engineering1 min de lectura
Friday Five #2: The exit code said zero
A deploy printed success while the old code kept running for 11 minutes. Five notes on trusting the tool less.
Rob Vasquez·
Five notes from the week. Short on purpose.
- Broke: A deploy printed success and exited 0, and my API kept running the old code for 11 minutes. The platform had rolled both machines back 85 seconds after the push because the image was too big to unpack.
- Changed: After every deploy I now check three things the deploy command can't tell me: the machine's event log for a revert, the image tag on the machine that is actually running, and a build hash on the health endpoint.
- Found: The same week, three pull requests in three different repos failed CI with the same blank error. I debugged two of them as separate bugs. The cause was the org's Actions billing cap. When one failure shows up in unrelated repos, I now check what they share before I read any code.
- Using: A build hash on the health endpoint. It's small to add and it answers "which version is really running" faster than any dashboard.
- Asking: Which green check do you trust the least?