completed_groups_release_first_value_bitmap takes 10 minutes under force_hash_collisions #27033

Is your feature request related to a problem or challenge?

aggregates::clustered_partial_stream::tests::completed_groups_release_first_value_bitmap (added in #26782) takes about 10 minutes in the cargo test hash collisions (amd64) merge-queue job. Without force_hash_collisions, it takes about 1 second.

The test aggregates 131,072 distinct Int64 keys in one batch. Under force_hash_collisions, GroupValuesPrimitive gives every key the hash 0, so each insert compares against all earlier keys. The test also runs the plan two times (without a limit and with a 16 KB limit).

The test checks that FIRST_VALUE(v ORDER BY v) releases its per-batch bitmap. Colliding hashes do not change what it checks.

Evidence from merge-queue runs:

• In the passing rerun for docs: Clarify Known Users criteria and clean up active project list #26926 (job), the physical-plan unit test binary took 609.89s. All other tests in the binary finished by 03:40:31. This test ran from about 03:40:07 to 03:50:17.
• The median time of the job went from 6.2 minutes (230 jobs from Oct 21 until fix: release per-batch accumulator scratch after a clustered aggregate emit under memory pressure #26782 merged on Oct 30) to 12.0 minutes (70 jobs after).
• All 5 failed attempts of the job since fix: release per-batch accumulator scratch after a clustered aggregate emit under memory pressure #26782 merged were spot runner shutdowns ("The runner has received a shutdown signal") while this test was running. Each one failed the merge-queue run for that PR, and a rerun passed:
  • minor: speedup one SPM unit test #26929: job
  • test: cover grouping set equivalences and constants #26898: job
  • fix: unparse Int8 casts as SIGNED for MySQL dialect #26883: job
  • fix: split BoundedWindowAggExec output into batch_size batches #26974: job
  • docs: Clarify Known Users criteria and clean up active project list #26926: job

Locally (--profile ci, single test), it takes 0.92s by default and 309s with force_hash_collisions.

Describe the solution you'd like

Skip the test under force_hash_collisions, as #26594 did for sort_after_group_by_reserves_shared_output_once. The regular cargo test job still runs it.

Describe alternatives you've considered

Use fewer groups with a lower memory limit. This keeps the test in the hash-collision job, but the bitmap must still be larger than what the active group needs, so it is a call for the test's author.

Additional context

Related to #26759.