Classification

A gRPC failure does not live where the HTTP classifier looks. Application errors travel in the grpc-status trailer on an HTTP 200, and transport failures surface as an RpcException after the HTTP classifier has judged that 200 a success.

GrpcResilience.Classifier is Classifier.Default plus one rule: read the StatusCode on an RpcException.

// GrpcResilience.Default is Resilience.Default with this classifier already on it.
var classifier = GrpcResilience.Classifier;

var unavailable = classifier.ClassifyException(new RpcException(new Status(StatusCode.Unavailable, "moving")));
var notFound = classifier.ClassifyException(new RpcException(new Status(StatusCode.NotFound, "no such order")));
var exhausted = classifier.ClassifyException(new RpcException(new Status(StatusCode.ResourceExhausted, "quota")));

Console.WriteLine(unavailable.Kind); // Transient
Console.WriteLine(notFound.Kind); // Permanent - an answer, not a failure
Console.WriteLine(exhausted.Kind); // Throttled - the dependency is defending itself

GrpcResilience.Default is Resilience.Default with that classifier already on it, so AddGrpcResilience() needs no classifier argument. For more information about verdicts, see Classification.

The shipped table

StatusCode	Verdict	Why
Unavailable	Transient	The transport could not reach the method. The canonical retryable status.
DeadlineExceeded	Transient	A ceiling this side or the peer set expired. A fresh attempt gets a fresh ceiling.
ResourceExhausted	Throttled	The dependency is defending itself - out of quota, out of memory, out of concurrency. One verdict buys the long backoff curve, no evidence against the breaker, and no charge to the retry budget.
Internal	Permanent	The server's own bug. Retrying multiplies load against something that is already broken.
Unauthenticated, PermissionDenied	Permanent	Credentials do not fix themselves on a retry. Refresh them in BeforeAttempt, which runs before each attempt.
InvalidArgument, NotFound, AlreadyExists, FailedPrecondition, OutOfRange, Unimplemented, DataLoss	Permanent	Answers, not failures - the same line the HTTP classifier takes with a 404.
Aborted	Permanent	A transaction conflict. Whether repeating one is safe depends on the store, so the conservative verdict ships. For an example of changing it, see the following section.
Cancelled	Permanent	The interceptor translates your cancellations before the classifier sees them, so an RpcException(Cancelled) that reaches it is a peer that hung up. Repeating a call the other end abandoned is a guess.
Anything else	Permanent	Classifier.Default does not retry what it does not recognize. Retrying a programming error converts a fast, clear failure into a slow, confusing one.

Change any of it

Every row is one line to override. The last rule registered for an exception type is the one that runs, so a rule of yours replaces the shipped one and can fall back to it:

// Aborted is a transaction conflict, and whether repeating one is safe depends on the store -
// so the shipped verdict is Permanent and this is how a store that wants it says so.
var policy = GrpcResilience.Default with
{
    Classify = GrpcResilience.Classifier.On<RpcException>(
        static e => e.StatusCode == StatusCode.Aborted
            ? Verdict.Transient
            : GrpcResilience.Classifier.ClassifyException(e)),
};

Aborted is the row most often worth changing. A transactional store that reports write conflicts as Aborted and expects the client to retry them is a real shape - it is just not the only one, and repeating a conflicting write against a store that does not expect it is worse than failing.

What the classifier does not see

Two things reach the caller without passing the classifier at all, and both are deliberate:

• Your own cancellation. A cancelled caller token is never a failure, is never retried, and no classifier can override it.
• Your own attempt timeout. An attempt that exceeds its ceiling produces an AttemptTimeoutException, judged by the executor rather than by a predicate. See Deadlines for how the interceptor keeps a gRPC DeadlineExceeded from being confused with one.

Retry pushback

gRPC carries retry pushback as a google.rpc.RetryInfo message inside the grpc-status-details-bin trailer. NResilience does not read it: doing so means taking a dependency on Google.Rpc and base64-decoding a details field, and the Throttled verdict already produces the long backoff curve without it.

If your dependency sends one and you want to honor it, write a classifier rule that returns Verdict.Throttled(retryAfter).