Reduce the cold project-worker cross-post delivery tail

The production-shaped e2e case Project worker processEventBatch receives events from every project stream and can cross-post commits a unique worker, appends to a fresh child stream, and proves the project-wide worker feed can cross-post that event. Two consecutive preview workflows exhausted its 30-second public stream-wait deadline before passing on the repository's one retry:

  • PR #2251's final preview: 59.18 seconds across both attempts.
  • PR #2253 Depot run 3jp43c0dbg: 62.79 seconds across both attempts.

The durable cold-build handoff preserves the event, but the combined commit, build, feed-delivery, and worker-start tail is too close to the former budget. The e2e wait is temporarily 100 seconds so this substantial coverage remains active without converting healthy delayed delivery into a retry.

Work#

  • Record separate timings for repo commit, coordinator handoff, bundle cache lookup or build, feed redelivery, dynamic Worker startup, cross-post append, and feed acknowledgement.
  • Correlate cold and warm samples with Worker traces and the existing build and stream telemetry under the normal fully parallel preview load.
  • Remove the dominant cold-path or contention latency without weakening the fresh-worker and fresh-child-stream coverage.

Exit criteria#

  • The case passes on its first attempt in at least 25 consecutive canonical preview runs.
  • Its cold-path p99 delivery completes within 30 seconds with no unexplained errors, stalled feeds, duplicate processing, or leaked builds.
  • Restore the public stream-wait deadline from 100 seconds to 30 seconds.

Was this page helpful?