Kafka消息的压缩机制

最近在做 AWS cost saving 的事情,对于 Kafka 消息集群,计划通过压缩消息来减少消息存储所占空间,从而达到减少 cost 的目的。本文将结合源码从 Kafka 支持的消息压缩类型、何时需要压缩、如何开启压缩、何处进行解压缩以及压缩原理来总结 Kafka 整个消息压缩机制。文中所涉及源码部分均来自于 Kafka 当前最新的 3.3.0-SNAPSHOT 版本。

Kafka支持的消息压缩类型

什么是 Kafka 的消息压缩

在谈消息压缩类型之前,我们先看下 Kafka 中关于消息压缩的定义是什么。

Kafka 官网 有这样一段解释:

此为 Kafka 中端到端的块压缩功能。如果启用,数据将由 producer 压缩,以压缩格式写入服务器,并由 consumer 解压缩。压缩将提高 consumer 的吞吐量,但需付出一定的解压成本。这在跨数据中心镜像数据时尤其有用。

也就是说,Kafka 的消息压缩是指将消息本身采用特定的压缩算法进行压缩并存储,待消费时再解压。

我们知道压缩就是用时间换空间,其基本理念是基于重复,将重复的片段编码为字典,字典的 key 为重复片段,value 为更短的代码,比如序列号,然后将原始内容中的片段用代码表示,达到缩短内容的效果,压缩后的内容则由字典和代码序列两部分组成。解压时根据字典和代码序列可无损地还原为原始内容。注:有损压缩不在此次讨论范围。

通常来讲,重复越多,压缩效果越好。比如 JSON 是 Kafka 消息中常用的序列化格式,单条消息内可能并没有多少重复片段,但如果是批量消息,则会有大量重复的字段名,批量中消息越多,则重复越多,这也是为什么 Kafka 更偏向块压缩,而不是单条消息压缩。

消息压缩类型

目前 Kafka 共支持四种主要的压缩类型:Gzip、Snappy、Lz4 和 Zstd。关于这几种压缩的特性,

压缩类型 压缩比率 CPU 使用率 压缩速度 带宽使用率
Gzip Highest Highest Slowest Lowest
Snappy Medium Moderate Moderate Medium
Lz4 Low Lowest Fastest Highest
Zstd Medium Moderate Moderate Medium

从上表可知,Snappy 在 CPU 使用率、压缩比、压缩速度和网络带宽使用率之间实现良好的平衡,我们最终也是采用的该类型进行压缩试点。这里值得一提的是,Zstd 是 Facebook 于 2016 年开源的新压缩算法,压缩率和压缩性能都不错,具有与 Snappy(Google 杰作)相似的特性,直到 Kafka 的 2.1.0 版本才引入支持。

针对这几种压缩本身的性能,Zstd GitHub 官方 公布了压测对比结果如下,

Compressor name Ratio Compression Decompress.
zstd 1.5.1 -1 2.887 530 MB/s 1700 MB/s
zlib 1.2.11 -1 2.743 95 MB/s 400 MB/s
brotli 1.0.9 -0 2.702 395 MB/s 450 MB/s
zstd 1.5.1 --fast=1 2.437 600 MB/s 2150 MB/s
zstd 1.5.1 --fast=3 2.239 670 MB/s 2250 MB/s
quicklz 1.5.0 -1 2.238 540 MB/s 760 MB/s
zstd 1.5.1 --fast=4 2.148 710 MB/s 2300 MB/s
lzo1x 2.10 -1 2.106 660 MB/s 845 MB/s
lz4 1.9.3 2.101 740 MB/s 4500 MB/s
lzf 3.6 -1 2.077 410 MB/s 830 MB/s
snappy 1.1.9 2.073 550 MB/s 1750 MB/s

可以看到 Zstd 可以通过压缩速度为代价获得更高的压缩比,二者之间的权衡可通过 --fast 参数灵活配置。

何时需要压缩

压缩是需要额外的 CPU 代价的,并且会带来一定的消息分发延迟,因而在压缩前要慎重考虑是否有必要。笔者认为需考虑以下几方面:

  • 压缩带来的磁盘空间和带宽节省远大于额外的 CPU 代价,这样的压缩是值得的。
  • 数据量足够大且具重复性。消息压缩是批量的,低频的数据流可能都无法填满一个批量,会影响压缩比。数据重复性越高,往往压缩效果越好,例如 JSON、XML 等结构化数据;但若数据不具重复性,例如文本都是唯一的 md5 或 UUID 之类,违背了压缩的重复性前提,压缩效果可能不会理想。
  • 系统对消息分发的延迟没有严苛要求,可容忍轻微的延迟增长。

如何开启压缩

Kafka 通过配置属性 compression.type 控制是否压缩。该属性在 producer 端和 broker 端各自都有一份,也就是说,我们可以选择在 producer 或 broker 端开启压缩,对应的应用场景各有不同。

在 Broker 端开启压缩

compression.type 属性

Broker 端的 compression.type 属性默认值为 producer,即直接继承 producer 端所发来消息的压缩方式,无论消息采用何种压缩或者不压缩,broker 都原样存储,这一点可以从如下代码片段看出:

class UnifiedLog(...) extends Logging with KafkaMetricsGroup {
  ...
  private def analyzeAndValidateRecords(records: MemoryRecords,
                                        origin: AppendOrigin,
                                        ignoreRecordSize: Boolean,
                                        leaderEpoch: Int): LogAppendInfo = {
    records.batches.forEach { batch =>
      ...
      val messageCodec = CompressionCodec.getCompressionCodec(batch.compressionType.id)
      if (messageCodec != NoCompressionCodec)
        sourceCodec = messageCodec
    }
    // Apply broker-side compression if any
    val targetCodec = BrokerCompressionCodec.getTargetCompressionCodec(config.compressionType, sourceCodec);
  }
}

object BrokerCompressionCodec {

  val brokerCompressionCodecs = List(UncompressedCodec, ZStdCompressionCodec, LZ4CompressionCodec, SnappyCompressionCodec, GZIPCompressionCodec, ProducerCompressionCodec)
  val brokerCompressionOptions: List[String] = brokerCompressionCodecs.map(codec => codec.name)

  def isValid(compressionType: String): Boolean = brokerCompressionOptions.contains(compressionType.toLowerCase(Locale.ROOT))

  def getCompressionCodec(compressionType: String): CompressionCodec = {
    compressionType.toLowerCase(Locale.ROOT) match {
      case UncompressedCodec.name => NoCompressionCodec
      case _ => CompressionCodec.getCompressionCodec(compressionType)
    }
  }

  def getTargetCompressionCodec(compressionType: String, producerCompression: CompressionCodec): CompressionCodec = {
    if (ProducerCompressionCodec.name.equals(compressionType))
      producerCompression
    else
      getCompressionCodec(compressionType)
  }
}

sourceCodecrecordBatch 上的编码,即表示从 producer 端发来的这批消息的编码。 targetCodec 为 broker 端配置的压缩编码,从函数 getTargetCompressionCodec 可以看出最终存储消息的目标编码是结合 broker 端的 compressionType 和 producer 端的 producerCompression 综合判断的:当 compressionTypeproducer 时直接采用 producer 端的 producerCompression,否则就采用 broker 端自身的编码设置 compressionType。从 brokerCompressionCodecs 的取值可看出,compression.type 的可选值为 [uncompressed, zstd, lz4, snappy, gzip, producer]。其中 uncompressednone 是等价的,producer 不用多说,其余四个则是标准的压缩类型。

broker 和 topic 两个级别

在 broker 端的压缩配置分为两个级别:全局的 broker 级别 和 局部的 topic 级别。顾名思义,如果配置的是 broker 级别,则对于该 Kafka 集群中所有的 topic 都是生效的。但如果 topic 级别配置了自己的压缩类型,则会覆盖 broker 全局的配置,以 topic 自己配置的为准。

broker 级别

要配置 broker 级别的压缩类型,可通过 configs 命令修改 compression.type 配置项取值。此处要使修改生效,是否需要重启 broker 取决于 Kafak 的版本,在 1.1.0 之前,任何配置项的改动都需要重启 broker 才生效,而从 1.1.0 版本开始,Kafka 引入了动态 broker 参数,将配置项分为三类:read-onlyper-brokercluster-wide,第一类跟原来一样需重启才生效,而后面两类都是动态生效的,只是影响范围不同,关于 Kafka 动态参数,以后单开博文介绍。从 官网 可以看到,compression.type 是属于 cluster-wide 的,如果是 1.1.0 及之后的版本,则无需重启 broker。

topic 级别

topic 的配置分为两部分,一部分是 topic 特有的,如 partitions 等,另一部分则是默认采用 broker 配置,但也可以覆盖。如果要定义 topic 级别的压缩,可以在 topic 创建时通过 --config 选项覆盖配置项 compression.type 的取值,命令如下:

sh bin/kafka-topics.sh --create --topic my-topic --replication-factor 1 --partitions 1 --config compression.type=snappy

当然也可以通过 configs 命令修改 topic 的 compression.type 取值,命令如下:

bin/kafka-configs.sh --entity-type topics --entity-name my-topic --alter --add-config compression.type=snappy

在 Producer 端压缩

compression.type 属性

跟 broker 端一样,producer 端的压缩配置属性依然是 compression.type,只不过默认值和可选值有所不同。默认值为 none,表示不压缩,可选值为枚举类 CompressionType 中所有实例对应 name 的列表。

开启压缩的方式

直接在代码层面更改 producer 的 config,示例如下。但需要注意的是,改完 config 之后,需要重启 producer 端的应用程序,压缩才会生效。

@Configuration
@EnableKafka
public class KafkaProducerConfig {
    @Bean
    public KafkaTemplate<byte[], byte[]> kafkaTemplate() {
        Map<String, Object> config = new HashMap<>();
        config.put(ProducerConfig.BOOTSTRAP_SERVERS_CONFIG, brokerServer);
        config.put(ProducerConfig.KEY_SERIALIZER_CLASS_CONFIG, keySerializer);
        config.put(ProducerConfig.VALUE_SERIALIZER_CLASS_CONFIG, valueSerializer);
        config.put(ProducerConfig.ACKS_CONFIG, "all");
        config.put(ProducerConfig.MAX_IN_FLIGHT_REQUESTS_PER_CONNECTION, "1");
        config.put(ProducerConfig.BATCH_SIZE_CONFIG, "16384");
        config.put(ProducerConfig.RETRIES_CONFIG, Integer.MAX_VALUE);
        config.put(ProducerConfig.RETRY_BACKOFF_MS_CONFIG, "3000");
        config.put(ProducerConfig.LINGER_MS_CONFIG, "1");
        ...
        // 开启 Snappy 压缩
        config.put(ProducerConfig.COMPRESSION_TYPE_CONFIG, CompressionType.SNAPPY.name);

        return new KafkaTemplate<>(new DefaultKafkaProducerFactory<>(config));
    }
}

压缩和解压的位置

何处会压缩

可能产生压缩的地方有两处:producer 端和 broker 端。

producer 端

producer 端发生压缩的唯一条件就是在 producer 端为属性 compression.type 配置了除 none 之外有效的压缩类型。此时,producer 在向所负责的所有 topics 发消息之前,都会将消息压缩处理。

broker 端

对于 broker 端,产生压缩的情况就复杂得多,这不仅取决于 broker 端自身的压缩编码 targetCodec 是否是需要压缩的类型,还取决于 targetCodec 跟 producer 端的 sourceCodec 是否相同,除此之外,还跟消息格式的 magic 版本有关。直接看代码,broker 端的消息读写是由 UnifiedLog 负责的,消息持久化的核心入口是 append 方法,代码如下:

class UnifiedLog(...) extends Logging with KafkaMetricsGroup {
  ...
  private def append(records: MemoryRecords,
                     origin: AppendOrigin,
                     interBrokerProtocolVersion: ApiVersion,
                     validateAndAssignOffsets: Boolean,
                     leaderEpoch: Int,
                     requestLocal: Option[RequestLocal],
                     ignoreRecordSize: Boolean): LogAppendInfo = {
    ...
    val appendInfo = analyzeAndValidateRecords(records, origin, ignoreRecordSize, leaderEpoch)

    // return if we have no valid messages or if this is a duplicate of the last appended entry
    if (appendInfo.shallowCount == 0) appendInfo
    else {

      // trim any invalid bytes or partial messages before appending it to the on-disk log
      var validRecords = trimInvalidBytes(records, appendInfo)

      // they are valid, insert them in the log
      lock synchronized {
        maybeHandleIOException(s"Error while appending records to $topicPartition in dir ${dir.getParent}") {
          localLog.checkIfMemoryMappedBufferClosed()
          if (validateAndAssignOffsets) {
            // assign offsets to the message set
            val offset = new LongRef(localLog.logEndOffset)
            appendInfo.firstOffset = Some(LogOffsetMetadata(offset.value))
            val now = time.milliseconds
            val validateAndOffsetAssignResult = try {
              LogValidator.validateMessagesAndAssignOffsets(validRecords,
                topicPartition,
                offset,
                time,
                now,
                appendInfo.sourceCodec,
                appendInfo.targetCodec,
                config.compact,
                config.recordVersion.value,
                config.messageTimestampType,
                config.messageTimestampDifferenceMaxMs,
                leaderEpoch,
                origin,
                interBrokerProtocolVersion,
                brokerTopicStats,
                requestLocal.getOrElse(throw new IllegalArgumentException(
                  "requestLocal should be defined if assignOffsets is true")))
            } catch {
              case e: IOException =>
                throw new KafkaException(s"Error validating messages while appending to log $name", e)
            }
            ...
          } else {
            // we are taking the offsets we are given
            ...
          }
          ...
          maybeDuplicate match {
            case Some(duplicate) =>
              ...
              localLog.append(appendInfo.lastOffset, appendInfo.maxTimestamp, appendInfo.offsetOfMaxTimestamp, validRecords)
              updateHighWatermarkWithLogEndOffset()
              ...
              trace(s"Appended message set with last offset: ${appendInfo.lastOffset}, " +
                s"first offset: ${appendInfo.firstOffset}, " +
                s"next offset: ${localLog.logEndOffset}, " +
                s"and messages: $validRecords")

              if (localLog.unflushedMessages >= config.flushInterval) flush(false)
          }
          appendInfo
        }
      }
    }
  }
}

可以看到,先是采用 analyzeAndValidateRecordsrecordBatch 的维度对批量消息整体做校验,比如 CRC、size 等,不会细化到单条消息,所以这里不会涉及解压。这一步通过之后,会采用 LogValidator.validateMessagesAndAssignOffsetsrecordBatch以及单条消息做进一步验证并为消息分配 offset该过程可能涉及解压。完成这一步之后,调用 localLog.append 方法将消息追加到本地日志,这一步才是真正的落盘。我们继续关注可能发生解压的 LogValidator 部分,代码如下:

private[log] object LogValidator extends Logging {
  private[log] def validateMessagesAndAssignOffsets(records: MemoryRecords,
                                                    topicPartition: TopicPartition,
                                                    offsetCounter: LongRef,
                                                    time: Time,
                                                    now: Long,
                                                    sourceCodec: CompressionCodec,
                                                    targetCodec: CompressionCodec,
                                                    compactedTopic: Boolean,
                                                    magic: Byte,
                                                    timestampType: TimestampType,
                                                    timestampDiffMaxMs: Long,
                                                    partitionLeaderEpoch: Int,
                                                    origin: AppendOrigin,
                                                    interBrokerProtocolVersion: ApiVersion,
                                                    brokerTopicStats: BrokerTopicStats,
                                                    requestLocal: RequestLocal): ValidationAndOffsetAssignResult = {
    if (sourceCodec == NoCompressionCodec && targetCodec == NoCompressionCodec) {
      // check the magic value
      if (!records.hasMatchingMagic(magic))
        convertAndAssignOffsetsNonCompressed(records, topicPartition, offsetCounter, compactedTopic, time, now, timestampType,
          timestampDiffMaxMs, magic, partitionLeaderEpoch, origin, brokerTopicStats)
      else
        // Do in-place validation, offset assignment and maybe set timestamp
        assignOffsetsNonCompressed(records, topicPartition, offsetCounter, now, compactedTopic, timestampType, timestampDiffMaxMs,
          partitionLeaderEpoch, origin, magic, brokerTopicStats)
    } else {
      validateMessagesAndAssignOffsetsCompressed(records, topicPartition, offsetCounter, time, now, sourceCodec,
        targetCodec, compactedTopic, magic, timestampType, timestampDiffMaxMs, partitionLeaderEpoch, origin,
        interBrokerProtocolVersion, brokerTopicStats, requestLocal)
    }
  }
  ...
}

从上可知,当 broker 端配置的压缩编码 targetCodec 与所收到的批量消息的压缩编码 sourceCodec 都为 none 即不压缩时,会再检查消息格式的版本,如果与 broker 端配置的版本不同,则需要先将原批量消息转换为目标版本 magic 对应格式的新批量消息,然后再在新批量消息中分配 offset;否则直接在原批量消息中就地分配 offset,此过程均不涉及解压缩。这里稍微解释下分配 offset 的逻辑,我们知道在 Kafka 中 offsetpartition 下每条消息的唯一标识,consumer 端也是根据 offset 来追踪消费进度,而 offset 的生成和写入则是在 broker 端,就是此处提到的 offset 分配。理论上说,broker 需要为每条消息都分配一个 offset 的,但在实践中,因为用的是 recordBatch,内部消息是顺序排列的且总记录数是知道的,而 recordBatch 本身会记录 baseOffset ,故通常只需设置 lastOffset即可。唯一的例外是,当因消息格式转换或解压缩而需要创建新的 recordBatch时,会调用 memoryRecordsBuilderappendWithOffset 方法为每一条消息记录分配 offset

targetCodecsourceCodec 至少有一个不为 none 即需要压缩时,情况就复杂一些,具体逻辑都在 validateMessagesAndAssignOffsetsCompressed方法中,

private[log] object LogValidator extends Logging {
  ...
  def validateMessagesAndAssignOffsetsCompressed(...): ValidationAndOffsetAssignResult = {
    ...
    // No in place assignment situation 1
    var inPlaceAssignment = sourceCodec == targetCodec

    var maxTimestamp = RecordBatch.NO_TIMESTAMP
    val expectedInnerOffset = new LongRef(0)
    val validatedRecords = new mutable.ArrayBuffer[Record]

    var uncompressedSizeInBytes = 0

    // Assume there's only one batch with compressed memory records; otherwise, return InvalidRecordException
    // One exception though is that with format smaller than v2, if sourceCodec is noCompression, then each batch is actually
    // a single record so we'd need to special handle it by creating a single wrapper batch that includes all the records
    val firstBatch = getFirstBatchAndMaybeValidateNoMoreBatches(records, sourceCodec)

    // No in place assignment situation 2 and 3: we only need to check for the first batch because:
    //  1. For most cases (compressed records, v2, for example), there's only one batch anyways.
    //  2. For cases that there may be multiple batches, all batches' magic should be the same.
    if (firstBatch.magic != toMagic || toMagic == RecordBatch.MAGIC_VALUE_V0)
      inPlaceAssignment = false

    // Do not compress control records unless they are written compressed
    if (sourceCodec == NoCompressionCodec && firstBatch.isControlBatch)
      inPlaceAssignment = true

    records.batches.forEach { batch =>
      validateBatch(topicPartition, firstBatch, batch, origin, toMagic, brokerTopicStats)
      uncompressedSizeInBytes += AbstractRecords.recordBatchHeaderSizeInBytes(toMagic, batch.compressionType())

      // if we are on version 2 and beyond, and we know we are going for in place assignment,
      // then we can optimize the iterator to skip key / value / headers since they would not be used at all
      val recordsIterator = if (inPlaceAssignment && firstBatch.magic >= RecordBatch.MAGIC_VALUE_V2)
        batch.skipKeyValueIterator(requestLocal.bufferSupplier)
      else
        batch.streamingIterator(requestLocal.bufferSupplier)

      try {
        val recordErrors = new ArrayBuffer[ApiRecordError](0)
        // this is a hot path and we want to avoid any unnecessary allocations.
        var batchIndex = 0
        recordsIterator.forEachRemaining { record =>
          val expectedOffset = expectedInnerOffset.getAndIncrement()
          val recordError = validateRecordCompression(batchIndex, record).orElse {
            validateRecord(batch, topicPartition, record, batchIndex, now,
              timestampType, timestampDiffMaxMs, compactedTopic, brokerTopicStats).orElse {
              if (batch.magic > RecordBatch.MAGIC_VALUE_V0 && toMagic > RecordBatch.MAGIC_VALUE_V0) {
                if (record.timestamp > maxTimestamp)
                  maxTimestamp = record.timestamp

                // Some older clients do not implement the V1 internal offsets correctly.
                // Historically the broker handled this by rewriting the batches rather
                // than rejecting the request. We must continue this handling here to avoid
                // breaking these clients.
                if (record.offset != expectedOffset)
                  inPlaceAssignment = false
              }
              None
            }
          }

          recordError match {
            case Some(e) => recordErrors += e
            case None =>
              uncompressedSizeInBytes += record.sizeInBytes()
              validatedRecords += record
          }
         batchIndex += 1
        }
        processRecordErrors(recordErrors)
      } finally {
        recordsIterator.close()
      }
    }

    if (!inPlaceAssignment) {
      val (producerId, producerEpoch, sequence, isTransactional) = {
        // note that we only reassign offsets for requests coming straight from a producer. For records with magic V2,
        // there should be exactly one RecordBatch per request, so the following is all we need to do. For Records
        // with older magic versions, there will never be a producer id, etc.
        val first = records.batches.asScala.head
        (first.producerId, first.producerEpoch, first.baseSequence, first.isTransactional)
      }
      buildRecordsAndAssignOffsets(toMagic, offsetCounter, time, timestampType, CompressionType.forId(targetCodec.codec),
        now, validatedRecords, producerId, producerEpoch, sequence, isTransactional, partitionLeaderEpoch,
        uncompressedSizeInBytes)
    } else {
      // we can update the batch only and write the compressed payload as is;
      // again we assume only one record batch within the compressed set
      val batch = records.batches.iterator.next()
      val lastOffset = offsetCounter.addAndGet(validatedRecords.size) - 1

      batch.setLastOffset(lastOffset)

      if (timestampType == TimestampType.LOG_APPEND_TIME)
        maxTimestamp = now

      if (toMagic >= RecordBatch.MAGIC_VALUE_V1)
        batch.setMaxTimestamp(timestampType, maxTimestamp)

      if (toMagic >= RecordBatch.MAGIC_VALUE_V2)
        batch.setPartitionLeaderEpoch(partitionLeaderEpoch)

      val recordConversionStats = new RecordConversionStats(uncompressedSizeInBytes, 0, 0)
      ValidationAndOffsetAssignResult(validatedRecords = records,
        maxTimestamp = maxTimestamp,
        shallowOffsetOfMaxTimestamp = lastOffset,
        messageSizeMaybeChanged = false,
        recordConversionStats = recordConversionStats)
    }
  }
  ...
}

可以看到,inPlaceAssignment 是用于标识是否可以原地修改 recordBatch 来分配 offset,有三种情况不能原地修改:

  • sourceCodec 和 targetCodec 不同,这个比较好理解,编码不同,构建目标 payload 时原 recordBatch 自然不能复用。
  • 目标消息格式版本 magic 与 broker 接收到的 recordBatchmagic 不同,此时需要消息格式转换,需要构建新的 recordBatch,这个跟第一种情况是一样的,无法复用原 recordBatch
  • 目标消息格式版本为 V0,因为老版本 V0 格式的消息,需要为每条消息重新分配绝对 offset,无法复用原 recordBatch

此时,inPlaceAssignment 为 false,直接走 buildRecordsAndAssignOffsets 逻辑来构建新的 recordBatch,此时是否压缩取决于 targetCodec,如果不为none,则此处会按照 targetCodec 编码进行压缩。

除了上述三种情况之外,都是可以原地修改,此时可以直接复用原 recordBatch来构建目标消息的 payload,此时不存在压缩处理。

何处会解压

可能发生解压的地方依然是两处:consumer 端和 broker 端。

consumer 端

consumer 端发生解压的唯一条件就是从 broker 端拉取到的消息是带压缩的。此时,consumer 会根据 recordBatchcompressionType 来对消息进行解压,具体细节后面源码分析部分会讲。

broker 端

broker 端是否发生解压取决于 producer 发过来的批量消息 recordBatch 是否是压缩的:如果 producer 开启了压缩,则会发生解压,否则不会。原因简单说下,在 broker 端持久化消息前,会对消息做各种验证,此时必然会迭代 recordBatch,而在迭代的过程中,会直接采用 recordBatch 上的 compressionType 对消息字节流进行处理,是否解压取决于 compressionType 是否是压缩类型。关于这点,可以在 LogValidatorvalidateMessagesAndAssignOffsets 方法实现中可以看到,在 convertAndAssignOffsetsNonCompressedassignOffsetsNonCompressedvalidateMessagesAndAssignOffsetsCompressed 三个不同的分支中,都会看到 records.batches.forEach {...} 的身影,而在后面的源码分析中会发现,在 recordBatch 的迭代器逻辑中,直接采用的 compressionType 的解压逻辑对消息字节流读取的。也就是说,如果 recordBatch 是压缩的 ,只要对其进行了迭代访问,则会自动触发解压逻辑。

压缩和解压原理

压缩和解压涉及到几个关键的类:CompressionTypeMemoryRecordsBuilderDefaultRecordBatchAbstractLegacyRecordBatch。其中 CompressionType 是压缩相关的枚举,集压缩定义和实现为一体;MemoryRecordsBuilder 是负责将新的消息数据写入内存 buffer,即调用 CompressionType 中的压缩逻辑 wrapForOutput 来写入消息;而 DefaultRecordBatchAbstractLegacyRecordBatch 则是负责读取消息数据,即调用 CompressionType 的解压逻辑 wrapForInput 将消息还原为无压缩数据。只不过二者区别是,前者是用于处理新版本格式的消息(即 magic >= 2),而后者则是处理老版本格式的消息(即 magic 为 0 或 1)。

CompressionType

在说 CompressionType 之前,我们先看下 CompressionCodec 这个 Scala 脚本。

CompressionCodec

部分源码如下,

...
case object GZIPCompressionCodec extends CompressionCodec with BrokerCompressionCodec {
  val codec = 1
  val name = "gzip"
}

case object SnappyCompressionCodec extends CompressionCodec with BrokerCompressionCodec {
  val codec = 2
  val name = "snappy"
}

case object LZ4CompressionCodec extends CompressionCodec with BrokerCompressionCodec {
  val codec = 3
  val name = "lz4"
}

case object ZStdCompressionCodec extends CompressionCodec with BrokerCompressionCodec {
  val codec = 4
  val name = "zstd"
}

case object NoCompressionCodec extends CompressionCodec with BrokerCompressionCodec {
  val codec = 0
  val name = "none"
}

case object UncompressedCodec extends BrokerCompressionCodec {
  val name = "uncompressed"
}

case object ProducerCompressionCodec extends BrokerCompressionCodec {
  val name = "producer"
}

该脚本定义了 GZIPCompressionCodec 等共 7 个 case object,可类比于 Java 中枚举,这些 case object 中的 name 集合则刚好覆盖了前文所提到的属性 compression.type 的所有可选值,包括 producer 端和 broker 端的。而与 name 绑定在一起的 codec 则是最终真正写入消息体的压缩编码,name 只是为了可读性友好。从上可知,压缩编码codec 的有效取值只有 0~4,分别对应 nonegzipsnappylz4zstd,而这五种取值恰好是 CompressionType 中定义的五种枚举常量。

由此可知,CompressionCodec是面向配置属性 compression.type的可选值的,并将数值化的压缩编码 codec 映射为可读性强的 name;而 CompressionType则是定义了与压缩编码对应的枚举常量,二者通过 name 关联。

CompressionType 源码

CompressionType 定义了与压缩编码对应的五种压缩类型枚举,并且通过用于压缩的 wrapForOutput和用于解压的 wrapForInput这两个抽象方法将每种压缩类型与对应的压缩实现绑定在一起,既避免了常规的 if-else 判断,也将压缩的定义与实现完全收敛到 CompressionType ,符合单一职责原则。其实类似这种优雅的设计在 JDK 中也能经常看到其身影,比如 TimeUnit。直接看源码,

public enum CompressionType {
    ...
    GZIP(1, "gzip", 1.0f) {
        @Override
        public OutputStream wrapForOutput(ByteBufferOutputStream buffer, byte messageVersion) {
            try {
                return new BufferedOutputStream(new GZIPOutputStream(buffer, 8 * 1024), 16 * 1024);
            } catch (Exception e) {
                throw new KafkaException(e);
            }
        }

        @Override
        public InputStream wrapForInput(ByteBuffer buffer, byte messageVersion, BufferSupplier decompressionBufferSupplier) {
            try {
                // Set output buffer (uncompressed) to 16 KB (none by default) and input buffer (compressed) to
                // 8 KB (0.5 KB by default) to ensure reasonable performance in cases where the caller reads a small
                // number of bytes (potentially a single byte)
                return new BufferedInputStream(new GZIPInputStream(new ByteBufferInputStream(buffer), 8 * 1024),
                        16 * 1024);
            } catch (Exception e) {
                throw new KafkaException(e);
            }
        }
    },
    ...
    ZSTD(4, "zstd", 1.0f) {
        @Override
        public OutputStream wrapForOutput(ByteBufferOutputStream buffer, byte messageVersion) {
            return ZstdFactory.wrapForOutput(buffer);
        }

        @Override
        public InputStream wrapForInput(ByteBuffer buffer, byte messageVersion, BufferSupplier decompressionBufferSupplier) {
            return ZstdFactory.wrapForInput(buffer, messageVersion, decompressionBufferSupplier);
        }
    };
    ...
    // Wrap bufferStream with an OutputStream that will compress data with this CompressionType.
    public abstract OutputStream wrapForOutput(ByteBufferOutputStream bufferStream, byte messageVersion);
    // Wrap buffer with an InputStream that will decompress data with this CompressionType.
    public abstract InputStream wrapForInput(ByteBuffer buffer, byte messageVersion, BufferSupplier decompressionBufferSupplier);

    ...
}

每种压缩类型对于 wrapForOutputwrapForInput 两方法的具体实现已经很清楚地阐述了压缩和解压的方式,感兴趣的朋友可以从该入口 step in 一探究竟。这里就不细述。当然这只是处理压缩最小的基本单元,为了搞清楚 Kafka 在何处使用它,还得继续看其他几个核心类。

在此之前,就上述源码,抛开本次主题,我还想谈几个值得学习借鉴的细节,

  1. SnappyZstd 都是用的 XXXFactory 静态方法来构建 Stream 对象,而其他的比如 Lz4 则都是直接通过 new 创建的对象。之所以这么做,我们进一步 step in 就会发现,对于 SnappyZstd,Kafka 都是直接依赖的第三方库,而其他的则是 JDK 或 Kafka 自己的实现。为了减少第三方库的副作用,通过此方式将第三方库的类的惰性加载做到极致,这也体现出作者对 Java 类加载时机的充分理解,很精致的处理
  2. GzipwrapForInput实现中,在 KAFKA-6430 这个 Improvement 提交中,input buffer 从 0.5 KB 调大到 8 KB,其目的就是能够在一次 Gzip 压缩中处理更多的字节,以获得更高的性能。至少,从 commit 的描述上看,throughput 能翻倍。
  3. 抽象方法 wrapForInput 中暴露的最后一个 BufferSupplier类型的参数 decompressionBufferSupplier,正如方法的参数说明所言,对于比较小的批量消息,如果在 wrapForInput 内部新建 buffer,那么每次方法调用都会新分配buffer,这可能比压缩处理本身更耗时,所以该参数给了一个选择的机会,在外面分配内存,然后方法内循环利用。在日常的编码中,对于循环中所需的空间,我也经常会思考是每次新建好还是先在外面分配,然后内部循环利用更好,case by case.

MemoryRecordsBuilder

public class MemoryRecordsBuilder implements AutoCloseable {
    ...
    // Used to append records, may compress data on the fly
    private DataOutputStream appendStream;
    ...

    public MemoryRecordsBuilder(ByteBufferOutputStream bufferStream,
                                byte magic,
                                CompressionType compressionType,
                                TimestampType timestampType,
                                long baseOffset,
                                long logAppendTime,
                                long producerId,
                                short producerEpoch,
                                int baseSequence,
                                boolean isTransactional,
                                boolean isControlBatch,
                                int partitionLeaderEpoch,
                                int writeLimit,
                                long deleteHorizonMs) {
        if (magic > RecordBatch.MAGIC_VALUE_V0 && timestampType == TimestampType.NO_TIMESTAMP_TYPE)
            throw new IllegalArgumentException("TimestampType must be set for magic > 0");
        if (magic < RecordBatch.MAGIC_VALUE_V2) {
            if (isTransactional)
                throw new IllegalArgumentException("Transactional records are not supported for magic " + magic);
            if (isControlBatch)
                throw new IllegalArgumentException("Control records are not supported for magic " + magic);
            if (compressionType == CompressionType.ZSTD)
                throw new IllegalArgumentException("ZStandard compression is not supported for magic " + magic);
            if (deleteHorizonMs != RecordBatch.NO_TIMESTAMP)
                throw new IllegalArgumentException("Delete horizon timestamp is not supported for magic " + magic);
        }
        ...
        this.appendStream = new DataOutputStream(compressionType.wrapForOutput(this.bufferStream, magic));
        ...
    }
  
    public void close() {
        ...
        if (numRecords == 0L) {
            buffer().position(initialPosition);
            builtRecords = MemoryRecords.EMPTY;
        } else {
            if (magic > RecordBatch.MAGIC_VALUE_V1)
                this.actualCompressionRatio = (float) writeDefaultBatchHeader() / this.uncompressedRecordsSizeInBytes;
            else if (compressionType != CompressionType.NONE)
                this.actualCompressionRatio = (float) writeLegacyCompressedWrapperHeader() / this.uncompressedRecordsSizeInBytes;

            ByteBuffer buffer = buffer().duplicate();
            buffer.flip();
            buffer.position(initialPosition);
            builtRecords = MemoryRecords.readableRecords(buffer.slice());
        }
    }
    ...
    private int writeDefaultBatchHeader() {
        ...
        DefaultRecordBatch.writeHeader(buffer, baseOffset, offsetDelta, size, magic, compressionType, timestampType,
                baseTimestamp, maxTimestamp, producerId, producerEpoch, baseSequence, isTransactional, isControlBatch,
                hasDeleteHorizonMs(), partitionLeaderEpoch, numRecords);

        buffer.position(pos);
        return writtenCompressed;
    }
    private int writeLegacyCompressedWrapperHeader() {
        ...
        int wrapperSize = pos - initialPosition - Records.LOG_OVERHEAD;
        int writtenCompressed = wrapperSize - LegacyRecord.recordOverhead(magic);
        AbstractLegacyRecordBatch.writeHeader(buffer, lastOffset, wrapperSize);

        long timestamp = timestampType == TimestampType.LOG_APPEND_TIME ? logAppendTime : maxTimestamp;
        LegacyRecord.writeCompressedRecordHeader(buffer, magic, wrapperSize, timestamp, compressionType, timestampType);

        buffer.position(pos);
        return writtenCompressed;
    }
}

可以看到,appendStream 是用于追加消息到内存 buffer 的,直接采用的 compressionType 的压缩逻辑来构建写入流的,如果此处 compressionType属于非 none 的有效压缩类型,则会产生压缩。此外,从上面 magic 的判断逻辑可知,消息的时间戳类型是从大版本 V1 开始支持的;而事务消息、控制消息、Zstd 压缩和 deleteHorizonMs都是从 V2 才开始支持的。这里的 V1V2 对应消息格式的版本,其中 V1 是从 0.10.0 版本开始引入的,在此之前都是 V0 版本,而 V2 则是从 0.11.0 版本开始引入,直到现在的最新版依然是 V2

close() 方法可以看出,MemoryRecordsBuilder 在构建 memoryRecords 时,会根据消息格式的版本高低,写入不同的 Header。对于新版消息,在 writeDefaultBatchHeader 方法中直接调用 DefaultRecordBatch.writeHeader(...)写入新版消息特定的 Header;而对于老版消息,则是在 writeLegacyCompressedWrapperHeader方法中调用 AbstractLegacyRecordBatch.writeHeaderLegacyRecord.writeCompressedRecordHeader 写入老版消息的 Header。虽然 Header 的格式各不相同,但我们在两种 Header 中都可以看到 compressionType 的身影,以此可见,Kafka 是允许多种版本的消息共存的,以及压缩与非压缩消息的共存,因为这些信息是保存在 recordBatch 上的,是批量消息级别。

DefaultRecordBatch

public class DefaultRecordBatch extends AbstractRecordBatch implements MutableRecordBatch {
    ...
    @Override
    public Iterator<Record> iterator() {
        if (count() == 0)
            return Collections.emptyIterator();

        if (!isCompressed())
            return uncompressedIterator();

        // for a normal iterator, we cannot ensure that the underlying compression stream is closed,
        // so we decompress the full record set here. Use cases which call for a lower memory footprint
        // can use `streamingIterator` at the cost of additional complexity
        try (CloseableIterator<Record> iterator = compressedIterator(BufferSupplier.NO_CACHING, false)) {
            List<Record> records = new ArrayList<>(count());
            while (iterator.hasNext())
                records.add(iterator.next());
            return records.iterator();
        }
    }
    ...
}

RecordBatch 是表示批量消息的接口,对于老版格式的消息(版本 V0V1),如果没有压缩,只会包含单条消息,否则可以包含多条;而新版格式消息(版本 V2 及以上)无论是否压缩,都是通常包含多条消息。且该接口中有一个 compressionType()方法来标识该 batch 的压缩类型,它会作为读消息时解压的判断依据。而上面的 DefaultRecordBatch 则是该接口的针对新版本格式消息的默认实现,它也实现了 Iterable<Record> 接口,因而 iterator() 是访问批量消息的核心逻辑,当 compressionType() 返回 none 时,表示不压缩,直接返回非压缩迭代器,此处跳过,当有压缩时,走的是压缩迭代器,具体实现如下,

    public DataInputStream recordInputStream(BufferSupplier bufferSupplier) {
        final ByteBuffer buffer = this.buffer.duplicate();
        buffer.position(RECORDS_OFFSET);
        return new DataInputStream(compressionType().wrapForInput(buffer, magic(), bufferSupplier));
    }

    private CloseableIterator<Record> compressedIterator(BufferSupplier bufferSupplier, boolean skipKeyValue) {
        final DataInputStream inputStream = recordInputStream(bufferSupplier);

        if (skipKeyValue) {
            // this buffer is used to skip length delimited fields like key, value, headers
            byte[] skipArray = new byte[MAX_SKIP_BUFFER_SIZE];
          
            return new StreamRecordIterator(inputStream) {
                ...
            }
        } else {
            ...  
        }
    }

我们可以看到,compressedIterator() 在构造 Stream 迭代器之前,调用了 recordInputStream(...),该方法中通过 compressionType 的解压逻辑对原数据进行了解压。

AbstractLegacyRecordBatch

public abstract class AbstractLegacyRecordBatch extends AbstractRecordBatch implements Record {
    ...
    CloseableIterator<Record> iterator(BufferSupplier bufferSupplier) {
        if (isCompressed())
            return new DeepRecordsIterator(this, false, Integer.MAX_VALUE, bufferSupplier);

        return new CloseableIterator<Record>() {
            private boolean hasNext = true;

            @Override
            public void close() {}

            @Override
            public boolean hasNext() {
                return hasNext;
            }

            @Override
            public Record next() {
                if (!hasNext)
                    throw new NoSuchElementException();
                hasNext = false;
                return AbstractLegacyRecordBatch.this;
            }

            @Override
            public void remove() {
                throw new UnsupportedOperationException();
            }
        };
    }
    ...
    private static class DeepRecordsIterator extends AbstractIterator<Record> implements CloseableIterator<Record> {
        private DeepRecordsIterator(AbstractLegacyRecordBatch wrapperEntry,
                                    boolean ensureMatchingMagic,
                                    int maxMessageSize,
                                    BufferSupplier bufferSupplier) {
            LegacyRecord wrapperRecord = wrapperEntry.outerRecord();
            this.wrapperMagic = wrapperRecord.magic();
            if (wrapperMagic != RecordBatch.MAGIC_VALUE_V0 && wrapperMagic != RecordBatch.MAGIC_VALUE_V1)
                throw new InvalidRecordException("Invalid wrapper magic found in legacy deep record iterator " + wrapperMagic);

            CompressionType compressionType = wrapperRecord.compressionType();
            if (compressionType == CompressionType.ZSTD)
                throw new InvalidRecordException("Invalid wrapper compressionType found in legacy deep record iterator " + wrapperMagic);
            ByteBuffer wrapperValue = wrapperRecord.value();
            if (wrapperValue == null)
                throw new InvalidRecordException("Found invalid compressed record set with null value (magic = " +
                        wrapperMagic + ")");

            InputStream stream = compressionType.wrapForInput(wrapperValue, wrapperRecord.magic(), bufferSupplier);
            ...
        }
    }
}

AbstractLegacyRecordBatch 跟前面的 DefaultRecordBatch 大同小异,同样也是 iterator() 入口,当开启了压缩时,返回压缩迭代器 DeepRecordsIterator,只是名字不同而已,迭代器内部依然是直接通过 compressionType 的解压逻辑对数据流进行解压。

原文首发于:https://www.yangbing.fun/2022/04/30/compression-mechanism-of-the-Kafka-message/

posted @ 2022-05-11 14:04  sherlockyb  阅读(1889)  评论(0编辑  收藏  举报