Takahiro Suzuki | d7bf820 | 2020-12-17 20:21:59 +0900 | [diff] [blame] | 1 | package sarama |
| 2 | |
| 3 | import ( |
| 4 | "encoding/binary" |
| 5 | "fmt" |
| 6 | "sync" |
| 7 | "time" |
| 8 | |
| 9 | "github.com/eapache/go-resiliency/breaker" |
| 10 | "github.com/eapache/queue" |
| 11 | ) |
| 12 | |
| 13 | // AsyncProducer publishes Kafka messages using a non-blocking API. It routes messages |
| 14 | // to the correct broker for the provided topic-partition, refreshing metadata as appropriate, |
| 15 | // and parses responses for errors. You must read from the Errors() channel or the |
| 16 | // producer will deadlock. You must call Close() or AsyncClose() on a producer to avoid |
| 17 | // leaks: it will not be garbage-collected automatically when it passes out of |
| 18 | // scope. |
| 19 | type AsyncProducer interface { |
| 20 | |
| 21 | // AsyncClose triggers a shutdown of the producer. The shutdown has completed |
| 22 | // when both the Errors and Successes channels have been closed. When calling |
| 23 | // AsyncClose, you *must* continue to read from those channels in order to |
| 24 | // drain the results of any messages in flight. |
| 25 | AsyncClose() |
| 26 | |
| 27 | // Close shuts down the producer and waits for any buffered messages to be |
| 28 | // flushed. You must call this function before a producer object passes out of |
| 29 | // scope, as it may otherwise leak memory. You must call this before calling |
| 30 | // Close on the underlying client. |
| 31 | Close() error |
| 32 | |
| 33 | // Input is the input channel for the user to write messages to that they |
| 34 | // wish to send. |
| 35 | Input() chan<- *ProducerMessage |
| 36 | |
| 37 | // Successes is the success output channel back to the user when Return.Successes is |
| 38 | // enabled. If Return.Successes is true, you MUST read from this channel or the |
| 39 | // Producer will deadlock. It is suggested that you send and read messages |
| 40 | // together in a single select statement. |
| 41 | Successes() <-chan *ProducerMessage |
| 42 | |
| 43 | // Errors is the error output channel back to the user. You MUST read from this |
| 44 | // channel or the Producer will deadlock when the channel is full. Alternatively, |
| 45 | // you can set Producer.Return.Errors in your config to false, which prevents |
| 46 | // errors to be returned. |
| 47 | Errors() <-chan *ProducerError |
| 48 | } |
| 49 | |
| 50 | // transactionManager keeps the state necessary to ensure idempotent production |
| 51 | type transactionManager struct { |
| 52 | producerID int64 |
| 53 | producerEpoch int16 |
| 54 | sequenceNumbers map[string]int32 |
| 55 | mutex sync.Mutex |
| 56 | } |
| 57 | |
| 58 | const ( |
| 59 | noProducerID = -1 |
| 60 | noProducerEpoch = -1 |
| 61 | ) |
| 62 | |
| 63 | func (t *transactionManager) getAndIncrementSequenceNumber(topic string, partition int32) int32 { |
| 64 | key := fmt.Sprintf("%s-%d", topic, partition) |
| 65 | t.mutex.Lock() |
| 66 | defer t.mutex.Unlock() |
| 67 | sequence := t.sequenceNumbers[key] |
| 68 | t.sequenceNumbers[key] = sequence + 1 |
| 69 | return sequence |
| 70 | } |
| 71 | |
| 72 | func newTransactionManager(conf *Config, client Client) (*transactionManager, error) { |
| 73 | txnmgr := &transactionManager{ |
| 74 | producerID: noProducerID, |
| 75 | producerEpoch: noProducerEpoch, |
| 76 | } |
| 77 | |
| 78 | if conf.Producer.Idempotent { |
| 79 | initProducerIDResponse, err := client.InitProducerID() |
| 80 | if err != nil { |
| 81 | return nil, err |
| 82 | } |
| 83 | txnmgr.producerID = initProducerIDResponse.ProducerID |
| 84 | txnmgr.producerEpoch = initProducerIDResponse.ProducerEpoch |
| 85 | txnmgr.sequenceNumbers = make(map[string]int32) |
| 86 | txnmgr.mutex = sync.Mutex{} |
| 87 | |
| 88 | Logger.Printf("Obtained a ProducerId: %d and ProducerEpoch: %d\n", txnmgr.producerID, txnmgr.producerEpoch) |
| 89 | } |
| 90 | |
| 91 | return txnmgr, nil |
| 92 | } |
| 93 | |
| 94 | type asyncProducer struct { |
| 95 | client Client |
| 96 | conf *Config |
| 97 | |
| 98 | errors chan *ProducerError |
| 99 | input, successes, retries chan *ProducerMessage |
| 100 | inFlight sync.WaitGroup |
| 101 | |
| 102 | brokers map[*Broker]*brokerProducer |
| 103 | brokerRefs map[*brokerProducer]int |
| 104 | brokerLock sync.Mutex |
| 105 | |
| 106 | txnmgr *transactionManager |
| 107 | } |
| 108 | |
| 109 | // NewAsyncProducer creates a new AsyncProducer using the given broker addresses and configuration. |
| 110 | func NewAsyncProducer(addrs []string, conf *Config) (AsyncProducer, error) { |
| 111 | client, err := NewClient(addrs, conf) |
| 112 | if err != nil { |
| 113 | return nil, err |
| 114 | } |
| 115 | return newAsyncProducer(client) |
| 116 | } |
| 117 | |
| 118 | // NewAsyncProducerFromClient creates a new Producer using the given client. It is still |
| 119 | // necessary to call Close() on the underlying client when shutting down this producer. |
| 120 | func NewAsyncProducerFromClient(client Client) (AsyncProducer, error) { |
| 121 | // For clients passed in by the client, ensure we don't |
| 122 | // call Close() on it. |
| 123 | cli := &nopCloserClient{client} |
| 124 | return newAsyncProducer(cli) |
| 125 | } |
| 126 | |
| 127 | func newAsyncProducer(client Client) (AsyncProducer, error) { |
| 128 | // Check that we are not dealing with a closed Client before processing any other arguments |
| 129 | if client.Closed() { |
| 130 | return nil, ErrClosedClient |
| 131 | } |
| 132 | |
| 133 | txnmgr, err := newTransactionManager(client.Config(), client) |
| 134 | if err != nil { |
| 135 | return nil, err |
| 136 | } |
| 137 | |
| 138 | p := &asyncProducer{ |
| 139 | client: client, |
| 140 | conf: client.Config(), |
| 141 | errors: make(chan *ProducerError), |
| 142 | input: make(chan *ProducerMessage), |
| 143 | successes: make(chan *ProducerMessage), |
| 144 | retries: make(chan *ProducerMessage), |
| 145 | brokers: make(map[*Broker]*brokerProducer), |
| 146 | brokerRefs: make(map[*brokerProducer]int), |
| 147 | txnmgr: txnmgr, |
| 148 | } |
| 149 | |
| 150 | // launch our singleton dispatchers |
| 151 | go withRecover(p.dispatcher) |
| 152 | go withRecover(p.retryHandler) |
| 153 | |
| 154 | return p, nil |
| 155 | } |
| 156 | |
| 157 | type flagSet int8 |
| 158 | |
| 159 | const ( |
| 160 | syn flagSet = 1 << iota // first message from partitionProducer to brokerProducer |
| 161 | fin // final message from partitionProducer to brokerProducer and back |
| 162 | shutdown // start the shutdown process |
| 163 | ) |
| 164 | |
| 165 | // ProducerMessage is the collection of elements passed to the Producer in order to send a message. |
| 166 | type ProducerMessage struct { |
| 167 | Topic string // The Kafka topic for this message. |
| 168 | // The partitioning key for this message. Pre-existing Encoders include |
| 169 | // StringEncoder and ByteEncoder. |
| 170 | Key Encoder |
| 171 | // The actual message to store in Kafka. Pre-existing Encoders include |
| 172 | // StringEncoder and ByteEncoder. |
| 173 | Value Encoder |
| 174 | |
| 175 | // The headers are key-value pairs that are transparently passed |
| 176 | // by Kafka between producers and consumers. |
| 177 | Headers []RecordHeader |
| 178 | |
| 179 | // This field is used to hold arbitrary data you wish to include so it |
| 180 | // will be available when receiving on the Successes and Errors channels. |
| 181 | // Sarama completely ignores this field and is only to be used for |
| 182 | // pass-through data. |
| 183 | Metadata interface{} |
| 184 | |
| 185 | // Below this point are filled in by the producer as the message is processed |
| 186 | |
| 187 | // Offset is the offset of the message stored on the broker. This is only |
| 188 | // guaranteed to be defined if the message was successfully delivered and |
| 189 | // RequiredAcks is not NoResponse. |
| 190 | Offset int64 |
| 191 | // Partition is the partition that the message was sent to. This is only |
| 192 | // guaranteed to be defined if the message was successfully delivered. |
| 193 | Partition int32 |
| 194 | // Timestamp can vary in behaviour depending on broker configuration, being |
| 195 | // in either one of the CreateTime or LogAppendTime modes (default CreateTime), |
| 196 | // and requiring version at least 0.10.0. |
| 197 | // |
| 198 | // When configured to CreateTime, the timestamp is specified by the producer |
| 199 | // either by explicitly setting this field, or when the message is added |
| 200 | // to a produce set. |
| 201 | // |
| 202 | // When configured to LogAppendTime, the timestamp assigned to the message |
| 203 | // by the broker. This is only guaranteed to be defined if the message was |
| 204 | // successfully delivered and RequiredAcks is not NoResponse. |
| 205 | Timestamp time.Time |
| 206 | |
| 207 | retries int |
| 208 | flags flagSet |
| 209 | expectation chan *ProducerError |
| 210 | sequenceNumber int32 |
| 211 | } |
| 212 | |
| 213 | const producerMessageOverhead = 26 // the metadata overhead of CRC, flags, etc. |
| 214 | |
| 215 | func (m *ProducerMessage) byteSize(version int) int { |
| 216 | var size int |
| 217 | if version >= 2 { |
| 218 | size = maximumRecordOverhead |
| 219 | for _, h := range m.Headers { |
| 220 | size += len(h.Key) + len(h.Value) + 2*binary.MaxVarintLen32 |
| 221 | } |
| 222 | } else { |
| 223 | size = producerMessageOverhead |
| 224 | } |
| 225 | if m.Key != nil { |
| 226 | size += m.Key.Length() |
| 227 | } |
| 228 | if m.Value != nil { |
| 229 | size += m.Value.Length() |
| 230 | } |
| 231 | return size |
| 232 | } |
| 233 | |
| 234 | func (m *ProducerMessage) clear() { |
| 235 | m.flags = 0 |
| 236 | m.retries = 0 |
| 237 | } |
| 238 | |
| 239 | // ProducerError is the type of error generated when the producer fails to deliver a message. |
| 240 | // It contains the original ProducerMessage as well as the actual error value. |
| 241 | type ProducerError struct { |
| 242 | Msg *ProducerMessage |
| 243 | Err error |
| 244 | } |
| 245 | |
| 246 | func (pe ProducerError) Error() string { |
| 247 | return fmt.Sprintf("kafka: Failed to produce message to topic %s: %s", pe.Msg.Topic, pe.Err) |
| 248 | } |
| 249 | |
| 250 | // ProducerErrors is a type that wraps a batch of "ProducerError"s and implements the Error interface. |
| 251 | // It can be returned from the Producer's Close method to avoid the need to manually drain the Errors channel |
| 252 | // when closing a producer. |
| 253 | type ProducerErrors []*ProducerError |
| 254 | |
| 255 | func (pe ProducerErrors) Error() string { |
| 256 | return fmt.Sprintf("kafka: Failed to deliver %d messages.", len(pe)) |
| 257 | } |
| 258 | |
| 259 | func (p *asyncProducer) Errors() <-chan *ProducerError { |
| 260 | return p.errors |
| 261 | } |
| 262 | |
| 263 | func (p *asyncProducer) Successes() <-chan *ProducerMessage { |
| 264 | return p.successes |
| 265 | } |
| 266 | |
| 267 | func (p *asyncProducer) Input() chan<- *ProducerMessage { |
| 268 | return p.input |
| 269 | } |
| 270 | |
| 271 | func (p *asyncProducer) Close() error { |
| 272 | p.AsyncClose() |
| 273 | |
| 274 | if p.conf.Producer.Return.Successes { |
| 275 | go withRecover(func() { |
| 276 | for range p.successes { |
| 277 | } |
| 278 | }) |
| 279 | } |
| 280 | |
| 281 | var errors ProducerErrors |
| 282 | if p.conf.Producer.Return.Errors { |
| 283 | for event := range p.errors { |
| 284 | errors = append(errors, event) |
| 285 | } |
| 286 | } else { |
| 287 | <-p.errors |
| 288 | } |
| 289 | |
| 290 | if len(errors) > 0 { |
| 291 | return errors |
| 292 | } |
| 293 | return nil |
| 294 | } |
| 295 | |
| 296 | func (p *asyncProducer) AsyncClose() { |
| 297 | go withRecover(p.shutdown) |
| 298 | } |
| 299 | |
| 300 | // singleton |
| 301 | // dispatches messages by topic |
| 302 | func (p *asyncProducer) dispatcher() { |
| 303 | handlers := make(map[string]chan<- *ProducerMessage) |
| 304 | shuttingDown := false |
| 305 | |
| 306 | for msg := range p.input { |
| 307 | if msg == nil { |
| 308 | Logger.Println("Something tried to send a nil message, it was ignored.") |
| 309 | continue |
| 310 | } |
| 311 | |
| 312 | if msg.flags&shutdown != 0 { |
| 313 | shuttingDown = true |
| 314 | p.inFlight.Done() |
| 315 | continue |
| 316 | } else if msg.retries == 0 { |
| 317 | if shuttingDown { |
| 318 | // we can't just call returnError here because that decrements the wait group, |
| 319 | // which hasn't been incremented yet for this message, and shouldn't be |
| 320 | pErr := &ProducerError{Msg: msg, Err: ErrShuttingDown} |
| 321 | if p.conf.Producer.Return.Errors { |
| 322 | p.errors <- pErr |
| 323 | } else { |
| 324 | Logger.Println(pErr) |
| 325 | } |
| 326 | continue |
| 327 | } |
| 328 | p.inFlight.Add(1) |
| 329 | } |
| 330 | |
| 331 | version := 1 |
| 332 | if p.conf.Version.IsAtLeast(V0_11_0_0) { |
| 333 | version = 2 |
| 334 | } else if msg.Headers != nil { |
| 335 | p.returnError(msg, ConfigurationError("Producing headers requires Kafka at least v0.11")) |
| 336 | continue |
| 337 | } |
| 338 | if msg.byteSize(version) > p.conf.Producer.MaxMessageBytes { |
| 339 | p.returnError(msg, ErrMessageSizeTooLarge) |
| 340 | continue |
| 341 | } |
| 342 | |
| 343 | handler := handlers[msg.Topic] |
| 344 | if handler == nil { |
| 345 | handler = p.newTopicProducer(msg.Topic) |
| 346 | handlers[msg.Topic] = handler |
| 347 | } |
| 348 | |
| 349 | handler <- msg |
| 350 | } |
| 351 | |
| 352 | for _, handler := range handlers { |
| 353 | close(handler) |
| 354 | } |
| 355 | } |
| 356 | |
| 357 | // one per topic |
| 358 | // partitions messages, then dispatches them by partition |
| 359 | type topicProducer struct { |
| 360 | parent *asyncProducer |
| 361 | topic string |
| 362 | input <-chan *ProducerMessage |
| 363 | |
| 364 | breaker *breaker.Breaker |
| 365 | handlers map[int32]chan<- *ProducerMessage |
| 366 | partitioner Partitioner |
| 367 | } |
| 368 | |
| 369 | func (p *asyncProducer) newTopicProducer(topic string) chan<- *ProducerMessage { |
| 370 | input := make(chan *ProducerMessage, p.conf.ChannelBufferSize) |
| 371 | tp := &topicProducer{ |
| 372 | parent: p, |
| 373 | topic: topic, |
| 374 | input: input, |
| 375 | breaker: breaker.New(3, 1, 10*time.Second), |
| 376 | handlers: make(map[int32]chan<- *ProducerMessage), |
| 377 | partitioner: p.conf.Producer.Partitioner(topic), |
| 378 | } |
| 379 | go withRecover(tp.dispatch) |
| 380 | return input |
| 381 | } |
| 382 | |
| 383 | func (tp *topicProducer) dispatch() { |
| 384 | for msg := range tp.input { |
| 385 | if msg.retries == 0 { |
| 386 | if err := tp.partitionMessage(msg); err != nil { |
| 387 | tp.parent.returnError(msg, err) |
| 388 | continue |
| 389 | } |
| 390 | } |
| 391 | // All messages being retried (sent or not) have already had their retry count updated |
| 392 | if tp.parent.conf.Producer.Idempotent && msg.retries == 0 { |
| 393 | msg.sequenceNumber = tp.parent.txnmgr.getAndIncrementSequenceNumber(msg.Topic, msg.Partition) |
| 394 | } |
| 395 | |
| 396 | handler := tp.handlers[msg.Partition] |
| 397 | if handler == nil { |
| 398 | handler = tp.parent.newPartitionProducer(msg.Topic, msg.Partition) |
| 399 | tp.handlers[msg.Partition] = handler |
| 400 | } |
| 401 | |
| 402 | handler <- msg |
| 403 | } |
| 404 | |
| 405 | for _, handler := range tp.handlers { |
| 406 | close(handler) |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | func (tp *topicProducer) partitionMessage(msg *ProducerMessage) error { |
| 411 | var partitions []int32 |
| 412 | |
| 413 | err := tp.breaker.Run(func() (err error) { |
| 414 | var requiresConsistency = false |
| 415 | if ep, ok := tp.partitioner.(DynamicConsistencyPartitioner); ok { |
| 416 | requiresConsistency = ep.MessageRequiresConsistency(msg) |
| 417 | } else { |
| 418 | requiresConsistency = tp.partitioner.RequiresConsistency() |
| 419 | } |
| 420 | |
| 421 | if requiresConsistency { |
| 422 | partitions, err = tp.parent.client.Partitions(msg.Topic) |
| 423 | } else { |
| 424 | partitions, err = tp.parent.client.WritablePartitions(msg.Topic) |
| 425 | } |
| 426 | return |
| 427 | }) |
| 428 | |
| 429 | if err != nil { |
| 430 | return err |
| 431 | } |
| 432 | |
| 433 | numPartitions := int32(len(partitions)) |
| 434 | |
| 435 | if numPartitions == 0 { |
| 436 | return ErrLeaderNotAvailable |
| 437 | } |
| 438 | |
| 439 | choice, err := tp.partitioner.Partition(msg, numPartitions) |
| 440 | |
| 441 | if err != nil { |
| 442 | return err |
| 443 | } else if choice < 0 || choice >= numPartitions { |
| 444 | return ErrInvalidPartition |
| 445 | } |
| 446 | |
| 447 | msg.Partition = partitions[choice] |
| 448 | |
| 449 | return nil |
| 450 | } |
| 451 | |
| 452 | // one per partition per topic |
| 453 | // dispatches messages to the appropriate broker |
| 454 | // also responsible for maintaining message order during retries |
| 455 | type partitionProducer struct { |
| 456 | parent *asyncProducer |
| 457 | topic string |
| 458 | partition int32 |
| 459 | input <-chan *ProducerMessage |
| 460 | |
| 461 | leader *Broker |
| 462 | breaker *breaker.Breaker |
| 463 | brokerProducer *brokerProducer |
| 464 | |
| 465 | // highWatermark tracks the "current" retry level, which is the only one where we actually let messages through, |
| 466 | // all other messages get buffered in retryState[msg.retries].buf to preserve ordering |
| 467 | // retryState[msg.retries].expectChaser simply tracks whether we've seen a fin message for a given level (and |
| 468 | // therefore whether our buffer is complete and safe to flush) |
| 469 | highWatermark int |
| 470 | retryState []partitionRetryState |
| 471 | } |
| 472 | |
| 473 | type partitionRetryState struct { |
| 474 | buf []*ProducerMessage |
| 475 | expectChaser bool |
| 476 | } |
| 477 | |
| 478 | func (p *asyncProducer) newPartitionProducer(topic string, partition int32) chan<- *ProducerMessage { |
| 479 | input := make(chan *ProducerMessage, p.conf.ChannelBufferSize) |
| 480 | pp := &partitionProducer{ |
| 481 | parent: p, |
| 482 | topic: topic, |
| 483 | partition: partition, |
| 484 | input: input, |
| 485 | |
| 486 | breaker: breaker.New(3, 1, 10*time.Second), |
| 487 | retryState: make([]partitionRetryState, p.conf.Producer.Retry.Max+1), |
| 488 | } |
| 489 | go withRecover(pp.dispatch) |
| 490 | return input |
| 491 | } |
| 492 | |
| 493 | func (pp *partitionProducer) backoff(retries int) { |
| 494 | var backoff time.Duration |
| 495 | if pp.parent.conf.Producer.Retry.BackoffFunc != nil { |
| 496 | maxRetries := pp.parent.conf.Producer.Retry.Max |
| 497 | backoff = pp.parent.conf.Producer.Retry.BackoffFunc(retries, maxRetries) |
| 498 | } else { |
| 499 | backoff = pp.parent.conf.Producer.Retry.Backoff |
| 500 | } |
| 501 | if backoff > 0 { |
| 502 | time.Sleep(backoff) |
| 503 | } |
| 504 | } |
| 505 | |
| 506 | func (pp *partitionProducer) dispatch() { |
| 507 | // try to prefetch the leader; if this doesn't work, we'll do a proper call to `updateLeader` |
| 508 | // on the first message |
| 509 | pp.leader, _ = pp.parent.client.Leader(pp.topic, pp.partition) |
| 510 | if pp.leader != nil { |
| 511 | pp.brokerProducer = pp.parent.getBrokerProducer(pp.leader) |
| 512 | pp.parent.inFlight.Add(1) // we're generating a syn message; track it so we don't shut down while it's still inflight |
| 513 | pp.brokerProducer.input <- &ProducerMessage{Topic: pp.topic, Partition: pp.partition, flags: syn} |
| 514 | } |
| 515 | |
| 516 | defer func() { |
| 517 | if pp.brokerProducer != nil { |
| 518 | pp.parent.unrefBrokerProducer(pp.leader, pp.brokerProducer) |
| 519 | } |
| 520 | }() |
| 521 | |
| 522 | for msg := range pp.input { |
| 523 | |
| 524 | if pp.brokerProducer != nil && pp.brokerProducer.abandoned != nil { |
| 525 | select { |
| 526 | case <-pp.brokerProducer.abandoned: |
| 527 | // a message on the abandoned channel means that our current broker selection is out of date |
| 528 | Logger.Printf("producer/leader/%s/%d abandoning broker %d\n", pp.topic, pp.partition, pp.leader.ID()) |
| 529 | pp.parent.unrefBrokerProducer(pp.leader, pp.brokerProducer) |
| 530 | pp.brokerProducer = nil |
| 531 | time.Sleep(pp.parent.conf.Producer.Retry.Backoff) |
| 532 | default: |
| 533 | // producer connection is still open. |
| 534 | } |
| 535 | } |
| 536 | |
| 537 | if msg.retries > pp.highWatermark { |
| 538 | // a new, higher, retry level; handle it and then back off |
| 539 | pp.newHighWatermark(msg.retries) |
| 540 | pp.backoff(msg.retries) |
| 541 | } else if pp.highWatermark > 0 { |
| 542 | // we are retrying something (else highWatermark would be 0) but this message is not a *new* retry level |
| 543 | if msg.retries < pp.highWatermark { |
| 544 | // in fact this message is not even the current retry level, so buffer it for now (unless it's a just a fin) |
| 545 | if msg.flags&fin == fin { |
| 546 | pp.retryState[msg.retries].expectChaser = false |
| 547 | pp.parent.inFlight.Done() // this fin is now handled and will be garbage collected |
| 548 | } else { |
| 549 | pp.retryState[msg.retries].buf = append(pp.retryState[msg.retries].buf, msg) |
| 550 | } |
| 551 | continue |
| 552 | } else if msg.flags&fin == fin { |
| 553 | // this message is of the current retry level (msg.retries == highWatermark) and the fin flag is set, |
| 554 | // meaning this retry level is done and we can go down (at least) one level and flush that |
| 555 | pp.retryState[pp.highWatermark].expectChaser = false |
| 556 | pp.flushRetryBuffers() |
| 557 | pp.parent.inFlight.Done() // this fin is now handled and will be garbage collected |
| 558 | continue |
| 559 | } |
| 560 | } |
| 561 | |
| 562 | // if we made it this far then the current msg contains real data, and can be sent to the next goroutine |
| 563 | // without breaking any of our ordering guarantees |
| 564 | |
| 565 | if pp.brokerProducer == nil { |
| 566 | if err := pp.updateLeader(); err != nil { |
| 567 | pp.parent.returnError(msg, err) |
| 568 | pp.backoff(msg.retries) |
| 569 | continue |
| 570 | } |
| 571 | Logger.Printf("producer/leader/%s/%d selected broker %d\n", pp.topic, pp.partition, pp.leader.ID()) |
| 572 | } |
| 573 | |
| 574 | pp.brokerProducer.input <- msg |
| 575 | } |
| 576 | } |
| 577 | |
| 578 | func (pp *partitionProducer) newHighWatermark(hwm int) { |
| 579 | Logger.Printf("producer/leader/%s/%d state change to [retrying-%d]\n", pp.topic, pp.partition, hwm) |
| 580 | pp.highWatermark = hwm |
| 581 | |
| 582 | // send off a fin so that we know when everything "in between" has made it |
| 583 | // back to us and we can safely flush the backlog (otherwise we risk re-ordering messages) |
| 584 | pp.retryState[pp.highWatermark].expectChaser = true |
| 585 | pp.parent.inFlight.Add(1) // we're generating a fin message; track it so we don't shut down while it's still inflight |
| 586 | pp.brokerProducer.input <- &ProducerMessage{Topic: pp.topic, Partition: pp.partition, flags: fin, retries: pp.highWatermark - 1} |
| 587 | |
| 588 | // a new HWM means that our current broker selection is out of date |
| 589 | Logger.Printf("producer/leader/%s/%d abandoning broker %d\n", pp.topic, pp.partition, pp.leader.ID()) |
| 590 | pp.parent.unrefBrokerProducer(pp.leader, pp.brokerProducer) |
| 591 | pp.brokerProducer = nil |
| 592 | } |
| 593 | |
| 594 | func (pp *partitionProducer) flushRetryBuffers() { |
| 595 | Logger.Printf("producer/leader/%s/%d state change to [flushing-%d]\n", pp.topic, pp.partition, pp.highWatermark) |
| 596 | for { |
| 597 | pp.highWatermark-- |
| 598 | |
| 599 | if pp.brokerProducer == nil { |
| 600 | if err := pp.updateLeader(); err != nil { |
| 601 | pp.parent.returnErrors(pp.retryState[pp.highWatermark].buf, err) |
| 602 | goto flushDone |
| 603 | } |
| 604 | Logger.Printf("producer/leader/%s/%d selected broker %d\n", pp.topic, pp.partition, pp.leader.ID()) |
| 605 | } |
| 606 | |
| 607 | for _, msg := range pp.retryState[pp.highWatermark].buf { |
| 608 | pp.brokerProducer.input <- msg |
| 609 | } |
| 610 | |
| 611 | flushDone: |
| 612 | pp.retryState[pp.highWatermark].buf = nil |
| 613 | if pp.retryState[pp.highWatermark].expectChaser { |
| 614 | Logger.Printf("producer/leader/%s/%d state change to [retrying-%d]\n", pp.topic, pp.partition, pp.highWatermark) |
| 615 | break |
| 616 | } else if pp.highWatermark == 0 { |
| 617 | Logger.Printf("producer/leader/%s/%d state change to [normal]\n", pp.topic, pp.partition) |
| 618 | break |
| 619 | } |
| 620 | } |
| 621 | } |
| 622 | |
| 623 | func (pp *partitionProducer) updateLeader() error { |
| 624 | return pp.breaker.Run(func() (err error) { |
| 625 | if err = pp.parent.client.RefreshMetadata(pp.topic); err != nil { |
| 626 | return err |
| 627 | } |
| 628 | |
| 629 | if pp.leader, err = pp.parent.client.Leader(pp.topic, pp.partition); err != nil { |
| 630 | return err |
| 631 | } |
| 632 | |
| 633 | pp.brokerProducer = pp.parent.getBrokerProducer(pp.leader) |
| 634 | pp.parent.inFlight.Add(1) // we're generating a syn message; track it so we don't shut down while it's still inflight |
| 635 | pp.brokerProducer.input <- &ProducerMessage{Topic: pp.topic, Partition: pp.partition, flags: syn} |
| 636 | |
| 637 | return nil |
| 638 | }) |
| 639 | } |
| 640 | |
| 641 | // one per broker; also constructs an associated flusher |
| 642 | func (p *asyncProducer) newBrokerProducer(broker *Broker) *brokerProducer { |
| 643 | var ( |
| 644 | input = make(chan *ProducerMessage) |
| 645 | bridge = make(chan *produceSet) |
| 646 | responses = make(chan *brokerProducerResponse) |
| 647 | ) |
| 648 | |
| 649 | bp := &brokerProducer{ |
| 650 | parent: p, |
| 651 | broker: broker, |
| 652 | input: input, |
| 653 | output: bridge, |
| 654 | responses: responses, |
| 655 | buffer: newProduceSet(p), |
| 656 | currentRetries: make(map[string]map[int32]error), |
| 657 | } |
| 658 | go withRecover(bp.run) |
| 659 | |
| 660 | // minimal bridge to make the network response `select`able |
| 661 | go withRecover(func() { |
| 662 | for set := range bridge { |
| 663 | request := set.buildRequest() |
| 664 | |
| 665 | response, err := broker.Produce(request) |
| 666 | |
| 667 | responses <- &brokerProducerResponse{ |
| 668 | set: set, |
| 669 | err: err, |
| 670 | res: response, |
| 671 | } |
| 672 | } |
| 673 | close(responses) |
| 674 | }) |
| 675 | |
| 676 | if p.conf.Producer.Retry.Max <= 0 { |
| 677 | bp.abandoned = make(chan struct{}) |
| 678 | } |
| 679 | |
| 680 | return bp |
| 681 | } |
| 682 | |
| 683 | type brokerProducerResponse struct { |
| 684 | set *produceSet |
| 685 | err error |
| 686 | res *ProduceResponse |
| 687 | } |
| 688 | |
| 689 | // groups messages together into appropriately-sized batches for sending to the broker |
| 690 | // handles state related to retries etc |
| 691 | type brokerProducer struct { |
| 692 | parent *asyncProducer |
| 693 | broker *Broker |
| 694 | |
| 695 | input chan *ProducerMessage |
| 696 | output chan<- *produceSet |
| 697 | responses <-chan *brokerProducerResponse |
| 698 | abandoned chan struct{} |
| 699 | |
| 700 | buffer *produceSet |
| 701 | timer <-chan time.Time |
| 702 | timerFired bool |
| 703 | |
| 704 | closing error |
| 705 | currentRetries map[string]map[int32]error |
| 706 | } |
| 707 | |
| 708 | func (bp *brokerProducer) run() { |
| 709 | var output chan<- *produceSet |
| 710 | Logger.Printf("producer/broker/%d starting up\n", bp.broker.ID()) |
| 711 | |
| 712 | for { |
| 713 | select { |
| 714 | case msg := <-bp.input: |
| 715 | if msg == nil { |
| 716 | bp.shutdown() |
| 717 | return |
| 718 | } |
| 719 | |
| 720 | if msg.flags&syn == syn { |
| 721 | Logger.Printf("producer/broker/%d state change to [open] on %s/%d\n", |
| 722 | bp.broker.ID(), msg.Topic, msg.Partition) |
| 723 | if bp.currentRetries[msg.Topic] == nil { |
| 724 | bp.currentRetries[msg.Topic] = make(map[int32]error) |
| 725 | } |
| 726 | bp.currentRetries[msg.Topic][msg.Partition] = nil |
| 727 | bp.parent.inFlight.Done() |
| 728 | continue |
| 729 | } |
| 730 | |
| 731 | if reason := bp.needsRetry(msg); reason != nil { |
| 732 | bp.parent.retryMessage(msg, reason) |
| 733 | |
| 734 | if bp.closing == nil && msg.flags&fin == fin { |
| 735 | // we were retrying this partition but we can start processing again |
| 736 | delete(bp.currentRetries[msg.Topic], msg.Partition) |
| 737 | Logger.Printf("producer/broker/%d state change to [closed] on %s/%d\n", |
| 738 | bp.broker.ID(), msg.Topic, msg.Partition) |
| 739 | } |
| 740 | |
| 741 | continue |
| 742 | } |
| 743 | |
| 744 | if bp.buffer.wouldOverflow(msg) { |
| 745 | if err := bp.waitForSpace(msg); err != nil { |
| 746 | bp.parent.retryMessage(msg, err) |
| 747 | continue |
| 748 | } |
| 749 | } |
| 750 | |
| 751 | if err := bp.buffer.add(msg); err != nil { |
| 752 | bp.parent.returnError(msg, err) |
| 753 | continue |
| 754 | } |
| 755 | |
| 756 | if bp.parent.conf.Producer.Flush.Frequency > 0 && bp.timer == nil { |
| 757 | bp.timer = time.After(bp.parent.conf.Producer.Flush.Frequency) |
| 758 | } |
| 759 | case <-bp.timer: |
| 760 | bp.timerFired = true |
| 761 | case output <- bp.buffer: |
| 762 | bp.rollOver() |
| 763 | case response := <-bp.responses: |
| 764 | bp.handleResponse(response) |
| 765 | } |
| 766 | |
| 767 | if bp.timerFired || bp.buffer.readyToFlush() { |
| 768 | output = bp.output |
| 769 | } else { |
| 770 | output = nil |
| 771 | } |
| 772 | } |
| 773 | } |
| 774 | |
| 775 | func (bp *brokerProducer) shutdown() { |
| 776 | for !bp.buffer.empty() { |
| 777 | select { |
| 778 | case response := <-bp.responses: |
| 779 | bp.handleResponse(response) |
| 780 | case bp.output <- bp.buffer: |
| 781 | bp.rollOver() |
| 782 | } |
| 783 | } |
| 784 | close(bp.output) |
| 785 | for response := range bp.responses { |
| 786 | bp.handleResponse(response) |
| 787 | } |
| 788 | |
| 789 | Logger.Printf("producer/broker/%d shut down\n", bp.broker.ID()) |
| 790 | } |
| 791 | |
| 792 | func (bp *brokerProducer) needsRetry(msg *ProducerMessage) error { |
| 793 | if bp.closing != nil { |
| 794 | return bp.closing |
| 795 | } |
| 796 | |
| 797 | return bp.currentRetries[msg.Topic][msg.Partition] |
| 798 | } |
| 799 | |
| 800 | func (bp *brokerProducer) waitForSpace(msg *ProducerMessage) error { |
| 801 | Logger.Printf("producer/broker/%d maximum request accumulated, waiting for space\n", bp.broker.ID()) |
| 802 | |
| 803 | for { |
| 804 | select { |
| 805 | case response := <-bp.responses: |
| 806 | bp.handleResponse(response) |
| 807 | // handling a response can change our state, so re-check some things |
| 808 | if reason := bp.needsRetry(msg); reason != nil { |
| 809 | return reason |
| 810 | } else if !bp.buffer.wouldOverflow(msg) { |
| 811 | return nil |
| 812 | } |
| 813 | case bp.output <- bp.buffer: |
| 814 | bp.rollOver() |
| 815 | return nil |
| 816 | } |
| 817 | } |
| 818 | } |
| 819 | |
| 820 | func (bp *brokerProducer) rollOver() { |
| 821 | bp.timer = nil |
| 822 | bp.timerFired = false |
| 823 | bp.buffer = newProduceSet(bp.parent) |
| 824 | } |
| 825 | |
| 826 | func (bp *brokerProducer) handleResponse(response *brokerProducerResponse) { |
| 827 | if response.err != nil { |
| 828 | bp.handleError(response.set, response.err) |
| 829 | } else { |
| 830 | bp.handleSuccess(response.set, response.res) |
| 831 | } |
| 832 | |
| 833 | if bp.buffer.empty() { |
| 834 | bp.rollOver() // this can happen if the response invalidated our buffer |
| 835 | } |
| 836 | } |
| 837 | |
| 838 | func (bp *brokerProducer) handleSuccess(sent *produceSet, response *ProduceResponse) { |
| 839 | // we iterate through the blocks in the request set, not the response, so that we notice |
| 840 | // if the response is missing a block completely |
| 841 | var retryTopics []string |
| 842 | sent.eachPartition(func(topic string, partition int32, pSet *partitionSet) { |
| 843 | if response == nil { |
| 844 | // this only happens when RequiredAcks is NoResponse, so we have to assume success |
| 845 | bp.parent.returnSuccesses(pSet.msgs) |
| 846 | return |
| 847 | } |
| 848 | |
| 849 | block := response.GetBlock(topic, partition) |
| 850 | if block == nil { |
| 851 | bp.parent.returnErrors(pSet.msgs, ErrIncompleteResponse) |
| 852 | return |
| 853 | } |
| 854 | |
| 855 | switch block.Err { |
| 856 | // Success |
| 857 | case ErrNoError: |
| 858 | if bp.parent.conf.Version.IsAtLeast(V0_10_0_0) && !block.Timestamp.IsZero() { |
| 859 | for _, msg := range pSet.msgs { |
| 860 | msg.Timestamp = block.Timestamp |
| 861 | } |
| 862 | } |
| 863 | for i, msg := range pSet.msgs { |
| 864 | msg.Offset = block.Offset + int64(i) |
| 865 | } |
| 866 | bp.parent.returnSuccesses(pSet.msgs) |
| 867 | // Duplicate |
| 868 | case ErrDuplicateSequenceNumber: |
| 869 | bp.parent.returnSuccesses(pSet.msgs) |
| 870 | // Retriable errors |
| 871 | case ErrInvalidMessage, ErrUnknownTopicOrPartition, ErrLeaderNotAvailable, ErrNotLeaderForPartition, |
| 872 | ErrRequestTimedOut, ErrNotEnoughReplicas, ErrNotEnoughReplicasAfterAppend: |
| 873 | if bp.parent.conf.Producer.Retry.Max <= 0 { |
| 874 | bp.parent.abandonBrokerConnection(bp.broker) |
| 875 | bp.parent.returnErrors(pSet.msgs, block.Err) |
| 876 | } else { |
| 877 | retryTopics = append(retryTopics, topic) |
| 878 | } |
| 879 | // Other non-retriable errors |
| 880 | default: |
| 881 | if bp.parent.conf.Producer.Retry.Max <= 0 { |
| 882 | bp.parent.abandonBrokerConnection(bp.broker) |
| 883 | } |
| 884 | bp.parent.returnErrors(pSet.msgs, block.Err) |
| 885 | } |
| 886 | }) |
| 887 | |
| 888 | if len(retryTopics) > 0 { |
| 889 | if bp.parent.conf.Producer.Idempotent { |
| 890 | err := bp.parent.client.RefreshMetadata(retryTopics...) |
| 891 | if err != nil { |
| 892 | Logger.Printf("Failed refreshing metadata because of %v\n", err) |
| 893 | } |
| 894 | } |
| 895 | |
| 896 | sent.eachPartition(func(topic string, partition int32, pSet *partitionSet) { |
| 897 | block := response.GetBlock(topic, partition) |
| 898 | if block == nil { |
| 899 | // handled in the previous "eachPartition" loop |
| 900 | return |
| 901 | } |
| 902 | |
| 903 | switch block.Err { |
| 904 | case ErrInvalidMessage, ErrUnknownTopicOrPartition, ErrLeaderNotAvailable, ErrNotLeaderForPartition, |
| 905 | ErrRequestTimedOut, ErrNotEnoughReplicas, ErrNotEnoughReplicasAfterAppend: |
| 906 | Logger.Printf("producer/broker/%d state change to [retrying] on %s/%d because %v\n", |
| 907 | bp.broker.ID(), topic, partition, block.Err) |
| 908 | if bp.currentRetries[topic] == nil { |
| 909 | bp.currentRetries[topic] = make(map[int32]error) |
| 910 | } |
| 911 | bp.currentRetries[topic][partition] = block.Err |
| 912 | if bp.parent.conf.Producer.Idempotent { |
| 913 | go bp.parent.retryBatch(topic, partition, pSet, block.Err) |
| 914 | } else { |
| 915 | bp.parent.retryMessages(pSet.msgs, block.Err) |
| 916 | } |
| 917 | // dropping the following messages has the side effect of incrementing their retry count |
| 918 | bp.parent.retryMessages(bp.buffer.dropPartition(topic, partition), block.Err) |
| 919 | } |
| 920 | }) |
| 921 | } |
| 922 | } |
| 923 | |
| 924 | func (p *asyncProducer) retryBatch(topic string, partition int32, pSet *partitionSet, kerr KError) { |
| 925 | Logger.Printf("Retrying batch for %v-%d because of %s\n", topic, partition, kerr) |
| 926 | produceSet := newProduceSet(p) |
| 927 | produceSet.msgs[topic] = make(map[int32]*partitionSet) |
| 928 | produceSet.msgs[topic][partition] = pSet |
| 929 | produceSet.bufferBytes += pSet.bufferBytes |
| 930 | produceSet.bufferCount += len(pSet.msgs) |
| 931 | for _, msg := range pSet.msgs { |
| 932 | if msg.retries >= p.conf.Producer.Retry.Max { |
| 933 | p.returnError(msg, kerr) |
| 934 | return |
| 935 | } |
| 936 | msg.retries++ |
| 937 | } |
| 938 | |
| 939 | // it's expected that a metadata refresh has been requested prior to calling retryBatch |
| 940 | leader, err := p.client.Leader(topic, partition) |
| 941 | if err != nil { |
| 942 | Logger.Printf("Failed retrying batch for %v-%d because of %v while looking up for new leader\n", topic, partition, err) |
| 943 | for _, msg := range pSet.msgs { |
| 944 | p.returnError(msg, kerr) |
| 945 | } |
| 946 | return |
| 947 | } |
| 948 | bp := p.getBrokerProducer(leader) |
| 949 | bp.output <- produceSet |
| 950 | } |
| 951 | |
| 952 | func (bp *brokerProducer) handleError(sent *produceSet, err error) { |
| 953 | switch err.(type) { |
| 954 | case PacketEncodingError: |
| 955 | sent.eachPartition(func(topic string, partition int32, pSet *partitionSet) { |
| 956 | bp.parent.returnErrors(pSet.msgs, err) |
| 957 | }) |
| 958 | default: |
| 959 | Logger.Printf("producer/broker/%d state change to [closing] because %s\n", bp.broker.ID(), err) |
| 960 | bp.parent.abandonBrokerConnection(bp.broker) |
| 961 | _ = bp.broker.Close() |
| 962 | bp.closing = err |
| 963 | sent.eachPartition(func(topic string, partition int32, pSet *partitionSet) { |
| 964 | bp.parent.retryMessages(pSet.msgs, err) |
| 965 | }) |
| 966 | bp.buffer.eachPartition(func(topic string, partition int32, pSet *partitionSet) { |
| 967 | bp.parent.retryMessages(pSet.msgs, err) |
| 968 | }) |
| 969 | bp.rollOver() |
| 970 | } |
| 971 | } |
| 972 | |
| 973 | // singleton |
| 974 | // effectively a "bridge" between the flushers and the dispatcher in order to avoid deadlock |
| 975 | // based on https://godoc.org/github.com/eapache/channels#InfiniteChannel |
| 976 | func (p *asyncProducer) retryHandler() { |
| 977 | var msg *ProducerMessage |
| 978 | buf := queue.New() |
| 979 | |
| 980 | for { |
| 981 | if buf.Length() == 0 { |
| 982 | msg = <-p.retries |
| 983 | } else { |
| 984 | select { |
| 985 | case msg = <-p.retries: |
| 986 | case p.input <- buf.Peek().(*ProducerMessage): |
| 987 | buf.Remove() |
| 988 | continue |
| 989 | } |
| 990 | } |
| 991 | |
| 992 | if msg == nil { |
| 993 | return |
| 994 | } |
| 995 | |
| 996 | buf.Add(msg) |
| 997 | } |
| 998 | } |
| 999 | |
| 1000 | // utility functions |
| 1001 | |
| 1002 | func (p *asyncProducer) shutdown() { |
| 1003 | Logger.Println("Producer shutting down.") |
| 1004 | p.inFlight.Add(1) |
| 1005 | p.input <- &ProducerMessage{flags: shutdown} |
| 1006 | |
| 1007 | p.inFlight.Wait() |
| 1008 | |
| 1009 | err := p.client.Close() |
| 1010 | if err != nil { |
| 1011 | Logger.Println("producer/shutdown failed to close the embedded client:", err) |
| 1012 | } |
| 1013 | |
| 1014 | close(p.input) |
| 1015 | close(p.retries) |
| 1016 | close(p.errors) |
| 1017 | close(p.successes) |
| 1018 | } |
| 1019 | |
| 1020 | func (p *asyncProducer) returnError(msg *ProducerMessage, err error) { |
| 1021 | msg.clear() |
| 1022 | pErr := &ProducerError{Msg: msg, Err: err} |
| 1023 | if p.conf.Producer.Return.Errors { |
| 1024 | p.errors <- pErr |
| 1025 | } else { |
| 1026 | Logger.Println(pErr) |
| 1027 | } |
| 1028 | p.inFlight.Done() |
| 1029 | } |
| 1030 | |
| 1031 | func (p *asyncProducer) returnErrors(batch []*ProducerMessage, err error) { |
| 1032 | for _, msg := range batch { |
| 1033 | p.returnError(msg, err) |
| 1034 | } |
| 1035 | } |
| 1036 | |
| 1037 | func (p *asyncProducer) returnSuccesses(batch []*ProducerMessage) { |
| 1038 | for _, msg := range batch { |
| 1039 | if p.conf.Producer.Return.Successes { |
| 1040 | msg.clear() |
| 1041 | p.successes <- msg |
| 1042 | } |
| 1043 | p.inFlight.Done() |
| 1044 | } |
| 1045 | } |
| 1046 | |
| 1047 | func (p *asyncProducer) retryMessage(msg *ProducerMessage, err error) { |
| 1048 | if msg.retries >= p.conf.Producer.Retry.Max { |
| 1049 | p.returnError(msg, err) |
| 1050 | } else { |
| 1051 | msg.retries++ |
| 1052 | p.retries <- msg |
| 1053 | } |
| 1054 | } |
| 1055 | |
| 1056 | func (p *asyncProducer) retryMessages(batch []*ProducerMessage, err error) { |
| 1057 | for _, msg := range batch { |
| 1058 | p.retryMessage(msg, err) |
| 1059 | } |
| 1060 | } |
| 1061 | |
| 1062 | func (p *asyncProducer) getBrokerProducer(broker *Broker) *brokerProducer { |
| 1063 | p.brokerLock.Lock() |
| 1064 | defer p.brokerLock.Unlock() |
| 1065 | |
| 1066 | bp := p.brokers[broker] |
| 1067 | |
| 1068 | if bp == nil { |
| 1069 | bp = p.newBrokerProducer(broker) |
| 1070 | p.brokers[broker] = bp |
| 1071 | p.brokerRefs[bp] = 0 |
| 1072 | } |
| 1073 | |
| 1074 | p.brokerRefs[bp]++ |
| 1075 | |
| 1076 | return bp |
| 1077 | } |
| 1078 | |
| 1079 | func (p *asyncProducer) unrefBrokerProducer(broker *Broker, bp *brokerProducer) { |
| 1080 | p.brokerLock.Lock() |
| 1081 | defer p.brokerLock.Unlock() |
| 1082 | |
| 1083 | p.brokerRefs[bp]-- |
| 1084 | if p.brokerRefs[bp] == 0 { |
| 1085 | close(bp.input) |
| 1086 | delete(p.brokerRefs, bp) |
| 1087 | |
| 1088 | if p.brokers[broker] == bp { |
| 1089 | delete(p.brokers, broker) |
| 1090 | } |
| 1091 | } |
| 1092 | } |
| 1093 | |
| 1094 | func (p *asyncProducer) abandonBrokerConnection(broker *Broker) { |
| 1095 | p.brokerLock.Lock() |
| 1096 | defer p.brokerLock.Unlock() |
| 1097 | |
| 1098 | bc, ok := p.brokers[broker] |
| 1099 | if ok && bc.abandoned != nil { |
| 1100 | close(bc.abandoned) |
| 1101 | } |
| 1102 | |
| 1103 | delete(p.brokers, broker) |
| 1104 | } |