Transient fault handling with gRPC retries

[!INCLUDE]

By James Newton-King

gRPC retries is a feature that allows gRPC clients to automatically retry failed calls. This article discusses how to configure a retry policy to make resilient, fault tolerant gRPC apps in .NET.

gRPC retries requires Grpc.Net.Client version 2.36.0 or later.

Transient fault handling

gRPC calls can be interrupted by transient faults. Transient faults include:

• Momentary loss of network connectivity.
• Temporary unavailability of a service.
• Timeouts due to server load.

When a gRPC call is interrupted, the client throws an RpcException with details about the error. The client app must catch the exception and choose how to handle the error.

var client = new Greeter.GreeterClient(channel);
try
{
    var response = await client.SayHelloAsync(
        new HelloRequest { Name = ".NET" });

    Console.WriteLine("From server: " + response.Message);
}
catch (RpcException ex)
{
    // Write logic to inspect the error and retry
    // if the error is from a transient fault.
}

Duplicating retry logic throughout an app is verbose and error prone. Fortunately the .NET gRPC client has a built-in support for automatic retries.

Configure a gRPC retry policy

A retry policy is configured once when a gRPC channel is created:

var defaultMethodConfig = new MethodConfig
{
    Names = { MethodName.Default },
    RetryPolicy = new RetryPolicy
    {
        MaxAttempts = 5,
        InitialBackoff = TimeSpan.FromSeconds(1),
        MaxBackoff = TimeSpan.FromSeconds(5),
        BackoffMultiplier = 1.5,
        RetryableStatusCodes = { StatusCode.Unavailable }
    }
};

var channel = GrpcChannel.ForAddress("https://localhost:5001", new GrpcChannelOptions
{
    ServiceConfig = new ServiceConfig { MethodConfigs = { defaultMethodConfig } }
});

The preceding code:

• Creates a MethodConfig. Retry policies can be configured per-method and methods are matched using the Names property. This method is configured with MethodName.Default, so it's applied to all gRPC methods called by this channel.
• Configures a retry policy. This policy instructs clients to automatically retry gRPC calls that fail with the status code Unavailable.
• Configures the created channel to use the retry policy by setting GrpcChannelOptions.ServiceConfig.

gRPC clients created with the channel will automatically retry failed calls:

var client = new Greeter.GreeterClient(channel);
var response = await client.SayHelloAsync(
    new HelloRequest { Name = ".NET" });

Console.WriteLine("From server: " + response.Message);

When retries are valid

Calls are retried when:

• The failing status code matches a value in RetryableStatusCodes.
• The previous number of attempts is less than MaxAttempts.
• The call hasn't been commited.
• The deadline hasn't been exceeded.

A gRPC call becomes committed in two scenarios:

• The client receives response headers. Response headers are sent by the server when ServerCallContext.WriteResponseHeadersAsync is called, or when the first message is written to the server response stream.
• The client's outgoing message (or messages if streaming) has exceeded the client's maximum buffer size. MaxRetryBufferSize and MaxRetryBufferPerCallSize are configured on the channel.

Committed calls won't retry, regardless of the status code or the previous number of attempts.

Streaming calls

Streaming calls can be used with gRPC retries, but there are important considerations when they are used together:

• Server streaming, bidirectional streaming: Streaming RPCs that return multiple messages from the server won't retry after the first message has been received. Apps must add additional logic to manually re-establish server and bidirectional streaming calls.
• Client streaming, bidirectional streaming: Streaming RPCs that send multiple messages to the server won't retry if the outgoing messages have exceeded the client's maximum buffer size. The maximum buffer size can be increased with configuration.

For more information, see When retries are valid.

Retry backoff delay

The backoff delay between retry attempts is configured with InitialBackoff, MaxBackoff, and BackoffMultiplier. More information about each option is available in the gRPC retry options section.

The actual delay between retry attempts is randomized. A randomized delay between 0 and the current backoff determines when the next retry attempt is made. Consider that even with exponential backoff configured, increasing the current backoff between attempts, the actual delay between attempts isn't always larger. The delay is randomized to prevent retries from multiple calls from clustering together and potentially overloading the server.

Detect retries with metadata

gRPC retries can be detected by the presence of grpc-previous-rpc-attempts metadata. The grpc-previous-rpc-attempts metadata:

• Is automatically added to retried calls and sent to the server.
• Value represents the number of preceding retry attempts.
• Value is always an integer.

Consider the following retry scenario:

1. Client makes a gRPC call to the server.
2. Server fails and returns a retriable status code response.
3. Client retries the gRPC call. Because there was one previous attempt, grpc-previous-rpc-attempts metadata has a value of 1. Metadata is sent to the server with the retry.
4. Server succeeds and returns OK.
5. Client reports success. grpc-previous-rpc-attempts is in the response metadata and has a value of 1.

The grpc-previous-rpc-attempts metadata is not present on the initial gRPC call, is 1 for the first retry, 2 for the second retry, and so on.

gRPC retry options

The following table describes options for configuring gRPC retry policies:

Hedging

Hedging is an alternative retry strategy. Hedging enables aggressively sending multiple copies of a single gRPC call without waiting for a response. Hedged gRPC calls may be executed multiple times on the server and the first successful result is used. It's important that hedging is only enabled for methods that are safe to execute multiple times without adverse effect.

Hedging has pros and cons when compared to retries:

• An advantage to hedging is it might return a successful result faster. It allows for multiple simultaneously gRPC calls and will complete when the first successful result is available.
• A disadvantage to hedging is it can be wasteful. Multiple calls could be made and all succeed. Only the first result is used and the rest are discarded.

Configure a gRPC hedging policy

A hedging policy is configured like a retry policy. Note that a hedging policy can't be combined with a retry policy.

var defaultMethodConfig = new MethodConfig
{
    Names = { MethodName.Default },
    HedgingPolicy = new HedgingPolicy
    {
        MaxAttempts = 5,
        NonFatalStatusCodes = { StatusCode.Unavailable }
    }
};

var channel = GrpcChannel.ForAddress("https://localhost:5001", new GrpcChannelOptions
{
    ServiceConfig = new ServiceConfig { MethodConfigs = { defaultMethodConfig } }
});

gRPC hedging options

The following table describes options for configuring gRPC hedging policies:

Additional resources

• xref:grpc/client
• Retry general guidance - Best practices for cloud applications