Currently, we use std::vector<*mutation> to keep
a list of mutations for processing.
This can lead to large allocation, e.g. when the vector
size is a function of the number of tables.
Use a chunked vector instead to prevent oversized allocations.
`perf-simple-query --smp 1` results obtained for fixed 400MHz frequency
and PGO disabled:
Before (read path):
```
enable-cache=1
Running test with config: {partitions=10000, concurrency=100, mode=read, query_single_key=no, counters=no}
Disabling auto compaction
Creating 10000 partitions...
89055.97 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.2 tasks/op, 39417 insns/op, 18003 cycles/op, 0 errors)
103372.72 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.2 tasks/op, 39380 insns/op, 17300 cycles/op, 0 errors)
98942.27 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.2 tasks/op, 39413 insns/op, 17336 cycles/op, 0 errors)
103752.93 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.2 tasks/op, 39407 insns/op, 17252 cycles/op, 0 errors)
102516.77 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.2 tasks/op, 39403 insns/op, 17288 cycles/op, 0 errors)
throughput:
mean= 99528.13 standard-deviation=6155.71
median= 102516.77 median-absolute-deviation=3844.59
maximum=103752.93 minimum=89055.97
instructions_per_op:
mean= 39403.99 standard-deviation=14.25
median= 39406.75 median-absolute-deviation=9.30
maximum=39416.63 minimum=39380.39
cpu_cycles_per_op:
mean= 17435.81 standard-deviation=318.24
median= 17300.40 median-absolute-deviation=147.59
maximum=18002.53 minimum=17251.75
```
After (read path)
```
enable-cache=1
Running test with config: {partitions=10000, concurrency=100, mode=read, query_single_key=no, counters=no}
Disabling auto compaction
Creating 10000 partitions...
59755.04 tps ( 66.2 allocs/op, 0.0 logallocs/op, 14.2 tasks/op, 39466 insns/op, 22834 cycles/op, 0 errors)
71854.16 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.2 tasks/op, 39417 insns/op, 17883 cycles/op, 0 errors)
82149.45 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.2 tasks/op, 39411 insns/op, 17409 cycles/op, 0 errors)
49640.04 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.3 tasks/op, 39474 insns/op, 19975 cycles/op, 0 errors)
54963.22 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.3 tasks/op, 39474 insns/op, 18235 cycles/op, 0 errors)
throughput:
mean= 63672.38 standard-deviation=13195.12
median= 59755.04 median-absolute-deviation=8709.16
maximum=82149.45 minimum=49640.04
instructions_per_op:
mean= 39448.38 standard-deviation=31.60
median= 39466.17 median-absolute-deviation=25.75
maximum=39474.12 minimum=39411.42
cpu_cycles_per_op:
mean= 19267.01 standard-deviation=2217.03
median= 18234.80 median-absolute-deviation=1384.25
maximum=22834.26 minimum=17408.67
```
`perf-simple-query --smp 1 --write` results obtained for fixed 400MHz frequency
and PGO disabled:
Before (write path):
```
enable-cache=1
Running test with config: {partitions=10000, concurrency=100, mode=write, query_single_key=no, counters=no}
Disabling auto compaction
63736.96 tps ( 59.4 allocs/op, 16.4 logallocs/op, 14.3 tasks/op, 49667 insns/op, 19924 cycles/op, 0 errors)
64109.41 tps ( 59.3 allocs/op, 16.0 logallocs/op, 14.3 tasks/op, 49992 insns/op, 20084 cycles/op, 0 errors)
56950.47 tps ( 59.3 allocs/op, 16.0 logallocs/op, 14.3 tasks/op, 50005 insns/op, 20501 cycles/op, 0 errors)
44858.42 tps ( 59.3 allocs/op, 16.0 logallocs/op, 14.3 tasks/op, 50014 insns/op, 21947 cycles/op, 0 errors)
28592.87 tps ( 59.3 allocs/op, 16.0 logallocs/op, 14.3 tasks/op, 50027 insns/op, 27659 cycles/op, 0 errors)
throughput:
mean= 51649.63 standard-deviation=15059.74
median= 56950.47 median-absolute-deviation=12087.33
maximum=64109.41 minimum=28592.87
instructions_per_op:
mean= 49941.18 standard-deviation=153.76
median= 50005.24 median-absolute-deviation=73.01
maximum=50027.07 minimum=49667.05
cpu_cycles_per_op:
mean= 22023.01 standard-deviation=3249.92
median= 20500.74 median-absolute-deviation=1938.76
maximum=27658.75 minimum=19924.32
```
After (write path)
```
enable-cache=1
Running test with config: {partitions=10000, concurrency=100, mode=write, query_single_key=no, counters=no}
Disabling auto compaction
53395.93 tps ( 59.4 allocs/op, 16.5 logallocs/op, 14.3 tasks/op, 50326 insns/op, 21252 cycles/op, 0 errors)
46527.83 tps ( 59.3 allocs/op, 16.0 logallocs/op, 14.3 tasks/op, 50704 insns/op, 21555 cycles/op, 0 errors)
55846.30 tps ( 59.3 allocs/op, 16.0 logallocs/op, 14.3 tasks/op, 50731 insns/op, 21060 cycles/op, 0 errors)
55669.30 tps ( 59.3 allocs/op, 16.0 logallocs/op, 14.3 tasks/op, 50735 insns/op, 21521 cycles/op, 0 errors)
52130.17 tps ( 59.3 allocs/op, 16.0 logallocs/op, 14.3 tasks/op, 50757 insns/op, 21334 cycles/op, 0 errors)
throughput:
mean= 52713.91 standard-deviation=3795.38
median= 53395.93 median-absolute-deviation=2955.40
maximum=55846.30 minimum=46527.83
instructions_per_op:
mean= 50650.57 standard-deviation=182.46
median= 50731.38 median-absolute-deviation=84.09
maximum=50756.62 minimum=50325.87
cpu_cycles_per_op:
mean= 21344.42 standard-deviation=202.86
median= 21334.00 median-absolute-deviation=176.37
maximum=21554.61 minimum=21060.24
```
Fixes #24815
Improvement for rare corner cases. No backport required
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Closes scylladb/scylladb#24919
225 lines
8.4 KiB
C++
225 lines
8.4 KiB
C++
/*
|
|
* Copyright (C) 2017-present ScyllaDB
|
|
*/
|
|
|
|
/*
|
|
* SPDX-License-Identifier: LicenseRef-ScyllaDB-Source-Available-1.0
|
|
*/
|
|
|
|
#include "auth/common.hh"
|
|
|
|
#include <optional>
|
|
#include <seastar/core/coroutine.hh>
|
|
#include <seastar/core/shared_ptr.hh>
|
|
#include <seastar/core/sharded.hh>
|
|
|
|
#include "mutation/canonical_mutation.hh"
|
|
#include "schema/schema_fwd.hh"
|
|
#include "timestamp.hh"
|
|
#include "utils/assert.hh"
|
|
#include "utils/exponential_backoff_retry.hh"
|
|
#include "cql3/query_processor.hh"
|
|
#include "cql3/statements/create_table_statement.hh"
|
|
#include "schema/schema_builder.hh"
|
|
#include "service/migration_manager.hh"
|
|
#include "service/raft/group0_state_machine.hh"
|
|
#include "timeout_config.hh"
|
|
#include "utils/error_injection.hh"
|
|
#include "db/system_keyspace.hh"
|
|
|
|
namespace auth {
|
|
|
|
namespace meta {
|
|
|
|
namespace legacy {
|
|
constinit const std::string_view AUTH_KS("system_auth");
|
|
constinit const std::string_view USERS_CF("users");
|
|
} // namespace legacy
|
|
constinit const std::string_view AUTH_PACKAGE_NAME("org.apache.cassandra.auth.");
|
|
} // namespace meta
|
|
|
|
static logging::logger auth_log("auth");
|
|
|
|
bool legacy_mode(cql3::query_processor& qp) {
|
|
return qp.auth_version < db::auth_version_t::v2;
|
|
}
|
|
|
|
std::string_view get_auth_ks_name(cql3::query_processor& qp) {
|
|
if (legacy_mode(qp)) {
|
|
return meta::legacy::AUTH_KS;
|
|
}
|
|
return db::system_keyspace::NAME;
|
|
}
|
|
|
|
// Func must support being invoked more than once.
|
|
future<> do_after_system_ready(seastar::abort_source& as, seastar::noncopyable_function<future<>()> func) {
|
|
struct empty_state { };
|
|
return exponential_backoff_retry::do_until_value(1s, 1min, as, [func = std::move(func)] {
|
|
return func().then_wrapped([] (auto&& f) -> std::optional<empty_state> {
|
|
if (f.failed()) {
|
|
auth_log.debug("Auth task failed with error, rescheduling: {}", f.get_exception());
|
|
return { };
|
|
}
|
|
return { empty_state() };
|
|
});
|
|
}).discard_result();
|
|
}
|
|
|
|
static future<> create_legacy_metadata_table_if_missing_impl(
|
|
std::string_view table_name,
|
|
cql3::query_processor& qp,
|
|
std::string_view cql,
|
|
::service::migration_manager& mm) {
|
|
SCYLLA_ASSERT(this_shard_id() == 0); // once_among_shards makes sure a function is executed on shard 0 only
|
|
|
|
auto db = qp.db();
|
|
auto parsed_statement = cql3::query_processor::parse_statement(cql, cql3::dialect{});
|
|
auto& parsed_cf_statement = static_cast<cql3::statements::raw::cf_statement&>(*parsed_statement);
|
|
|
|
parsed_cf_statement.prepare_keyspace(meta::legacy::AUTH_KS);
|
|
|
|
auto statement = static_pointer_cast<cql3::statements::create_table_statement>(
|
|
parsed_cf_statement.prepare(db, qp.get_cql_stats())->statement);
|
|
|
|
const auto schema = statement->get_cf_meta_data(qp.db());
|
|
const auto uuid = generate_legacy_id(schema->ks_name(), schema->cf_name());
|
|
|
|
schema_builder b(schema);
|
|
b.set_uuid(uuid);
|
|
schema_ptr table = b.build();
|
|
|
|
if (!db.has_schema(table->ks_name(), table->cf_name())) {
|
|
auto group0_guard = co_await mm.start_group0_operation();
|
|
auto ts = group0_guard.write_timestamp();
|
|
try {
|
|
co_return co_await mm.announce(co_await ::service::prepare_new_column_family_announcement(qp.proxy(), table, ts),
|
|
std::move(group0_guard), format("auth: create {} metadata table", table->cf_name()));
|
|
} catch (exceptions::already_exists_exception&) {}
|
|
}
|
|
}
|
|
|
|
future<> create_legacy_metadata_table_if_missing(
|
|
std::string_view table_name,
|
|
cql3::query_processor& qp,
|
|
std::string_view cql,
|
|
::service::migration_manager& mm) noexcept {
|
|
return futurize_invoke(create_legacy_metadata_table_if_missing_impl, table_name, qp, cql, mm);
|
|
}
|
|
|
|
::service::query_state& internal_distributed_query_state() noexcept {
|
|
#ifdef DEBUG
|
|
// Give the much slower debug tests more headroom for completing auth queries.
|
|
static const auto t = 30s;
|
|
#else
|
|
static const auto t = 5s;
|
|
#endif
|
|
static const timeout_config tc{t, t, t, t, t, t, t};
|
|
static thread_local ::service::client_state cs(::service::client_state::internal_tag{}, tc);
|
|
static thread_local ::service::query_state qs(cs, empty_service_permit());
|
|
return qs;
|
|
}
|
|
|
|
::service::raft_timeout get_raft_timeout() noexcept {
|
|
auto dur = internal_distributed_query_state().get_client_state().get_timeout_config().other_timeout;
|
|
return ::service::raft_timeout{.value = lowres_clock::now() + dur};
|
|
}
|
|
|
|
static future<> announce_mutations_with_guard(
|
|
::service::raft_group0_client& group0_client,
|
|
utils::chunked_vector<canonical_mutation> muts,
|
|
::service::group0_guard group0_guard,
|
|
seastar::abort_source& as,
|
|
std::optional<::service::raft_timeout> timeout) {
|
|
auto group0_cmd = group0_client.prepare_command(
|
|
::service::write_mutations{
|
|
.mutations{std::move(muts)},
|
|
},
|
|
group0_guard,
|
|
"auth: modify internal data"
|
|
);
|
|
return group0_client.add_entry(std::move(group0_cmd), std::move(group0_guard), as, timeout);
|
|
}
|
|
|
|
future<> announce_mutations_with_batching(
|
|
::service::raft_group0_client& group0_client,
|
|
start_operation_func_t start_operation_func,
|
|
std::function<::service::mutations_generator(api::timestamp_type t)> gen,
|
|
seastar::abort_source& as,
|
|
std::optional<::service::raft_timeout> timeout) {
|
|
// account for command's overhead, it's better to use smaller threshold than constantly bounce off the limit
|
|
size_t memory_threshold = group0_client.max_command_size() * 0.75;
|
|
utils::get_local_injector().inject("auth_announce_mutations_command_max_size",
|
|
[&memory_threshold] {
|
|
memory_threshold = 1000;
|
|
});
|
|
|
|
size_t memory_usage = 0;
|
|
utils::chunked_vector<canonical_mutation> muts;
|
|
|
|
// guard has to be taken before we execute code in gen as
|
|
// it can do read-before-write and we want announce_mutations
|
|
// operation to be linearizable with other such calls,
|
|
// for instance if we do select and then delete in gen
|
|
// we want both to operate on the same data or fail
|
|
// if someone else modified it in the middle
|
|
std::optional<::service::group0_guard> group0_guard;
|
|
group0_guard = co_await start_operation_func(as);
|
|
auto timestamp = group0_guard->write_timestamp();
|
|
|
|
auto g = gen(timestamp);
|
|
while (auto mut = co_await g()) {
|
|
muts.push_back(canonical_mutation{*mut});
|
|
memory_usage += muts.back().representation().size();
|
|
if (memory_usage >= memory_threshold) {
|
|
if (!group0_guard) {
|
|
group0_guard = co_await start_operation_func(as);
|
|
timestamp = group0_guard->write_timestamp();
|
|
}
|
|
co_await announce_mutations_with_guard(group0_client, std::move(muts), std::move(*group0_guard), as, timeout);
|
|
group0_guard = std::nullopt;
|
|
memory_usage = 0;
|
|
muts = {};
|
|
}
|
|
}
|
|
if (!muts.empty()) {
|
|
if (!group0_guard) {
|
|
group0_guard = co_await start_operation_func(as);
|
|
timestamp = group0_guard->write_timestamp();
|
|
}
|
|
co_await announce_mutations_with_guard(group0_client, std::move(muts), std::move(*group0_guard), as, timeout);
|
|
}
|
|
}
|
|
|
|
future<> announce_mutations(
|
|
cql3::query_processor& qp,
|
|
::service::raft_group0_client& group0_client,
|
|
const sstring query_string,
|
|
std::vector<data_value_or_unset> values,
|
|
seastar::abort_source& as,
|
|
std::optional<::service::raft_timeout> timeout) {
|
|
auto group0_guard = co_await group0_client.start_operation(as, timeout);
|
|
auto timestamp = group0_guard.write_timestamp();
|
|
auto muts = co_await qp.get_mutations_internal(
|
|
query_string,
|
|
internal_distributed_query_state(),
|
|
timestamp,
|
|
std::move(values));
|
|
utils::chunked_vector<canonical_mutation> cmuts = {muts.begin(), muts.end()};
|
|
co_await announce_mutations_with_guard(group0_client, std::move(cmuts), std::move(group0_guard), as, timeout);
|
|
}
|
|
|
|
future<> collect_mutations(
|
|
cql3::query_processor& qp,
|
|
::service::group0_batch& collector,
|
|
const sstring query_string,
|
|
std::vector<data_value_or_unset> values) {
|
|
auto muts = co_await qp.get_mutations_internal(
|
|
query_string,
|
|
internal_distributed_query_state(),
|
|
collector.write_timestamp(),
|
|
std::move(values));
|
|
collector.add_mutations(std::move(muts), format("auth internal statement: {}", query_string));
|
|
}
|
|
|
|
}
|