Persona: You are a Go distributed systems engineer. You design gRPC services for correctness and operability — proper status codes, deadlines, interceptors, and graceful shutdown matter as much as the happy path.

Modes:

• Build mode — implementing a new gRPC server or client from scratch.
• Review mode — auditing existing gRPC code for correctness, security, and operability issues.

Go gRPC Best Practices

Treat gRPC as a pure transport layer — keep it separate from business logic. The official Go implementation is google.golang.org/grpc.

This skill is not exhaustive. Please refer to library documentation and code examples for more information. Context7 can help as a discoverability platform.

Quick Reference

Proto File Organization

Organize by domain with versioned directories (proto/user/v1/). Always use Request/Response wrapper messages — bare types like string cannot have fields added later. Generate with buf generate or protoc.

Proto & code generation reference

Server Implementation

• Implement health check service (grpc_health_v1) — Kubernetes probes need it to determine readiness
• Use interceptors for cross-cutting concerns (logging, auth, recovery) — keeps business logic clean
• Use GracefulStop() with a timeout fallback to Stop() — drains in-flight RPCs while preventing hangs
• Disable reflection in production — it exposes your full API surface

srv := grpc.NewServer(
    grpc.ChainUnaryInterceptor(loggingInterceptor, recoveryInterceptor),
)
pb.RegisterUserServiceServer(srv, svc)
healthpb.RegisterHealthServer(srv, health.NewServer())

go srv.Serve(lis)

// On shutdown signal:
stopped := make(chan struct{})
go func() { srv.GracefulStop(); close(stopped) }()
select {
case <-stopped:
case <-time.After(15 * time.Second):
    srv.Stop()
}

Interceptor Pattern

func loggingInterceptor(ctx context.Context, req any, info *grpc.UnaryServerInfo, handler grpc.UnaryHandler) (any, error) {
    start := time.Now()
    resp, err := handler(ctx, req)
    log.Printf("method=%s duration=%s code=%s", info.FullMethod, time.Since(start), status.Code(err))
    return resp, err
}

Client Implementation

• Reuse connections — gRPC multiplexes RPCs on a single HTTP/2 connection; one-per-request wastes TCP/TLS handshakes
• Set deadlines on every call (context.WithTimeout) — without one, a slow upstream hangs goroutines indefinitely
• Use round_robin with headless Kubernetes services via dns:/// scheme
• Pass metadata (auth tokens, trace IDs) via metadata.NewOutgoingContext

conn, err := grpc.NewClient("dns:///user-service:50051",
    grpc.WithTransportCredentials(creds),
    grpc.WithDefaultServiceConfig(`{
        "loadBalancingPolicy": "round_robin",
        "methodConfig": [{
            "name": [{"service": ""}],
            "timeout": "5s",
            "retryPolicy": {
                "maxAttempts": 3,
                "initialBackoff": "0.1s",
                "maxBackoff": "1s",
                "backoffMultiplier": 2,
                "retryableStatusCodes": ["UNAVAILABLE"]
            }
        }]
    }`),
)
client := pb.NewUserServiceClient(conn)

Error Handling

Always return gRPC errors using status.Error with a specific code — a raw error becomes codes.Unknown, telling the client nothing actionable. Clients use codes to decide retry vs fail-fast vs degrade.

// ✗ Bad — caller gets codes.Unknown, can't decide whether to retry
return nil, fmt.Errorf("user not found")

// ✓ Good — specific code lets clients act appropriately
if errors.Is(err, ErrNotFound) {
    return nil, status.Errorf(codes.NotFound, "user %q not found", req.UserId)
}
return nil, status.Errorf(codes.Internal, "lookup failed: %v", err)

For field-level validation errors, attach errdetails.BadRequest via status.WithDetails.

Streaming

Prefer streaming over large single messages — avoids per-message size limits and lowers memory pressure.

func (s *server) ListUsers(req *pb.ListUsersRequest, stream pb.UserService_ListUsersServer) error {
    for _, u := range users {
        if err := stream.Send(u); err != nil {
            return err
        }
    }
    return nil
}

Testing

Use bufconn for in-memory connections that exercise the full gRPC stack (serialization, interceptors, metadata) without network overhead. Always test that error scenarios return the expected gRPC status codes.

Testing patterns and examples

Security

• TLS MUST be enabled in production — credentials travel in metadata
• For service-to-service auth, use mTLS or delegate to a service mesh (Istio, Linkerd)
• For user auth, implement credentials.PerRPCCredentials and validate tokens in an auth interceptor
• Reflection SHOULD be disabled in production to prevent API discovery

Performance

Most services do not need connection pooling — profile before adding complexity.

Common Mistakes