Abstract

A method is disclosed for verifying a replaced service instance before public traffic is restored to it, while the public edge serves a maintenance page. After the rollout completes, a smoke gate runs its checks from inside the new instance: through the container orchestrator's remote-execution facility it issues HTTP requests to the application over the loopback interface. Each request carries an in-instance caller marker, which the application's authentication layer, absent any credential, honours only when the raw transport peer of the connection is a loopback address, never a client address derived from a forwarded-for header, and then attaches a synthetic service principal. Authenticated routes, including an end-to-end language-model round trip, are therefore exercised without any stored credential. Because the instance mounts a volume that attaches to one node at a time, the driver first pins the replacement to the node holding the volume by making every other eligible node unschedulable for the deploy window, and later reverses only the changes it made. Automatic rollback is taken only on a positive rollout-failure signal, never on silence or timeout, and runs before any pin is lifted, so a pinned rollback instance is confined to the holder node. On a smoke failure the deploy driver itself neither rolls back nor, by default, restores traffic: it keeps the maintenance page up, pages an operator and records the deploy as failed, and an independent watchdog lifts the page once an expiry recorded at engagement has passed, so the service cannot be stranded behind maintenance.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS