perAttemptRecvTimeout in RetryPolicy does nothing

[vlad1147]

What version of gRPC-Java are you using?

1.76.0 (but it's the same problem in the latest release too)

What is your environment?

Android app with gprc-java and grpc-android libs

What did you expect to see?

#8301 added perAttemptRecvTimeoutNanos to the RetryPolicy but the problem that it's never used so it's not possible to set a timeout for an attempt within the gRPC call. Not sure but might be related to this issue #1943

What did you see instead?

I would expect RetriableStream to use perAttemptRecvTimeoutNanos from RetryPolicy so makeRetryDecision will actually retry when an attempt takes longer than specified in perAttemptRecvTimeoutNanos

Steps to reproduce the bug

Can be reproduced with any channel and RetryPolicy that has perAttemptRecvTimeout.

A bit of context why do I even need it

I'm looking for a way to detect and recover from a black hole gRPC connection that happens in my app.
I have logs from production with DEADLINE_EXCEEDED exception that either has waiting_for_connection or remote_addr=/10.0.2.2:8443 (it's from local test but the point that there is a remote_addr in production). As far as I understand in the first case the channel is in CONNECTING state and in the second is in READY but neither of them can succeed and backend has no errors. I reproduced the issue locally by using toxiproxy and simulated a black hole connection. My idea is to pass perAttemptRecvTimeout in RetryPolicy so I can rely on gRPC internal retry mechanism and when an attempt fails with DEADLINE_EXCEEDED to call channel.enterIdle (same as AndroidChannelBuilder does when detects the change in network to force a new connection for the next call) in ClientStreamFactory (same idea as for refreshing an expired auth token but without CallCredentials as was suggested here and it actually works in our app #7345 (comment))

object : ClientStreamTracer.Factory() {
        override fun newClientStreamTracer(
            info: ClientStreamTracer.StreamInfo,
            headers: Metadata,
        ): ClientStreamTracer = object : ClientStreamTracer() {
            override fun streamClosed(status: Status) {
                if (status.code == Status.Code.DEADLINE_EXCEEDED) {
                    channel.enterIdle()
                }
            }
        }
    }

I also considered keepAlive option but it doesn't seems to work as smooth as the idea above and if I'm not wrong - it will not recover from a black hole connection when the channel is in CONNECTING because it's only sent for an established connection so the channel has to be in READY state. And also it will drain battery and spam the backend.

Also I thought to have a retry interceptor but I'm afraid it's a fragile approach that previously caused us a lot of crashes when it was attempted for token refresh and retry afterwards. (Probably it's possible to implement it flawlessly but it likely it will be difficult to understand and easy to break in the future and therefore not considered).

And as a last resort it's also possible to use try/catch and retry on the call-site and manage the channel there but it requires to change every single call-site in the app and can be easily forgotten for new calls. So not a sustainable option.

Just wanted to share what I already thought of and I believe the first suggestion that uses perAttemptRecvTimeout is the best among listed approaches. I would be happy to hear if there are other possible robust ways to fix it.

[kannanjgithub]

We should make the retry attempt honor the perAttemptRecvTimeout setting from the retry policy.