Introduce grpc health check in etcd client

[chaochn47]

What would you like to be added?

Background

gRPC Health checks are used to probe whether the server is able to handle RPCs. A server may choose to reply “unhealthy” because it is not ready to take requests, it is shutting down or some other reason. The client can act accordingly if the response is not received within some time window or the response says unhealthy in it.

ref.

1. https://github.com/grpc/grpc/blob/master/doc/health-checking.md
2. https://github.com/grpc/proposal/blob/master/A17-client-side-health-checking.md

#8121 added basic grpc health service only on server side since etcd v3.3.

etcd/server/etcdserver/api/v3rpc/grpc.go

Lines 81 to 86 in 6220174

	// server should register all the services manually
	// use empty service name for all etcd services' health status,
	// see https://github.com/grpc/grpc/blob/master/doc/health-checking.md for more
	hsrv := health.NewServer()
	hsrv.SetServingStatus("", healthpb.HealthCheckResponse_SERVING)
	healthpb.RegisterHealthServer(grpcServer, hsrv)

Problem

In a multi etcd server endpoints scenario, etcd client only fails over to the other endpoint when existing connection/channel is in not in Ready state. However, etcd client does not know about if etcd server can handle RPCs.

For example

1. defrag stops the process completely for noticeable duration #9222
2. Removed etcd member failed to stop on stuck disk #14338

It needs a comprehensive design and testing.

Placeholder google doc etcd client grpc health check copied from the KEP template.

Why is this needed?

Improve etcd availability by failing over to other healthy etcd endpoints

[chaochn47]

@serathius @ahrtr @wenjiaswe @ptabor @etcd-io/members for discussion if this feature is wanted.

[fuweid]

Hi @chaochn47 , based on your PR #16278, it seems like that the PR marks not_serving status because of defrag. It should be applied to any maintenance APIs, is it correct? My only concern is that based on that https://github.com/grpc/grpc-go/blob/d7f45cdf9ae720256d328fcdb6356ae58122a7f6/health/client.go#L104, it might impact the on-going requests(still in trasfering data) and force to recreate connection. (I haven't verified it yet, update it later)

[serathius]

For me this duplicates #16007. We need single comprehensive design for etcd probing. As part of this work we can update the grpc probes.

[chaochn47]

Thanks @fuweid for the review!

based on your PR #16278, it seems like that the PR marks not_serving status because of defrag. It should be applied to any maintenance APIs, is it correct?

Online Defrag is one of the use cases that server reports "unhealthy" to client to fail over the requests. I do not plan to support any other maintenance APIs now. #16278 is just a POC and health server should not exposed to maintenance server, while to defrag API only ideally.

My only concern is that based on that https://github.com/grpc/grpc-go/blob/d7f45cdf9ae720256d328fcdb6356ae58122a7f6/health/client.go#L104, it might impact the on-going requests(still in trasfering data) and force to recreate connection.

I won't be too concerned about this failure mode since it is not called out in https://github.com/grpc/proposal/blob/master/A17-client-side-health-checking.md and grpc/proposal#97. But I will validate with adding some testing cases in https://github.com/grpc/grpc-go/blob/master/test/healthcheck_test.go in case it's an unexpected golang specific implementation issue.

[chaochn47]

Thanks @serathius for taking a look!

For me this duplicates #16007. We need single comprehensive design for etcd probing. As part of this work we can update the grpc probes.

I don't agree that it's a duplicate of #16007. I believe /livez and /readyz concepts are brought from https://kubernetes.io/docs/tasks/configure-pod-container/configure-liveness-readiness-startup-probes/ and it's a polling health check mechanism. It is designed for etcd administration to determine when to restart/replace etcd with a new process / container or when to route the client traffic to etcd server.

However, gRPC client side health check replies on etcd (gRPC) server pushes its health status to client for requests fail-over via a server side streaming API. It is designed only for round-robin policy and will improve etcd availability.

I do agree some of the server health status determination logic can be re-used but the two proposals don't share the same motivation. For example, we definitely do not want terminating etcd server when online defrag is active but expect client to route requests to other etcd servers.

@wenjiaswe @logicalhan please drop a comment in case I misunderstood the proposal #16007, thx~

[serathius]

However, gRPC client side health check replies on etcd (gRPC) server pushes its health status to client for requests fail-over via a server side streaming API. It is designed only for round-robin policy and will improve etcd availability.

Pull push, just changes who has the control over validation period. Server or client.

What's important here is definition of etcd health. I expect both types of health checking will end up with pretty similar definition of health.

[chaochn47]

Thanks. I think we are on the same page now.

Do you still think it's a duplicate of #16007? Will it cover how to consume the health status change?

I would like the etcd client to fail over to other healthy endpoints considering etcd health, not just connection is up or down.

[chaochn47]

/cc @jyotima

[wenjiaswe]

gRPC Health checks are used to probe whether the server is able to handle RPCs.

That sounds like one case of not ready for /readyz, is my understanding correct?

[chaochn47]

Yeah I think so but I want this thread focus on etcd client fail over by introducing gRPC health checks.

I suggest we move the conversation about etcd failure modes classification into which probes to #16007

[serathius]

This issue can be left opened, as a followup to introduce grpc client health checks when we figure out proper health definitions.

[ahrtr]

It's a valid feature to me, and I agree we should design it together with the Livez/Readyz feature, because the key is how to define health (e.g. get stuck on write for a long time > a threshold).

Also two detailed comments:

• We plan to get rid of defragmentation in 3.7 in Support customizing the rebalance threshold bbolt#422. [But I do not get time to do a PoC so far :( ]
• The PR gRPC health server sets serving status to NOT_SERVING on defrag #16278 is still depending on an experimental gRPC API WithResolvers, but we want to remove the dependency on the experimental gRPC API in Remove the dependency on grpc-go's experimental API (e.g. resolver & balancer) #15145