Symptom
It worked in CI and died on deploy
You noticed
Every check passed. The deploy finished. The service never came up, or came up and returned 500 on the first real request, because of an environment variable, a port, or a secret that only existed on the laptop.
The doubt
does not run where deployed. It worked on the laptop and in CI. In the target environment a config value is missing, a port is wrong, or the service never comes up.
Commonly misread as closing it: Integration Tests.
Point this at it
Smoke tests, as a shell script of four curl calls the deploy job runs against the URL it just deployed: GET /health, POST one record, GET it back, DELETE it. Any non-2xx fails the script, and a failed script halts the rollout or rolls it back. The run takes under a second.
The part that matters is the target. Point it at the artifact that just went out, in the environment it went out to. A smoke run against staging while production stays old, or against a cached edge, is a green reading about a system nobody is serving.
Reading it
- Red: the deployed artifact is dead on that path in that environment. Re-run once to rule out bad credentials or a wrong URL; two consecutive failures are the artifact. Roll back, then diagnose, and keep the status code, body, and response time the failing check returned.
- Green: the system is alive on those four paths, here. A pass is necessary for the rollout to continue and sufficient for almost nothing else.
- It does not prove correctness on any path, behavior under load, or what happens when something the service depends on stops answering.
If the doubt survives
If the smoke run is green and the service goes down later in the same environment, it came up fine and fell over when a dependency got slow or disappeared. That is falls over with a dependency, and the next play is for it.
Next play: One slow dependency took everything down