Tasks

Each one is a manifest to write or repair, graded on gates the engine can answer by looking at what it recorded rather than by reading your YAML. Every failed request carries the reason it failed, stamped when it was emitted.

• The quarter that was one

  Rollouts · warm-up

  `checkout` runs four replicas and needs three of them serving at all times. The rollout below is about to replace the image. Change the manifest so the service never drops below three ready replicas — and do not simply run more of them.

• Ready, and unable to serve

  Probes · real

  `checkout` takes about fifteen seconds after start before it can serve a request — connection pool, cache warm-up. Its readiness probe answers long before that, so the rollout sends traffic to pods that cannot serve it. Fix the manifest so the rollout costs no failed requests.

• The rollout that will never finish

  Rollouts · real

  `checkout` has been rolling out a new image for twenty minutes. Every pod is healthy, every probe is passing, no container has restarted, and the service is serving every request. It is also still entirely on the old version. Make the rollout able to proceed, without reducing how many replicas stay available.

• The requests that were already in flight

  Traffic and endpoints · real

  Every rollout of `checkout` drops a handful of requests. The pods are healthy, the new ones come up fine, nothing restarts, and the endpoints are updated the moment a pod starts shutting down. Find what is costing the requests and fix it. Your change will be rolled out once to take effect, and the rollout after that is the one that is measured.

• Half a cluster, and nothing fits

  Scheduling · real

  `checkout` needs four replicas. The cluster has two nodes with two cores each and is barely using them, and the rollout will not place its pods. The service genuinely needs about 250 millicores per replica under load. Make it fit, without giving it less headroom than it has now.

• It is not the database

  Pressure · hard

  `checkout` has got slow under load. Latency is up, callers are timing out, and the team is looking at the database — which is answering in single-digit milliseconds. The nodes are barely half used. Find what is actually holding it back and fix it, without asking for more replicas.