A stale Aside sync timer and socket outlive their test file in a shared Bun isolate
ci agents · case
Symptom
Section titled “Symptom”A multi-file bun test --isolate batch on windows-latest intermittently
stalls or crashes after tests/server/local-aside-sync-capability.test.ts
starts its third listener. One attempt reached the 480-second process timeout
(job 107037821321); the next ended with:
panic(main thread): Segmentation fault at address 0x18The panic occurred in job 107043425787. Every file passed alone, and the multi-file batch also passed on another run. Treat the segmentation fault as a Bun runtime bug triggered by leaked handles; no native stack was captured to establish a deeper crash cause.
Run the multi-file isolated test batch containing
tests/server/local-aside-sync-capability.test.ts on Bun 1.4.0 and
windows-latest. The observed failure was intermittent: the timeout and panic
occurred on consecutive attempts, while each file passed alone and another
batch run passed. Green run 35828289232 passed both Aside tests in a multi-file
isolate process, but its shards had been rebalanced, so it was not an exact
replay of the failing batch.
Two ownership gaps let resources outlive the test file inside the shared Bun process:
- An
AbortSignal.timeout()deadline for an Aside sync stayed armed after the sync succeeded. - The direct-local HTTP request settled before its destroyed socket emitted
close. The socket and a timer could therefore remain active after the test file had finished.
The timeout and panic are two observed batch outcomes around that leaked resource lifetime. The panic is attributed to Bun being triggered by leaked handles, not to an inferred native stack or a proved internal crash path.
PR #5634 clears the deadline on every exit, keeps an absolute timer that bounds
request and socket teardown, settles exactly once if close never arrives or
destroy() throws, and awaits server.stop(true) during teardown.
Workaround
Section titled “Workaround”Make the request own its deadline and teardown. Clear the deadline on every
exit; use one absolute bound for both the request and socket shutdown; make
settlement idempotent across response, error, timeout, and close events; and
await server.stop(true) before the test file completes.
Unsafe workarounds:
- Raising the 480-second batch timeout lets leaked resources remain alive even longer and hides the ownership defect.
- Running the test file alone avoids the shared-process boundary that exposed the leak; it does not close the resources on success.
- Reordering files only changes when the leak is exposed.