Retries for specific gRPC errors/error codes

[vaibhavsawhney]

As I can see that we used to have gRPC Response Classifiers in Linkerd 1.x through which we could decide which type of errors/ error codes we want to retry our request for if service profile is enabled.

I tried finding out if we can do something similar to this in Linkerd 2.x but I couldn't.
The main motive of using this was to not retry for errors which are irrecoverable in nature but Linkerd retries for them as well.

I could find Response Match which can be used to define status codes for considering a response as failure and success. But then I looked into an issue #3376 raised last year around the same for which a fix was done later #3729. Though this fix has a fixed set of error codes that will be retried and cannot be changed according to the client's preference.

I wanted to ask if there is any further development in this area. Is there a way to define error codes that should be retried?

[grampelberg]

@adleong we can do HTTP error codes in service profiles but not gRPC error codes right?

[vaibhavsawhney]

Yes that's what the issue #3376 says and I too couldn't get any other way for gRPC.

[olix0r]

We currently only support retries for requests that have no payloads; and gRPC requests necessarily have payloads.

I think we could reasonably support retrying requests that fit into some small buffer capacity... Ideally no more than a few KB.

#4366 would also cover many UNAVAILABLE cases automatically.

[vaibhavsawhney]

Thank you both for the response. I guess this discussion can be concluded by saying that we cannot do cherry-picking of status codes in gRPC to retry the linked requests currently.

Let me know if I have understood this correctly and I will mark this as the answer and close this discussion.