A sane Kubernetes deploy checklist

A deploy goes wrong in ordinary ways. The pod has no limit and starves a node. The tag is latest, so a rollback pulls whatever that word means on the day. The process is up, and readiness never says so. None of this needs a story about a particular outage. It is the list to walk before the manifest is applied.

On the manifest

• CPU and memory requests and limits are set. A pod with no limit can take the node with it.
• The image tag is a version you can roll back to, not latest.
• Liveness and readiness probes exist. Readiness fails when the process is up but not ready to serve.
• The container is not privileged, and it does not mount the container socket.
• Secrets are not baked into the image, and they are not written in plain text in the manifest.
• The service account, the namespace and the network policy are the ones you meant.

Before you call it deployed

• You know the command that rolls it back, and you have run that command once on a deploy that did not matter.
• There is a signal, written down, that tells you the change worked. If the only signal is that nobody has complained, you do not know yet.
• Someone other than the person who wrote the manifest can read the probes and say what healthy means.

Check the YAML without a cluster

The Kubernetes YAML linter flags a missing limit, a latest tag, a privileged container and a missing probe. The manifest is checked and then discarded.