Deflake SlidingConnectionHealTest unknown_pos leg on testAndroidHostTest under parallel load #73

New issue

Closed

opened 2026-08-10 14:32:39 +02:00 by thecrealm · 0 comments

thecrealm commented

2026-08-10 14:32:39 +02:00

Owner

During the 2026-08-10 local gate for #68, unknown_pos_restarts_the_connection_and_the_initial_payload_marks_the_room timed out (30s virtual time, TimeoutCancellationException at SlidingConnectionHealTest.kt:102) on :katrix-testkit:testAndroidHostTest while the jvmTest variant passed in the same build; it passes deterministically in isolation (--rerun-tasks). Load-dependent — same family as the D22 transient-5xx flake seen in CI run #122. Investigate the virtual-time polling loop (while cursor != fresh: delay(50)) vs MockHomeserver idleSyncDelay under a contended scheduler; consider awaiting a flow instead of polling, per the D36 crash-honest deflake pattern.

During the 2026-08-10 local gate for #68, `unknown_pos_restarts_the_connection_and_the_initial_payload_marks_the_room` timed out (30s virtual time, TimeoutCancellationException at SlidingConnectionHealTest.kt:102) on :katrix-testkit:testAndroidHostTest while the jvmTest variant passed in the same build; it passes deterministically in isolation (--rerun-tasks). Load-dependent — same family as the D22 transient-5xx flake seen in CI run #122. Investigate the virtual-time polling loop (`while cursor != fresh: delay(50)`) vs MockHomeserver idleSyncDelay under a contended scheduler; consider awaiting a flow instead of polling, per the D36 crash-honest deflake pattern.

thecrealm added the

label

2026-08-10 14:32:49 +02:00

thecrealm referenced this issue

2026-08-10 14:42:47 +02:00

Element verification e2e: try dropping the 15s device-list warmup in favor of the D37 re-request #72

thecrealm added

now

and removed

labels

2026-08-11 12:29:10 +02:00

thecrealm referenced this issue from a commit