Health & readiness

The engine exposes two unauthenticated probes so an orchestrator or load balancer can tell whether to restart an instance and whether to route traffic to it.

Both bypass authentication - probes carry no API key - and are not written to the request log.

Liveness: GET /healthz

Returns 200 with {"status":"ok"} as long as the HTTP listener is up. It deliberately checks no dependency: a transient database blip should not cause an orchestrator to kill and restart a healthy process. Use it for a Kubernetes livenessProbe.

curl -i $DURABLEX_ENGINE_URL/healthz
# 200 OK
# {"status":"ok"}

Readiness: GET /readyz

Returns 200 with {"status":"ready"} only when every dependency probe passes and the engine is not draining; otherwise 503 with {"status":"not_ready"}. Use it for a Kubernetes readinessProbe so the load balancer stops routing to an instance that cannot serve.

The checks map reports each probed dependency:

curl -s $DURABLEX_ENGINE_URL/readyz
# {"status":"ready","checks":{"store":"ok"}}

Readiness probes only what is actually a network dependency:

• store - the datastore, always probed.
• queue / limiter - probed only when backed by a network service; in-process defaults have no network to check and never appear in checks.

A failed dependency is marked unavailable and flips the response to 503. The detailed connection error is written to the engine log, not the response body, since /readyz is unauthenticated:

# {"status":"not_ready","checks":{"store":"ok","queue":"unavailable"}}

During shutdown

On SIGINT/SIGTERM the engine begins a graceful drain. For the whole drain window /readyz returns 503 with {"status":"not_ready","draining":true} so the load balancer removes the instance from rotation before in-flight requests finish, rather than sending it new traffic mid-shutdown.

Kubernetes example

livenessProbe:
  httpGet:
    path: /healthz
    port: 6770
  periodSeconds: 10
readinessProbe:
  httpGet:
    path: /readyz
    port: 6770
  periodSeconds: 10