Use HTTP probes for Ray readiness and liviness probes

[andrewsykim]

Why are these changes needed?

HTTP probes are considered lighter-weight than exec probes. However, exec probes have the advantage of doing multiple health checks. In KubeRay, we use exec probes to execute "wget" commands against multiple endpoints. Use of exec probes seems to be causing some issues, as shown in #2264 and from KubeRay scalability testing.

This PR explores using HTTP probes instead. This PR needs more consideration as using HTTP probes means we can only health check 1 end point per probe. Marking WIP for now until that quesiton is resolved.

Related issue number

Checks

• I've made sure the tests are passing.
• Testing Strategy
  • Unit tests
  • Manual tests
  • This PR is not tested :(

[andrewsykim]

Using HTTP probes means we can only query 1 endpoint per probe now. For head pod this would /api/gcs_healthz and for worker pod it would be api/local_raylet_healthz. I'm not sure if not health checking api/local_raylet_healthz in the head pod is problematic, it would depend on what whether /api/gcs_healthz incorporates raylet health in some way as well

[YQ-Wang]

We also face this issue when the workload is high.

[kevin85421]

@andrewsykim do we still need this PR after #2353 has been merged?

[andrewsykim]

I think we should still consider use of HTTP probes, they are significantly ligher weight. I haven't root caused the issue I'm seeing, but increasing the timeout did not fully resolve the issue I'm seeing where exec probes cause high load

[joshhvulcan]

I have encountered some bizarre behavior with the exec probes that I think would be solved with http probes. The biggest issue may actually be a k8s bug though. The ray-head container had died but the autoscaler container was still running so the pod was kept alive by k8s. The probe was failing because the exec was failing and k8s took no action because the probe failed to fail? (lol). Anyway, http likely probes would have done the needful here.

Readiness probe errored: rpc error: code = Unknown desc = failed to exec in container: failed to start exec "03e693753f58930cd9bf004e047ff1cf7c26afd30ea916cbe0d291e130ea9d27": OCI runtime exec failed: exec failed: cannot exec in a stopped container: unknown

[andrewsykim]

I'm still in favor of this change and I would like to see it merged for v1.3. However, using http probes means we can only probe 1 HTTP endpoint per container. Specifically for the Head pod, it means probing only the dashboard endpoint and not the raylet agent. Are w okay with that change? @kevin85421 @joshhvulcan

[joshhvulcan]

I think an http probe on the dashboard would be sufficient for the failures we have experienced.

[andrewsykim]

PR updated, PTAL