· Reliability · 6 min read
Readiness probes, graceful shutdowns, and boring rollouts
By Ethan Cole, Co-Founder & CTO
Zero-downtime deploys rarely fail because of the scheduler. They fail because an app ignored SIGTERM or received traffic before its dependencies were warm.
A rollout that drops a few dozen requests in the middle of the afternoon usually points to one of two gaps: traffic shifted to a container before it was ready, or an old container exited while requests were still in flight.
By default, Nimbus wires HTTP health checks on the path you specify (such as /healthz) into every deploy. New containers come up alongside existing replicas and receive zero production traffic until consecutive readiness probes succeed.
If the new build fails its probe or crashes during boot, the rollout halts automatically and the prior revision keeps serving without manual intervention.
On shutdown, containers receive SIGTERM with a grace window to finish active responses and close database connections cleanly before being removed from service.
Boring production deploys are not built from heroic runbooks. They come from enforcing probe checks and clean drains on every single push.