I’m investigating an unexpected Cloud Run revision startup timeout despite successful TCP startup probes on two separate instances.
Environment:
- Cloud Run (fully managed), asia-northeast1
- Python 3.12, Uvicorn, FastMCP
- Container port 8080
- TCP startup probe
- 1 vCPU, 512 MiB
Observed events on October 23, 2026 (UTC):
Instance A:
- 03:06:47.940726 — Uvicorn started
- 03:06:47.941257 — TCP startup probe succeeded
- 03:06:48.045006 — Uvicorn began graceful shutdown
Instance B:
- 03:06:47.939349 — Uvicorn started
- 03:06:47.939814 — TCP startup probe succeeded
- 03:10:50.228624 — Uvicorn began graceful shutdown
Revision-level events:
- 03:10:47.237562 — Startup probe timeout
- 03:10:48.390356 — Revision marked Retired
These events belong to two distinct instances.
The previous revision was healthy with the same Cloud Run configuration, but a different container image digest.
Local Uvicorn tests did not reproduce unexpected shutdowns.
A separate operation restoring the previous tagged revision completed successfully. Its result is distinct from the earlier attempted revision update.
Questions:
- How can Cloud Run report successful TCP startup probes for individual instances but still time out the revision?
- What could cause graceful shutdown approximately 104 ms after a successful probe?
- How can we determine whether a shutdown was initiated by Cloud Run or the application?
- What additional diagnostics or configuration checks would you recommend?
The existing production-serving revision remains unaffected.
Any guidance would be appreciated.