The flaky test 03518_alter_logical_race stresses concurrent ALTER/INSERT against a ReplicatedMergeTree table. Its MODIFY COLUMN ... UInt64 step is a data-rewriting mutation. On a busy CI host (it fails on Stateless tests (arm_binary, parallel), ~7 times/day on master) that mutation gets starved by the two MergeTree throttles that defer mutations in favour of merges: max_replicated_mutations_in_queue (default 8) and number_of_free_entries_in_pool_to_execute_mutation (default 20). Under the test's heavy concurrent INSERT load there are always merges to run, so the data part of the very first MODIFY never gets a turn.

Replicated metadata alters must finish in strict version order, and an alter does not leave the alter-version chain until its data mutation completes. So a starved data mutation at the head of the chain blocks every later MODIFY's metadata change, which then waits (with the default alter_sync = 1) until the table is dropped at teardown. The timeout wrapping each statement fires first and fails the test with a spurious TIMEOUT / NO ALTER PROGRESS, even though nothing is hung — the mutation just never gets scheduled.

The server log of the failing run shows this exactly: mutation 0000000000 (alter version 3) is created but never executes (won't select new parts to merge or mutate because queued mutations reached max_replicated_mutations_in_queue (8)), and the next MODIFY's metadata entry is rejected 309044 times with Cannot execute alter metadata ... because another alter 3 must be executed before.

The fix raises max_replicated_mutations_in_queue and sets number_of_free_entries_in_pool_to_execute_mutation = 0 on the test table so the data mutation is no longer deferred. The two ALTER workers issue their MODIFYs synchronously, so in-flight mutations stay naturally low; this only removes the head-of-chain stall, it does not let mutations pile up. Verified by running the test repeatedly under saturating CPU load: it now completes real ALTER cycles in ~12–15s instead of hanging to the 180s hard cap.

CI report: https://s3.amazonaws.com/clickhouse-test-reports/json.html?REF=master&sha=2a3a502d2ed0af65bb0a5c91223c91c1b2e28047&name_0=MasterCI&name_1=Stateless%20tests%20%28arm_binary%2C%20parallel%29

Changelog category (leave one):

• CI Fix or Improvement (changelog entry is not required)

Changelog entry (a user-readable short description of the changes that goes into CHANGELOG.md):

Not for changelog — fixes flakiness of the stateless test 03518_alter_logical_race.

Documentation entry for user-facing changes

• Documentation is written (mandatory for new features)

🤖 Generated with Claude Code