Fix ListQueues: iterator leak, unbounded CQL round-trips, and PageState after Close

[mykaul]

What changed?

• Set explicit CQL PageSize for Cassandra/Scylla queue reads.
• Fix Cassandra/Scylla iterator lifecycle: close iterators on processing errors and preserve close errors.
• Fix GetTaskQueuesByBuildId to recreate iterators with PageState and consume all pages. There is no artificial 10,000-result cap.
• Fix ListQueues paging for under-filled ALLOW FILTERING pages, while bounding each request to 10 CQL round-trips and returning NextPageToken when more results remain.
• Capture PageState before closing iterators.
• Pre-allocate the ListQueues result slice.
• Add regression coverage for queue-name query errors and task-queue mappings spanning multiple pages.

Why?

Cassandra/Scylla can return under-filled pages for ALLOW FILTERING queries, so a single CQL page does not necessarily satisfy the requested page size. Manual page traversal must also preserve the page token before iterator close. Iterator errors must not leak resources or be discarded.

Validation

• ScyllaDB: TestCassandraTaskQueueUserDataSuite
• ScyllaDB: TestCassandraQueueV2Persistence
• make lint-code

The 10-page ListQueues bound is intentional because that API is paginated; GetTaskQueuesByBuildId remains unbounded and returns all mappings as required by its API contract.

[CLAassistant]

All committers have signed the CLA.

[prathyushpv]

Code review

Found 1 issue:

1. ListQueues inner scan loop can return more rows than PageSize. The CQL page size is set to request.PageSize on every iteration (line 492), but after the first page partially fills queues, subsequent pages can still return up to request.PageSize rows. The inner loop (lines 494-524) appends all of them without checking len(queues) >= request.PageSize. For example: if PageSize=100 and the first page returns 90 rows, the second page fetches another 100, producing 190 total. Fix: either use request.PageSize - len(queues) as the CQL page size, or add a len(queues) >= request.PageSize break inside the inner loop.

https://github.com/temporalio/temporal/blob/b6fdcdbe31f900c54eb75f8a13ed9d641630ec4b/common/persistence/cassandra/queue_v2_store.go#L487-L524

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[mykaul]

Code review

Found 1 issue:

1. ListQueues inner scan loop can return more rows than PageSize. The CQL page size is set to request.PageSize on every iteration (line 492), but after the first page partially fills queues, subsequent pages can still return up to request.PageSize rows. The inner loop (lines 494-524) appends all of them without checking len(queues) >= request.PageSize. For example: if PageSize=100 and the first page returns 90 rows, the second page fetches another 100, producing 190 total. Fix: either use request.PageSize - len(queues) as the CQL page size, or add a len(queues) >= request.PageSize break inside the inner loop.

https://github.com/temporalio/temporal/blob/b6fdcdbe31f900c54eb75f8a13ed9d641630ec4b/common/persistence/cassandra/queue_v2_store.go#L487-L524

🤖 Generated with Claude Code

• If this code review was useful, please react with 👍. Otherwise, react with 👎.

Fixed with the first option: PageSize is now set to request.PageSize - len(queues) .
Also rebased on top of origin/main.
Ran go test -tags test_dep -run TestCassandraQueueV2 on the change.

[mykaul]

v2 — Review feedback addressed + additional fixes

Review comments addressed:

1. Loop condition — changed to for len(taskQueues) < maxTaskQueuesByBuildIdResults
2. Break instead of early return — inner loop now uses break when hitting cap
3. Don't ignore iter.Close() error — uses errors.Join(err, iter.Close()) on all early-return paths

Additional fixes found during implementation:

4. Recreate iterator per page with PageState(pageToken) — gocql does not auto-paginate when PageState is used manually; the old code silently returned only the first page
5. Capture PageState() before Close() — Close() invalidates page state in gocql
6. Same iter.Close() error handling fix applied to ListQueues — had the same _ = iter.Close() pattern
7. Simplified pageToken logic in ListQueues — removed redundant if/else
8. SQL backends: added ORDER BY task_queue_name LIMIT 10000 — aligns MySQL, PostgreSQL, and SQLite to the same deterministic truncation behavior as Cassandra/ScyllaDB
9. Comment + CONSIDER item documenting the shared 10k cap across all backends

Rebased on origin/main, squashed into a single commit.

[mykaul]

8. SQL backends: added ORDER BY task_queue_name LIMIT 10000 — aligns MySQL, PostgreSQL, and SQLite to the same deterministic truncation behavior as Cassandra/ScyllaDB

Reviewers - take a note of the above please.

[mykaul]

ping - is there anything I need to do to move this fix forward?

[rodrigozhou]

The PR lgtm.
Just wondering if there's any place that assumes the function would return all task queues. Since your changes you limit to 10k, it could potentially break something.
cc: @dnr

[rodrigozhou]

Is this different from len(pageToken) == 0?

[mykaul]

Indeed, fixed.

[mykaul]

The PR lgtm. Just wondering if there's any place that assumes the function would return all task queues. Since your changes you limit to 10k, it could potentially break something. cc: @dnr

OK, reverted this 10K, now just using paging.

[mykaul]

Updated PR description!