state), timeout);
     });
 }
@@ -4849,7 +4859,7 @@ void storage_proxy::init_messaging_service() {
         });
     };
 
-    auto receive_mutation_handler = [] (const rpc::client_info& cinfo, rpc::opt_time_point t, frozen_mutation in, std::vector<gms::inet_address> forward,
+    auto receive_mutation_handler = [] (smp_service_group smp_grp, const rpc::client_info& cinfo, rpc::opt_time_point t, frozen_mutation in, std::vector<gms::inet_address> forward,
             gms::inet_address reply_to, unsigned shard, storage_proxy::response_id_type response_id, rpc::optional<std::optional<tracing::trace_info>> trace_info) {
         tracing::trace_state_ptr trace_state_ptr;
         auto src_addr = netw::messaging_service::get_source(cinfo);
@@ -4857,9 +4867,9 @@ void storage_proxy::init_messaging_service() {
         utils::UUID schema_version = in.schema_version();
         return handle_write(src_addr, t, schema_version, std::move(in), std::move(forward), reply_to, shard, response_id,
                 trace_info ? *trace_info : std::nullopt,
-                /* apply_fn */ [] (shared_ptr<storage_proxy>& p, tracing::trace_state_ptr tr_state, schema_ptr s, const frozen_mutation& m,
+                /* apply_fn */ [smp_grp] (shared_ptr<storage_proxy>& p, tracing::trace_state_ptr tr_state, schema_ptr s, const frozen_mutation& m,
                         clock_type::time_point timeout) {
-                    return p->mutate_locally(std::move(s), m, std::move(tr_state), db::commitlog::force_sync::no, timeout);
+                    return p->mutate_locally(std::move(s), m, std::move(tr_state), db::commitlog::force_sync::no, timeout, smp_grp);
                 },
                 /* forward_fn */ [] (netw::messaging_service::msg_addr addr, clock_type::time_point timeout, const frozen_mutation& m,
                         gms::inet_address reply_to, unsigned shard, response_id_type response_id,
@@ -4868,8 +4878,8 @@ void storage_proxy::init_messaging_service() {
                     return ms.send_mutation(addr, timeout, m, {}, reply_to, shard, response_id, std::move(trace_info));
                 });
     };
-    ms.register_mutation(receive_mutation_handler);
-    ms.register_hint_mutation(receive_mutation_handler);
+    ms.register_mutation(std::bind_front<>(receive_mutation_handler, _write_smp_service_group));
+    ms.register_hint_mutation(std::bind_front<>(receive_mutation_handler, _hints_write_smp_service_group));
 
     ms.register_paxos_learn([] (const rpc::client_info& cinfo, rpc::opt_time_point t, paxos::proposal decision,
             std::vector<gms::inet_address> forward, gms::inet_address reply_to, unsigned shard,
@@ -5112,18 +5122,22 @@ void storage_proxy::init_messaging_service() {
 future<> storage_proxy::uninit_messaging_service() {
     auto& ms = netw::get_local_messaging_service();
     return when_all_succeed(
+        ms.unregister_counter_mutation(),
         ms.unregister_mutation(),
+        ms.unregister_hint_mutation(),
         ms.unregister_mutation_done(),
         ms.unregister_mutation_failed(),
         ms.unregister_read_data(),
         ms.unregister_read_mutation_data(),
         ms.unregister_read_digest(),
         ms.unregister_truncate(),
+        ms.unregister_get_schema_version(),
         ms.unregister_paxos_prepare(),
         ms.unregister_paxos_accept(),
         ms.unregister_paxos_learn(),
         ms.unregister_paxos_prune()
     ).discard_result();
+
 }
 
 future<rpc::tuple<foreign_ptr<lw_shared_ptr<reconcilable_result>>, cache_temperature>>
@@ -5217,8 +5231,7 @@ future<> storage_proxy::drain_on_shutdown() {
 
 future<>
 storage_proxy::stop() {
-    // FIXME: hints manager should be stopped here but it seems like this function is never called
-    return uninit_messaging_service();
+    return make_ready_future<>();
 }
 
 }

service/storage_proxy.hh

@@ -166,6 +166,7 @@ public:
         size_t available_memory;
         smp_service_group read_smp_service_group = default_smp_service_group();
         smp_service_group write_smp_service_group = default_smp_service_group();
+        smp_service_group hints_write_smp_service_group = default_smp_service_group();
         // Write acknowledgments might not be received on the correct shard, and
         // they need a separate smp_service_group to prevent an ABBA deadlock
         // with writes.
@@ -256,6 +257,7 @@ private:
     locator::token_metadata& _token_metadata;
     smp_service_group _read_smp_service_group;
     smp_service_group _write_smp_service_group;
+    smp_service_group _hints_write_smp_service_group;
     smp_service_group _write_ack_smp_service_group;
     response_id_type _next_response_id;
     response_handlers_map _response_handlers;
@@ -314,7 +316,6 @@ private:
 
     cdc_stats _cdc_stats;
 private:
-    future<> uninit_messaging_service();
     future<coordinator_query_result> query_singular(lw_shared_ptr<query::read_command> cmd,
             dht::partition_range_vector&& partition_ranges,
             db::consistency_level cl,
@@ -469,13 +470,31 @@ public:
         return next;
     }
     void init_messaging_service();
+    future<> uninit_messaging_service();
 
+private:
     // Applies mutation on this node.
     // Resolves with timed_out_error when timeout is reached.
-    future<> mutate_locally(const mutation& m, tracing::trace_state_ptr tr_state, clock_type::time_point timeout = clock_type::time_point::max());
+    future<> mutate_locally(const mutation& m, tracing::trace_state_ptr tr_state, db::commitlog::force_sync sync, clock_type::time_point timeout, smp_service_group smp_grp);
     // Applies mutation on this node.
     // Resolves with timed_out_error when timeout is reached.
-    future<> mutate_locally(const schema_ptr&, const frozen_mutation& m, tracing::trace_state_ptr tr_state, db::commitlog::force_sync sync, clock_type::time_point timeout = clock_type::time_point::max());
+    future<> mutate_locally(const schema_ptr&, const frozen_mutation& m, tracing::trace_state_ptr tr_state, db::commitlog::force_sync sync, clock_type::time_point timeout,
+            smp_service_group smp_grp);
+    // Applies mutations on this node.
+    // Resolves with timed_out_error when timeout is reached.
+    future<> mutate_locally(std::vector<mutation> mutation, tracing::trace_state_ptr tr_state, clock_type::time_point timeout, smp_service_group smp_grp);
+
+public:
+    // Applies mutation on this node.
+    // Resolves with timed_out_error when timeout is reached.
+    future<> mutate_locally(const mutation& m, tracing::trace_state_ptr tr_state, db::commitlog::force_sync sync, clock_type::time_point timeout = clock_type::time_point::max()) {
+        return mutate_locally(m, tr_state, sync, timeout, _write_smp_service_group);
+    }
+    // Applies mutation on this node.
+    // Resolves with timed_out_error when timeout is reached.
+    future<> mutate_locally(const schema_ptr& s, const frozen_mutation& m, tracing::trace_state_ptr tr_state, db::commitlog::force_sync sync, clock_type::time_point timeout = clock_type::time_point::max()) {
+        return mutate_locally(s, m, tr_state, sync, timeout, _write_smp_service_group);
+    }
     // Applies mutations on this node.
     // Resolves with timed_out_error when timeout is reached.
     future<> mutate_locally(std::vector<mutation> mutation, tracing::trace_state_ptr tr_state, clock_type::time_point timeout = clock_type::time_point::max());

service/storage_service.cc

@@ -369,6 +369,9 @@ void storage_service::prepare_to_join(std::vector<inet_address> loaded_endpoints
         app_states.emplace(gms::application_state::CDC_STREAMS_TIMESTAMP, versioned_value::cdc_streams_timestamp(_cdc_streams_ts));
         app_states.emplace(gms::application_state::STATUS, versioned_value::normal(my_tokens));
     }
+    if (replacing_a_node_with_same_ip || replacing_a_node_with_diff_ip) {
+        app_states.emplace(gms::application_state::TOKENS, versioned_value::tokens(_bootstrap_tokens));
+    }
     slogger.info("Starting up server gossip");
 
     auto generation_number = db::system_keyspace::increment_and_get_generation().get0();
@@ -698,6 +701,8 @@ bool storage_service::do_handle_cdc_generation_intercept_nonfatal_errors(db_cloc
         throw cdc_generation_handling_nonfatal_exception(e.what());
     } catch (exceptions::unavailable_exception& e) {
         throw cdc_generation_handling_nonfatal_exception(e.what());
+    } catch (exceptions::read_failure_exception& e) {
+        throw cdc_generation_handling_nonfatal_exception(e.what());
     } catch (...) {
         const auto ep = std::current_exception();
         if (is_timeout_exception(ep)) {
@@ -890,12 +895,14 @@ future<> storage_service::check_and_repair_cdc_streams() {
             cdc_log.error("Aborting CDC generation repair due to missing STATUS");
             return;
         }
+        // Update _cdc_streams_ts first, so that do_handle_cdc_generation (which will get called due to the status update)
+        // won't try to update the gossiper, which would result in a deadlock inside add_local_application_state
+        _cdc_streams_ts = new_streams_ts;
         _gossiper.add_local_application_state({
                 { gms::application_state::CDC_STREAMS_TIMESTAMP, versioned_value::cdc_streams_timestamp(new_streams_ts) },
                 { gms::application_state::STATUS, *status }
         }).get();
         db::system_keyspace::update_cdc_streams_timestamp(new_streams_ts).get();
-        _cdc_streams_ts = new_streams_ts;
     });
 }
 
@@ -1884,9 +1891,11 @@ future<std::map<gms::inet_address, float>> storage_service::effective_ownership(
         return do_with(dht::token::describe_ownership(ss._token_metadata.sorted_tokens()),
                 ss._token_metadata.get_topology().get_datacenter_endpoints(),
                 std::map<gms::inet_address, float>(),
-                [&ss, keyspace_name](const std::map<token, float>& token_ownership, std::unordered_map<sstring,
+                std::move(keyspace_name),
+                [&ss](const std::map<token, float>& token_ownership, std::unordered_map<sstring,
                         std::unordered_set<gms::inet_address>>& datacenter_endpoints,
-                        std::map<gms::inet_address, float>& final_ownership) {
+                        std::map<gms::inet_address, float>& final_ownership,
+                        sstring& keyspace_name) {
             return do_for_each(datacenter_endpoints, [&ss, &keyspace_name, &final_ownership, &token_ownership](std::pair<sstring,std::unordered_set<inet_address>>&& endpoints) mutable {
                 return do_with(std::unordered_set<inet_address>(endpoints.second), [&ss, &keyspace_name, &final_ownership, &token_ownership](const std::unordered_set<inet_address>& endpoints_map) mutable {
                     return do_for_each(endpoints_map, [&ss, &keyspace_name, &final_ownership, &token_ownership](const gms::inet_address& endpoint) mutable {

sstables/compaction.cc

@@ -602,7 +602,7 @@ private:
         // - add support to merge summary (message: Partition merge counts were {%s}.).
         // - there is no easy way, currently, to know the exact number of total partitions.
         // By the time being, using estimated key count.
-        sstring formatted_msg = fmt::format("{} sstables to [{}]. {} to {} (~{} of original) in {}ms = {}. " \
+        sstring formatted_msg = fmt::format("{} sstables to [{}]. {} to {} (~{}% of original) in {}ms = {}. " \
             "~{} total partitions merged to {}.",
             _info->sstables, new_sstables_msg, pretty_printed_data_size(_info->start_size), pretty_printed_data_size(_info->end_size), int(ratio * 100),
             std::chrono::duration_cast<std::chrono::milliseconds>(duration).count(), pretty_printed_throughput(_info->end_size, duration),
@@ -1236,11 +1236,8 @@ private:
     // return estimated partitions per sstable for a given shard
     uint64_t partitions_per_sstable(shard_id s) const {
         uint64_t estimated_sstables = std::max(uint64_t(1), uint64_t(ceil(double(_estimation_per_shard[s].estimated_size) / _max_sstable_size)));
-        // As we adjust this estimate downwards from the compaction strategy, it can get to 0 so
-        // make sure we're returning at least 1.
-        return std::max(uint64_t(1),
-                std::min(uint64_t(ceil(double(_estimation_per_shard[s].estimated_partitions) / estimated_sstables)),
-                _cf.get_compaction_strategy().adjust_partition_estimate(_ms_metadata, _estimation_per_shard[s].estimated_partitions)));
+        return std::min(uint64_t(ceil(double(_estimation_per_shard[s].estimated_partitions) / estimated_sstables)),
+                _cf.get_compaction_strategy().adjust_partition_estimate(_ms_metadata, _estimation_per_shard[s].estimated_partitions));
     }
 public:
     resharding_compaction(column_family& cf, sstables::compaction_descriptor descriptor)

sstables/compaction_backlog_manager.hh

@@ -92,6 +92,9 @@ public:
     void transfer_ongoing_charges(compaction_backlog_tracker& new_bt, bool move_read_charges = true);
     void revert_charges(sstables::shared_sstable sst);
 private:
+    // Returns true if this SSTable can be added or removed from the tracker.
+    bool sstable_belongs_to_tracker(const sstables::shared_sstable& sst);
+
     void disable() {
         _disabled = true;
         _ongoing_writes = {};

sstables/compaction_manager.cc

@@ -218,7 +218,7 @@ std::vector<sstables::shared_sstable> compaction_manager::get_candidates(const c
     auto& cs = cf.get_compaction_strategy();
 
     // Filter out sstables that are being compacted.
-    for (auto& sst : cf.candidates_for_compaction()) {
+    for (auto& sst : cf.non_staging_sstables()) {
         if (_compacting_sstables.count(sst)) {
             continue;
         }
@@ -708,8 +708,8 @@ future<> compaction_manager::rewrite_sstables(column_family* cf, sstables::compa
     return task->compaction_done.get_future().then([task] {});
 }
 
-static bool needs_cleanup(const sstables::shared_sstable& sst,
-                   const dht::token_range_vector& owned_ranges,
+bool needs_cleanup(const sstables::shared_sstable& sst,
+                   const dht::token_range_vector& sorted_owned_ranges,
                    schema_ptr s) {
     auto first = sst->get_first_partition_key();
     auto last = sst->get_last_partition_key();
@@ -717,29 +717,40 @@ static bool needs_cleanup(const sstables::shared_sstable& sst,
     auto last_token = dht::get_token(*s, last);
     dht::token_range sst_token_range = dht::token_range::make(first_token, last_token);
 
+    auto r = std::lower_bound(sorted_owned_ranges.begin(), sorted_owned_ranges.end(), first_token,
+            [] (const range<dht::token>& a, const dht::token& b) {
+        // check that range a is before token b.
+        return a.after(b, dht::token_comparator());
+    });
+
     // return true iff sst partition range isn't fully contained in any of the owned ranges.
-    for (auto& r : owned_ranges) {
-        if (r.contains(sst_token_range, dht::token_comparator())) {
+    if (r != sorted_owned_ranges.end()) {
+        if (r->contains(sst_token_range, dht::token_comparator())) {
             return false;
         }
     }
     return true;
 }
 
-future<> compaction_manager::perform_cleanup(column_family* cf) {
+future<> compaction_manager::perform_cleanup(database& db, column_family* cf) {
     if (check_for_cleanup(cf)) {
         throw std::runtime_error(format("cleanup request failed: there is an ongoing cleanup on {}.{}",
             cf->schema()->ks_name(), cf->schema()->cf_name()));
     }
-    return rewrite_sstables(cf, sstables::compaction_options::make_cleanup(), [this] (const table& table) {
-        auto schema = table.schema();
-        auto owned_ranges = service::get_local_storage_service().get_local_ranges(schema->ks_name());
+    return seastar::async([this, cf, &db] {
+        auto schema = cf->schema();
+        auto& rs = db.find_keyspace(schema->ks_name()).get_replication_strategy();
+        auto sorted_owned_ranges = rs.get_ranges_in_thread(utils::fb_utilities::get_broadcast_address());
         auto sstables = std::vector<sstables::shared_sstable>{};
-        const auto candidates = table.candidates_for_compaction();
-        std::copy_if(candidates.begin(), candidates.end(), std::back_inserter(sstables), [&owned_ranges, schema] (const sstables::shared_sstable& sst) {
-            return owned_ranges.empty() || needs_cleanup(sst, owned_ranges, schema);
+        const auto candidates = get_candidates(*cf);
+        std::copy_if(candidates.begin(), candidates.end(), std::back_inserter(sstables), [&sorted_owned_ranges, schema] (const sstables::shared_sstable& sst) {
+            seastar::thread::maybe_yield();
+            return sorted_owned_ranges.empty() || needs_cleanup(sst, sorted_owned_ranges, schema);
         });
         return sstables;
+    }).then([this, cf] (std::vector<sstables::shared_sstable> sstables) {
+        return rewrite_sstables(cf, sstables::compaction_options::make_cleanup(),
+                [sstables = std::move(sstables)] (const table&) { return sstables; });
     });
 }
 
@@ -754,7 +765,7 @@ future<> compaction_manager::perform_sstable_upgrade(column_family* cf, bool exc
         return cf->run_with_compaction_disabled([this, cf, &tables, exclude_current_version] {
             auto last_version = cf->get_sstables_manager().get_highest_supported_format();
 
-            for (auto& sst : cf->candidates_for_compaction()) {
+            for (auto& sst : get_candidates(*cf)) {
                 // if we are a "normal" upgrade, we only care about
                 // tables with older versions, but potentially
                 // we are to actually rewrite everything. (-a)
@@ -779,8 +790,8 @@ future<> compaction_manager::perform_sstable_upgrade(column_family* cf, bool exc
 
 // Submit a column family to be scrubbed and wait for its termination.
 future<> compaction_manager::perform_sstable_scrub(column_family* cf, bool skip_corrupted) {
-    return rewrite_sstables(cf, sstables::compaction_options::make_scrub(skip_corrupted), [] (const table& cf) {
-        return cf.candidates_for_compaction();
+    return rewrite_sstables(cf, sstables::compaction_options::make_scrub(skip_corrupted), [this] (const table& cf) {
+        return get_candidates(cf);
     });
 }
 
@@ -857,7 +868,7 @@ double compaction_backlog_tracker::backlog() const {
 }
 
 void compaction_backlog_tracker::add_sstable(sstables::shared_sstable sst) {
-    if (_disabled) {
+    if (_disabled || !sstable_belongs_to_tracker(sst)) {
         return;
     }
     _ongoing_writes.erase(sst);
@@ -870,7 +881,7 @@ void compaction_backlog_tracker::add_sstable(sstables::shared_sstable sst) {
 }
 
 void compaction_backlog_tracker::remove_sstable(sstables::shared_sstable sst) {
-    if (_disabled) {
+    if (_disabled || !sstable_belongs_to_tracker(sst)) {
         return;
     }
 
@@ -883,6 +894,10 @@ void compaction_backlog_tracker::remove_sstable(sstables::shared_sstable sst) {
     }
 }
 
+bool compaction_backlog_tracker::sstable_belongs_to_tracker(const sstables::shared_sstable& sst) {
+    return !sst->requires_view_building();
+}
+
 void compaction_backlog_tracker::register_partially_written_sstable(sstables::shared_sstable sst, backlog_write_progress_manager& wp) {
     if (_disabled) {
         return;

sstables/compaction_manager.hh

@@ -205,7 +205,7 @@ public:
     // Cleanup is about discarding keys that are no longer relevant for a
     // given sstable, e.g. after node loses part of its token range because
     // of a newly added node.
-    future<> perform_cleanup(column_family* cf);
+    future<> perform_cleanup(database& db, column_family* cf);
 
     // Submit a column family to be upgraded and wait for its termination.
     future<> perform_sstable_upgrade(column_family* cf, bool exclude_current_version);
@@ -271,3 +271,5 @@ public:
     friend class compaction_weight_registration;
 };
 
+bool needs_cleanup(const sstables::shared_sstable& sst, const dht::token_range_vector& owned_ranges, schema_ptr s);
+

sstables/compaction_strategy.cc

@@ -438,8 +438,8 @@ std::unique_ptr<sstable_set_impl> leveled_compaction_strategy::make_sstable_set(
     return std::make_unique<partitioned_sstable_set>(std::move(schema));
 }
 
-std::unique_ptr<sstable_set_impl> make_partitioned_sstable_set(schema_ptr schema, bool use_level_metadata) {
-    return std::make_unique<partitioned_sstable_set>(std::move(schema), use_level_metadata);
+sstable_set make_partitioned_sstable_set(schema_ptr schema, lw_shared_ptr<sstable_list> all, bool use_level_metadata) {
+    return sstables::sstable_set(std::make_unique<partitioned_sstable_set>(schema, use_level_metadata), schema, std::move(all));
 }
 
 compaction_descriptor compaction_strategy_impl::get_major_compaction_job(column_family& cf, std::vector<sstables::shared_sstable> candidates) {

sstables/index_reader.hh

@@ -453,9 +453,16 @@ private:
                     auto indexes = std::move(entries_reader->_consumer.indexes);
                     return entries_reader->_context.close().then([indexes = std::move(indexes), ex = std::move(ex)] () mutable {
                         if (ex) {
-                            std::rethrow_exception(std::move(ex));
+                            return do_with(std::move(indexes), [ex = std::move(ex)] (index_list& indexes) mutable {
+                                return parallel_for_each(indexes, [] (index_entry& ie) mutable {
+                                    return ie.close_pi_stream();
+                                }).then_wrapped([ex = std::move(ex)] (future<>&& fut) mutable {
+                                    fut.ignore_ready_future();
+                                    return make_exception_future<index_list>(std::move(ex));
+                                });
+                            });
                         }
-                        return std::move(indexes);
+                        return make_ready_future<index_list>(std::move(indexes));
                     });
 
                 });

sstables/leveled_compaction_strategy.cc

@@ -178,7 +178,13 @@ leveled_compaction_strategy::get_reshaping_job(std::vector<shared_sstable> input
 
     size_t offstrategy_threshold = std::max(schema->min_compaction_threshold(), 4);
     size_t max_sstables = std::max(schema->max_compaction_threshold(), int(offstrategy_threshold));
-    unsigned tolerance = mode == reshape_mode::strict ? 0 : leveled_manifest::leveled_fan_out * 2;
+    auto tolerance = [mode] (unsigned level) -> unsigned {
+        if (mode == reshape_mode::strict) {
+            return 0;
+        }
+        constexpr unsigned fan_out = leveled_manifest::leveled_fan_out;
+        return std::max(double(fan_out), std::ceil(std::pow(fan_out, level) * 0.1));
+    };
 
     if (level_info[0].size() > offstrategy_threshold) {
         level_info[0].resize(std::min(level_info[0].size(), max_sstables));
@@ -193,7 +199,7 @@ leveled_compaction_strategy::get_reshaping_job(std::vector<shared_sstable> input
         }
         max_filled_level = std::max(max_filled_level, level);
 
-        if (!is_disjoint(level_info[level], tolerance)) {
+        if (!is_disjoint(level_info[level], tolerance(level))) {
             leveled_manifest::logger.warn("Turns out that level {} is not disjoint, so compacting everything on behalf of {}.{}", level, schema->ks_name(), schema->cf_name());
             // Unfortunately no good limit to limit input size to max_sstables for LCS major
             compaction_descriptor desc(std::move(input), std::optional<sstables::sstable_set>(), iop, max_filled_level, _max_sstable_size_in_mb * 1024 * 1024);

sstables/mc/writer.cc

@@ -741,6 +741,11 @@ public:
         , _run_identifier(cfg.run_identifier)
         , _write_regular_as_static(cfg.correctly_serialize_static_compact_in_mc && s.is_static_compact_table())
     {
+        // This can be 0 in some cases, which is albeit benign, can wreak havoc
+        // in lower-level writer code, so clamp it to [1, +inf) here, which is
+        // exactly what callers used to do anyway.
+        estimated_partitions = std::max(uint64_t(1), estimated_partitions);
+
         _sst.generate_toc(_schema.get_compressor_params().get_compressor(), _schema.bloom_filter_fp_chance());
         _sst.write_toc(_pc);
         _sst.create_data().get();

sstables/size_tiered_compaction_strategy.cc

@@ -27,7 +27,7 @@
 namespace sstables {
 
 std::vector<std::pair<sstables::shared_sstable, uint64_t>>
-size_tiered_compaction_strategy::create_sstable_and_length_pairs(const std::vector<sstables::shared_sstable>& sstables) const {
+size_tiered_compaction_strategy::create_sstable_and_length_pairs(const std::vector<sstables::shared_sstable>& sstables) {
 
     std::vector<std::pair<sstables::shared_sstable, uint64_t>> sstable_length_pairs;
     sstable_length_pairs.reserve(sstables.size());
@@ -43,7 +43,7 @@ size_tiered_compaction_strategy::create_sstable_and_length_pairs(const std::vect
 }
 
 std::vector<std::vector<sstables::shared_sstable>>
-size_tiered_compaction_strategy::get_buckets(const std::vector<sstables::shared_sstable>& sstables) const {
+size_tiered_compaction_strategy::get_buckets(const std::vector<sstables::shared_sstable>& sstables, size_tiered_compaction_strategy_options options) {
     // sstables sorted by size of its data file.
     auto sorted_sstables = create_sstable_and_length_pairs(sstables);
 
@@ -64,8 +64,8 @@ size_tiered_compaction_strategy::get_buckets(const std::vector<sstables::shared_
         for (auto it = buckets.begin(); it != buckets.end(); it++) {
             size_t old_average_size = it->first;
 
-            if ((size > (old_average_size * _options.bucket_low) && size < (old_average_size * _options.bucket_high)) ||
-                    (size < _options.min_sstable_size && old_average_size < _options.min_sstable_size)) {
+            if ((size > (old_average_size * options.bucket_low) && size < (old_average_size * options.bucket_high)) ||
+                    (size < options.min_sstable_size && old_average_size < options.min_sstable_size)) {
                 auto bucket = std::move(it->second);
                 size_t total_size = bucket.size() * old_average_size;
                 size_t new_average_size = (total_size + size) / (bucket.size() + 1);
@@ -97,6 +97,11 @@ size_tiered_compaction_strategy::get_buckets(const std::vector<sstables::shared_
     return bucket_list;
 }
 
+std::vector<std::vector<sstables::shared_sstable>>
+size_tiered_compaction_strategy::get_buckets(const std::vector<sstables::shared_sstable>& sstables) const {
+    return get_buckets(sstables, _options);
+}
+
 std::vector<sstables::shared_sstable>
 size_tiered_compaction_strategy::most_interesting_bucket(std::vector<std::vector<sstables::shared_sstable>> buckets,
         unsigned min_threshold, unsigned max_threshold)
@@ -176,23 +181,28 @@ size_tiered_compaction_strategy::get_sstables_for_compaction(column_family& cfs,
     return sstables::compaction_descriptor();
 }
 
+int64_t size_tiered_compaction_strategy::estimated_pending_compactions(const std::vector<sstables::shared_sstable>& sstables,
+        int min_threshold, int max_threshold, size_tiered_compaction_strategy_options options) {
+    int64_t n = 0;
+    for (auto& bucket : get_buckets(sstables, options)) {
+        if (bucket.size() >= size_t(min_threshold)) {
+            n += std::ceil(double(bucket.size()) / max_threshold);
+        }
+    }
+    return n;
+}
+
 int64_t size_tiered_compaction_strategy::estimated_pending_compactions(column_family& cf) const {
     int min_threshold = cf.min_compaction_threshold();
     int max_threshold = cf.schema()->max_compaction_threshold();
     std::vector<sstables::shared_sstable> sstables;
-    int64_t n = 0;
 
     sstables.reserve(cf.sstables_count());
     for (auto& entry : *cf.get_sstables()) {
         sstables.push_back(entry);
     }
 
-    for (auto& bucket : get_buckets(sstables)) {
-        if (bucket.size() >= size_t(min_threshold)) {
-            n += std::ceil(double(bucket.size()) / max_threshold);
-        }
-    }
-    return n;
+    return estimated_pending_compactions(sstables, min_threshold, max_threshold, _options);
 }
 
 std::vector<sstables::shared_sstable>

sstables/size_tiered_compaction_strategy.hh

@@ -116,9 +116,11 @@ class size_tiered_compaction_strategy : public compaction_strategy_impl {
     compaction_backlog_tracker _backlog_tracker;
 
     // Return a list of pair of shared_sstable and its respective size.
-    std::vector<std::pair<sstables::shared_sstable, uint64_t>> create_sstable_and_length_pairs(const std::vector<sstables::shared_sstable>& sstables) const;
+    static std::vector<std::pair<sstables::shared_sstable, uint64_t>> create_sstable_and_length_pairs(const std::vector<sstables::shared_sstable>& sstables);
 
     // Group files of similar size into buckets.
+    static std::vector<std::vector<sstables::shared_sstable>> get_buckets(const std::vector<sstables::shared_sstable>& sstables, size_tiered_compaction_strategy_options options);
+
     std::vector<std::vector<sstables::shared_sstable>> get_buckets(const std::vector<sstables::shared_sstable>& sstables) const;
 
     // Maybe return a bucket of sstables to compact
@@ -154,6 +156,8 @@ public:
 
     virtual compaction_descriptor get_sstables_for_compaction(column_family& cfs, std::vector<sstables::shared_sstable> candidates) override;
 
+    static int64_t estimated_pending_compactions(const std::vector<sstables::shared_sstable>& sstables,
+        int min_threshold, int max_threshold, size_tiered_compaction_strategy_options options);
     virtual int64_t estimated_pending_compactions(column_family& cf) const override;
 
     virtual compaction_strategy_type type() const {

sstables/sstable_set.hh

@@ -101,7 +101,7 @@ public:
     incremental_selector make_incremental_selector() const;
 };
 
-std::unique_ptr<sstable_set_impl> make_partitioned_sstable_set(schema_ptr schema, bool use_level_metadata = true);
+sstable_set make_partitioned_sstable_set(schema_ptr schema, lw_shared_ptr<sstable_list> all, bool use_level_metadata = true);
 
 std::ostream& operator<<(std::ostream& os, const sstables::sstable_run& run);
 

sstables/sstables.cc

@@ -2012,6 +2012,11 @@ components_writer::components_writer(sstable& sst, const schema& s, file_writer&
     , _tombstone_written(false)
     , _range_tombstones(s)
 {
+    // This can be 0 in some cases, which is albeit benign, can wreak havoc
+    // in lower-level writer code, so clamp it to [1, +inf) here, which is
+    // exactly what callers used to do anyway.
+    estimated_partitions = std::max(uint64_t(1), estimated_partitions);
+
     _sst._components->filter = utils::i_filter::get_filter(estimated_partitions, _schema.bloom_filter_fp_chance(), utils::filter_format::k_l_format);
     _sst._pi_write.desired_block_size = cfg.promoted_index_block_size;
     _sst._correctly_serialize_non_compound_range_tombstones = cfg.correctly_serialize_non_compound_range_tombstones;

sstables/time_window_compaction_strategy.cc

@@ -125,4 +125,194 @@ time_window_compaction_strategy::get_reshaping_job(std::vector<shared_sstable> i
     return compaction_descriptor();
 }
 
+compaction_descriptor
+time_window_compaction_strategy::get_sstables_for_compaction(column_family& cf, std::vector<shared_sstable> candidates) {
+    auto gc_before = gc_clock::now() - cf.schema()->gc_grace_seconds();
+
+    if (candidates.empty()) {
+        return compaction_descriptor();
+    }
+
+    // Find fully expired SSTables. Those will be included no matter what.
+    std::unordered_set<shared_sstable> expired;
+
+    if (db_clock::now() - _last_expired_check > _options.expired_sstable_check_frequency) {
+        clogger.debug("TWCS expired check sufficiently far in the past, checking for fully expired SSTables");
+        expired = get_fully_expired_sstables(cf, candidates, gc_before);
+        _last_expired_check = db_clock::now();
+    } else {
+        clogger.debug("TWCS skipping check for fully expired SSTables");
+    }
+
+    if (!expired.empty()) {
+        auto is_expired = [&] (const shared_sstable& s) { return expired.find(s) != expired.end(); };
+        candidates.erase(boost::remove_if(candidates, is_expired), candidates.end());
+    }
+
+    auto compaction_candidates = get_next_non_expired_sstables(cf, std::move(candidates), gc_before);
+    if (!expired.empty()) {
+        compaction_candidates.insert(compaction_candidates.end(), expired.begin(), expired.end());
+    }
+    return compaction_descriptor(std::move(compaction_candidates), cf.get_sstable_set(), service::get_local_compaction_priority());
+}
+
+time_window_compaction_strategy::bucket_compaction_mode
+time_window_compaction_strategy::compaction_mode(const bucket_t& bucket, timestamp_type bucket_key,
+        timestamp_type now, size_t min_threshold) const {
+    // STCS will also be performed on older window buckets, to avoid a bad write and
+    // space amplification when something like read repair cause small updates to
+    // those past windows.
+
+    if (bucket.size() >= 2 && !is_last_active_bucket(bucket_key, now) && _recent_active_windows.contains(bucket_key)) {
+        return bucket_compaction_mode::major;
+    } else if (bucket.size() >= size_t(min_threshold)) {
+        return bucket_compaction_mode::size_tiered;
+    }
+    return bucket_compaction_mode::none;
+}
+
+std::vector<shared_sstable>
+time_window_compaction_strategy::get_next_non_expired_sstables(column_family& cf,
+        std::vector<shared_sstable> non_expiring_sstables, gc_clock::time_point gc_before) {
+    auto most_interesting = get_compaction_candidates(cf, non_expiring_sstables);
+
+    if (!most_interesting.empty()) {
+        return most_interesting;
+    }
+
+    // if there is no sstable to compact in standard way, try compacting single sstable whose droppable tombstone
+    // ratio is greater than threshold.
+    auto e = boost::range::remove_if(non_expiring_sstables, [this, &gc_before] (const shared_sstable& sst) -> bool {
+        return !worth_dropping_tombstones(sst, gc_before);
+    });
+    non_expiring_sstables.erase(e, non_expiring_sstables.end());
+    if (non_expiring_sstables.empty()) {
+        return {};
+    }
+    auto it = boost::min_element(non_expiring_sstables, [] (auto& i, auto& j) {
+        return i->get_stats_metadata().min_timestamp < j->get_stats_metadata().min_timestamp;
+    });
+    return { *it };
+}
+
+std::vector<shared_sstable>
+time_window_compaction_strategy::get_compaction_candidates(column_family& cf, std::vector<shared_sstable> candidate_sstables) {
+    auto p = get_buckets(std::move(candidate_sstables), _options);
+    // Update the highest window seen, if necessary
+    _highest_window_seen = std::max(_highest_window_seen, p.second);
+
+    update_estimated_compaction_by_tasks(p.first, cf.min_compaction_threshold(), cf.schema()->max_compaction_threshold());
+
+    return newest_bucket(std::move(p.first), cf.min_compaction_threshold(), cf.schema()->max_compaction_threshold(),
+        _options.sstable_window_size, _highest_window_seen, _stcs_options);
+}
+
+timestamp_type
+time_window_compaction_strategy::get_window_lower_bound(std::chrono::seconds sstable_window_size, timestamp_type timestamp) {
+    using namespace std::chrono;
+    auto timestamp_in_sec = duration_cast<seconds>(microseconds(timestamp)).count();
+
+    // mask out window size from timestamp to get lower bound of its window
+    auto window_lower_bound_in_sec = seconds(timestamp_in_sec - (timestamp_in_sec % sstable_window_size.count()));
+
+    return timestamp_type(duration_cast<microseconds>(window_lower_bound_in_sec).count());
+}
+
+std::pair<std::map<timestamp_type, std::vector<shared_sstable>>, timestamp_type>
+time_window_compaction_strategy::get_buckets(std::vector<shared_sstable> files, time_window_compaction_strategy_options& options) {
+    std::map<timestamp_type, std::vector<shared_sstable>> buckets;
+
+    timestamp_type max_timestamp = 0;
+    // Create map to represent buckets
+    // For each sstable, add sstable to the time bucket
+    // Where the bucket is the file's max timestamp rounded to the nearest window bucket
+    for (auto&& f : files) {
+        timestamp_type ts = to_timestamp_type(options.timestamp_resolution, f->get_stats_metadata().max_timestamp);
+        timestamp_type lower_bound = get_window_lower_bound(options.sstable_window_size, ts);
+        buckets[lower_bound].push_back(std::move(f));
+        max_timestamp = std::max(max_timestamp, lower_bound);
+    }
+
+    return std::make_pair(std::move(buckets), max_timestamp);
+}
+
+static std::ostream& operator<<(std::ostream& os, const std::map<timestamp_type, std::vector<shared_sstable>>& buckets) {
+    os << "  buckets = {\n";
+    for (auto& bucket : buckets | boost::adaptors::reversed) {
+        os << format("    key={}, size={}\n", bucket.first, bucket.second.size());
+    }
+    os << "  }\n";
+    return os;
+}
+
+std::vector<shared_sstable>
+time_window_compaction_strategy::newest_bucket(std::map<timestamp_type, std::vector<shared_sstable>> buckets,
+        int min_threshold, int max_threshold, std::chrono::seconds sstable_window_size, timestamp_type now,
+        size_tiered_compaction_strategy_options& stcs_options) {
+    clogger.debug("time_window_compaction_strategy::newest_bucket:\n  now {}\n{}", now, buckets);
+
+    for (auto&& key_bucket : buckets | boost::adaptors::reversed) {
+        auto key = key_bucket.first;
+        auto& bucket = key_bucket.second;
+
+        if (is_last_active_bucket(key, now)) {
+            _recent_active_windows.insert(key);
+        }
+        switch (compaction_mode(bucket, key, now, min_threshold)) {
+        case bucket_compaction_mode::size_tiered: {
+            // If we're in the newest bucket, we'll use STCS to prioritize sstables.
+            auto stcs_interesting_bucket = size_tiered_compaction_strategy::most_interesting_bucket(bucket, min_threshold, max_threshold, stcs_options);
+
+            // If the tables in the current bucket aren't eligible in the STCS strategy, we'll skip it and look for other buckets
+            if (!stcs_interesting_bucket.empty()) {
+                clogger.debug("bucket size {} >= 2, key {}, performing STCS on what's here", bucket.size(), key);
+                return stcs_interesting_bucket;
+            }
+            break;
+        }
+        case bucket_compaction_mode::major:
+            _recent_active_windows.erase(key);
+            clogger.debug("bucket size {} >= 2 and not in current bucket, key {}, compacting what's here", bucket.size(), key);
+            return trim_to_threshold(std::move(bucket), max_threshold);
+        default:
+            clogger.debug("No compaction necessary for bucket size {} , key {}, now {}", bucket.size(), key, now);
+            break;
+        }
+    }
+    return {};
+}
+
+std::vector<shared_sstable>
+time_window_compaction_strategy::trim_to_threshold(std::vector<shared_sstable> bucket, int max_threshold) {
+    auto n = std::min(bucket.size(), size_t(max_threshold));
+    // Trim the largest sstables off the end to meet the maxThreshold
+    boost::partial_sort(bucket, bucket.begin() + n, [] (auto& i, auto& j) {
+        return i->ondisk_data_size() < j->ondisk_data_size();
+    });
+    bucket.resize(n);
+    return bucket;
+}
+
+void time_window_compaction_strategy::update_estimated_compaction_by_tasks(std::map<timestamp_type, std::vector<shared_sstable>>& tasks,
+                                                                           int min_threshold, int max_threshold) {
+    int64_t n = 0;
+    timestamp_type now = _highest_window_seen;
+
+    for (auto& task : tasks) {
+        const bucket_t& bucket = task.second;
+        timestamp_type bucket_key = task.first;
+
+        switch (compaction_mode(bucket, bucket_key, now, min_threshold)) {
+        case bucket_compaction_mode::size_tiered:
+            n += size_tiered_compaction_strategy::estimated_pending_compactions(bucket, min_threshold, max_threshold, _stcs_options);
+            break;
+        case bucket_compaction_mode::major:
+            n++;
+        default:
+            break;
+        }
+    }
+    _estimated_remaining_tasks = n;
+}
+
 }

sstables/time_window_compaction_strategy.hh

@@ -141,6 +141,8 @@ class time_window_compaction_strategy : public compaction_strategy_impl {
     int64_t _estimated_remaining_tasks = 0;
     db_clock::time_point _last_expired_check;
     timestamp_type _highest_window_seen;
+    // Keep track of all recent active windows that still need to be compacted into a single SSTable
+    std::unordered_set<timestamp_type> _recent_active_windows;
     size_tiered_compaction_strategy_options _stcs_options;
     compaction_backlog_tracker _backlog_tracker;
 public:
@@ -149,37 +151,11 @@ public:
     // Better co-locate some windows into the same sstables than OOM.
     static constexpr uint64_t max_data_segregation_window_count = 100;
 
+    using bucket_t = std::vector<shared_sstable>;
+    enum class bucket_compaction_mode { none, size_tiered, major };
 public:
     time_window_compaction_strategy(const std::map<sstring, sstring>& options);
-    virtual compaction_descriptor get_sstables_for_compaction(column_family& cf, std::vector<shared_sstable> candidates) override {
-        auto gc_before = gc_clock::now() - cf.schema()->gc_grace_seconds();
-
-        if (candidates.empty()) {
-            return compaction_descriptor();
-        }
-
-        // Find fully expired SSTables. Those will be included no matter what.
-        std::unordered_set<shared_sstable> expired;
-
-        if (db_clock::now() - _last_expired_check > _options.expired_sstable_check_frequency) {
-            clogger.debug("TWCS expired check sufficiently far in the past, checking for fully expired SSTables");
-            expired = get_fully_expired_sstables(cf, candidates, gc_before);
-            _last_expired_check = db_clock::now();
-        } else {
-            clogger.debug("TWCS skipping check for fully expired SSTables");
-        }
-
-        if (!expired.empty()) {
-            auto is_expired = [&] (const shared_sstable& s) { return expired.find(s) != expired.end(); };
-            candidates.erase(boost::remove_if(candidates, is_expired), candidates.end());
-        }
-
-        auto compaction_candidates = get_next_non_expired_sstables(cf, std::move(candidates), gc_before);
-        if (!expired.empty()) {
-            compaction_candidates.insert(compaction_candidates.end(), expired.begin(), expired.end());
-        }
-        return compaction_descriptor(std::move(compaction_candidates), cf.get_sstable_set(), service::get_local_compaction_priority());
-    }
+    virtual compaction_descriptor get_sstables_for_compaction(column_family& cf, std::vector<shared_sstable> candidates) override;
 private:
     static timestamp_type
     to_timestamp_type(time_window_compaction_strategy_options::timestamp_resolutions resolution, int64_t timestamp_from_sstable) {
@@ -193,114 +169,36 @@ private:
         };
     }
 
+    // Returns true if bucket is the last, most active one.
+    bool is_last_active_bucket(timestamp_type bucket_key, timestamp_type now) const {
+        return bucket_key >= now;
+    }
+
+    // Returns which compaction type should be performed on a given window bucket.
+    bucket_compaction_mode
+    compaction_mode(const bucket_t& bucket, timestamp_type bucket_key, timestamp_type now, size_t min_threshold) const;
+
     std::vector<shared_sstable>
-    get_next_non_expired_sstables(column_family& cf, std::vector<shared_sstable> non_expiring_sstables, gc_clock::time_point gc_before) {
-        auto most_interesting = get_compaction_candidates(cf, non_expiring_sstables);
+    get_next_non_expired_sstables(column_family& cf, std::vector<shared_sstable> non_expiring_sstables, gc_clock::time_point gc_before);
 
-        if (!most_interesting.empty()) {
-            return most_interesting;
-        }
-
-        // if there is no sstable to compact in standard way, try compacting single sstable whose droppable tombstone
-        // ratio is greater than threshold.
-        auto e = boost::range::remove_if(non_expiring_sstables, [this, &gc_before] (const shared_sstable& sst) -> bool {
-            return !worth_dropping_tombstones(sst, gc_before);
-        });
-        non_expiring_sstables.erase(e, non_expiring_sstables.end());
-        if (non_expiring_sstables.empty()) {
-            return {};
-        }
-        auto it = boost::min_element(non_expiring_sstables, [] (auto& i, auto& j) {
-            return i->get_stats_metadata().min_timestamp < j->get_stats_metadata().min_timestamp;
-        });
-        return { *it };
-    }
-
-    std::vector<shared_sstable> get_compaction_candidates(column_family& cf, std::vector<shared_sstable> candidate_sstables) {
-        auto p = get_buckets(std::move(candidate_sstables), _options);
-        // Update the highest window seen, if necessary
-        _highest_window_seen = std::max(_highest_window_seen, p.second);
-
-        update_estimated_compaction_by_tasks(p.first, cf.min_compaction_threshold());
-
-        return newest_bucket(std::move(p.first), cf.min_compaction_threshold(), cf.schema()->max_compaction_threshold(),
-            _options.sstable_window_size, _highest_window_seen, _stcs_options);
-    }
+    std::vector<shared_sstable> get_compaction_candidates(column_family& cf, std::vector<shared_sstable> candidate_sstables);
 public:
     // Find the lowest timestamp for window of given size
     static timestamp_type
-    get_window_lower_bound(std::chrono::seconds sstable_window_size, timestamp_type timestamp) {
-        using namespace std::chrono;
-        auto timestamp_in_sec = duration_cast<seconds>(microseconds(timestamp)).count();
-
-        // mask out window size from timestamp to get lower bound of its window
-        auto window_lower_bound_in_sec = seconds(timestamp_in_sec - (timestamp_in_sec % sstable_window_size.count()));
-
-        return timestamp_type(duration_cast<microseconds>(window_lower_bound_in_sec).count());
-    }
+    get_window_lower_bound(std::chrono::seconds sstable_window_size, timestamp_type timestamp);
 
     // Group files with similar max timestamp into buckets.
     // @return A pair, where the left element is the bucket representation (map of timestamp to sstablereader),
     // and the right is the highest timestamp seen
     static std::pair<std::map<timestamp_type, std::vector<shared_sstable>>, timestamp_type>
-    get_buckets(std::vector<shared_sstable> files, time_window_compaction_strategy_options& options) {
-        std::map<timestamp_type, std::vector<shared_sstable>> buckets;
+    get_buckets(std::vector<shared_sstable> files, time_window_compaction_strategy_options& options);
 
-        timestamp_type max_timestamp = 0;
-        // Create map to represent buckets
-        // For each sstable, add sstable to the time bucket
-        // Where the bucket is the file's max timestamp rounded to the nearest window bucket
-        for (auto&& f : files) {
-            timestamp_type ts = to_timestamp_type(options.timestamp_resolution, f->get_stats_metadata().max_timestamp);
-            timestamp_type lower_bound = get_window_lower_bound(options.sstable_window_size, ts);
-            buckets[lower_bound].push_back(std::move(f));
-            max_timestamp = std::max(max_timestamp, lower_bound);
-        }
-
-        return std::make_pair(std::move(buckets), max_timestamp);
-    }
-
-    static std::vector<shared_sstable>
+    std::vector<shared_sstable>
     newest_bucket(std::map<timestamp_type, std::vector<shared_sstable>> buckets, int min_threshold, int max_threshold,
-            std::chrono::seconds sstable_window_size, timestamp_type now, size_tiered_compaction_strategy_options& stcs_options) {
-        // If the current bucket has at least minThreshold SSTables, choose that one.
-        // For any other bucket, at least 2 SSTables is enough.
-        // In any case, limit to maxThreshold SSTables.
-
-        for (auto&& key_bucket : buckets | boost::adaptors::reversed) {
-            auto key = key_bucket.first;
-            auto& bucket = key_bucket.second;
-
-            clogger.trace("Key {}, now {}", key, now);
-
-            if (bucket.size() >= size_t(min_threshold) && key >= now) {
-                // If we're in the newest bucket, we'll use STCS to prioritize sstables
-                auto stcs_interesting_bucket = size_tiered_compaction_strategy::most_interesting_bucket(bucket, min_threshold, max_threshold, stcs_options);
-
-                // If the tables in the current bucket aren't eligible in the STCS strategy, we'll skip it and look for other buckets
-                if (!stcs_interesting_bucket.empty()) {
-                    return stcs_interesting_bucket;
-                }
-            } else if (bucket.size() >= 2 && key < now) {
-                clogger.debug("bucket size {} >= 2 and not in current bucket, compacting what's here", bucket.size());
-                return trim_to_threshold(std::move(bucket), max_threshold);
-            } else {
-                clogger.debug("No compaction necessary for bucket size {} , key {}, now {}", bucket.size(), key, now);
-            }
-        }
-        return {};
-    }
+            std::chrono::seconds sstable_window_size, timestamp_type now, size_tiered_compaction_strategy_options& stcs_options);
 
     static std::vector<shared_sstable>
-    trim_to_threshold(std::vector<shared_sstable> bucket, int max_threshold) {
-        auto n = std::min(bucket.size(), size_t(max_threshold));
-        // Trim the largest sstables off the end to meet the maxThreshold
-        boost::partial_sort(bucket, bucket.begin() + n, [] (auto& i, auto& j) {
-            return i->ondisk_data_size() < j->ondisk_data_size();
-        });
-        bucket.resize(n);
-        return bucket;
-    }
+    trim_to_threshold(std::vector<shared_sstable> bucket, int max_threshold);
 
     static int64_t
     get_window_for(const time_window_compaction_strategy_options& options, api::timestamp_type ts) {
@@ -312,23 +210,8 @@ public:
         return timestamp_type(std::chrono::duration_cast<std::chrono::microseconds>(options.get_sstable_window_size()).count());
     }
 private:
-    void update_estimated_compaction_by_tasks(std::map<timestamp_type, std::vector<shared_sstable>>& tasks, int min_threshold) {
-        int64_t n = 0;
-        timestamp_type now = _highest_window_seen;
-
-        for (auto task : tasks) {
-            auto key = task.first;
-
-            // For current window, make sure it's compactable
-            auto count = task.second.size();
-            if (key >= now && count >= size_t(min_threshold)) {
-                n++;
-            } else if (key < now && count >= 2) {
-                n++;
-            }
-        }
-        _estimated_remaining_tasks = n;
-    }
+    void update_estimated_compaction_by_tasks(std::map<timestamp_type, std::vector<shared_sstable>>& tasks,
+        int min_threshold, int max_threshold);
 
     friend class time_window_backlog_tracker;
 public:

streaming/stream_session.cc

@@ -229,7 +229,7 @@ void stream_session::init_messaging_service_handler() {
                             schema_ptr s = reader.schema();
                             auto& pc = service::get_local_streaming_priority();
 
-                            return sst->write_components(std::move(reader), std::max(1ul, adjusted_estimated_partitions), s,
+                            return sst->write_components(std::move(reader), adjusted_estimated_partitions, s,
                                                          cf->get_sstables_manager().configure_writer(),
                                                          encoding_stats{}, pc).then([sst] {
                                 return sst->open_data();
@@ -317,6 +317,15 @@ void stream_session::init_messaging_service_handler() {
     });
 }
 
+future<> stream_session::uninit_messaging_service_handler() {
+    return when_all_succeed(
+        ms().unregister_prepare_message(),
+        ms().unregister_prepare_done_message(),
+        ms().unregister_stream_mutation_fragments(),
+        ms().unregister_stream_mutation_done(),
+        ms().unregister_complete_message()).discard_result();
+}
+
 distributed<database>* stream_session::_db;
 distributed<db::system_distributed_keyspace>* stream_session::_sys_dist_ks;
 distributed<db::view::view_update_generator>* stream_session::_view_update_generator;
@@ -340,9 +349,13 @@ future<> stream_session::init_streaming_service(distributed<database>& db, distr
     // });
     return get_stream_manager().start().then([] {
         gms::get_local_gossiper().register_(get_local_stream_manager().shared_from_this());
-        return _db->invoke_on_all([] (auto& db) {
-            init_messaging_service_handler();
-        });
+        return smp::invoke_on_all([] { init_messaging_service_handler(); });
+    });
+}
+
+future<> stream_session::uninit_streaming_service() {
+    return smp::invoke_on_all([] {
+        return uninit_messaging_service_handler();
     });
 }
 

streaming/stream_session.hh

@@ -142,6 +142,7 @@ private:
     using token = dht::token;
     using ring_position = dht::ring_position;
     static void init_messaging_service_handler();
+    static future<> uninit_messaging_service_handler();
     static distributed<database>* _db;
     static distributed<db::system_distributed_keyspace>* _sys_dist_ks;
     static distributed<db::view::view_update_generator>* _view_update_generator;
@@ -152,6 +153,7 @@ public:
     static database& get_local_db() { return _db->local(); }
     static distributed<database>& get_db() { return *_db; };
     static future<> init_streaming_service(distributed<database>& db, distributed<db::system_distributed_keyspace>& sys_dist_ks, distributed<db::view::view_update_generator>& view_update_generator);
+    static future<> uninit_streaming_service();
 public:
     /**
      * Streaming endpoint.

table.cc

@@ -23,7 +23,6 @@
 #include "sstables/sstables.hh"
 #include "sstables/sstables_manager.hh"
 #include "service/priority_manager.hh"
-#include "db/view/view_updating_consumer.hh"
 #include "db/schema_tables.hh"
 #include "cell_locking.hh"
 #include "mutation_fragment.hh"
@@ -326,6 +325,32 @@ flat_mutation_reader make_range_sstable_reader(schema_ptr s,
             fwd_mr);
 }
 
+flat_mutation_reader make_restricted_range_sstable_reader(schema_ptr s,
+        reader_permit permit,
+        lw_shared_ptr<sstables::sstable_set> sstables,
+        const dht::partition_range& pr,
+        const query::partition_slice& slice,
+        const io_priority_class& pc,
+        tracing::trace_state_ptr trace_state,
+        streamed_mutation::forwarding fwd,
+        mutation_reader::forwarding fwd_mr,
+        sstables::read_monitor_generator& monitor_generator)
+{
+    auto ms = mutation_source([sstables=std::move(sstables), &monitor_generator] (
+            schema_ptr s,
+            reader_permit permit,
+            const dht::partition_range& pr,
+            const query::partition_slice& slice,
+            const io_priority_class& pc,
+            tracing::trace_state_ptr trace_state,
+            streamed_mutation::forwarding fwd,
+            mutation_reader::forwarding fwd_mr) {
+        return make_range_sstable_reader(std::move(s), std::move(permit), std::move(sstables), pr, slice, pc,
+                std::move(trace_state), fwd, fwd_mr, monitor_generator);
+    });
+    return make_restricted_flat_reader(std::move(ms), std::move(s), std::move(permit), pr, slice, pc, std::move(trace_state), fwd, fwd_mr);
+}
+
 flat_mutation_reader
 table::make_sstable_reader(schema_ptr s,
                                    reader_permit permit,
@@ -1396,7 +1421,7 @@ future<std::unordered_set<sstring>> table::get_sstables_by_partition_key(const s
             [this] (std::unordered_set<sstring>& filenames, lw_shared_ptr<sstables::sstable_set::incremental_selector>& sel, partition_key& pk) {
         return do_with(dht::decorated_key(dht::decorate_key(*_schema, pk)),
                 [this, &filenames, &sel, &pk](dht::decorated_key& dk) mutable {
-            auto sst = sel->select(dk).sstables;
+            const auto& sst = sel->select(dk).sstables;
             auto hk = sstables::sstable::make_hashed_key(*_schema, dk.key());
 
             return do_for_each(sst, [this, &filenames, &dk, hk = std::move(hk)] (std::vector<sstables::shared_sstable>::const_iterator::reference s) mutable {
@@ -1425,7 +1450,7 @@ std::vector<sstables::shared_sstable> table::select_sstables(const dht::partitio
     return _sstables->select(range);
 }
 
-std::vector<sstables::shared_sstable> table::candidates_for_compaction() const {
+std::vector<sstables::shared_sstable> table::non_staging_sstables() const {
     return boost::copy_range<std::vector<sstables::shared_sstable>>(*get_sstables()
             | boost::adaptors::filtered([this] (auto& sst) {
         return !_sstables_need_rewrite.count(sst->generation()) && !_sstables_staging.count(sst->generation());
@@ -1958,6 +1983,11 @@ void table::set_schema(schema_ptr s) {
     }
     _schema = std::move(s);
 
+    for (auto&& v : _views) {
+        v->view_info()->set_base_info(
+            v->view_info()->make_base_dependent_view_info(*_schema));
+    }
+
     set_compaction_strategy(_schema->compaction_strategy());
     trigger_compaction();
 }
@@ -1969,7 +1999,8 @@ static std::vector<view_ptr>::iterator find_view(std::vector<view_ptr>& views, c
 }
 
 void table::add_or_update_view(view_ptr v) {
-    v->view_info()->initialize_base_dependent_fields(*schema());
+    v->view_info()->set_base_info(
+        v->view_info()->make_base_dependent_view_info(*_schema));
     auto existing = find_view(_views, v);
     if (existing != _views.end()) {
         *existing = std::move(v);
@@ -2022,7 +2053,7 @@ static size_t memory_usage_of(const std::vector<frozen_mutation_and_schema>& ms)
  * @return a future resolving to the mutations to apply to the views, which can be empty.
  */
 future<> table::generate_and_propagate_view_updates(const schema_ptr& base,
-        std::vector<view_ptr>&& views,
+        std::vector<db::view::view_and_base>&& views,
         mutation&& m,
         flat_mutation_reader_opt existings,
         tracing::trace_state_ptr tr_state,
@@ -2134,7 +2165,7 @@ table::local_base_lock(
  * @return a future that resolves when the updates have been acknowledged by the view replicas
  */
 future<> table::populate_views(
-        std::vector<view_ptr> views,
+        std::vector<db::view::view_and_base> views,
         dht::token base_token,
         flat_mutation_reader&& reader,
         gc_clock::time_point now) {
@@ -2505,7 +2536,7 @@ future<row_locker::lock_holder> table::do_push_view_replica_updates(const schema
     utils::get_local_injector().inject("table_push_view_replica_updates_stale_time_point", [&now] {
         now -= 10s;
     });
-    auto views = affected_views(base, m, now);
+    auto views = db::view::with_base_info_snapshot(affected_views(base, m, now));
     if (views.empty()) {
         return make_ready_future<row_locker::lock_holder>();
     }
@@ -2588,16 +2619,3 @@ table::as_mutation_source_excluding(std::vector<sstables::shared_sstable>& ssts)
         return this->make_reader_excluding_sstables(std::move(s), std::move(permit), ssts, range, slice, pc, std::move(trace_state), fwd, fwd_mr);
     });
 }
-
-stop_iteration db::view::view_updating_consumer::consume_end_of_partition() {
-    if (_as->abort_requested()) {
-        return stop_iteration::yes;
-    }
-    try {
-        auto lock_holder = _table->stream_view_replica_updates(_schema, std::move(*_m), db::no_timeout, _excluded_sstables).get();
-    } catch (...) {
-        tlogger.warn("Failed to push replica updates for table {}.{}: {}", _schema->ks_name(), _schema->cf_name(), std::current_exception());
-    }
-    _m.reset();
-    return stop_iteration::no;
-}

test.py

@@ -447,6 +447,9 @@ async def run_test(test, options, gentle_kill=False, env=dict()):
                 env=dict(os.environ,
                          UBSAN_OPTIONS=":".join(filter(None, UBSAN_OPTIONS)),
                          ASAN_OPTIONS=":".join(filter(None, ASAN_OPTIONS)),
+                         # TMPDIR env variable is used by any seastar/scylla
+                         # test for directory to store test temporary data.
+                         TMPDIR=os.path.join(options.tmpdir, test.mode),
                          **env,
                          ),
                 preexec_fn=os.setsid,

test/alternator/run

@@ -28,8 +28,8 @@ fi
 SCYLLA_IP=127.1.$(($$ >> 8 & 255)).$(($$ & 255))
 echo "Running Scylla on $SCYLLA_IP"
 
-tmp_dir=/tmp/alternator-test-$$
-mkdir $tmp_dir
+tmp_dir="$(readlink -e ${TMPDIR-/tmp})"/alternator-test-$$
+mkdir "$tmp_dir"
 
 # We run the cleanup() function on exit for any reason - successful finish
 # of the script, an error (since we have "set -e"), or a signal.
@@ -76,7 +76,7 @@ done
 # argv[0] isn't good enough - because killall inspects the actual executable
 # filename in /proc/<pid>/stat. So we need to name the executable differently.
 # Luckily, using a symbolic link is good enough.
-SCYLLA_LINK=$tmp_dir/test_scylla
+SCYLLA_LINK="$tmp_dir"/test_scylla
 ln -s "$SCYLLA" "$SCYLLA_LINK"
 
 "$SCYLLA_LINK" --options-file "$source_path/conf/scylla.yaml" \

test/alternator/test_expected.py

@@ -522,6 +522,15 @@ def test_update_expected_1_null(test_table_s):
             Expected={'a': {'ComparisonOperator': 'NULL', 'AttributeValueList': [2]}}
         )
 
+# When ComparisonOperator = "NULL", AttributeValueList should be empty if it
+# exists, but as this test verifies, it may also be missing completely.
+def test_update_expected_1_null_missing_list(test_table_s):
+    p = random_string()
+    test_table_s.update_item(Key={'p': p},
+        AttributeUpdates={'a': {'Value': 2, 'Action': 'PUT'}},
+        Expected={'a': {'ComparisonOperator': 'NULL'}})
+    assert test_table_s.get_item(Key={'p': p}, ConsistentRead=True)['Item']['a'] == 2
+
 # Tests for Expected with ComparisonOperator = "CONTAINS":
 def test_update_expected_1_contains(test_table_s):
     # true cases. CONTAINS can be used for two unrelated things: check substrings

test/alternator/test_table.py

@@ -308,3 +308,17 @@ def test_list_tables_wrong_limit(dynamodb):
     # lower limit (min. 1) is imposed by boto3 library checks
     with pytest.raises(ClientError, match='ValidationException'):
         dynamodb.meta.client.list_tables(Limit=101)
+
+# Even before Alternator gains support for configuring server-side encryption
+# ("encryption at rest") with CreateTable's SSESpecification option, we should
+# support the option "Enabled=false" which is the default, and means the server
+# takes care of whatever server-side encryption is done, on its own.
+# Reproduces issue #7031.
+def test_table_sse_off(dynamodb):
+    # If StreamSpecification is given, but has StreamEnabled=false, it's as
+    # if StreamSpecification was missing, and fine. No other attribues are
+    # necessary.
+    table = create_test_table(dynamodb, SSESpecification = {'Enabled': False},
+        KeySchema=[{ 'AttributeName': 'p', 'KeyType': 'HASH' }],
+        AttributeDefinitions=[{ 'AttributeName': 'p', 'AttributeType': 'S' }]);
+    table.delete();

test/alternator/test_tag.py

@@ -25,7 +25,7 @@ import pytest
 from botocore.exceptions import ClientError
 import re
 import time
-from util import multiset, create_test_table
+from util import multiset, create_test_table, test_table_name
 
 def delete_tags(table, arn):
     got = table.meta.client.list_tags_of_resource(ResourceArn=arn)
@@ -156,6 +156,56 @@ def test_tag_resource_write_isolation_values(scylla_only, test_table):
     with pytest.raises(ClientError, match='ValidationException'):
         test_table.meta.client.tag_resource(ResourceArn=arn, Tags=[{'Key':'system:write_isolation', 'Value':'bah'}])
 
+# Test that if trying to create a table with forbidden tags (in this test,
+# a list of tags longer than the maximum allowed of 50 tags), the table
+# is not created at all.
+def test_too_long_tags_from_creation(dynamodb):
+    # The feature of creating a table already with tags was only added to
+    # DynamoDB in April 2019, and to the botocore library in version 1.12.136
+    # so older versions of the library cannot run this test.
+    import botocore
+    from distutils.version import LooseVersion
+    if (LooseVersion(botocore.__version__) < LooseVersion('1.12.136')):
+        pytest.skip("Botocore version 1.12.136 or above required to run this test")
+    name = test_table_name()
+    # Setting 100 tags is not allowed, the following table creation should fail:
+    with pytest.raises(ClientError, match='ValidationException'):
+        dynamodb.create_table(TableName=name,
+            BillingMode='PAY_PER_REQUEST',
+            KeySchema=[{ 'AttributeName': 'p', 'KeyType': 'HASH' }],
+            AttributeDefinitions=[{ 'AttributeName': 'p', 'AttributeType': 'S' }],
+            Tags=[{'Key': str(i), 'Value': str(i)} for i in range(100)])
+    # After the table creation failed, the table should not exist.
+    with pytest.raises(ClientError, match='ResourceNotFoundException'):
+        dynamodb.meta.client.describe_table(TableName=name)
+
+# This test is similar to the above, but uses another case of forbidden tags -
+# here an illegal value for the system::write_isolation tag. This is a
+# scylla_only test because only Alternator checks the validity of the
+# system::write_isolation tag.
+# Reproduces issue #6809, where the table creation appeared to fail, but it
+# was actually created (without the tag).
+def test_forbidden_tags_from_creation(scylla_only, dynamodb):
+    # The feature of creating a table already with tags was only added to
+    # DynamoDB in April 2019, and to the botocore library in version 1.12.136
+    # so older versions of the library cannot run this test.
+    import botocore
+    from distutils.version import LooseVersion
+    if (LooseVersion(botocore.__version__) < LooseVersion('1.12.136')):
+        pytest.skip("Botocore version 1.12.136 or above required to run this test")
+    name = test_table_name()
+    # It is not allowed to set the system:write_isolation to "dog", so the
+    # following table creation should fail:
+    with pytest.raises(ClientError, match='ValidationException'):
+        dynamodb.create_table(TableName=name,
+            BillingMode='PAY_PER_REQUEST',
+            KeySchema=[{ 'AttributeName': 'p', 'KeyType': 'HASH' }],
+            AttributeDefinitions=[{ 'AttributeName': 'p', 'AttributeType': 'S' }],
+            Tags=[{'Key': 'system:write_isolation', 'Value': 'dog'}])
+    # After the table creation failed, the table should not exist.
+    with pytest.raises(ClientError, match='ResourceNotFoundException'):
+        dynamodb.meta.client.describe_table(TableName=name)
+
 # Test checking that unicode tags are allowed
 @pytest.mark.xfail(reason="unicode tags not yet supported")
 def test_tag_resource_unicode(test_table):

test/boost/batchlog_manager_test.cc

@@ -64,7 +64,7 @@ SEASTAR_TEST_CASE(test_execute_batch) {
             auto version = netw::messaging_service::current_version;
             auto bm = bp.get_batch_log_mutation_for({ m }, s->id(), version, db_clock::now() - db_clock::duration(3h));
 
-            return qp.proxy().mutate_locally(bm, tracing::trace_state_ptr()).then([&bp] () mutable {
+            return qp.proxy().mutate_locally(bm, tracing::trace_state_ptr(), db::commitlog::force_sync::no).then([&bp] () mutable {
                 return bp.count_all_batches().then([](auto n) {
                     BOOST_CHECK_EQUAL(n, 1);
                 }).then([&bp] () mutable {

test/boost/big_decimal_test.cc

@@ -157,6 +157,13 @@ BOOST_AUTO_TEST_CASE(test_big_decimal_div) {
     test_div("-0.25", 10, "-0.02");
     test_div("-0.26", 10, "-0.03");
     test_div("-10E10", 3, "-3E10");
+
+    // Document a small oddity, 1e1 has -1 decimal places, so dividing
+    // it by 2 produces 0. This is not the behavior in cassandra, but
+    // scylla doesn't expose arithmetic operations, so this doesn't
+    // seem to be visible from CQL.
+    test_div("10", 2, "5");
+    test_div("1e1", 2, "0e1");
 }
 
 BOOST_AUTO_TEST_CASE(test_big_decimal_assignadd) {

test/boost/castas_fcts_test.cc

@@ -142,6 +142,19 @@ SEASTAR_TEST_CASE(test_decimal_to_bigint) {
     });
 }
 
+SEASTAR_TEST_CASE(test_decimal_to_float) {
+    return do_with_cql_env_thread([&](auto& e) {
+        e.execute_cql("CREATE TABLE test (key text primary key, value decimal)").get();
+        e.execute_cql("INSERT INTO test (key, value) VALUES ('k1', 10)").get();
+        e.execute_cql("INSERT INTO test (key, value) VALUES ('k2', 1e1)").get();
+        auto v = e.execute_cql("SELECT key, CAST(value as float) from test").get0();
+        assert_that(v).is_rows().with_rows_ignore_order({
+            {{serialized("k1")}, {serialized(float(10))}},
+            {{serialized("k2")}, {serialized(float(10))}},
+        });
+    });
+}
+
 SEASTAR_TEST_CASE(test_varint_to_bigint) {
     return do_with_cql_env_thread([&](auto& e) {
         e.execute_cql("CREATE TABLE test (key text primary key, value varint)").get();

test/boost/cql_query_test.cc

@@ -3479,10 +3479,13 @@ SEASTAR_TEST_CASE(test_select_with_mixed_order_table) {
 }
 
 uint64_t
-run_and_examine_cache_stat_change(cql_test_env& e, uint64_t cache_tracker::stats::*metric, std::function<void (cql_test_env& e)> func) {
+run_and_examine_cache_read_stats_change(cql_test_env& e, std::string_view cf_name, std::function<void (cql_test_env& e)> func) {
     auto read_stat = [&] {
-        auto local_read_metric = [metric] (database& db) { return db.row_cache_tracker().get_stats().*metric; };
-        return e.db().map_reduce0(local_read_metric, uint64_t(0), std::plus<uint64_t>()).get0();
+        return e.db().map_reduce0([&cf_name] (const database& db) {
+            auto& t = db.find_column_family("ks", cf_name);
+            auto& stats = t.get_row_cache().stats();
+            return stats.reads_with_misses.count() + stats.reads_with_no_misses.count();
+        }, uint64_t(0), std::plus<uint64_t>()).get0();
     };
     auto before = read_stat();
     func(e);
@@ -3493,11 +3496,11 @@ run_and_examine_cache_stat_change(cql_test_env& e, uint64_t cache_tracker::stats
 SEASTAR_TEST_CASE(test_cache_bypass) {
     return do_with_cql_env_thread([] (cql_test_env& e) {
         e.execute_cql("CREATE TABLE t (k int PRIMARY KEY)").get();
-        auto with_cache = run_and_examine_cache_stat_change(e, &cache_tracker::stats::reads, [] (cql_test_env& e) {
+        auto with_cache = run_and_examine_cache_read_stats_change(e, "t", [] (cql_test_env& e) {
             e.execute_cql("SELECT * FROM t").get();
         });
         BOOST_REQUIRE(with_cache >= smp::count);  // scan may make multiple passes per shard
-        auto without_cache = run_and_examine_cache_stat_change(e, &cache_tracker::stats::reads, [] (cql_test_env& e) {
+        auto without_cache = run_and_examine_cache_read_stats_change(e, "t", [] (cql_test_env& e) {
             e.execute_cql("SELECT * FROM t BYPASS CACHE").get();
         });
         BOOST_REQUIRE_EQUAL(without_cache, 0);
@@ -4583,3 +4586,30 @@ SEASTAR_TEST_CASE(test_internal_alter_table_on_a_distributed_table) {
         });
     });
 }
+
+SEASTAR_TEST_CASE(test_impossible_where) {
+    return do_with_cql_env_thread([] (cql_test_env& e) {
+        cquery_nofail(e, "CREATE TABLE t(p int PRIMARY KEY, r int)");
+        cquery_nofail(e, "INSERT INTO  t(p,r) VALUES (0, 0)");
+        cquery_nofail(e, "INSERT INTO  t(p,r) VALUES (1, 10)");
+        cquery_nofail(e, "INSERT INTO  t(p,r) VALUES (2, 20)");
+        require_rows(e, "SELECT * FROM t WHERE r>10 AND r<10 ALLOW FILTERING", {});
+        require_rows(e, "SELECT * FROM t WHERE r>=10 AND r<=0 ALLOW FILTERING", {});
+
+        cquery_nofail(e, "CREATE TABLE t2(p int, c int, PRIMARY KEY(p, c)) WITH CLUSTERING ORDER BY (c DESC)");
+        cquery_nofail(e, "INSERT INTO  t2(p,c) VALUES (0, 0)");
+        cquery_nofail(e, "INSERT INTO  t2(p,c) VALUES (1, 10)");
+        cquery_nofail(e, "INSERT INTO  t2(p,c) VALUES (2, 20)");
+        require_rows(e, "SELECT * FROM t2 WHERE c>10 AND c<10 ALLOW FILTERING", {});
+        require_rows(e, "SELECT * FROM t2 WHERE c>=10 AND c<=0 ALLOW FILTERING", {});
+    });
+}
+
+SEASTAR_TEST_CASE(test_counter_column_added_into_non_counter_table) {
+    return do_with_cql_env_thread([] (cql_test_env& e) {
+        cquery_nofail(e, "CREATE TABLE t (pk int, ck int, PRIMARY KEY(pk, ck))");
+
+        BOOST_REQUIRE_THROW(e.execute_cql("ALTER TABLE t ADD \"c\" counter;").get(),
+                exceptions::configuration_exception);
+    });
+}

test/boost/filtering_test.cc

@@ -1134,6 +1134,9 @@ SEASTAR_TEST_CASE(test_filtering) {
                 { int32_type->decompose(8), int32_type->decompose(3) },
                 { int32_type->decompose(9), int32_type->decompose(3) },
             });
+            require_rows(e, "SELECT k FROM cf WHERE k=12 AND (m,n)>=(4,0) ALLOW FILTERING;", {
+                    { int32_type->decompose(12), int32_type->decompose(4), int32_type->decompose(5)},
+                });
         }
 
         // test filtering on clustering keys

test/boost/mutation_reader_test.cc

@@ -870,8 +870,6 @@ SEASTAR_TEST_CASE(reader_selector_fast_forwarding_test) {
     });
 }
 
-static const std::size_t new_reader_base_cost{16 * 1024};
-
 sstables::shared_sstable create_sstable(sstables::test_env& env, simple_schema& sschema, const sstring& path) {
     std::vector<mutation> mutations;
     mutations.reserve(1 << 14);
@@ -2588,6 +2586,7 @@ SEASTAR_THREAD_TEST_CASE(test_queue_reader) {
 
         BOOST_REQUIRE_THROW(handle.push(partition_end{}).get(), std::runtime_error);
         BOOST_REQUIRE_THROW(handle.push_end_of_stream(), std::runtime_error);
+        BOOST_REQUIRE_THROW(fill_buffer_fut.get(), broken_promise);
     }
 
     // Abandoned handle aborts, move-assignment
@@ -2850,3 +2849,488 @@ SEASTAR_THREAD_TEST_CASE(test_manual_paused_evictable_reader_is_mutation_source)
 
     run_mutation_source_tests(make_populate);
 }
+
+namespace {
+
+std::deque<mutation_fragment> copy_fragments(const schema& s, const std::deque<mutation_fragment>& o) {
+    std::deque<mutation_fragment> buf;
+    for (const auto& mf : o) {
+        buf.emplace_back(s, mf);
+    }
+    return buf;
+}
+
+flat_mutation_reader create_evictable_reader_and_evict_after_first_buffer(
+        schema_ptr schema,
+        reader_permit permit,
+        const dht::partition_range& prange,
+        const query::partition_slice& slice,
+        std::deque<mutation_fragment> first_buffer,
+        position_in_partition_view last_fragment_position,
+        std::deque<mutation_fragment> second_buffer,
+        size_t max_buffer_size) {
+    class factory {
+        schema_ptr _schema;
+        std::optional<std::deque<mutation_fragment>> _first_buffer;
+        std::optional<std::deque<mutation_fragment>> _second_buffer;
+        size_t _max_buffer_size;
+
+    private:
+        std::optional<std::deque<mutation_fragment>> copy_buffer(const std::optional<std::deque<mutation_fragment>>& o) {
+            if (!o) {
+                return {};
+            }
+            return copy_fragments(*_schema, *o);
+        }
+
+    public:
+        factory(schema_ptr schema, std::deque<mutation_fragment> first_buffer, std::deque<mutation_fragment> second_buffer, size_t max_buffer_size)
+            : _schema(std::move(schema)), _first_buffer(std::move(first_buffer)), _second_buffer(std::move(second_buffer)), _max_buffer_size(max_buffer_size) {
+        }
+
+        factory(const factory& o)
+            : _schema(o._schema)
+            , _first_buffer(copy_buffer(o._first_buffer))
+            , _second_buffer(copy_buffer(o._second_buffer)) {
+        }
+        factory(factory&& o) = default;
+
+        flat_mutation_reader operator()(
+                schema_ptr s,
+                reader_permit permit,
+                const dht::partition_range& range,
+                const query::partition_slice& slice,
+                const io_priority_class& pc,
+                tracing::trace_state_ptr trace_state,
+                streamed_mutation::forwarding fwd_sm,
+                mutation_reader::forwarding fwd_mr) {
+            BOOST_REQUIRE(s == _schema);
+            if (_first_buffer) {
+                auto buf = *std::exchange(_first_buffer, {});
+                auto rd = make_flat_mutation_reader_from_fragments(_schema, std::move(buf));
+                rd.set_max_buffer_size(_max_buffer_size);
+                return rd;
+            }
+            if (_second_buffer) {
+                auto buf = *std::exchange(_second_buffer, {});
+                auto rd = make_flat_mutation_reader_from_fragments(_schema, std::move(buf));
+                rd.set_max_buffer_size(_max_buffer_size);
+                return rd;
+            }
+            return make_empty_flat_reader(_schema);
+        }
+    };
+    auto ms = mutation_source(factory(schema, std::move(first_buffer), std::move(second_buffer), max_buffer_size));
+
+    auto [rd, handle] = make_manually_paused_evictable_reader(
+            std::move(ms),
+            schema,
+            permit,
+            prange,
+            slice,
+            seastar::default_priority_class(),
+            nullptr,
+            mutation_reader::forwarding::yes);
+
+    rd.set_max_buffer_size(max_buffer_size);
+
+    rd.fill_buffer(db::no_timeout).get0();
+
+    const auto eq_cmp = position_in_partition::equal_compare(*schema);
+    BOOST_REQUIRE(rd.is_buffer_full());
+    BOOST_REQUIRE(eq_cmp(rd.buffer().back().position(), last_fragment_position));
+    BOOST_REQUIRE(!rd.is_end_of_stream());
+
+    rd.detach_buffer();
+
+    handle.pause();
+
+    while(permit.semaphore().try_evict_one_inactive_read());
+
+    return std::move(rd);
+}
+
+}
+
+SEASTAR_THREAD_TEST_CASE(test_evictable_reader_trim_range_tombstones) {
+    reader_concurrency_semaphore semaphore(reader_concurrency_semaphore::no_limits{}, get_name());
+    simple_schema s;
+
+    const auto pkey = s.make_pkey();
+    size_t max_buffer_size = 512;
+    const int first_ck = 100;
+    const int second_buffer_ck = first_ck + 100;
+
+    size_t mem_usage = 0;
+
+    std::deque<mutation_fragment> first_buffer;
+    first_buffer.emplace_back(partition_start{pkey, {}});
+    mem_usage = first_buffer.back().memory_usage(*s.schema());
+    for (int i = 0; i < second_buffer_ck; ++i) {
+        first_buffer.emplace_back(s.make_row(s.make_ckey(i++), "v"));
+        mem_usage += first_buffer.back().memory_usage(*s.schema());
+    }
+    const auto last_fragment_position = position_in_partition(first_buffer.back().position());
+    max_buffer_size = mem_usage;
+    first_buffer.emplace_back(s.make_row(s.make_ckey(second_buffer_ck), "v"));
+
+    std::deque<mutation_fragment> second_buffer;
+    second_buffer.emplace_back(partition_start{pkey, {}});
+    mem_usage = second_buffer.back().memory_usage(*s.schema());
+    second_buffer.emplace_back(s.make_range_tombstone(query::clustering_range::make_ending_with(s.make_ckey(second_buffer_ck + 10))));
+    int ckey = second_buffer_ck;
+    while (mem_usage <= max_buffer_size) {
+        second_buffer.emplace_back(s.make_row(s.make_ckey(ckey++), "v"));
+        mem_usage += second_buffer.back().memory_usage(*s.schema());
+    }
+    second_buffer.emplace_back(partition_end{});
+
+    auto rd = create_evictable_reader_and_evict_after_first_buffer(s.schema(), semaphore.make_permit(), query::full_partition_range,
+            s.schema()->full_slice(), std::move(first_buffer), last_fragment_position, std::move(second_buffer), max_buffer_size);
+
+    rd.fill_buffer(db::no_timeout).get();
+
+    const auto tri_cmp = position_in_partition::tri_compare(*s.schema());
+
+    BOOST_REQUIRE(tri_cmp(last_fragment_position, rd.peek_buffer().position()) < 0);
+}
+
+namespace {
+
+void check_evictable_reader_validation_is_triggered(
+        std::string_view test_name,
+        std::string_view error_prefix, // empty str if no exception is expected
+        schema_ptr schema,
+        reader_permit permit,
+        const dht::partition_range& prange,
+        const query::partition_slice& slice,
+        std::deque<mutation_fragment> first_buffer,
+        position_in_partition_view last_fragment_position,
+        std::deque<mutation_fragment> second_buffer,
+        size_t max_buffer_size) {
+
+    testlog.info("check_evictable_reader_validation_is_triggered(): checking {} test case: {}", error_prefix.empty() ? "positive" : "negative", test_name);
+
+    auto rd = create_evictable_reader_and_evict_after_first_buffer(std::move(schema), std::move(permit), prange, slice, std::move(first_buffer),
+            last_fragment_position, std::move(second_buffer), max_buffer_size);
+
+    const bool fail = !error_prefix.empty();
+
+    try {
+        rd.fill_buffer(db::no_timeout).get0();
+    } catch (std::runtime_error& e) {
+        if (fail) {
+            if (error_prefix == std::string_view(e.what(), error_prefix.size())) {
+                testlog.trace("Expected exception caught: {}", std::current_exception());
+                return;
+            } else {
+                BOOST_FAIL(fmt::format("Exception with unexpected message caught: {}", std::current_exception()));
+            }
+        } else {
+            BOOST_FAIL(fmt::format("Unexpected exception caught: {}", std::current_exception()));
+        }
+    }
+    if (fail) {
+        BOOST_FAIL(fmt::format("Expected exception not thrown"));
+    }
+}
+
+}
+
+SEASTAR_THREAD_TEST_CASE(test_evictable_reader_self_validation) {
+    set_abort_on_internal_error(false);
+    auto reset_on_internal_abort = defer([] {
+        set_abort_on_internal_error(true);
+    });
+
+    reader_concurrency_semaphore semaphore(reader_concurrency_semaphore::no_limits{}, get_name());
+    simple_schema s;
+
+    auto pkeys = s.make_pkeys(4);
+    std::ranges::sort(pkeys, dht::decorated_key::less_comparator(s.schema()));
+
+    size_t max_buffer_size = 512;
+    const int first_ck = 100;
+    const int second_buffer_ck = first_ck + 100;
+    const int last_ck = second_buffer_ck + 100;
+
+    static const char partition_error_prefix[] = "maybe_validate_partition_start(): validation failed";
+    static const char position_in_partition_error_prefix[] = "validate_position_in_partition(): validation failed";
+    static const char trim_range_tombstones_error_prefix[] = "maybe_trim_range_tombstone(): validation failed";
+
+    const auto prange = dht::partition_range::make(
+            dht::partition_range::bound(pkeys[1], true),
+            dht::partition_range::bound(pkeys[2], true));
+
+    const auto ckrange = query::clustering_range::make(
+            query::clustering_range::bound(s.make_ckey(first_ck), true),
+            query::clustering_range::bound(s.make_ckey(last_ck), true));
+
+    const auto slice = partition_slice_builder(*s.schema()).with_range(ckrange).build();
+
+    std::deque<mutation_fragment> first_buffer;
+    first_buffer.emplace_back(partition_start{pkeys[1], {}});
+    size_t mem_usage = first_buffer.back().memory_usage(*s.schema());
+    for (int i = 0; i < second_buffer_ck; ++i) {
+        first_buffer.emplace_back(s.make_row(s.make_ckey(i++), "v"));
+        mem_usage += first_buffer.back().memory_usage(*s.schema());
+    }
+    max_buffer_size = mem_usage;
+    auto last_fragment_position = position_in_partition(first_buffer.back().position());
+    first_buffer.emplace_back(s.make_row(s.make_ckey(second_buffer_ck), "v"));
+
+    auto make_second_buffer = [&s, &max_buffer_size, second_buffer_ck] (dht::decorated_key pkey, std::optional<int> first_ckey = {},
+            bool inject_range_tombstone = false) mutable {
+        auto ckey = first_ckey ? *first_ckey : second_buffer_ck;
+        std::deque<mutation_fragment> second_buffer;
+        second_buffer.emplace_back(partition_start{std::move(pkey), {}});
+        size_t mem_usage = second_buffer.back().memory_usage(*s.schema());
+        if (inject_range_tombstone) {
+            second_buffer.emplace_back(s.make_range_tombstone(query::clustering_range::make_ending_with(s.make_ckey(last_ck))));
+        }
+        while (mem_usage <= max_buffer_size) {
+            second_buffer.emplace_back(s.make_row(s.make_ckey(ckey++), "v"));
+            mem_usage += second_buffer.back().memory_usage(*s.schema());
+        }
+        second_buffer.emplace_back(partition_end{});
+        return second_buffer;
+    };
+
+    //
+    // Continuing the same partition
+    //
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey < _last_pkey; pkey ∉ prange",
+            partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[0]),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey == _last_pkey",
+            "",
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[1]),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey == _last_pkey; position_in_partition ∉ ckrange (<)",
+            position_in_partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[1], first_ck - 10),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey == _last_pkey; position_in_partition ∉ ckrange (<); start with trimmable range-tombstone",
+            position_in_partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[1], first_ck - 10, true),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey == _last_pkey; position_in_partition ∉ ckrange; position_in_partition < _next_position_in_partition",
+            position_in_partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[1], second_buffer_ck - 2),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey == _last_pkey; position_in_partition ∉ ckrange; position_in_partition < _next_position_in_partition; start with trimmable range-tombstone",
+            position_in_partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[1], second_buffer_ck - 2, true),
+            max_buffer_size);
+
+    {
+        auto second_buffer = make_second_buffer(pkeys[1], second_buffer_ck);
+        second_buffer[1] = s.make_range_tombstone(query::clustering_range::make_ending_with(s.make_ckey(second_buffer_ck - 10)));
+        check_evictable_reader_validation_is_triggered(
+                "pkey == _last_pkey; end(range_tombstone) < _next_position_in_partition",
+                trim_range_tombstones_error_prefix,
+                s.schema(),
+                semaphore.make_permit(),
+                prange,
+                slice,
+                copy_fragments(*s.schema(), first_buffer),
+                last_fragment_position,
+                std::move(second_buffer),
+                max_buffer_size);
+    }
+
+    {
+        auto second_buffer = make_second_buffer(pkeys[1], second_buffer_ck);
+        second_buffer[1] = s.make_range_tombstone(query::clustering_range::make_ending_with(s.make_ckey(second_buffer_ck + 10)));
+        check_evictable_reader_validation_is_triggered(
+                "pkey == _last_pkey; end(range_tombstone) > _next_position_in_partition",
+                "",
+                s.schema(),
+                semaphore.make_permit(),
+                prange,
+                slice,
+                copy_fragments(*s.schema(), first_buffer),
+                last_fragment_position,
+                std::move(second_buffer),
+                max_buffer_size);
+    }
+
+    {
+        auto second_buffer = make_second_buffer(pkeys[1], second_buffer_ck);
+        second_buffer[1] = s.make_range_tombstone(query::clustering_range::make_starting_with(s.make_ckey(last_ck + 10)));
+        check_evictable_reader_validation_is_triggered(
+                "pkey == _last_pkey; start(range_tombstone) ∉ ckrange (>)",
+                position_in_partition_error_prefix,
+                s.schema(),
+                semaphore.make_permit(),
+                prange,
+                slice,
+                copy_fragments(*s.schema(), first_buffer),
+                last_fragment_position,
+                std::move(second_buffer),
+                max_buffer_size);
+    }
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey == _last_pkey; position_in_partition ∈ ckrange",
+            "",
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[1], second_buffer_ck),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey == _last_pkey; position_in_partition ∉ ckrange (>)",
+            position_in_partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[1], last_ck + 10),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey > _last_pkey; pkey ∈ pkrange",
+            partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[2]),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey > _last_pkey; pkey ∉ pkrange",
+            partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[3]),
+            max_buffer_size);
+
+    //
+    // Continuing from next partition
+    //
+
+    first_buffer.clear();
+
+    first_buffer.emplace_back(partition_start{pkeys[1], {}});
+    mem_usage = first_buffer.back().memory_usage(*s.schema());
+    for (int i = 0; i < second_buffer_ck; ++i) {
+        first_buffer.emplace_back(s.make_row(s.make_ckey(i++), "v"));
+        mem_usage += first_buffer.back().memory_usage(*s.schema());
+    }
+    first_buffer.emplace_back(partition_end{});
+    mem_usage += first_buffer.back().memory_usage(*s.schema());
+    last_fragment_position = position_in_partition(first_buffer.back().position());
+    max_buffer_size = mem_usage;
+    first_buffer.emplace_back(partition_start{pkeys[2], {}});
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey < _last_pkey; pkey ∉ pkrange",
+            partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[0]),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey == _last_pkey",
+            partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[1]),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey > _last_pkey; pkey ∈ pkrange",
+            "",
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[2]),
+            max_buffer_size);
+
+    check_evictable_reader_validation_is_triggered(
+            "pkey > _last_pkey; pkey ∉ pkrange",
+            partition_error_prefix,
+            s.schema(),
+            semaphore.make_permit(),
+            prange,
+            slice,
+            copy_fragments(*s.schema(), first_buffer),
+            last_fragment_position,
+            make_second_buffer(pkeys[3]),
+            max_buffer_size);
+}

test/boost/querier_cache_test.cc

@@ -769,3 +769,27 @@ SEASTAR_THREAD_TEST_CASE(test_immediate_evict_on_insert) {
 
     fut.get();
 }
+
+namespace {
+
+class inactive_read : public reader_concurrency_semaphore::inactive_read {
+public:
+    virtual void evict() override {
+    }
+};
+
+}
+
+SEASTAR_THREAD_TEST_CASE(test_unique_inactive_read_handle) {
+    reader_concurrency_semaphore sem1(reader_concurrency_semaphore::no_limits{}, "sem1");
+    reader_concurrency_semaphore sem2(reader_concurrency_semaphore::no_limits{}, ""); // to see the message for an unnamed semaphore
+
+    auto sem1_h1 = sem1.register_inactive_read(std::make_unique<inactive_read>());
+    auto sem2_h1 = sem2.register_inactive_read(std::make_unique<inactive_read>());
+
+    // Sanity check that lookup still works with empty handle.
+    BOOST_REQUIRE(!sem1.unregister_inactive_read(reader_concurrency_semaphore::inactive_read_handle{}));
+
+    BOOST_REQUIRE_THROW(sem1.unregister_inactive_read(std::move(sem2_h1)), std::runtime_error);
+    BOOST_REQUIRE_THROW(sem2.unregister_inactive_read(std::move(sem1_h1)), std::runtime_error);
+}

test/boost/sstable_datafile_test.cc

@@ -3168,6 +3168,8 @@ SEASTAR_TEST_CASE(time_window_strategy_correctness_test) {
             sstables.push_back(make_sstable_containing(sst_gen, {std::move(mut)}));
         }
 
+        std::map<sstring, sstring> options;
+        time_window_compaction_strategy twcs(options);
         std::map<api::timestamp_type, std::vector<shared_sstable>> buckets;
 
         // We'll put 3 sstables into the newest bucket
@@ -3177,13 +3179,13 @@ SEASTAR_TEST_CASE(time_window_strategy_correctness_test) {
         }
         sstables::size_tiered_compaction_strategy_options stcs_options;
         auto now = api::timestamp_clock::now().time_since_epoch().count();
-        auto new_bucket = time_window_compaction_strategy::newest_bucket(buckets, 4, 32, duration_cast<seconds>(hours(1)),
+        auto new_bucket = twcs.newest_bucket(buckets, 4, 32, duration_cast<seconds>(hours(1)),
             time_window_compaction_strategy::get_window_lower_bound(duration_cast<seconds>(hours(1)), now), stcs_options);
         // incoming bucket should not be accepted when it has below the min threshold SSTables
         BOOST_REQUIRE(new_bucket.empty());
 
         now = api::timestamp_clock::now().time_since_epoch().count();
-        new_bucket = time_window_compaction_strategy::newest_bucket(buckets, 2, 32, duration_cast<seconds>(hours(1)),
+        new_bucket = twcs.newest_bucket(buckets, 2, 32, duration_cast<seconds>(hours(1)),
             time_window_compaction_strategy::get_window_lower_bound(duration_cast<seconds>(hours(1)), now), stcs_options);
         // incoming bucket should be accepted when it is larger than the min threshold SSTables
         BOOST_REQUIRE(!new_bucket.empty());
@@ -3218,13 +3220,88 @@ SEASTAR_TEST_CASE(time_window_strategy_correctness_test) {
         }
 
         now = api::timestamp_clock::now().time_since_epoch().count();
-        new_bucket = time_window_compaction_strategy::newest_bucket(buckets, 4, 32, duration_cast<seconds>(hours(1)),
+        new_bucket = twcs.newest_bucket(buckets, 4, 32, duration_cast<seconds>(hours(1)),
             time_window_compaction_strategy::get_window_lower_bound(duration_cast<seconds>(hours(1)), now), stcs_options);
         // new bucket should be trimmed to max threshold of 32
         BOOST_REQUIRE(new_bucket.size() == size_t(32));
     });
 }
 
+// Check that TWCS will only perform size-tiered on the current window and also
+// the past windows that were already previously compacted into a single SSTable.
+SEASTAR_TEST_CASE(time_window_strategy_size_tiered_behavior_correctness) {
+    using namespace std::chrono;
+
+    return test_env::do_with_async([] (test_env& env) {
+        storage_service_for_tests ssft;
+        auto s = schema_builder("tests", "time_window_strategy")
+                .with_column("id", utf8_type, column_kind::partition_key)
+                .with_column("value", int32_type).build();
+
+        auto tmp = tmpdir();
+        auto sst_gen = [&env, s, &tmp, gen = make_lw_shared<unsigned>(1)] () mutable {
+            return env.make_sstable(s, tmp.path().string(), (*gen)++, la, big);
+        };
+
+        auto make_insert = [&] (partition_key key, api::timestamp_type t) {
+            mutation m(s, key);
+            m.set_clustered_cell(clustering_key::make_empty(), bytes("value"), data_value(int32_t(1)), t);
+            return m;
+        };
+
+        std::map<sstring, sstring> options;
+        sstables::size_tiered_compaction_strategy_options stcs_options;
+        time_window_compaction_strategy twcs(options);
+        std::map<api::timestamp_type, std::vector<shared_sstable>> buckets; // windows
+        int min_threshold = 4;
+        int max_threshold = 32;
+        auto window_size = duration_cast<seconds>(hours(1));
+
+        auto add_new_sstable_to_bucket = [&] (api::timestamp_type ts, api::timestamp_type window_ts) {
+            auto key = partition_key::from_exploded(*s, {to_bytes("key" + to_sstring(ts))});
+            auto mut = make_insert(std::move(key), ts);
+            auto sst = make_sstable_containing(sst_gen, {std::move(mut)});
+            auto bound = time_window_compaction_strategy::get_window_lower_bound(window_size, window_ts);
+            buckets[bound].push_back(std::move(sst));
+        };
+
+        api::timestamp_type current_window_ts = api::timestamp_clock::now().time_since_epoch().count();
+        api::timestamp_type past_window_ts = current_window_ts - duration_cast<microseconds>(seconds(2L * 3600L)).count();
+
+        // create 1 sstable into past time window and let the strategy know about it
+        add_new_sstable_to_bucket(0, past_window_ts);
+
+        auto now = time_window_compaction_strategy::get_window_lower_bound(window_size, past_window_ts);
+
+        // past window cannot be compacted because it has a single SSTable
+        BOOST_REQUIRE(twcs.newest_bucket(buckets, min_threshold, max_threshold, window_size, now, stcs_options).size() == 0);
+
+        // create min_threshold-1 sstables into current time window
+        for (api::timestamp_type t = 0; t < min_threshold - 1; t++) {
+            add_new_sstable_to_bucket(t, current_window_ts);
+        }
+        // add 1 sstable into past window.
+        add_new_sstable_to_bucket(1, past_window_ts);
+
+        now = time_window_compaction_strategy::get_window_lower_bound(window_size, current_window_ts);
+
+        // past window can now be compacted into a single SSTable because it was the previous current (active) window.
+        // current window cannot be compacted because it has less than min_threshold SSTables
+        BOOST_REQUIRE(twcs.newest_bucket(buckets, min_threshold, max_threshold, window_size, now, stcs_options).size() == 2);
+
+        // now past window cannot be compacted again, because it was already compacted into a single SSTable, now it switches to STCS mode.
+        BOOST_REQUIRE(twcs.newest_bucket(buckets, min_threshold, max_threshold, window_size, now, stcs_options).size() == 0);
+
+        // make past window contain more than min_threshold similar-sized SSTables, allowing it to be compacted again.
+        for (api::timestamp_type t = 2; t < min_threshold; t++) {
+            add_new_sstable_to_bucket(t, past_window_ts);
+        }
+
+        // now past window can be compacted again because it switched to STCS mode and has more than min_threshold SSTables.
+        BOOST_REQUIRE(twcs.newest_bucket(buckets, min_threshold, max_threshold, window_size, now, stcs_options).size() == size_t(min_threshold));
+    });
+}
+
 SEASTAR_TEST_CASE(test_promoted_index_read) {
     // create table promoted_index_read (
     //        pk int,
@@ -4720,8 +4797,8 @@ SEASTAR_TEST_CASE(sstable_scrub_test) {
 
             table->add_sstable_and_update_cache(sst).get();
 
-            BOOST_REQUIRE(table->candidates_for_compaction().size() == 1);
-            BOOST_REQUIRE(table->candidates_for_compaction().front() == sst);
+            BOOST_REQUIRE(table->non_staging_sstables().size() == 1);
+            BOOST_REQUIRE(table->non_staging_sstables().front() == sst);
 
             auto verify_fragments = [&] (sstables::shared_sstable sst, const std::vector<mutation_fragment>& mfs) {
                 auto r = assert_that(sst->as_mutation_source().make_reader(schema, tests::make_permit()));
@@ -4742,7 +4819,7 @@ SEASTAR_TEST_CASE(sstable_scrub_test) {
             // We expect the scrub with skip_corrupted=false to stop on the first invalid fragment.
             compaction_manager.perform_sstable_scrub(table.get(), false).get();
 
-            BOOST_REQUIRE(table->candidates_for_compaction().size() == 1);
+            BOOST_REQUIRE(table->non_staging_sstables().size() == 1);
             verify_fragments(sst, corrupt_fragments);
 
             testlog.info("Scrub with --skip-corrupted=true");
@@ -4750,9 +4827,9 @@ SEASTAR_TEST_CASE(sstable_scrub_test) {
             // We expect the scrub with skip_corrupted=true to get rid of all invalid data.
             compaction_manager.perform_sstable_scrub(table.get(), true).get();
 
-            BOOST_REQUIRE(table->candidates_for_compaction().size() == 1);
-            BOOST_REQUIRE(table->candidates_for_compaction().front() != sst);
-            verify_fragments(table->candidates_for_compaction().front(), scrubbed_fragments);
+            BOOST_REQUIRE(table->non_staging_sstables().size() == 1);
+            BOOST_REQUIRE(table->non_staging_sstables().front() != sst);
+            verify_fragments(table->non_staging_sstables().front(), scrubbed_fragments);
         });
     }, test_cfg);
 }
@@ -5847,3 +5924,156 @@ SEASTAR_TEST_CASE(test_bug_6472) {
         return make_ready_future<>();
     });
 }
+
+SEASTAR_TEST_CASE(sstable_needs_cleanup_test) {
+    test_env env;
+    auto s = make_lw_shared(schema({}, some_keyspace, some_column_family,
+        {{"p1", utf8_type}}, {}, {}, {}, utf8_type));
+
+    auto tokens = token_generation_for_current_shard(10);
+
+    auto sst_gen = [&env, s, gen = make_lw_shared<unsigned>(1)] (sstring first, sstring last) mutable {
+        return sstable_for_overlapping_test(env, s, (*gen)++, first, last);
+    };
+    auto token = [&] (size_t index) -> dht::token {
+        return tokens[index].second;
+    };
+    auto key_from_token = [&] (size_t index) -> sstring {
+        return tokens[index].first;
+    };
+    auto token_range = [&] (size_t first, size_t last) -> dht::token_range {
+        return dht::token_range::make(token(first), token(last));
+    };
+
+    {
+        auto local_ranges = { token_range(0, 9) };
+        auto sst = sst_gen(key_from_token(0), key_from_token(9));
+        BOOST_REQUIRE(!needs_cleanup(sst, local_ranges, s));
+    }
+
+    {
+        auto local_ranges = { token_range(0, 1), token_range(3, 4), token_range(5, 6) };
+
+        auto sst = sst_gen(key_from_token(0), key_from_token(1));
+        BOOST_REQUIRE(!needs_cleanup(sst, local_ranges, s));
+
+        auto sst2 = sst_gen(key_from_token(2), key_from_token(2));
+        BOOST_REQUIRE(needs_cleanup(sst2, local_ranges, s));
+
+        auto sst3 = sst_gen(key_from_token(0), key_from_token(6));
+        BOOST_REQUIRE(needs_cleanup(sst3, local_ranges, s));
+
+        auto sst5 = sst_gen(key_from_token(7), key_from_token(7));
+        BOOST_REQUIRE(needs_cleanup(sst5, local_ranges, s));
+    }
+
+    return make_ready_future<>();
+}
+
+SEASTAR_TEST_CASE(test_twcs_partition_estimate) {
+    return test_setup::do_with_tmp_directory([] (test_env& env, sstring tmpdir_path) {
+        auto builder = schema_builder("tests", "test_bug_6472")
+                .with_column("id", utf8_type, column_kind::partition_key)
+                .with_column("cl", int32_type, column_kind::clustering_key)
+                .with_column("value", int32_type);
+        builder.set_compaction_strategy(sstables::compaction_strategy_type::time_window);
+        std::map<sstring, sstring> opts = {
+            { time_window_compaction_strategy_options::COMPACTION_WINDOW_UNIT_KEY, "HOURS" },
+            { time_window_compaction_strategy_options::COMPACTION_WINDOW_SIZE_KEY, "1" },
+        };
+        builder.set_compaction_strategy_options(opts);
+        builder.set_gc_grace_seconds(0);
+        auto s = builder.build();
+
+        const auto rows_per_partition = 200;
+
+        auto sst_gen = [&env, s, tmpdir_path, gen = make_lw_shared<unsigned>(1)] () mutable {
+            return env.make_sstable(s, tmpdir_path, (*gen)++, la, big);
+        };
+
+        auto next_timestamp = [] (int sstable_idx, int ck_idx) {
+            using namespace std::chrono;
+            auto window = hours(sstable_idx * rows_per_partition + ck_idx);
+            return (gc_clock::now().time_since_epoch() - duration_cast<microseconds>(window)).count();
+        };
+
+        auto tokens = token_generation_for_shard(4, this_shard_id(), test_db_config.murmur3_partitioner_ignore_msb_bits(), smp::count);
+
+        auto make_sstable = [&] (int sstable_idx) {
+            static thread_local int32_t value = 1;
+
+            auto key_str = tokens[sstable_idx].first;
+            auto key = partition_key::from_exploded(*s, {to_bytes(key_str)});
+
+            mutation m(s, key);
+            for (auto ck = 0; ck < rows_per_partition; ++ck) {
+                auto c_key = clustering_key::from_exploded(*s, {int32_type->decompose(value++)});
+                m.set_clustered_cell(c_key, bytes("value"), data_value(int32_t(value)), next_timestamp(sstable_idx, ck));
+            }
+            return make_sstable_containing(sst_gen, {m});
+        };
+
+        auto cm = make_lw_shared<compaction_manager>();
+        column_family::config cfg = column_family_test_config();
+        cfg.datadir = tmpdir_path;
+        cfg.enable_disk_writes = true;
+        cfg.enable_commitlog = false;
+        cfg.enable_cache = false;
+        cfg.enable_incremental_backups = false;
+        auto tracker = make_lw_shared<cache_tracker>();
+        cell_locker_stats cl_stats;
+        auto cf = make_lw_shared<column_family>(s, cfg, column_family::no_commitlog(), *cm, cl_stats, *tracker);
+        cf->mark_ready_for_writes();
+        cf->start();
+
+        std::vector<shared_sstable> sstables_spanning_many_windows = {
+            make_sstable(0),
+            make_sstable(1),
+            make_sstable(2),
+            make_sstable(3),
+        };
+
+        auto ret = compact_sstables(sstables::compaction_descriptor(sstables_spanning_many_windows,
+                    cf->get_sstable_set(), default_priority_class()), *cf, sst_gen, replacer_fn_no_op()).get0();
+        // The real test here is that we don't assert() in
+        // sstables::prepare_summary() with the compact_sstables() call above,
+        // this is only here as a sanity check.
+        BOOST_REQUIRE_EQUAL(ret.new_sstables.size(), std::min(sstables_spanning_many_windows.size() * rows_per_partition,
+                    sstables::time_window_compaction_strategy::max_data_segregation_window_count));
+        return make_ready_future<>();
+    });
+}
+
+SEASTAR_TEST_CASE(test_zero_estimated_partitions) {
+    return test_setup::do_with_tmp_directory([] (test_env& env, sstring tmpdir_path) {
+        simple_schema ss;
+        auto s = ss.schema();
+
+        auto pk = ss.make_pkey(make_local_key(s));
+        auto mut = mutation(s, pk);
+        ss.add_row(mut, ss.make_ckey(0), "val");
+
+        for (const auto version : all_sstable_versions) {
+            testlog.info("version={}", sstables::to_string(version));
+
+            auto mr = flat_mutation_reader_from_mutations({mut});
+
+            auto sst = env.make_sstable(s, tmpdir_path, 0, version, big);
+            sstable_writer_config cfg = test_sstables_manager.configure_writer();
+            sst->write_components(std::move(mr), 0, s, cfg, encoding_stats{}).get();
+            sst->load().get();
+
+            auto sst_mr = sst->as_mutation_source().make_reader(s, tests::make_permit(), query::full_partition_range, s->full_slice());
+            auto sst_mut = read_mutation_from_flat_mutation_reader(sst_mr, db::no_timeout).get0();
+
+            // The real test here is that we don't assert() in
+            // sstables::prepare_summary() with the write_components() call above,
+            // this is only here as a sanity check.
+            BOOST_REQUIRE(sst_mr.is_buffer_empty());
+            BOOST_REQUIRE(sst_mr.is_end_of_stream());
+            BOOST_REQUIRE_EQUAL(mut, sst_mut);
+        }
+
+        return make_ready_future<>();
+    });
+}

test/boost/stall_free_test.cc

@@ -0,0 +1,55 @@
+/*
+ * Copyright (C) 2020 ScyllaDB
+ */
+
+/*
+ * This file is part of Scylla.
+ *
+ * Scylla is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * Scylla is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with Scylla.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include <seastar/testing/thread_test_case.hh>
+#include "utils/stall_free.hh"
+
+SEASTAR_THREAD_TEST_CASE(test_merge1) {
+    std::list<int> l1{1, 2, 5, 8};
+    std::list<int> l2{3};
+    std::list<int> expected{1,2,3,5,8};
+    utils::merge_to_gently(l1, l2, std::less<int>());
+    BOOST_CHECK(l1 == expected);
+}
+
+SEASTAR_THREAD_TEST_CASE(test_merge2) {
+    std::list<int> l1{1};
+    std::list<int> l2{3, 5, 6};
+    std::list<int> expected{1,3,5,6};
+    utils::merge_to_gently(l1, l2, std::less<int>());
+    BOOST_CHECK(l1 == expected);
+}
+
+SEASTAR_THREAD_TEST_CASE(test_merge3) {
+    std::list<int> l1{};
+    std::list<int> l2{3, 5, 6};
+    std::list<int> expected{3,5,6};
+    utils::merge_to_gently(l1, l2, std::less<int>());
+    BOOST_CHECK(l1 == expected);
+}
+
+SEASTAR_THREAD_TEST_CASE(test_merge4) {
+    std::list<int> l1{1};
+    std::list<int> l2{};
+    std::list<int> expected{1};
+    utils::merge_to_gently(l1, l2, std::less<int>());
+    BOOST_CHECK(l1 == expected);
+}

test/boost/view_build_test.cc

@@ -26,12 +26,16 @@
 #include "db/system_keyspace.hh"
 
 #include <seastar/testing/test_case.hh>
+#include <seastar/testing/thread_test_case.hh>
 #include "test/lib/cql_test_env.hh"
 #include "test/lib/cql_assertions.hh"
 #include "test/lib/sstable_utils.hh"
 #include "schema_builder.hh"
 #include "service/priority_manager.hh"
 #include "test/lib/test_services.hh"
+#include "test/lib/data_model.hh"
+#include "test/lib/log.hh"
+#include "utils/ranges.hh"
 
 using namespace std::literals::chrono_literals;
 
@@ -421,23 +425,49 @@ SEASTAR_TEST_CASE(test_view_update_generator) {
         auto& view_update_generator = e.local_view_update_generator();
         auto s = test_table_schema();
 
+        std::vector<shared_sstable> ssts;
+
+        lw_shared_ptr<table> t = e.local_db().find_column_family("ks", "t").shared_from_this();
+
+        auto write_to_sstable = [&] (mutation m) {
+            auto sst = t->make_streaming_staging_sstable();
+            sstables::sstable_writer_config sst_cfg = test_sstables_manager.configure_writer();
+            auto& pc = service::get_local_streaming_priority();
+
+            sst->write_components(flat_mutation_reader_from_mutations({m}), 1ul, s, sst_cfg, {}, pc).get();
+            sst->open_data().get();
+            t->add_sstable_and_update_cache(sst).get();
+            return sst;
+        };
+
         auto key = partition_key::from_exploded(*s, {to_bytes(key1)});
         mutation m(s, key);
         auto col = s->get_column_definition("v");
         for (int i = 1024; i < 1280; ++i) {
             auto& row = m.partition().clustered_row(*s, clustering_key::from_exploded(*s, {to_bytes(fmt::format("c{}", i))}));
             row.cells().apply(*col, atomic_cell::make_live(*col->type, 2345, col->type->decompose(sstring(fmt::format("v{}", i)))));
+            // Scatter the data in a bunch of different sstables, so we
+            // can test the registration semaphore of the view update
+            // generator
+            if (!(i % 10)) {
+                ssts.push_back(write_to_sstable(std::exchange(m, mutation(s, key))));
+            }
         }
-        lw_shared_ptr<table> t = e.local_db().find_column_family("ks", "t").shared_from_this();
+        ssts.push_back(write_to_sstable(std::move(m)));
 
-        auto sst = t->make_streaming_staging_sstable();
-        sstables::sstable_writer_config sst_cfg = test_sstables_manager.configure_writer();
-        auto& pc = service::get_local_streaming_priority();
+        BOOST_REQUIRE_EQUAL(view_update_generator.available_register_units(), db::view::view_update_generator::registration_queue_size);
 
-        sst->write_components(flat_mutation_reader_from_mutations({m}), 1ul, s, sst_cfg, {}, pc).get();
-        sst->open_data().get();
-        t->add_sstable_and_update_cache(sst).get();
-        view_update_generator.register_staging_sstable(sst, t).get();
+        parallel_for_each(ssts.begin(), ssts.begin() + 10, [&] (shared_sstable& sst) {
+            return view_update_generator.register_staging_sstable(sst, t);
+        }).get();
+
+        BOOST_REQUIRE_EQUAL(view_update_generator.available_register_units(), db::view::view_update_generator::registration_queue_size);
+
+        parallel_for_each(ssts.begin() + 10, ssts.end(), [&] (shared_sstable& sst) {
+            return view_update_generator.register_staging_sstable(sst, t);
+        }).get();
+
+        BOOST_REQUIRE_EQUAL(view_update_generator.available_register_units(), db::view::view_update_generator::registration_queue_size);
 
         eventually([&, key1, key2] {
             auto msg = e.execute_cql(fmt::format("SELECT * FROM t WHERE p = '{}'", key1)).get0();
@@ -464,5 +494,373 @@ SEASTAR_TEST_CASE(test_view_update_generator) {
 
             }
         });
+
+        BOOST_REQUIRE_EQUAL(view_update_generator.available_register_units(), db::view::view_update_generator::registration_queue_size);
     });
 }
+
+SEASTAR_THREAD_TEST_CASE(test_view_update_generator_deadlock) {
+    cql_test_config test_cfg;
+    auto& db_cfg = *test_cfg.db_config;
+
+    db_cfg.enable_cache(false);
+    db_cfg.enable_commitlog(false);
+
+    test_cfg.dbcfg.emplace();
+    test_cfg.dbcfg->available_memory = memory::stats().total_memory();
+    test_cfg.dbcfg->statement_scheduling_group = seastar::create_scheduling_group("statement", 1000).get0();
+    test_cfg.dbcfg->streaming_scheduling_group = seastar::create_scheduling_group("streaming", 200).get0();
+
+    do_with_cql_env([] (cql_test_env& e) -> future<> {
+        e.execute_cql("create table t (p text, c text, v text, primary key (p, c))").get();
+        e.execute_cql("create materialized view tv as select * from t "
+                      "where p is not null and c is not null and v is not null "
+                      "primary key (v, c, p)").get();
+
+        auto msb = e.local_db().get_config().murmur3_partitioner_ignore_msb_bits();
+        auto key1 = token_generation_for_shard(1, this_shard_id(), msb).front().first;
+
+        for (auto i = 0; i < 1024; ++i) {
+            e.execute_cql(fmt::format("insert into t (p, c, v) values ('{}', 'c{}', 'x')", key1, i)).get();
+        }
+
+        // We need data on the disk so that the pre-image reader is forced to go to disk.
+        e.db().invoke_on_all([] (database& db) {
+            return db.flush_all_memtables();
+        }).get();
+
+        auto& view_update_generator = e.local_view_update_generator();
+        auto s = test_table_schema();
+
+        lw_shared_ptr<table> t = e.local_db().find_column_family("ks", "t").shared_from_this();
+
+        auto key = partition_key::from_exploded(*s, {to_bytes(key1)});
+        mutation m(s, key);
+        auto col = s->get_column_definition("v");
+        const auto filler_val_size = 4 * 1024;
+        const auto filler_val = sstring(filler_val_size, 'a');
+        for (int i = 0; i < 1024; ++i) {
+            auto& row = m.partition().clustered_row(*s, clustering_key::from_exploded(*s, {to_bytes(fmt::format("c{}", i))}));
+            row.cells().apply(*col, atomic_cell::make_live(*col->type, 2345, col->type->decompose(filler_val)));
+        }
+
+        auto sst = t->make_streaming_staging_sstable();
+        sstables::sstable_writer_config sst_cfg = test_sstables_manager.configure_writer();
+        auto& pc = service::get_local_streaming_priority();
+
+        sst->write_components(flat_mutation_reader_from_mutations({m}), 1ul, s, sst_cfg, {}, pc).get();
+        sst->open_data().get();
+        t->add_sstable_and_update_cache(sst).get();
+
+        auto& sem = *with_scheduling_group(e.local_db().get_streaming_scheduling_group(), [&] () {
+            return &e.local_db().make_query_class_config().semaphore;
+        }).get0();
+
+        // consume all units except what is needed to admit a single reader.
+        sem.consume(sem.initial_resources() - reader_resources{1, new_reader_base_cost});
+
+        testlog.info("res = [.count={}, .memory={}]", sem.available_resources().count, sem.available_resources().memory);
+
+        BOOST_REQUIRE_EQUAL(sem.get_inactive_read_stats().permit_based_evictions, 0);
+
+        view_update_generator.register_staging_sstable(sst, t).get();
+
+        eventually_true([&] {
+            return sem.get_inactive_read_stats().permit_based_evictions > 0;
+        });
+
+        return make_ready_future<>();
+    }, std::move(test_cfg)).get();
+}
+
+// Test that registered sstables (and semaphore units) are not leaked when
+// sstables are register *while* a batch of sstables are processed.
+SEASTAR_THREAD_TEST_CASE(test_view_update_generator_register_semaphore_unit_leak) {
+    cql_test_config test_cfg;
+    auto& db_cfg = *test_cfg.db_config;
+
+    db_cfg.enable_cache(false);
+    db_cfg.enable_commitlog(false);
+
+    do_with_cql_env([] (cql_test_env& e) -> future<> {
+        e.execute_cql("create table t (p text, c text, v text, primary key (p, c))").get();
+        e.execute_cql("create materialized view tv as select * from t "
+                      "where p is not null and c is not null and v is not null "
+                      "primary key (v, c, p)").get();
+
+        auto msb = e.local_db().get_config().murmur3_partitioner_ignore_msb_bits();
+        auto key1 = token_generation_for_shard(1, this_shard_id(), msb).front().first;
+
+        for (auto i = 0; i < 1024; ++i) {
+            e.execute_cql(fmt::format("insert into t (p, c, v) values ('{}', 'c{}', 'x')", key1, i)).get();
+        }
+
+        // We need data on the disk so that the pre-image reader is forced to go to disk.
+        e.db().invoke_on_all([] (database& db) {
+            return db.flush_all_memtables();
+        }).get();
+
+        auto& view_update_generator = e.local_view_update_generator();
+        auto s = test_table_schema();
+
+        lw_shared_ptr<table> t = e.local_db().find_column_family("ks", "t").shared_from_this();
+
+        auto key = partition_key::from_exploded(*s, {to_bytes(key1)});
+
+        auto make_sstable = [&] {
+            mutation m(s, key);
+            auto col = s->get_column_definition("v");
+            const auto val = sstring(10, 'a');
+            for (int i = 0; i < 1024; ++i) {
+                auto& row = m.partition().clustered_row(*s, clustering_key::from_exploded(*s, {to_bytes(fmt::format("c{}", i))}));
+                row.cells().apply(*col, atomic_cell::make_live(*col->type, 2345, col->type->decompose(val)));
+            }
+
+            auto sst = t->make_streaming_staging_sstable();
+            sstables::sstable_writer_config sst_cfg = test_sstables_manager.configure_writer();
+            auto& pc = service::get_local_streaming_priority();
+
+            sst->write_components(flat_mutation_reader_from_mutations({m}), 1ul, s, sst_cfg, {}, pc).get();
+            sst->open_data().get();
+            t->add_sstable_and_update_cache(sst).get();
+            return sst;
+        };
+
+        std::vector<sstables::shared_sstable> prepared_sstables;
+
+        // We need 2 * N + 1 sstables. N should be at least 5 (number of units
+        // on the register semaphore) + 1 (just to make sure the returned future
+        // blocks). While the initial batch is processed we want to register N
+        // more sstables, + 1 to detect the leak (N units will be returned from
+        // the initial batch). See below for more details.
+        const auto num_sstables = (view_update_generator.available_register_units() + 1) * 2 + 1;
+        for (auto i = 0; i < num_sstables; ++i) {
+            prepared_sstables.push_back(make_sstable());
+        }
+
+        // First batch: register N sstables.
+        while (view_update_generator.available_register_units()) {
+            auto fut = view_update_generator.register_staging_sstable(std::move(prepared_sstables.back()), t);
+            prepared_sstables.pop_back();
+            BOOST_REQUIRE(fut.available());
+        }
+
+        // Make sure we consumed all units and thus the register future blocks.
+        auto fut1 = view_update_generator.register_staging_sstable(std::move(prepared_sstables.back()), t);
+        prepared_sstables.pop_back();
+        BOOST_REQUIRE(!fut1.available());
+
+        std::vector<future<>> futures;
+        futures.reserve(prepared_sstables.size());
+
+        // While the first batch is processed, concurrently register the
+        // remaining N + 1 sstables, yielding in-between so the first batch
+        // processing can progress.
+        while (!prepared_sstables.empty()) {
+            thread::yield();
+            futures.emplace_back(view_update_generator.register_staging_sstable(std::move(prepared_sstables.back()), t));
+            prepared_sstables.pop_back();
+        }
+
+        // Make sure the first batch is processed.
+        fut1.get();
+
+        auto fut_res = when_all_succeed(futures.begin(), futures.end());
+
+        // Watchdog timer which will break out of the deadlock and fail the test.
+        timer watchdog_timer([&view_update_generator] {
+            // Re-start it so stop() on shutdown doesn't crash.
+            (void)view_update_generator.stop().then([&] { return view_update_generator.start(); });
+        });
+
+        watchdog_timer.arm(std::chrono::seconds(60));
+
+        // Wait on the second batch, will fail if the watchdog timer fails.
+        fut_res.get();
+
+        watchdog_timer.cancel();
+
+        return make_ready_future<>();
+
+    }, std::move(test_cfg)).get();
+}
+
+SEASTAR_THREAD_TEST_CASE(test_view_update_generator_buffering) {
+    using partition_size_map = std::map<dht::decorated_key, size_t, dht::ring_position_less_comparator>;
+
+    class consumer_verifier {
+        schema_ptr _schema;
+        reader_permit _permit;
+        const partition_size_map& _partition_rows;
+        std::vector<mutation>& _collected_muts;
+        bool& _failed;
+        std::unique_ptr<row_locker> _rl;
+        std::unique_ptr<row_locker::stats> _rl_stats;
+        clustering_key::less_compare _less_cmp;
+        const size_t _max_rows_soft;
+        const size_t _max_rows_hard;
+        size_t _buffer_rows = 0;
+
+    private:
+        static size_t rows_in_limit(size_t l) {
+            const size_t _100kb = 100 * 1024;
+            // round up
+            return l / _100kb + std::min(size_t(1), l % _100kb);
+        }
+
+        static size_t rows_in_mut(const mutation& m) {
+            return std::distance(m.partition().clustered_rows().begin(), m.partition().clustered_rows().end());
+        }
+
+        void check(mutation mut) {
+            // First we check that we would be able to create a reader, even
+            // though the staging reader consumed all resources.
+            auto fut = _permit.wait_admission(new_reader_base_cost, db::timeout_clock::now());
+            BOOST_REQUIRE(!fut.failed());
+            auto res_units = fut.get0();
+
+            const size_t current_rows = rows_in_mut(mut);
+            const auto total_rows = _partition_rows.at(mut.decorated_key());
+            _buffer_rows += current_rows;
+
+            testlog.trace("consumer_verifier::check(): key={}, rows={}/{}, _buffer={}",
+                    partition_key::with_schema_wrapper(*_schema, mut.key()),
+                    current_rows,
+                    total_rows,
+                    _buffer_rows);
+
+            BOOST_REQUIRE(current_rows);
+            BOOST_REQUIRE(current_rows <= _max_rows_hard);
+            BOOST_REQUIRE(_buffer_rows <= _max_rows_hard);
+
+            // The current partition doesn't have all of its rows yet, verify
+            // that the new mutation contains the next rows for the same
+            // partition
+            if (!_collected_muts.empty() && rows_in_mut(_collected_muts.back()) < _partition_rows.at(_collected_muts.back().decorated_key())) {
+                BOOST_REQUIRE(_collected_muts.back().decorated_key().equal(*mut.schema(), mut.decorated_key()));
+                const auto& previous_ckey = (--_collected_muts.back().partition().clustered_rows().end())->key();
+                const auto& next_ckey = mut.partition().clustered_rows().begin()->key();
+                BOOST_REQUIRE(_less_cmp(previous_ckey, next_ckey));
+                mutation_application_stats stats;
+                _collected_muts.back().partition().apply(*_schema, mut.partition(), *mut.schema(), stats);
+            // The new mutation is a new partition.
+            } else {
+                if (!_collected_muts.empty()) {
+                    BOOST_REQUIRE(!_collected_muts.back().decorated_key().equal(*mut.schema(), mut.decorated_key()));
+                }
+                _collected_muts.push_back(std::move(mut));
+            }
+
+            if (_buffer_rows >= _max_rows_hard) { // buffer flushed on hard limit
+                _buffer_rows = 0;
+                testlog.trace("consumer_verifier::check(): buffer ends on hard limit");
+            } else if (_buffer_rows >= _max_rows_soft) { // buffer flushed on soft limit
+                _buffer_rows = 0;
+                testlog.trace("consumer_verifier::check(): buffer ends on soft limit");
+            }
+        }
+
+    public:
+        consumer_verifier(schema_ptr schema, reader_permit permit, const partition_size_map& partition_rows, std::vector<mutation>& collected_muts, bool& failed)
+            : _schema(std::move(schema))
+            , _permit(std::move(permit))
+            , _partition_rows(partition_rows)
+            , _collected_muts(collected_muts)
+            , _failed(failed)
+            , _rl(std::make_unique<row_locker>(_schema))
+            , _rl_stats(std::make_unique<row_locker::stats>())
+            , _less_cmp(*_schema)
+            , _max_rows_soft(rows_in_limit(db::view::view_updating_consumer::buffer_size_soft_limit))
+            , _max_rows_hard(rows_in_limit(db::view::view_updating_consumer::buffer_size_hard_limit))
+        { }
+
+        future<row_locker::lock_holder> operator()(mutation mut) {
+            try {
+                check(std::move(mut));
+            } catch (...) {
+                testlog.error("consumer_verifier::operator(): caught unexpected exception {}", std::current_exception());
+                _failed |= true;
+            }
+            return _rl->lock_pk(_collected_muts.back().decorated_key(), true, db::no_timeout, *_rl_stats);
+        }
+    };
+
+    reader_concurrency_semaphore sem(1, new_reader_base_cost, get_name());
+
+    auto schema = schema_builder("ks", "cf")
+            .with_column("pk", int32_type, column_kind::partition_key)
+            .with_column("ck", int32_type, column_kind::clustering_key)
+            .with_column("v", bytes_type)
+            .build();
+
+    const auto blob_100kb = bytes(100 * 1024, bytes::value_type(0xab));
+    const abort_source as;
+
+    const auto partition_size_sets = std::vector<std::vector<int>>{{12}, {8, 4}, {8, 16}, {22}, {8, 8, 8, 8}, {8, 8, 8, 16, 8}, {8, 20, 16, 16}, {50}, {21}, {21, 2}};
+    const auto max_partition_set_size = std::ranges::max_element(partition_size_sets, [] (const std::vector<int>& a, const std::vector<int>& b) { return a.size() < b.size(); })->size();
+    auto pkeys = ranges::to<std::vector<dht::decorated_key>>(std::views::iota(size_t{0}, max_partition_set_size) | std::views::transform([schema] (int i) {
+        return dht::decorate_key(*schema, partition_key::from_single_value(*schema, int32_type->decompose(data_value(i))));
+    }));
+    std::ranges::sort(pkeys, dht::ring_position_less_comparator(*schema));
+
+    for (auto partition_sizes_100kb : partition_size_sets) {
+        testlog.debug("partition_sizes_100kb={}", partition_sizes_100kb);
+        partition_size_map partition_rows{dht::ring_position_less_comparator(*schema)};
+        std::vector<mutation> muts;
+        auto pk = 0;
+        for (auto partition_size_100kb : partition_sizes_100kb) {
+            auto mut_desc = tests::data_model::mutation_description(pkeys.at(pk++).key().explode(*schema));
+            for (auto ck = 0; ck < partition_size_100kb; ++ck) {
+                mut_desc.add_clustered_cell({int32_type->decompose(data_value(ck))}, "v", tests::data_model::mutation_description::value(blob_100kb));
+            }
+            muts.push_back(mut_desc.build(schema));
+            partition_rows.emplace(muts.back().decorated_key(), partition_size_100kb);
+        }
+
+        std::ranges::sort(muts, [less = dht::ring_position_less_comparator(*schema)] (const mutation& a, const mutation& b) {
+            return less(a.decorated_key(), b.decorated_key());
+        });
+
+        auto permit = sem.make_permit();
+
+        auto mt = make_lw_shared<memtable>(schema);
+        for (const auto& mut : muts) {
+            mt->apply(mut);
+        }
+
+        auto ms = mutation_source([mt] (
+                    schema_ptr s,
+                    reader_permit permit,
+                    const dht::partition_range& pr,
+                    const query::partition_slice& ps,
+                    const io_priority_class& pc,
+                    tracing::trace_state_ptr ts,
+                    streamed_mutation::forwarding fwd_ms,
+                    mutation_reader::forwarding fwd_mr) {
+            return make_restricted_flat_reader(mt->as_data_source(), s, std::move(permit), pr, ps, pc, std::move(ts), fwd_ms, fwd_mr);
+        });
+        auto [staging_reader, staging_reader_handle] = make_manually_paused_evictable_reader(
+                std::move(ms),
+                schema,
+                permit,
+                query::full_partition_range,
+                schema->full_slice(),
+                service::get_local_streaming_priority(),
+                nullptr,
+                ::mutation_reader::forwarding::no);
+
+        std::vector<mutation> collected_muts;
+        bool failed = false;
+
+        staging_reader.consume_in_thread(db::view::view_updating_consumer(schema, as, staging_reader_handle,
+                    consumer_verifier(schema, permit, partition_rows, collected_muts, failed)), db::no_timeout);
+
+        BOOST_REQUIRE(!failed);
+
+        BOOST_REQUIRE_EQUAL(muts.size(), collected_muts.size());
+        for (size_t i = 0; i < muts.size(); ++i) {
+            testlog.trace("compare mutation {}", i);
+            BOOST_REQUIRE_EQUAL(muts[i], collected_muts[i]);
+        }
+    }
+}

test/boost/view_schema_test.cc

@@ -118,6 +118,25 @@ SEASTAR_TEST_CASE(test_access_and_schema) {
     });
 }
 
+SEASTAR_TEST_CASE(test_column_dropped_from_base) {
+    return do_with_cql_env_thread([] (auto& e) {
+        e.execute_cql("create table cf (p int, c ascii, a int, v int, primary key (p, c));").get();
+        e.execute_cql("create materialized view vcf as select p, c, v from cf "
+                      "where v is not null and p is not null and c is not null "
+                      "primary key (v, p, c)").get();
+        e.execute_cql("alter table cf drop a;").get();
+        e.execute_cql("insert into cf (p, c, v) values (0, 'foo', 1);").get();
+        eventually([&] {
+        auto msg = e.execute_cql("select v from vcf").get0();
+        assert_that(msg).is_rows()
+            .with_size(1)
+            .with_row({
+                {int32_type->decompose(1)}
+            });
+        });
+    });
+}
+
 SEASTAR_TEST_CASE(test_updates) {
     return do_with_cql_env_thread([] (auto& e) {
         e.execute_cql("create table base (k int, v int, primary key (k));").get();