Flaky: GalaxyDriverFixRound2Tests.A_fetch_completing_after_init_populates_host_statuses (10 s timeout under load) #856

New Issue

2026-11-11T07:55:31-05:00

dohertj2 commented

2026-11-11 07:55:31 -05:00

Failed once in the full-suite run on 03e51457 (10 s, during a serial run after heavier suites); passed 5/5 alone and 3/3 full Galaxy.Tests reruns. No Galaxy code changed in that batch. Capture with trx under load and apply the presence-vs-absence rule / self-diagnosing timeouts as in #843.

Failed once in the full-suite run on `03e51457` (10 s, during a serial run after heavier suites); passed 5/5 alone and 3/3 full Galaxy.Tests reruns. No Galaxy code changed in that batch. Capture with trx under load and apply the presence-vs-absence rule / self-diagnosing timeouts as in #843.

dohertj2 added the type:test priority:p3 labels 2026-11-11 07:55:31 -05:00

dohertj2 referenced this issue

2026-11-11 10:01:34 -05:00

Galaxy.Tests: remaining short fixed waits fail under heavy parallel load #858

dohertj2 referenced this issue

2026-11-11 10:01:34 -05:00

Galaxy: can an MXAccess initial value arrive on the event stream before the driver registers the new item handle? #859

dohertj2 commented

2026-11-11 11:28:48 -05:00

Fixed (988046cb, 90de3891). Pushed to master in e857135e. Root cause: a real visibility race — PerPlatformProbeWatcher.WatchedPlatforms read the handle map without the lock, so a test could see a platform before its probe binding was in the registry and its ScanState event was dropped as stale (11/700 runs under parallel load). Production dispatch was already ordered safely; the property now reads under the lock (PerPlatformProbeWatcherVisibilityTests, 0/1400 after). A second flake found under the same load (GalaxyDriverShutdownTests, fake's cancellation callback could be unregistered before running) fixed by signalling from the iterator's finally. Follow-ups: #858, #859.

Fixed (`988046cb`, `90de3891`). Pushed to master in `e857135e`. Root cause: a real visibility race — `PerPlatformProbeWatcher.WatchedPlatforms` read the handle map without the lock, so a test could see a platform before its probe binding was in the registry and its ScanState event was dropped as stale (11/700 runs under parallel load). Production dispatch was already ordered safely; the property now reads under the lock (`PerPlatformProbeWatcherVisibilityTests`, 0/1400 after). A second flake found under the same load (`GalaxyDriverShutdownTests`, fake's cancellation callback could be unregistered before running) fixed by signalling from the iterator's `finally`. Follow-ups: #858, #859.

dohertj2 closed this issue

2026-11-11 11:28:49 -05:00

dohertj2 referenced this issue from a commit

2026-11-11 11:28:50 -05:00

fix(galaxy): WatchedPlatforms cannot show a platform before its probe binding is routable (#856)

dohertj2 referenced this issue from a commit

2026-11-11 11:28:50 -05:00

test(galaxy): shutdown stream-cancel signal from the iterator's exit, not a token registration (#856)

dohertj2 referenced this issue from a commit

2026-11-11 11:28:50 -05:00

Merge #856 Galaxy probe-watcher visibility race (test-only reachable) + shutdown test flake

dohertj2 referenced this issue from a commit