nikesh.krishnan | 6dd882b | 2023-03-14 10:02:41 +0530 | [diff] [blame] | 1 | // Copyright 2016 Michal Witkowski. All Rights Reserved. |
| 2 | // See LICENSE for licensing terms. |
| 3 | |
| 4 | package grpc_retry |
| 5 | |
| 6 | import ( |
| 7 | "context" |
| 8 | "fmt" |
| 9 | "io" |
| 10 | "sync" |
| 11 | "time" |
| 12 | |
| 13 | "github.com/grpc-ecosystem/go-grpc-middleware/util/metautils" |
| 14 | "golang.org/x/net/trace" |
| 15 | "google.golang.org/grpc" |
| 16 | "google.golang.org/grpc/codes" |
| 17 | "google.golang.org/grpc/metadata" |
| 18 | "google.golang.org/grpc/status" |
| 19 | ) |
| 20 | |
| 21 | const ( |
| 22 | AttemptMetadataKey = "x-retry-attempty" |
| 23 | ) |
| 24 | |
| 25 | // UnaryClientInterceptor returns a new retrying unary client interceptor. |
| 26 | // |
| 27 | // The default configuration of the interceptor is to not retry *at all*. This behaviour can be |
| 28 | // changed through options (e.g. WithMax) on creation of the interceptor or on call (through grpc.CallOptions). |
| 29 | func UnaryClientInterceptor(optFuncs ...CallOption) grpc.UnaryClientInterceptor { |
| 30 | intOpts := reuseOrNewWithCallOptions(defaultOptions, optFuncs) |
| 31 | return func(parentCtx context.Context, method string, req, reply interface{}, cc *grpc.ClientConn, invoker grpc.UnaryInvoker, opts ...grpc.CallOption) error { |
| 32 | grpcOpts, retryOpts := filterCallOptions(opts) |
| 33 | callOpts := reuseOrNewWithCallOptions(intOpts, retryOpts) |
| 34 | // short circuit for simplicity, and avoiding allocations. |
| 35 | if callOpts.max == 0 { |
| 36 | return invoker(parentCtx, method, req, reply, cc, grpcOpts...) |
| 37 | } |
| 38 | var lastErr error |
| 39 | for attempt := uint(0); attempt < callOpts.max; attempt++ { |
| 40 | if err := waitRetryBackoff(attempt, parentCtx, callOpts); err != nil { |
| 41 | return err |
| 42 | } |
| 43 | callCtx := perCallContext(parentCtx, callOpts, attempt) |
| 44 | lastErr = invoker(callCtx, method, req, reply, cc, grpcOpts...) |
| 45 | // TODO(mwitkow): Maybe dial and transport errors should be retriable? |
| 46 | if lastErr == nil { |
| 47 | return nil |
| 48 | } |
| 49 | logTrace(parentCtx, "grpc_retry attempt: %d, got err: %v", attempt, lastErr) |
| 50 | if isContextError(lastErr) { |
| 51 | if parentCtx.Err() != nil { |
| 52 | logTrace(parentCtx, "grpc_retry attempt: %d, parent context error: %v", attempt, parentCtx.Err()) |
| 53 | // its the parent context deadline or cancellation. |
| 54 | return lastErr |
| 55 | } else if callOpts.perCallTimeout != 0 { |
| 56 | // We have set a perCallTimeout in the retry middleware, which would result in a context error if |
| 57 | // the deadline was exceeded, in which case try again. |
| 58 | logTrace(parentCtx, "grpc_retry attempt: %d, context error from retry call", attempt) |
| 59 | continue |
| 60 | } |
| 61 | } |
| 62 | if !isRetriable(lastErr, callOpts) { |
| 63 | return lastErr |
| 64 | } |
| 65 | } |
| 66 | return lastErr |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | // StreamClientInterceptor returns a new retrying stream client interceptor for server side streaming calls. |
| 71 | // |
| 72 | // The default configuration of the interceptor is to not retry *at all*. This behaviour can be |
| 73 | // changed through options (e.g. WithMax) on creation of the interceptor or on call (through grpc.CallOptions). |
| 74 | // |
| 75 | // Retry logic is available *only for ServerStreams*, i.e. 1:n streams, as the internal logic needs |
| 76 | // to buffer the messages sent by the client. If retry is enabled on any other streams (ClientStreams, |
| 77 | // BidiStreams), the retry interceptor will fail the call. |
| 78 | func StreamClientInterceptor(optFuncs ...CallOption) grpc.StreamClientInterceptor { |
| 79 | intOpts := reuseOrNewWithCallOptions(defaultOptions, optFuncs) |
| 80 | return func(parentCtx context.Context, desc *grpc.StreamDesc, cc *grpc.ClientConn, method string, streamer grpc.Streamer, opts ...grpc.CallOption) (grpc.ClientStream, error) { |
| 81 | grpcOpts, retryOpts := filterCallOptions(opts) |
| 82 | callOpts := reuseOrNewWithCallOptions(intOpts, retryOpts) |
| 83 | // short circuit for simplicity, and avoiding allocations. |
| 84 | if callOpts.max == 0 { |
| 85 | return streamer(parentCtx, desc, cc, method, grpcOpts...) |
| 86 | } |
| 87 | if desc.ClientStreams { |
| 88 | return nil, status.Errorf(codes.Unimplemented, "grpc_retry: cannot retry on ClientStreams, set grpc_retry.Disable()") |
| 89 | } |
| 90 | |
| 91 | var lastErr error |
| 92 | for attempt := uint(0); attempt < callOpts.max; attempt++ { |
| 93 | if err := waitRetryBackoff(attempt, parentCtx, callOpts); err != nil { |
| 94 | return nil, err |
| 95 | } |
| 96 | callCtx := perCallContext(parentCtx, callOpts, 0) |
| 97 | |
| 98 | var newStreamer grpc.ClientStream |
| 99 | newStreamer, lastErr = streamer(callCtx, desc, cc, method, grpcOpts...) |
| 100 | if lastErr == nil { |
| 101 | retryingStreamer := &serverStreamingRetryingStream{ |
| 102 | ClientStream: newStreamer, |
| 103 | callOpts: callOpts, |
| 104 | parentCtx: parentCtx, |
| 105 | streamerCall: func(ctx context.Context) (grpc.ClientStream, error) { |
| 106 | return streamer(ctx, desc, cc, method, grpcOpts...) |
| 107 | }, |
| 108 | } |
| 109 | return retryingStreamer, nil |
| 110 | } |
| 111 | |
| 112 | logTrace(parentCtx, "grpc_retry attempt: %d, got err: %v", attempt, lastErr) |
| 113 | if isContextError(lastErr) { |
| 114 | if parentCtx.Err() != nil { |
| 115 | logTrace(parentCtx, "grpc_retry attempt: %d, parent context error: %v", attempt, parentCtx.Err()) |
| 116 | // its the parent context deadline or cancellation. |
| 117 | return nil, lastErr |
| 118 | } else if callOpts.perCallTimeout != 0 { |
| 119 | // We have set a perCallTimeout in the retry middleware, which would result in a context error if |
| 120 | // the deadline was exceeded, in which case try again. |
| 121 | logTrace(parentCtx, "grpc_retry attempt: %d, context error from retry call", attempt) |
| 122 | continue |
| 123 | } |
| 124 | } |
| 125 | if !isRetriable(lastErr, callOpts) { |
| 126 | return nil, lastErr |
| 127 | } |
| 128 | } |
| 129 | return nil, lastErr |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | // type serverStreamingRetryingStream is the implementation of grpc.ClientStream that acts as a |
| 134 | // proxy to the underlying call. If any of the RecvMsg() calls fail, it will try to reestablish |
| 135 | // a new ClientStream according to the retry policy. |
| 136 | type serverStreamingRetryingStream struct { |
| 137 | grpc.ClientStream |
| 138 | bufferedSends []interface{} // single message that the client can sen |
| 139 | receivedGood bool // indicates whether any prior receives were successful |
| 140 | wasClosedSend bool // indicates that CloseSend was closed |
| 141 | parentCtx context.Context |
| 142 | callOpts *options |
| 143 | streamerCall func(ctx context.Context) (grpc.ClientStream, error) |
| 144 | mu sync.RWMutex |
| 145 | } |
| 146 | |
| 147 | func (s *serverStreamingRetryingStream) setStream(clientStream grpc.ClientStream) { |
| 148 | s.mu.Lock() |
| 149 | s.ClientStream = clientStream |
| 150 | s.mu.Unlock() |
| 151 | } |
| 152 | |
| 153 | func (s *serverStreamingRetryingStream) getStream() grpc.ClientStream { |
| 154 | s.mu.RLock() |
| 155 | defer s.mu.RUnlock() |
| 156 | return s.ClientStream |
| 157 | } |
| 158 | |
| 159 | func (s *serverStreamingRetryingStream) SendMsg(m interface{}) error { |
| 160 | s.mu.Lock() |
| 161 | s.bufferedSends = append(s.bufferedSends, m) |
| 162 | s.mu.Unlock() |
| 163 | return s.getStream().SendMsg(m) |
| 164 | } |
| 165 | |
| 166 | func (s *serverStreamingRetryingStream) CloseSend() error { |
| 167 | s.mu.Lock() |
| 168 | s.wasClosedSend = true |
| 169 | s.mu.Unlock() |
| 170 | return s.getStream().CloseSend() |
| 171 | } |
| 172 | |
| 173 | func (s *serverStreamingRetryingStream) Header() (metadata.MD, error) { |
| 174 | return s.getStream().Header() |
| 175 | } |
| 176 | |
| 177 | func (s *serverStreamingRetryingStream) Trailer() metadata.MD { |
| 178 | return s.getStream().Trailer() |
| 179 | } |
| 180 | |
| 181 | func (s *serverStreamingRetryingStream) RecvMsg(m interface{}) error { |
| 182 | attemptRetry, lastErr := s.receiveMsgAndIndicateRetry(m) |
| 183 | if !attemptRetry { |
| 184 | return lastErr // success or hard failure |
| 185 | } |
| 186 | // We start off from attempt 1, because zeroth was already made on normal SendMsg(). |
| 187 | for attempt := uint(1); attempt < s.callOpts.max; attempt++ { |
| 188 | if err := waitRetryBackoff(attempt, s.parentCtx, s.callOpts); err != nil { |
| 189 | return err |
| 190 | } |
| 191 | callCtx := perCallContext(s.parentCtx, s.callOpts, attempt) |
| 192 | newStream, err := s.reestablishStreamAndResendBuffer(callCtx) |
| 193 | if err != nil { |
| 194 | // Retry dial and transport errors of establishing stream as grpc doesn't retry. |
| 195 | if isRetriable(err, s.callOpts) { |
| 196 | continue |
| 197 | } |
| 198 | return err |
| 199 | } |
| 200 | |
| 201 | s.setStream(newStream) |
| 202 | attemptRetry, lastErr = s.receiveMsgAndIndicateRetry(m) |
| 203 | //fmt.Printf("Received message and indicate: %v %v\n", attemptRetry, lastErr) |
| 204 | if !attemptRetry { |
| 205 | return lastErr |
| 206 | } |
| 207 | } |
| 208 | return lastErr |
| 209 | } |
| 210 | |
| 211 | func (s *serverStreamingRetryingStream) receiveMsgAndIndicateRetry(m interface{}) (bool, error) { |
| 212 | s.mu.RLock() |
| 213 | wasGood := s.receivedGood |
| 214 | s.mu.RUnlock() |
| 215 | err := s.getStream().RecvMsg(m) |
| 216 | if err == nil || err == io.EOF { |
| 217 | s.mu.Lock() |
| 218 | s.receivedGood = true |
| 219 | s.mu.Unlock() |
| 220 | return false, err |
| 221 | } else if wasGood { |
| 222 | // previous RecvMsg in the stream succeeded, no retry logic should interfere |
| 223 | return false, err |
| 224 | } |
| 225 | if isContextError(err) { |
| 226 | if s.parentCtx.Err() != nil { |
| 227 | logTrace(s.parentCtx, "grpc_retry parent context error: %v", s.parentCtx.Err()) |
| 228 | return false, err |
| 229 | } else if s.callOpts.perCallTimeout != 0 { |
| 230 | // We have set a perCallTimeout in the retry middleware, which would result in a context error if |
| 231 | // the deadline was exceeded, in which case try again. |
| 232 | logTrace(s.parentCtx, "grpc_retry context error from retry call") |
| 233 | return true, err |
| 234 | } |
| 235 | } |
| 236 | return isRetriable(err, s.callOpts), err |
| 237 | } |
| 238 | |
| 239 | func (s *serverStreamingRetryingStream) reestablishStreamAndResendBuffer( |
| 240 | callCtx context.Context, |
| 241 | ) (grpc.ClientStream, error) { |
| 242 | s.mu.RLock() |
| 243 | bufferedSends := s.bufferedSends |
| 244 | s.mu.RUnlock() |
| 245 | newStream, err := s.streamerCall(callCtx) |
| 246 | if err != nil { |
| 247 | logTrace(callCtx, "grpc_retry failed redialing new stream: %v", err) |
| 248 | return nil, err |
| 249 | } |
| 250 | for _, msg := range bufferedSends { |
| 251 | if err := newStream.SendMsg(msg); err != nil { |
| 252 | logTrace(callCtx, "grpc_retry failed resending message: %v", err) |
| 253 | return nil, err |
| 254 | } |
| 255 | } |
| 256 | if err := newStream.CloseSend(); err != nil { |
| 257 | logTrace(callCtx, "grpc_retry failed CloseSend on new stream %v", err) |
| 258 | return nil, err |
| 259 | } |
| 260 | return newStream, nil |
| 261 | } |
| 262 | |
| 263 | func waitRetryBackoff(attempt uint, parentCtx context.Context, callOpts *options) error { |
| 264 | var waitTime time.Duration = 0 |
| 265 | if attempt > 0 { |
| 266 | waitTime = callOpts.backoffFunc(parentCtx, attempt) |
| 267 | } |
| 268 | if waitTime > 0 { |
| 269 | logTrace(parentCtx, "grpc_retry attempt: %d, backoff for %v", attempt, waitTime) |
| 270 | timer := time.NewTimer(waitTime) |
| 271 | select { |
| 272 | case <-parentCtx.Done(): |
| 273 | timer.Stop() |
| 274 | return contextErrToGrpcErr(parentCtx.Err()) |
| 275 | case <-timer.C: |
| 276 | } |
| 277 | } |
| 278 | return nil |
| 279 | } |
| 280 | |
| 281 | func isRetriable(err error, callOpts *options) bool { |
| 282 | errCode := status.Code(err) |
| 283 | if isContextError(err) { |
| 284 | // context errors are not retriable based on user settings. |
| 285 | return false |
| 286 | } |
| 287 | for _, code := range callOpts.codes { |
| 288 | if code == errCode { |
| 289 | return true |
| 290 | } |
| 291 | } |
| 292 | return false |
| 293 | } |
| 294 | |
| 295 | func isContextError(err error) bool { |
| 296 | code := status.Code(err) |
| 297 | return code == codes.DeadlineExceeded || code == codes.Canceled |
| 298 | } |
| 299 | |
| 300 | func perCallContext(parentCtx context.Context, callOpts *options, attempt uint) context.Context { |
| 301 | ctx := parentCtx |
| 302 | if callOpts.perCallTimeout != 0 { |
| 303 | ctx, _ = context.WithTimeout(ctx, callOpts.perCallTimeout) |
| 304 | } |
| 305 | if attempt > 0 && callOpts.includeHeader { |
| 306 | mdClone := metautils.ExtractOutgoing(ctx).Clone().Set(AttemptMetadataKey, fmt.Sprintf("%d", attempt)) |
| 307 | ctx = mdClone.ToOutgoing(ctx) |
| 308 | } |
| 309 | return ctx |
| 310 | } |
| 311 | |
| 312 | func contextErrToGrpcErr(err error) error { |
| 313 | switch err { |
| 314 | case context.DeadlineExceeded: |
| 315 | return status.Error(codes.DeadlineExceeded, err.Error()) |
| 316 | case context.Canceled: |
| 317 | return status.Error(codes.Canceled, err.Error()) |
| 318 | default: |
| 319 | return status.Error(codes.Unknown, err.Error()) |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | func logTrace(ctx context.Context, format string, a ...interface{}) { |
| 324 | tr, ok := trace.FromContext(ctx) |
| 325 | if !ok { |
| 326 | return |
| 327 | } |
| 328 | tr.LazyPrintf(format, a...) |
| 329 | } |