Flaky live test: GetConfigAsync_Should_Read_Back_The_Shell_UpdateConfig_Wrote never sees config.updated (Windows net48) #440

Activity

1. Blind-Striker

   changed the title [-]Flaky live test: GetConfigAsync_Should_Read_Back_The_Shell_UpdateConfig_Wrote never sees config.updated (Windows net11.0)[/-] [+]Flaky live test: GetConfigAsync_Should_Read_Back_The_Shell_UpdateConfig_Wrote never sees config.updated (Windows net48)[/+] on Oct 17, 2026
2. Blind-Striker commented on Oct 17, 2026

   

   Blind-Striker

   on Oct 17, 2026

   CollaboratorAuthor

   More actions

   Root cause: cross-test config residue on the shared server (test isolation, not product)

   Measured from the TRX of merge-group run 39026886983, plus three green master runs (39022428987, 39022831816, 39025318239):

   • On net48, UpdateConfigAsync_Should_Persist_And_Clear_The_Shell always runs before GetConfigAsync_Should_Read_Back_The_Shell_UpdateConfig_Wrote. Both share one SimulatedDriveServerFixture, which ParallelConstraintKeys.ServerProcess serializes.
   • In the red run, Persist took 4.0 s (11:04:19.74–23.73) and Read_Back started 130 ms after it ended. In green runs Persist finishes in under 0.5 s.
   • On net10.0 and net11.0, Read_Back runs first, and those legs pass. Neither the SSE path nor net48 is at fault: net48 is just the only leg where Read_Back follows Persist, and this time the runner was slow.

   Mechanism, read in upstream packages/core/src/config.ts at the pin (lines 316-328 and 380-401):

   • experimental.config.update writes the global file and only requests a reload. The request goes through a sliding PubSub of capacity 1 and a 100 ms debounce, and reloadLock serializes the reloads.
   • A reload publishes config.updated only when the loaded entries change (isDeepStrictEqual).
   • Persist writes the shell, then null, and checks only the file, never the loaded state. So the test ends with reloads still in flight.
   • When its set-reload is slow, the loaded state becomes the shell value and config.updated is published before Read_Back subscribes. Its skill.updated reactions arrive later: these are the four events Read_Back's probe saw.
   • The reload left pending by Persist's clear then reads the file only after Read_Back has written the same shell. It sees no change and publishes nothing, and neither does Read_Back's own reload. The result is a 60 s timeout.

   Upstream behaves as designed. The defect is that one test leaves the shared server's configuration unsettled, while the other assumes a settled baseline.

   No other test writes this server's global configuration: ConfigClientLiveTests holds the only UpdateConfigAsync calls, and no test writes into the run root's config directory at test time.

   Fix plan:

   1. A deterministic red repro first: write the shell, wait for config.updated, then write null and the shell back to back within 100 ms. Expected: no config.updated.
   2. Then fold Persist's file assertions into Read_Back, so one test owns the whole write, settle, clear, settle cycle.

   No retry and no wider bound.
3. Blind-Striker

   closed this as completedin #443on Oct 17, 2026