Groene AI 5b2a7dfbaf Fix residual flakiness in test_overcommit_tracker/test_user_overcommit ...

The test fires concurrent queries A (low overcommit denominator, should be
killed) and B (high denominator, should survive) under a shared user memory
limit.  After PR #101574 tuned the parameters, a ~1% residual failure
remained where ALL B queries were also killed (2 master failures in 12 days,
~6 PR failures/week).

Root cause: both A and B used numbers(5000000) (~40 MB each).  With 20 B
queries, their collective memory (800 MB) far exceeded the 300 MB user
limit.  After the overcommit tracker killed all A queries, B's total still
exceeded the limit, causing B queries to also be killed.

Fix:
- Reduce B query data from numbers(5000000) to numbers(1000000) (~8 MB each)
- Reduce QUERY_COUNT from 40 to 20

After the fix, B's collective memory (10 × 8 MB = 80 MB) is well within
the 300 MB limit once A queries are killed, so at least one B query can
finish.  Total memory demand (10 × 40 + 10 × 8 = 480 MB) still comfortably
exceeds the limit to trigger the overcommit tracker.

The test intent is preserved: it still verifies that queries with low
memory_overcommit_ratio_denominator are killed first while queries with
high denominator survive under memory pressure.

Validated with 50 consecutive passes locally.

2026-04-14 09:29:43 +00:00

..

configs/users.d

tests: attempt to fix test_overcommit_tracker flakiness under amd_llvm_coverage

2026-03-31 16:42:45 +02:00

__init__.py

…

test.py

Fix residual flakiness in test_overcommit_tracker/test_user_overcommit

2026-04-14 09:29:43 +00:00