Groene AIandClaude Opus 5.5 156e637aa0 Disable jemalloc's automatic decay for the corrected node of test_MemoryTracking ...

`test_correction_follows_measurement` needs its baseline sample to hold no
retained pages, and takes it right after `SYSTEM JEMALLOC PURGE`. That purge is
not guaranteed to reach every arena: `pac_decay_to_limit` returns without
purging an arena whose `decay->purging` flag is set
(`contrib/jemalloc/src/pac.c:581`), and jemalloc's background thread holds that
flag while it decays the pages the previous test's ten queries freed 0.6 s
earlier. On master `a0bb275ec76f` the first purge left about 334 MB that the
second purge removed (MemoryTracking before/after/purged 498/430/164 MB, the
plain counter constant at 113 MB), the query reused those pages, and
`gap_after - gap_before > 64 MiB` failed.

With `dirty_decay_ms:-1` nothing purges dirty pages automatically
(`pac_maybe_decay_purge` returns before `pac_decay_to_limit` on a negative
decay), so on this node they leave only through forced full purges: both
purges of the test take full effect, and the "after" sample keeps every freed
page instead of depending on the 5 s decay. Same approach as
`test_dirty_pages_force_purge`. `instance_env_variables=True` keeps the variable
out of the cluster-wide env file, so `node` still runs with the default decay.

Locally, with CI's 16 per-CPU arenas, 4 of 150 first purges left 108-652 MB of
dirty pages with the default decay and 0 of 150 with `dirty_decay_ms:-1`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

2026-10-02 01:24:57 +00:00

..

configs

…

__init__.py

…

test.py

Disable jemalloc's automatic decay for the corrected node of test_MemoryTracking

2026-10-02 01:24:57 +00:00