Zero-Downtime Deployment

Zero-downtime: achieved when new instances start and pass health checks before old ones are terminated. Orchestrators handle the gradual traffic shift so no requests are dropped during the transition.

Graceful shutdown: when a container or process receives SIGTERM, it stops accepting new requests, waits for active ones to complete, then exits. Without it, in-flight requests are abruptly terminated during deployments.

Backward-compatible migrations: during a rolling deploy, old and new app versions run together. Adding a non-nullable column with no default, or renaming a column, breaks the old version. Migrations should be expand-then-contract: add new, keep old, remove old later.

Readiness probe: Kubernetes routes traffic only to pods whose readiness probe passes. During rolling updates, new pods receive traffic only once ready, ensuring users are never sent to a pod still warming up.

Connection draining: when an instance is deregistered (during a deploy or scale-in), the load balancer stops routing new requests to it but waits a configurable period for existing connections to finish before fully removing it.