Problem
-------
Secondary indexes are implemented via materialized views under the
hood. The way an index behaves is determined by the configuration
of the view. Currently, it can be modified by performing the CQL
statement `ALTER MATERIALIZED VIEW` on it. However, that raises some
concerns.
Consider, for instance, the following scenario:
1. The user creates a secondary index on a table.
2. In parallel, the user performs writes to the base table.
3. The user modifies the underlying materialized view, e.g. by setting
the `synchronous_updates` to `true` [1].
Some of the writes that happened before step 3 used the default value
of the property (which is `false`). That had an actual consequence
on what happened later on: the view updates were performed
asynchronously. Only after step 3 had finished did it change.
Unfortunately, as of now, there is no way to avoid a situation like
that. Whenever the user wants to configure a secondary index they're
creating, they need to do it in another schema change. Since it's
not always possible to control how the database is manipulated in
the meantime, it leads to problems like the one described.
That's not all, though. The fact that it's not possible to configure
secondary indexes is inconsistent with other schema entities. When
it comes to tables or materialized views, the user always have a means
to set some or even all of the properties during their creation.
Solution
--------
The solution to this problem is extending the `CREATE INDEX` CQL
statement by view properties. The syntax is of form:
```
> CREATE INDEX <index name>
> .. ON <keyspace>.<table> (<columns>)
> .. WITH <properties>
```
where `<properties>` corresponds to both index-specific and view
properties [2, 3]. View properties can only be used with indexes
implemented with materialized views; for example, it will be impossible
to create a vector index when specifying any view property (see
examples below).
When a view property is provided, it will be applied when creating the
underlying materialized view. The behavior should be similar to how
other CQL statements responsible for creating schema entities work.
High-level implementation strategy
----------------------------------
1. Make auxiliary changes.
2. Introduce data structures representing the new set of index
properties: both index-specific and those corresponding to the
underlying view.
3. Extend `CREATE INDEX` to accept view properties.
4. Extend `DESCRIBE INDEX` and other `DESCRIBE` statements to include
view properties in their output.
User documentation is also updated at the steps to reflect the
corresponding changes.
Implementation considerations
-----------------------------
There are a number of schema properties that are now obsolete. They're
accepted by other CQL statements, but they have no effect. They
include:
* `index_interval`
* `replicate_on_write`
* `populate_io_cache_on_flush`
* `read_repair_chance`
* `dclocal_read_repair_chance`
If the user tries to create a secondary index specifying any of those
keywords, the statement will fail with an appropriate error (see
examples below).
Unlike materialized views, we forbid specifying the clustering order
when creating a secondary index [4]. This limitation may be lifted
later on, but it's a detail that may or may not prove troublesome. It's
better to postpone covering it to when we have a better perspective on
the consequences it would bring.
Examples
--------
Good examples
```
> CREATE INDEX idx ON ks.t (v);
> CREATE INDEX idx ON ks.t (v) WITH comment = 'ok view property';
> CREATE INDEX idx ON ks.t (v)
.. WITH comment = 'multiple view properties are ok'
.. AND synchronous_updates = true;
> CREATE INDEX idx ON ks.t (v)
.. WITH comment = 'default value ok'
.. AND synchronous_updates = false;
```
Bad examples
```
> CREATE INDEX idx ON ks.t (v) WITH replicate_on_write = true;
SyntaxException: Unknown property 'replicate_on_write'
> CREATE INDEX idx ON ks.t (v)
.. WITH OPTIONS = {'option1': 'value1'}
.. AND comment = 'some text';
InvalidRequest: Error from server: code=2200 [Invalid query]
message="Cannot specify options for a non-CUSTOM index"
> CREATE CUSTOM INDEX idx ON ks.t (v)
.. WITH OPTIONS = {'option1': 'value1'}
.. AND comment = 'some text';
InvalidRequest: Error from server: code=2200 [Invalid query]
message="CUSTOM index requires specifying the index class"
> CREATE CUSTOM INDEX idx ON ks.t (v)
.. USING 'vector_index'
.. WITH OPTIONS = {'option1': 'value1'}
.. AND comment = 'some text';
InvalidRequest: Error from server: code=2200 [Invalid query]
message="You cannot use view properties with a vector index"
> CREATE INDEX idx ON ks.t (v) WITH CLUSTERING ORDER BY (v ASC);
InvalidRequest: Error from server: code=2200 [Invalid query]
message="Indexes do not allow for specifying the clustering order"
```
and so on. For more examples, see the relevant tests.
References:
[1] https://docs.scylladb.com/manual/branch-2025.4/cql/cql-extensions.html#synchronous-materialized-views
[2] https://docs.scylladb.com/manual/branch-2025.4/cql/secondary-indexes.html#create-index
[3] https://docs.scylladb.com/manual/branch-2025.4/cql/mv.html#mv-options
[4] https://docs.scylladb.com/manual/branch-2025.4/cql/dml/select.html#ordering-clauseFixesscylladb/scylladb#16454
Backport: not needed. This is an enhancement.
Closesscylladb/scylladb#24977
* github.com:scylladb/scylladb:
cql3: Extend DESC INDEX by view properties
cql3: Forbid using CLUSTERING ORDER BY when creating index
cql3: Extend CREATE INDEX by MV properties
cql3/statements/create_index_statement: Allow for view options
cql3/statements/create_index_statement: Rename member
cql3/statements/index_prop_defs: Re-introduce index_prop_defs
cql3/statements/property_definitions: Add extract_property()
cql3/statements/index_prop_defs.cc: Add namespace
cql3/statements/index_prop_defs.hh: Rename type
cql3/statements/view_prop_defs.cc: Move validation logic into file
cql3/statements: Introduce view_prop_defs.{hh,cc}
cql3/statements/create_view_statement.cc: Move validation of ID
schema/schema.hh: Do not include index_prop_defs.hh
Read timeouts are a common occurence and they typically occur when the replica is overloaded. So throwing exceptions for read timeouts is very harmful. Be careful not to thow exceptions while propagating them up the future chain. Add a test to enfore and detect regressions.
Fixes: scylladb/scylladb#25062
Improvement, normally not a backport candidate, but we may decide to backport if customer(s) are found to suffer from this.
Closesscylladb/scylladb#25068
* github.com:scylladb/scylladb:
reader_permit: remove check_abort()
test/boost/database_test: add test for read timeout exceptions
sstables/mx/reader: don't throw exceptions on the read-path
readers/multishard: don't throw exceptions on the read-path
replica/table: don't throw exceptions on the read-path
multishard_mutation_query: fix indentation
multishard_mutation_query: don't throw exceptions on the read-path
service/storage_proxy: don't throw exceptions on the full-scan path
cql3/query_processor: don't throw exceptions on the read-path
reader_permit: add get_abort_exception()
This method can cause performance regressions if used in the wrong place
-- namely if it is used to abort reads by throwing the abort exception.
Exceptions should be propagated during reads without throwing them,
otherwise they cause extra CPU load, making a bad situation worse.
Remove this method, so it doesn't accidentally get more users, migrate
remaining users to get_abort_exception().
Read timeouts shouldn't trigger exceptions thrown, exceptions should be
solely propagated via futures, otherwise they put extra strain on the
system at the worst possible time: when it is overload already enough
that reads started to time out.
The test covers both single partition reads and full scans, with two
scenarios:
* timeout while the read is queued
* timeout when the read is already ongoing
We add a test that validates that indexed queries
do not throw a warning related to vector search paging
Fixes: SCYLLADB-248
Closesscylladb/scylladb#28077
Create and drop view operations are currently performed on all shards, and their execution is not fully serialized. On slower processors this can lead to interleavings that leave stale entries in `system.scylla_views_build`
A problematic sequence looks like this:
* `on_create_view()` runs on shard 0 → entries for shard 0 and shard 1 are created
* `on_drop_view()` runs on shard 0 → entry for shard 0 is removed
* `on_create_view()` runs on shard 1 → entries for shard 0 and shard 1 are created again
* `on_drop_view()` runs on shard 1 → entry for shard 1 is removed, while the shard 0 entry remains
This results in a leftover row in `system.scylla_views_builds_in_progress`, causing `view_build_test.cc` to get stuck indefinitely in an eventual state and eventually be terminated by CI.
This patch fixes the issue by fully serializing all view create and drop operations through shard 0. Shard 0 becomes the single execution point and notifies other shards to perform their work in order. Requests originating.
new process:
- view_builder::on_create_view(...) runs only on shard 0 and kicks off dispatch_create_view(...) in the background.
- dispatch_create_view(...) (shard 0) first checks should_ignore_tablet_keyspace(...) and returns early if needed.
- dispatch_create_view(...) calls handle_seed_view_build_progress(...) on shard 0. That:
- writes the global “build progress” row across all shards via _sys_ks.register_view_for_building_for_all_shards(...).
- After seeding, dispatch_create_view(...) broadcasts to all shards with container().invoke_on_all(...).
- Each shard runs handle_create_view_local(...), which:
- waits for pending base writes/streams, flushes the base,
- resets the reader to the current token and adds the new view,
- handles errors and triggers _build_step to continue processing.
Drop view
- view_builder::on_drop_view(...) runs only on shard 0 and kicks off dispatch_drop_view(...) in the background.
- dispatch_drop_view(...) (shard 0) first checks should_ignore_tablet_keyspace(...) and returns early if needed.
- It broadcasts handle_drop_view_local(...) to all shards with invoke_on_all(...).
- Each shard runs handle_drop_view_local(...), which:
- removes the view from local build state (_base_to_build_step and _built_views) by scanning existing steps,
- ignores missing keyspace cases.
- After all shards finish local cleanup, shard 0 runs handle_drop_view_global_cleanup(...), which:
- removes global build progress, built‑view state, and view build status in system tables,
Shutdown
- drain() waits on _view_notification_sem before _sem so in‑flight dispatches finish before bookkeeping is halted.
In addition, the test is adjusted to remove the long eventual wait (596.52s / 30 iterations) and instead rely on the default wait of 17 iterations (~4.37 minutes), eliminating unnecessary delays while preserving correctness.
Fixes: https://github.com/scylladb/scylladb/issues/27898
Backport: not required as the problem happens on master
Closesscylladb/scylladb#27929
Fixes#27992
When doing a commit log oversized allocation, we lock out all other writers by grabbing
the _request_controller semaphore fully (max capacity).
We thereafter assert that the semaphore is in fact zero. However, due to how things work
with the bookkeep here, the semaphore can in fact become negative (some paths will not
actually wait for the semaphore, because this could deadlock).
Thus, if, after we grab the semaphore and execution actually returns to us (task schedule),
new_buffer via segment::allocate is called (due to a non-fully-full segment), we might
in fact grab the segment overhead from zero, resulting in a negative semaphore.
The same problem applies later when we try to sanity check the return of our permits.
Fix is trivial, just accept less-than-zero values, and take same possible ltz-value
into account in exit check (returning units)
Added whitebox (special callback interface for sync) unit test that provokes/creates
the race condition explicitly (and reliably).
Closesscylladb/scylladb#27998
VECTOR_SEARCH_INDEXING permission didn't work on cdc tables as we mistakenly checked for vector indexes on the cdc table insted of the base.
This patch fixes that and adds a test that validates this behavior.
Fixes: VECTOR-476
Closesscylladb/scylladb#28050
Currently, database::truncate_table_on_all_shards calls the table::can_flush only on the coordinator shard
and therefore it may miss shards with dirty data if the coordinator shard happens to have empty memtables, leading to clearing the memtables with dirty data rather than flushing them.
This change fixes that by making flush safe to be called, even if the memtable list is empty, and calling it on every shard that can flush (i.e. seal_immediate_fn is engaged).
Also, change database_test::do_with_some_data is use random keys instead of hard-coded key names, to reproduce this issue with `snapshot_list_contains_dropped_tables`.
Fixes#27639
* The issue exists since forever and might cause data loss due to wrongly clearing the memtable, so it needs backport to all live versions
Closesscylladb/scylladb#27643
* github.com:scylladb/scylladb:
test: database_test: do_with_some_data: randomize keys
database: truncate_table_on_all_shards: drop outdated TODO comment
database: truncate_table_on_all_shards: consider can_flush on all shards
memtable_list: unify can_flush and may_flush
test: database_test: add test_flush_empty_table_waits_on_outstanding_flush
replica: table, storage_group, compaction_group: add needs_flush
test: database_test: do_with_some_data_in_thread: accept void callback function
This patch adds tablet repair progress report support so that the user
could use the /task_manager/task_status API to query the progress.
In order to support this, a new system table is introduced to record the
user request related info, i.e, start of the request and end of the
request.
The progress is accurate when tablet split or merge happens in the
middle of the request, since the tokens of the tablet are recorded when
the request is started and when repair of each tablet is finished. The
original tablet repair is considered as finished when the finished
ranges cover the original tablet token ranges.
After this patch, the /task_manager/task_status API will report correct
progress_total and progress_completed.
Fixes#22564Fixes#26896Closesscylladb/scylladb#27679
With randomized keys, and since we're inserting only 2 keys,
it is possible that they would end up owned only by a single shard,
reproducing #27639 in snapshot_list_contains_dropped_tables.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Many test cases already assume `func` is being called a seastar
thread and although the function they pass returns a (ready) future,
it serves no purpose other than to conform to the interface.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Currently, tablet allocation intentionally ignores current load (
introduced by the commit #1e407ab) which could cause identical shard
selection when allocating a small number of tablets in the same topology.
When a tablet allocator is asked to allocate N tablets (where N is smaller
than the number of shards on a node), it selects the first N lowest shards.
If multiple such tables are created, each allocator run picks the same
shards, leading to tablet imbalance across shards.
This change initializes the load sketch with the current shard load,
scaled into the [0,1] range, ensuring allocation still remains even
while starting from globally least-loaded shards.
Fixes https://github.com/scylladb/scylladb/issues/27620Closesscylladb/scylladb#27802
The semaphore has detection and protection against regular resource
leaks, where some resources go unaccounted for and are not released by
the time the semaphore is destroyed. There is no detection or protection
against negative leaks: where resources are "made up" of thin air. This
kind of leaks looks benign at first sight, a few extra resources won't
hurt anyone so long as this is a small amount. But turns out that even a
single extra count resource can defeat a very important anti-deadlock
protection in can_admit_read(): the special case which admits a new
permit regardless of memory resources, when all original count resources
all available. This check uses ==, so if resource > original, the
protection is defeated indefinitely. Instead of just changing == to >=,
we add detection of such negative leaks to signal(), via
on_internal_error_noexcept().
At this time I still don't now how this negative leak happens (the code
doesn't confess), with this detection, hopefully we'll get a clue from
tests or the field. Note that on_internal_error_noexcept() will not
generate a coredump, unless ScyllaDB is explicitely configured to do so.
In production, it will just generate an error log with a backtrace.
The detection also clams the _resources to _initial_resources, to
prevent any damage from the negativae leak.
I just noticed that there is no unit test for the deadlock protection
described above, so one is added in this PR, even if only loosely
related to the rest of the patch.
Fixes: SCYLLADB-163
Closesscylladb/scylladb#27764
This reverts commit 1bb897c7ca, reversing
changes made to 954f2cbd2f. It makes
incompatible changes to the object storage configuration format, breaking
tests [1]. It's likely that it doesn't break any production configuration,
but we can't be sure.
Fixes#27966Closesscylladb/scylladb#27969
The boost test view_schema_test.cc::node_view_update_backlog can be
flaky if the test machine has a hiccup of 100ms, and this patch fixes
it:
The test is a unit test for db::view::node_update_backlog, which is
supposed to cache the backlog calculation for a given interval. The
test asks to cache the backlog for 100ms, and then without sleeping
at all tries to fetch a value again and expect the unchanged cached
value to be returned. However, if the test run experiences a context
switch of 100ms, it can fail, and it did once as reported in #27876.
The fix is to change the interval in this test from 100ms to something
much larger, like 10 seconds. We don't sleep this amount - we just need
the second fetch to happen *before* 10 seconds has passed, so there's
no harm in using a very large interval.
However, the second half of this test wants to check that after the
interval is over, we do get a new backlog calculation. So for the
second half of this test we can and should use a shorter backlog -
e.g., 10ms. We don't care if the test machine is slow or context switched,
for this half of the test we want to to sleep *more* than 10ms, and
that's easy.
The fixed test is faster than the old one (10ms instead of 100ms) and
more reliable on a shared test machine.
Fixes#27876.
Signed-off-by: Nadav Har'El <nyh@scylladb.com>
Closesscylladb/scylladb#27878
Currently, the tablet load balancer performs capacity based balancing by collecting the gross disk capacity of the nodes, and computes balance assuming that all tablet sizes are the same.
This change introduces size-based load balancing. The load balancer does not assume identical tablet sizes any more, and computes load based on actual tablet sizes.
The size-based load balancer computes the difference between the most and least loaded nodes in the balancing set (nodes in DC, or nodes in a rack in case of `rf-rack-valid-keyspaces`) and stops further balancing if this difference is bellow the config option `size_based_balance_threshold_percentage`.
This config option does not apply to the absolute load, but instead to the percentage of how much the most loaded node is more loaded than the least loaded node:
`delta = (most_loaded - least_loaded) / most_loaded`
If this delta is smaller then the config threshold, the balancer will consider the nodes balanced.
This PR is a part of a series of PRs which are based on top of each other.
- First part for tablet size collection via load_stats: #26035
- Second part reconcile load_stats: #26152
- The third part for load_sketch changes: #26153
- The fourth part which performs tablet load balancing based on tablet size: #26254
- The fifth part changes the load balancing simulator: #26438
This is a new feature, backport is not needed.
Fixes#26254Closesscylladb/scylladb#26254
* github.com:scylladb/scylladb:
test, load balancing: add test for table balance
load_balancer: add cluster feature for size based balancing
load_balancer: implement size-based load balancing
config: add size based load balancing config params
load_stats: use trinfo to decide how to reconcile tablet size
load_sketch: use tablet sizes in load computation
load_stats: add get_tablet_size_in_transition()
Add a `simple_value_with_expiry` utility class, which functions like
a `std::optional` with added timeout. When emplacing a value, user
needs to provide timeout, after which value expires (in which case
the `simple_value_with_expiry` object behaves as if was never set
at all).
Add boost tests for the new class.
This change adds a boost test which validates the resulting table
balance of size based load balancing. The threshold was set to a
conservative 1.5 overcommit to avoid flakyness.
This changes introduces tablet size based load balancing. It is an
extension of capacity based balancing with the addition of actual tablet
sizes.
It computes the difference between the most and least loaded nodes in
the DC and stops further balancing if this difference is bellow the
config option size_based_balance_threshold_percentage.
This config option does not apply to the absolute load, but instead to
the percentage of how much the most loaded node is more loaded than the
least loaded node:
delta = (most_loaded - least_loaded) / most_loaded
If this delta is smaller then the config threshold, the balancer will
consider the nodes balanced.
The test boost/error_injection_test.cc::test_inject_future_disabled
checks what happens when a sleep injection is *disabled*: The test
has a 10-millisecond-sleep injection and measures how much it takes.
The test expects it to take less than 10 milliseconds - in fact it
should take almost zero. But this is not guaranteed - on a slow debug
build and an overcommitted server this do-nothing injection can take
some time, and in one run (#27798) it took 14 milliseconds - and the
test failed.
The solution is easy - make the sleep-that-doesn't-happen much longer -
e.g., 10 whole seconds. Since this sleep still doesn't happen, we
expect the injection to return in less - much less - than 10 seconds.
This 10 seconds is so ridiculously high we don't expect the do-nothing
injection to take 10 seconds, not even a ridiculously busy test machine.
Fixes#27798
Signed-off-by: Nadav Har'El <nyh@scylladb.com>
Closesscylladb/scylladb#27874
This caused concurrent writers to operate on the same file, leading to file corruption. In some cases, this manifested as test failures and intermittent std::bad_alloc exceptions.
Change Description
This change ensures that each test instance uses a unique filename for downloaded bucket files.
By isolating file writes per test execution, concurrent runs no longer interfere with each other.
Fixes: #27824
backport not required
Closesscylladb/scylladb#27843
To configure S3 storage, one needs to do
```
object_storage_endpoints:
- name: s3.us-east-1.amazonaws.com
port: 443
https: true
aws_region: us-east-1
```
and for GCS it's
```
object_storage_endpoints:
- name: https://storage.googleapis.com:433
type: gs
credentials_file: <gcp account credentials json file>
```
This PR updates the S3 part to look like
```
object_storage_endpoints:
- name: https://s3.us-east-1.amazonaws.com:443
aws_region: us-east-1
```
fixes: #26570
Not-yet released feature, no need to backport. Old configs are not accepted any longer. If it's needed, then this decision needs to be revised.
Closesscylladb/scylladb#27360
* github.com:scylladb/scylladb:
object_storage: Temporarily handle pure endpoint addresses as endpoints
code: Remove dangling mentions of s3::endpoint_config
docs: Update docs according to new endpoints config option format
object_storage: Create s3 client with "extended" endpoint name
test: Add named constants for test_get_object_store_endpoints endpoint names
s3/storage: Tune config updating
sstable: Shuffle args for s3_client_wrapper
Due to the recent changes in the vector store service,
the service needs to read two of the system tables
to function correctly. This was not accounted for
when the new permission was added. This patch fixes that
by allowing these tables (group0_history and versions)
to be read with the VECTOR_SEARCH_INDEXING permission.
We also add a test that validates this behavior.
Fixes: SCYLLADB-73
Closesscylladb/scylladb#27546
Split prepare can run concurrently with repair.
Consider this:
1) split prepare starts
2) incremental repair starts
3) split prepare finishes
4) incremental repair produces unsplit sstable
5) split is not happening on sstable produced by repair
5.1) that sstable is not marked as repaired yet
5.2) might belong to repairing set (has compaction disabled)
6) split executes
7) repairing or repaired set has unsplit sstable
If split was acked to coordinator (meaning prepare phase finished),
repair must make sure that all sstables produced by it are split.
It's not happening today with incremental repair because it disables
split on sstables belonging to repairing group. And there's a window
where sstables produced by repair belong to that group.
To solve the problem, we want the invariant where all sealed sstables
will be split.
To achieve this, streaming consumers are patched to produce unsealed
sstable, and the new variant add_new_sstable_and_update_cache() will
take care of splitting the sstable while it's unsealed.
If no split is needed, the new sstable will be sealed and attached.
This solution was also needed to interact nicely with out of space
prevention too. If disk usage is critical, split must not happen on
restart, and the invariant aforementioned allows for it, since any
unsplit sstable left unsealed will be discarded on restart.
The streaming consumer will fail if disk usage is critical too.
The reason interposer consumer doesn't fully solve the problem is
because incremental repair can start before split, and the sstable
being produced when split decision was emitted must be split before
attached. So we need a solution which covers both scenarios.
Fixes#26041.
Fixes#27414.
Should be backported to 2025.4 that contains incremental repair
Closesscylladb/scylladb#26528
* github.com:scylladb/scylladb:
test: Add reproducer for split vs intra-node migration race
test: Verify split failure on behalf of repair during critical disk utilization
test: boost: Add failure_when_adding_new_sstable_test
test: Add reproducer for split vs incremental repair race condition
compaction: Fail split of new sstable if manager is disabled
replica: Don't split in do_add_sstable_and_update_cache()
streaming: Leave sstables unsealed until attached to the table
replica: Wire add_new_sstables_and_update_cache() into intra-node streaming
replica: Wire add_new_sstable_and_update_cache() into file streaming consumer
replica: Wire add_new_sstable_and_update_cache() into streaming consumer
replica: Document old add_sstable_and_update_cache() variants
replica: Introduce add_new_sstables_and_update_cache()
replica: Introduce add_new_sstable_and_update_cache()
replica: Account for sstables being added before ACKing split
replica: Remove repair read lock from maybe_split_new_sstable()
compaction: Preserve state of input sstable in maybe_split_new_sstable()
Rename maybe_split_sstable() to maybe_split_new_sstable()
sstables: Allow storage::snapshot() to leave destination sstable unsealed
sstables: Add option to leave sstable unsealed in the stream sink
test: Verify unsealed sstable can be compacted
sstables: Allow unsealed sstable to be loaded
sstables: Restore sstable_writer_config::leave_unsealed
When learning a schema that has a linked cdc schema, we need to learn
also the cdc schema, and at the end the schema should point to the
learned cdc schema.
This is needed because the linked cdc schema is used for generating cdc
mutations, and when we process the mutations later it is assumed in some
places that the mutation's schema has a schema registry entry.
We fix a scenario where we could end up with a schema that points to a
cdc schema that doesn't have a schema registry entry. This could happen
for example if the schema is loaded before it is learned, so when we
learn it we see that it already has an entry. In that case, we need to
set the cdc schema to the learned cdc schema as well, because it could
have been loaded previously with a cdc schema that was not learned.
Fixesscylladb/scylladb#27610Closesscylladb/scylladb#27704
If a keyspace has a numeric replication factor in a DC and rf < #racks,
then the replicas of tablets in this keyspace can be distributed among
all racks in the DC (different for each tablet). With rack list, we need all
tablet replicas to be placed on the same racks. Hence, the conversion
requires tablet co-location.
After this series, the conversion can be done using ALTER KEYSPACE
statement. The statement that does this conversion in any DC is not
allowed to change a rf in any DC. So, if we have dc1 and dc2 with 3 racks
each and a keyspace ks then with a single ALTER KEYSPACE we can do:
- {dc1 : 2} -> {dc1 : [r1, r2]};
- {dc1 : 2, dc2: 2} -> {dc1 : [r1, r2], dc2: [r2,r3]};
- {dc1 : 2, dc2: 2} -> {dc1 : [r1, r2], dc2: 2}
- {dc1 : 2} -> {dc1 : 2, dc2 : [r1]}
But we cannot do:
- {dc1 : 2} -> {dc1 : [r1, r2, r3]};
- {dc1 : 1, dc2 : [r1, r2] → dc1: [r1], dc2: [r1].
In order to do the co-locations rf change request is paused. Tablet
load balancer examines the paused rf change requests and schedules
necessary tablet migrations. During the process of co-location, no other
cross-rack migration is allowed.
Load balancer checks whether any paused rf change request is
ready to be resumed. If so, it puts the request back to global topology
request queue.
While an rf change request for a keyspace is running, any other rf change
of this keyspace will fail.
Fixes: #26398.
New feature, no backport
Closesscylladb/scylladb#27279
* github.com:scylladb/scylladb:
test: add est_rack_list_conversion_with_two_replicas_in_rack
test: test creating tablet_rack_list_colocation_plan
test: add test_numeric_rf_to_rack_list_conversion test
tasks: service: add global_topology_request_virtual_task
cql3: statements: allow altering from numeric rf to rack list
service: topology_coordinator: pause keyspace_rf_change request
service: implement make_rack_list_colocation_plan
service: add tablet_rack_list_colocation_plan
cql3: reject concurrent alter of the same keyspace
test: check paused rf change requests persistence
db: service: add paused_rf_change_requests to system.topology
service: pass topology and system_keyspace to load_balancer ctor
service: tablet_allocator: extract load updates
service: tablet_allocator: extract ensure_node
tasks, system_keyspace: Introduce get_topology_request_entry_opt()
node_ops: Drop get_pending_ids()
node_ops: Drop redundant get_status_helper()
This test was observed to fail multiple times recently in promotion,
because there were successful reads. The failure only reproduces on
arm64, it doesn't reproduce on x86.
The suspected reason is that the data set is too close to the edge,
where all reads fail due to too high memory consumption. Reduce the
number of sstables used by this test to 54 (from 64).
Fixes: #27248Closesscylladb/scylladb#27650
Allow altering from numeric replication factor to rack list. Ensure
that a single ALTER KEYSPACE statement doesn't try to both convert
to rack list and change rf.
Pass a pointer to service::topology and db::system_keyspace to load
balancer. It will be used in the following patches to create
rack_list_colocation plan.
We're extending the logic of DESCRIBE INDEX to include properties of the
underlying materialized view. Tests are provided to ensure the
implementation works as intended.
This reverts commit 866c96f536, reversing
changes made to 367633270a.
This change caused all longevities to fail, with a crash in parsing
scylla-metadata. The investigation is still ongoing, with no quick fix
in sight yet.
Fixes: #27496Closesscylladb/scylladb#27518
The batchlog table contains an entry for each logged batch that is processed by the local node as coordinator. These entries are typically very short lived, they are inserted when the batch is processed and deleted immediately after the batch is successfully applied.
When a table has `tombstone_gc = {'mode': 'repair'}` enabled, every repair has to flush all hints and batchlogs, so that we can be certain that there is no live data in any of these, older than the last repair. Since batches can contain member queries from any number of tables, the whole batchlog has to be flushed, even if repair-mode tombstone-gc is enabled for a single table.
Flushing the batchlog table happens by doing a batchlog replay. This involves reading the entire content of this table, and attempting to replay+delete any live entries (that are old enough to be replayed). Under normal operating circumstances, 99%+ of the content of the batchlog table is partition tombstones. Because of this, scanning the content of this table has to process thousands to millions of tombstones. This was observed to require up to 20 minutes to finish, causing repairs to slow down to a crawl, as the batchlog-flush has to be repeated at the end of the repair of each token-range.
When trying to address this problem, the first idea was that we should expedite the garbage-collection of these accumulated tombstones. This experiment failed, see https://github.com/scylladb/scylladb/pull/23752. The commitlog proved to be an impossible to bypass barrier, preventing quick garbage-collection of tombstones. So long as a single commit-log segment is alive, holding content from the batchlog table, all tombstones written after are blocked from GC.
The second approach, represented by this PR, is to not rely in tombstone GC to reduce the tombstone amount. Instead restructure the table such that a single higher-order tombstone can be used to shadow and allow for the eviction of the myriads of individual batchlog entry tombstones. This is realized by reorganizing the batchlog table such that individual batches are rows, not partitions.
This new schema is introduced by the new `system.batchlog_v2` table, introduced by this PR:
CREATE TABLE system.batchlog_v2 (
version int,
stage int,
shard int,
written_at timestamp,
id uuid,
data blob,
PRIMARY KEY ((version, stage, shard), written_at, id));
The new schema organization has the following goals:
1) Make post-replay batchlog cleanup possible with a simple range-tombstone. This allows dropping the individual dead batchlog entries, as they are shadowed by a higher level tombstone. This enables dropping tombstones without tombstone GC.
2) To make the above possible, introduce the stage key component: batchlog entries that fail the first replay attempt, are moved to the failed_replay stage, so the initial stage can be cleaned up safely.
3) Spread out the data among Scylla shards, via the batchlog shard column.
4) Make batchlog entries ordered by the batchlog create time (id). This allows for selecting batchlogs to replay, without post-filtering of batchlogs that are too young to be replayed.
Fixes: https://github.com/scylladb/scylladb/issues/23358
This is an improvement, normally not a backport-candidate. We might override this and backport to allow wider use of `tombstone_gc: {'mode': 'repair'}`.
Closesscylladb/scylladb#26671
* github.com:scylladb/scylladb:
db/config: change batchlog_replay_cleanup_after_replays default to 1
test/boost/batchlog_manager_test: add test for batchlog cleanup
replica/mutation_dump: always set position weight for clustering positions
service/storage_proxy: s/batch_replay_throw/storage_proxy_fail_replay_batch/
test/lib: introduce error_injection.hh
utils/error_injection: add debug log to disable() and disable_all()
test/lib/cql_test_env: forward config to batchlog
test/lib/cql_test_env: add batch type to execute_batch()
test/lib/cql_assertions: add with_size(predicate) overload
test/lib/cql_assertions: add source location to fail messages
test/lib/cql_assertions: columns_assertions: add assert_for_columns_of_each_row()
test/lib/cql_assertions: rows_assertions::assert_for_columns_of_row(): add index bound check
test/lib/cql_assertions: columns_assertions: add T* with_typed_column() overload
db/batchlog_manager: config: s/write_timeout/reply_timeot/
db,service: switch to system.batchlog_v2
db/system_keyspace: introduce system.batchlog_v2
service,db: extract generation of batchlog delete mutation
service,db: extract get_batchlog_mutation_for() from storage-proxy
db/batchlog_manager: only consider propagation delay with tombstone-gc=repair
db/batchlog_manager: don't drop entire batch if one mutations' table was dropped
data_dictionary: table: add get_truncation_time()
db/batchlog_manager: batch(): replace map_reduce() with simple loop
db/batchlog_manager: finish coroutinizing replay_all_failed_batches
db/batchlog_manager: improve replayAllFailedBatches logs
This patch series contains the following changes:
- Incorporation of `crypt_sha512.c` from musl to out codebase
- Conversion of `crypt_sha512.c` to C++ and coroutinization
- Coroutinization of `auth::passwords::check`
- Enabling use of `__crypt_sha512` orignated from `crypt_sha512.c` for
computing SHA 512 passwords of length <=255
- Addition of yielding in the aforementioned hashing implementation.
The alien thread was a solution for reactor stalls caused by indivisible
password‑hashing tasks (https://github.com/scylladb/scylladb/issues/24524).
However, because there is only one alien thread, overall hashing throughput was reduced
(see, e.g., https://github.com/scylladb/scylla-enterprise/issues/5711). To address this,
the alien‑thread solution is reverted, and a hashing implementation
with yielding is introduced in this patch series.
Before this patch series, ScyllaDB used SHA-512 hashing provided
by the `crypt_r` function, which in our case meant using the implementation
from the `libxcrypt` library. Adding yielding to this `libxcrypt`
implementation is problematic, both due to licensing (LGPL) and because the
implementation is split into many functions across multiple files. In
contrast, the SHA-512 implementation from `musl libc` has a more
permissive license and is concise, which makes it easier to incorporate
into the ScyllaDB codebase.
The performance of this solution was compared with the previous
implementation that used one alien thread and the implementation
after the alien thread was reverted. The results (median) of
`perf-cql-raw` with `--connection-per-request 1 --smp 10` parameters
are as follows:
- Alien thread: 41.5 new connections/s per shard
- Reverted alien thread: 244.1 new connections/s per shard
- This commit (yielding in hashing): 198.4 new connections/s per shard
The roughly 20% performance deterioration compared to
the old implementation without the alien thread comes from the fact
that the new hashing algorithm implemented in `utils/crypt_sha512.cc`
performs an expensive self-verification and stack cleanup.
On the other hand, with smp=10 the current implementation achieves
roughly 5x higher throughput than the alien thread. In addition,
due to yielding added in this commit, the algorithm is expected
to provide similar protection from stalls as the alien thread did.
In a test that in parallel started a cassandra-stress workload and
created thousands of new connections using python-driver, the values of
`scylla_reactor_stalls_count` metric were as follows:
- Alien thread: 109 stalls/shard total
- Reverted alien thread: 13186 stalls/shard total
- This commit (yielding in hashing): 149 stalls/shard total
Similarly, the `scylla_scheduler_time_spent_on_task_quota_violations_ms`
values were:
- Alien thread: 1087 ms/shard total
- Reverted alien thread: 72839 ms/shard total
- This commit (yielding in hashing): 1623 ms/shard total
To summarize, yielding during hashing computations achieves similar
throughput to the old solution without the alien thread but also
prevents stalls similarly to the alien thread.
Fixes: scylladb/scylladb#26859
Refs: scylladb/scylla-enterprise#5711
No automatic backport. After this PR is completed, the alien thread should be rather reverted from older branches (2025.2-2025.4 because on 2025.1 it's already removed). Backporting of the other commits needs further discussion.
Closesscylladb/scylladb#26860
* github.com:scylladb/scylladb:
test/boost: add too_long_password to auth_passwords_test
test/boost: add same_hashes_as_crypt_r to auth_passwords_test
auth: utils: add yielding to crypt_sha512
auth: change return type of passwords::check to future
auth: remove code duplication in verify_scheme
test/boost: coroutinize auth_passwords_test
utils: coroutinize crypt_sha512
utils: make crypt_sha512.cc to compile
utils: license: import crypt_sha512.c from musl to the project
Revert "auth: move passwords::check call to alien thread"
This is crucial with MVs, since the splitting must preserve the state of
the original sstable. We want the sstable to be in staging dir, so it's
excluded when calculating the diff for performing pushes to view
replicas.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Since the function must only be used on new sstables, it should
be renamed to something describing its usage should be restricted.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
This is crucial for splitting before sealing the sstable produced by
repair. This way, unsplit sstables won't be left on disk sealed.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
This reverts commit 8192f45e84.
The merge exposed a bug where truncate (via drop) fails and causes Raft
errors, leading to schema inconsistencies across nodes. This results in
test_table_drop_with_auto_snapshot failures with 'Keyspace test does not exist'
errors.
The specific problematic change was in commit 19b6207f which modified
truncate_table_on_all_shards to set use_sstable_identifier = true. This
causes exceptions during truncate that are not properly handled, leading
to Raft applier fiber stopping and nodes losing schema synchronization.
This reverts commit faad0167d7. It causes
a regression in
test_two_tablets_concurrent_repair_and_migration_repair_writer_level
in debug mode (with ~5%-10% probability).
Fixes#27510.
Closesscylladb/scylladb#27560
The test documents the current behavior of hashing algorithms that
fail if the passphrase has 512 bytes or more.
Moreover, it documents the behavior of the current bcrypt
implementation that compares only the first 72 bytes of the password.
Although we don't typically use bcrypt for password hashing, it is
possible to insert such a hash using
`CREATE ROLE ... WITH HASHED PASSWORD ...`.
Refs: scylladb/scylladb#26842
