Three Deployments in caverns-8tinqw. None is healthy, and every image is correct. The faults are in what each container is given.

kubectl get deploy,pods,svc -n caverns-8tinqw

Deployment       symptom                   source of truth
market-api       CrashLoopBackOff          Secret db-creds, key name
market-api       0/2 READY, no endpoints   readinessProbe.httpGet.port
market-web       RESTARTS rising           livenessProbe.httpGet.path
market-worker    OOMKilled / 137           resources.limits.memory
market-worker    queue backend: memory://  ConfigMap market-config, key name
all three        QoS BestEffort            resources.requests

Fix them in any order.

market-api

kubectl logs deploy/market-api names a variable the container cannot find. It comes from Secret db-creds. Compare the key in env[0].valueFrom.secretKeyRef with the keys in kubectl get secret db-creds -o yaml. Delete the Secret and create it with the URL under the right key (password s3cret).

The readiness probe dials port 9090 and fails after one miss. The container listens on 8080. Apply the Deployment with the probe on port 8080 and keep the secretKeyRef env. The new rollout also makes the Pods read the fixed Secret.

market-web

kubectl describe pod shows the Liveness: line and Unhealthy events with a 404. nginx serves /. Apply the Deployment with the probe on /.

market-worker

Last State shows OOMKilled. Its working set is about 96Mi. Raise the limit with kubectl set resources and set a memory request.

The worker reads QUEUE_URL from ConfigMap market-config, which stores the value under a different key. Recreate the ConfigMap with QUEUE_URL (keep MARKET_NAME) and restart the worker.

Requests on every Deployment

No Pod may be BestEffort. Set CPU and memory requests on market-web, market-api and market-worker:

kubectl set resources deploy/market-web -n caverns-8tinqw --requests=cpu=50m,memory=32Mi --limits=memory=64Mi
kubectl get pods -n caverns-8tinqw -o jsonpath='{range .items[*]}{.metadata.name} {.status.qosClass}{"\n"}{end}'

Keep the worker's memory limit at 128Mi or more.

The last objective is a question: which probe failure removes a Pod from a Service without restarting it?