Kafka源码分析

1 消息处理入口

以下是Kafka消息处理的入口，即客户端发送到服务端消息处理方法。

/**
   * Top-level method that handles all requests and multiplexes to the right api
   */
  def handle(request: RequestChannel.Request) {
    try{
      trace("Handling request:%s from connection %s;securityProtocol:%s,principal:%s".
        format(request.requestObj, request.connectionId, request.securityProtocol, request.session.principal))
      request.requestId match {
        case RequestKeys.ProduceKey => handleProducerRequest(request)
        case RequestKeys.FetchKey => handleFetchRequest(request)
        case RequestKeys.OffsetsKey => handleOffsetRequest(request)
        case RequestKeys.MetadataKey => handleTopicMetadataRequest(request)
        case RequestKeys.LeaderAndIsrKey => handleLeaderAndIsrRequest(request)
        case RequestKeys.StopReplicaKey => handleStopReplicaRequest(request)
        case RequestKeys.UpdateMetadataKey => handleUpdateMetadataRequest(request)
        case RequestKeys.ControlledShutdownKey => handleControlledShutdownRequest(request)
        case RequestKeys.OffsetCommitKey => handleOffsetCommitRequest(request)
        case RequestKeys.OffsetFetchKey => handleOffsetFetchRequest(request)
        case RequestKeys.GroupCoordinatorKey => handleGroupCoordinatorRequest(request)
        case RequestKeys.JoinGroupKey => handleJoinGroupRequest(request)
        case RequestKeys.HeartbeatKey => handleHeartbeatRequest(request)
        case RequestKeys.LeaveGroupKey => handleLeaveGroupRequest(request)
        case RequestKeys.SyncGroupKey => handleSyncGroupRequest(request)
        case RequestKeys.DescribeGroupsKey => handleDescribeGroupRequest(request)
        case RequestKeys.ListGroupsKey => handleListGroupsRequest(request)
        case requestId => throw new KafkaException("Unknown api code " + requestId)
      }
    } catch {
      case e: Throwable =>
        if ( request.requestObj != null)
          request.requestObj.handleError(e, requestChannel, request)
        else {
          val response = request.body.getErrorResponse(request.header.apiVersion, e)
          val respHeader = new ResponseHeader(request.header.correlationId)

          /* If request doesn't have a default error response, we just close the connection.
             For example, when produce request has acks set to 0 */
          if (response == null)
            requestChannel.closeConnection(request.processor, request)
          else
            requestChannel.sendResponse(new Response(request, new ResponseSend(request.connectionId, respHeader, response)))
        }
        error("error when handling request %s".format(request.requestObj), e)
    } finally
      request.apiLocalCompleteTimeMs = SystemTime.milliseconds
  }

2 内存中offset信息来源

/**
   * The most important guarantee that this API provides is that it should never return a stale offset. i.e., it either
   * returns the current offset or it begins to sync the cache from the log (and returns an error code).
   */
  def getOffsets(group: String, topicPartitions: Seq[TopicAndPartition]): Map[TopicAndPartition, OffsetMetadataAndError] = {
    trace("Getting offsets %s for group %s.".format(topicPartitions, group))

    if (isGroupLocal(group)) {
      if (topicPartitions.isEmpty) {
        // Return offsets for all partitions owned by this consumer group. (this only applies to consumers that commit offsets to Kafka.)
        offsetsCache.filter(_._1.group == group).map { case(groupTopicPartition, offsetAndMetadata) =>
          (groupTopicPartition.topicPartition, OffsetMetadataAndError(offsetAndMetadata.offset, offsetAndMetadata.metadata, ErrorMapping.NoError))
        }.toMap
      } else {
        topicPartitions.map { topicAndPartition =>
          val groupTopicPartition = GroupTopicPartition(group, topicAndPartition)
          (groupTopicPartition.topicPartition, getOffset(groupTopicPartition))
        }.toMap
      }
    } else {
      debug("Could not fetch offsets for group %s (not offset coordinator).".format(group))
      topicPartitions.map { topicAndPartition =>
        val groupTopicPartition = GroupTopicPartition(group, topicAndPartition)
        (groupTopicPartition.topicPartition, OffsetMetadataAndError.NotCoordinatorForGroup)
      }.toMap
    }
  }

从上面代码中可以看出，拉取的offset是从offsetsCache中获取。而在提交offset以及初始化group是会将对应的offset信息加入到该缓存中。

//该方法是在commitoffset中执行
/**
   * Store offsets by appending it to the replicated log and then inserting to cache
   */
  def prepareStoreOffsets(groupId: String,
                          consumerId: String,
                          generationId: Int,
                          offsetMetadata: immutable.Map[TopicAndPartition, OffsetAndMetadata],
                          responseCallback: immutable.Map[TopicAndPartition, Short] => Unit): DelayedStore = {
    // first filter out partitions with offset metadata size exceeding limit
    val filteredOffsetMetadata = offsetMetadata.filter { case (topicAndPartition, offsetAndMetadata) =>
      validateOffsetMetadataLength(offsetAndMetadata.metadata)
    }

    // construct the message set to append
    val messages = filteredOffsetMetadata.map { case (topicAndPartition, offsetAndMetadata) =>
      new Message(
        key = GroupMetadataManager.offsetCommitKey(groupId, topicAndPartition.topic, topicAndPartition.partition),
        bytes = GroupMetadataManager.offsetCommitValue(offsetAndMetadata)
      )
    }.toSeq

    val offsetTopicPartition = TopicAndPartition(GroupCoordinator.GroupMetadataTopicName, partitionFor(groupId))

    val offsetsAndMetadataMessageSet = Map(offsetTopicPartition ->
      new ByteBufferMessageSet(config.offsetsTopicCompressionCodec, messages:_*))

    // set the callback function to insert offsets into cache after log append completed
    def putCacheCallback(responseStatus: Map[TopicAndPartition, ProducerResponseStatus]) {
      // the append response should only contain the topics partition
      if (responseStatus.size != 1 || ! responseStatus.contains(offsetTopicPartition))
        throw new IllegalStateException("Append status %s should only have one partition %s"
          .format(responseStatus, offsetTopicPartition))

      // construct the commit response status and insert
      // the offset and metadata to cache if the append status has no error
      val status = responseStatus(offsetTopicPartition)

      val responseCode =
        if (status.error == ErrorMapping.NoError) {
          filteredOffsetMetadata.foreach { case (topicAndPartition, offsetAndMetadata) =>
            //将offset信息加入到缓存中
            putOffset(GroupTopicPartition(groupId, topicAndPartition), offsetAndMetadata)
          }
          ErrorMapping.NoError
        } else {
          debug("Offset commit %s from group %s consumer %s with generation %d failed when appending to log due to %s"
            .format(filteredOffsetMetadata, groupId, consumerId, generationId, ErrorMapping.exceptionNameFor(status.error)))

          // transform the log append error code to the corresponding the commit status error code
          if (status.error == ErrorMapping.UnknownTopicOrPartitionCode)
            ErrorMapping.ConsumerCoordinatorNotAvailableCode
          else if (status.error == ErrorMapping.NotLeaderForPartitionCode)
            ErrorMapping.NotCoordinatorForConsumerCode
          else if (status.error == ErrorMapping.MessageSizeTooLargeCode
            || status.error == ErrorMapping.MessageSetSizeTooLargeCode
            || status.error == ErrorMapping.InvalidFetchSizeCode)
            Errors.INVALID_COMMIT_OFFSET_SIZE.code
          else
            status.error
        }


      // compute the final error codes for the commit response
      val commitStatus = offsetMetadata.map { case (topicAndPartition, offsetAndMetadata) =>
        if (validateOffsetMetadataLength(offsetAndMetadata.metadata))
          (topicAndPartition, responseCode)
        else
          (topicAndPartition, ErrorMapping.OffsetMetadataTooLargeCode)
      }

      // finally trigger the callback logic passed from the API layer
      responseCallback(commitStatus)
    }

    DelayedStore(offsetsAndMetadataMessageSet, putCacheCallback)
  }

//该方法启动时异步执行
/**
   * Asynchronously read the partition from the offsets topic and populate the cache
   */
  def loadGroupsForPartition(offsetsPartition: Int,
                             onGroupLoaded: GroupMetadata => Unit) {
    val topicPartition = TopicAndPartition(GroupCoordinator.GroupMetadataTopicName, offsetsPartition)
    scheduler.schedule(topicPartition.toString, loadGroupsAndOffsets)

    def loadGroupsAndOffsets() {
      info("Loading offsets and group metadata from " + topicPartition)

      loadingPartitions synchronized {
        if (loadingPartitions.contains(offsetsPartition)) {
          info("Offset load from %s already in progress.".format(topicPartition))
          return
        } else {
          loadingPartitions.add(offsetsPartition)
        }
      }

      val startMs = SystemTime.milliseconds
      try {
        replicaManager.logManager.getLog(topicPartition) match {
          case Some(log) =>
            var currOffset = log.logSegments.head.baseOffset
            val buffer = ByteBuffer.allocate(config.loadBufferSize)
            // loop breaks if leader changes at any time during the load, since getHighWatermark is -1
            inWriteLock(offsetExpireLock) {
              val loadedGroups = mutable.Map[String, GroupMetadata]()
              val removedGroups = mutable.Set[String]()

              while (currOffset < getHighWatermark(offsetsPartition) && !shuttingDown.get()) {
                buffer.clear()
                val messages = log.read(currOffset, config.loadBufferSize).messageSet.asInstanceOf[FileMessageSet]
                messages.readInto(buffer, 0)
                val messageSet = new ByteBufferMessageSet(buffer)
                messageSet.foreach { msgAndOffset =>
                  require(msgAndOffset.message.key != null, "Offset entry key should not be null")
                  val baseKey = GroupMetadataManager.readMessageKey(msgAndOffset.message.key)

                  if (baseKey.isInstanceOf[OffsetKey]) {
                    // load offset
                    val key = baseKey.key.asInstanceOf[GroupTopicPartition]
                    if (msgAndOffset.message.payload == null) {
                      if (offsetsCache.remove(key) != null)
                        trace("Removed offset for %s due to tombstone entry.".format(key))
                      else
                        trace("Ignoring redundant tombstone for %s.".format(key))
                    } else {
                      // special handling for version 0:
                      // set the expiration time stamp as commit time stamp + server default retention time
                      val value = GroupMetadataManager.readOffsetMessageValue(msgAndOffset.message.payload)
                      //添加offset信息到缓存中
                      putOffset(key, value.copy (
                        expireTimestamp = {
                          if (value.expireTimestamp == org.apache.kafka.common.requests.OffsetCommitRequest.DEFAULT_TIMESTAMP)
                            value.commitTimestamp + config.offsetsRetentionMs
                          else
                            value.expireTimestamp
                        }
                      ))
                      trace("Loaded offset %s for %s.".format(value, key))
                    }
                  } else {
                    // load group metadata
                    val groupId = baseKey.key.asInstanceOf[String]
                    val groupMetadata = GroupMetadataManager.readGroupMessageValue(groupId, msgAndOffset.message.payload)
                    if (groupMetadata != null) {
                      trace(s"Loaded group metadata for group ${groupMetadata.groupId} with generation ${groupMetadata.generationId}")
                      removedGroups.remove(groupId)
                      loadedGroups.put(groupId, groupMetadata)
                    } else {
                      loadedGroups.remove(groupId)
                      removedGroups.add(groupId)
                    }
                  }

                  currOffset = msgAndOffset.nextOffset
                }
              }

              loadedGroups.values.foreach { group =>
                val currentGroup = addGroup(group)
                if (group != currentGroup)
                  debug(s"Attempt to load group ${group.groupId} from log with generation ${group.generationId} failed " +
                    s"because there is already a cached group with generation ${currentGroup.generationId}")
                else
                  onGroupLoaded(group)
              }

              removedGroups.foreach { groupId =>
                val group = groupsCache.get(groupId)
                if (group != null)
                  throw new IllegalStateException(s"Unexpected unload of acitve group ${group.groupId} while " +
                    s"loading partition ${topicPartition}")
              }
            }

            if (!shuttingDown.get())
              info("Finished loading offsets from %s in %d milliseconds."
                .format(topicPartition, SystemTime.milliseconds - startMs))
          case None =>
            warn("No log found for " + topicPartition)
        }
      }
      catch {
        case t: Throwable =>
          error("Error in loading offsets from " + topicPartition, t)
      }
      finally {
        loadingPartitions synchronized {
          ownedPartitions.add(offsetsPartition)
          loadingPartitions.remove(offsetsPartition)
        }
      }
    }
  }

3 Offset Commit实现

当消费端消费消息是会将offset提交，即offset提交信息，broker端把接收到的offset提交信息当做一个正常的生产请求，对offset请求的处理和正常的生产者请求处理方式是一样的。下面是内置生产者的一些属性：

• property.type=sync
• request.required.acks=-1
• key.serializer.class=StringEncoder

一旦将数据追加到leader的本地日志中，并且所有replicas都赶上leader，leader检查生产请求是“offset topic”，(因为broker端的处理逻辑针对offset请求和普通生产请求是一样的，如果是offset请求，还需要有不同的处理分支)，它就会要求offset manager添加这个offset（对于延迟的生产请求，更新操作会在延迟的生产请求被完成的时候）。因为设置了acks=-1，只有当这些offsets成功地复制到ISR的所有brokers，才会被提交给offset manager。

offset-commit.png

4 Offset Fetch实现

消费端在启动时会向broker端请求offset信息，一个Offset请求中包含多个topic-partitions，在consumer客户端中根据缓存的metadata信息区分哪些partition到哪个broker上请求，在返回中会根据不同状态反馈，如当前broker正在加载offset，则返回Loading状态。

对”offsets topic”的某些partition而言,broker状态发生改变,即被当做partition的leader或者follower时,LeaderAndIsrRequest请求会被触发:

• 如果broker是”offsets topic”中一些partitions的leader, broker会读取指定partition的logs文件,
  并将offsets加载到offset table缓存中.

  1. 任何对这些partition的提交请求仍然会更新offsets表.我们会防止日志文件中过期的offsets覆盖最近的提交请求的offsets.

  2. 被”offsets topic”中partition管理的offset抓取请求的keys直到加载成功之前是不会被使用的.
     broker会返回OffsetLoadingCode的OffsetFetchResponse给消费者客户端.

• 如果broker是follower: 和其他正常的kafka topic一样,follower会从leader中抓取数据.
  由于follower的offset manager不再负责partitions,它们会在cleanup方法被调用时清理数据.

5 kafka的文件存储

在Kafka中，消息是按Topic组织的。

• Partition：topic物理上的分组，一个topic可以划分为多个partition，每个partition是一个有序的队列。
• Segment：partition物理上由多个segment组成
• offset：每个partition都由一系列有序的、不可变的消息组成，这些消息被连续的追加到partition中。partition中每个消息都由一个连续的序列号叫做offset，用于partition唯一标识一条消息。

├── data0
│   ├── cleaner-offset-checkpoint
│   ├── client_mblogduration-35
│   │   ├── 00000000000004909731.index
│   │   ├── 00000000000004909731.log // 1G文件--Segment
│   │   ├── 00000000000005048975.index // 数字是Offset
│   │   ├── 00000000000005048975.log
│   ├── client_mblogduration-37
│   │   ├── 00000000000004955629.index
│   │   ├── 00000000000004955629.log
│   │   ├── 00000000000005098290.index
│   │   ├── 00000000000005098290.log
│   ├── __consumer_offsets-33
│   │   ├── 00000000000000105157.index
│   │   └── 00000000000000105157.log
│   ├── meta.properties
│   ├── recovery-point-offset-checkpoint
│   └── replication-offset-checkpoint

• cleaner-offset-checkpoint:存了每个log的最后清理offset
• meta.properties: broker.id信息
• recovery-point-offset-checkpoint:表示已经刷写到磁盘的记录。recoveryPoint以下的数据都是已经刷到磁盘上的了。
• replication-offset-checkpoint: 用来存储每个replica的HighWatermark的(high watermark (HW)，表示已经被commited的message，HW以下的数据都是各个replicas间同步的，一致的。)

6 Leader和Follower同步机制

举例说明，我们假设有一个topic，单分区，副本因子是2，即一个leader副本和一个follower副本。我们看下当producer发送一条消息时，broker端的副本到底会发生什么事情以及分区HW是如何被更新的。

下图是初始状态，我们稍微解释一下：初始时leader和follower的HW和LEO都是0(严格来说源代码会初始化LEO为-1，不过这不影响之后的讨论)。leader中的remote LEO指的就是leader端保存的follower LEO，也被初始化成0。此时，producer没有发送任何消息给leader，而follower已经开始不断地给leader发送FETCH请求了，但因为没有数据因此什么都不会发生。值得一提的是，follower发送过来的FETCH请求因为无数据而暂时会被寄存到leader端的purgatory中，待500ms(replica.fetch.wait.max.ms参数)超时后会强制完成。倘若在寄存期间producer端发送过来数据，那么会Kafka会自动唤醒该FETCH请求，让leader继续处理之。

KafkaC1.png

第一种情况：follower发送FETCH请求在leader处理完PRODUCE请求之后,producer给该topic分区发送了一条消息。此时的状态如下图所示:

KafkaC2.png

如图所示，leader接收到PRODUCE请求主要做两件事情：

• 把消息写入写底层log（同时也就自动地更新了leader的LEO）
• 尝试更新leader HW值（前面leader副本何时更新HW值一节中的第三个条件触发）。我们已经假设此时follower尚未发送FETCH请求，那么leader端保存的remote LEO依然是0，因此leader会比较它自己的LEO值和remote LEO值，发现最小值是0，与当前HW值相同，故不会更新分区HW值

所以，PRODUCE请求处理完成后leader端的HW值依然是0，而LEO是1，remote LEO是1。假设此时follower发送了FETCH请求(或者说follower早已发送了FETCH请求，只不过在broker的请求队列中排队)，那么状态变更如下图所示：

KafkaC3.png

本例中当follower发送FETCH请求时，leader端的处理依次是：

• 读取底层log数据
• 更新remote LEO = 0（为什么是0？ 因为此时follower还没有写入这条消息。leader如何确认follower还未写入呢？这是通过follower发来的FETCH请求中的fetch offset来确定的）
• 尝试更新分区HW——此时leader LEO = 1，remote LEO = 0，故分区HW值= min(leader LEO, follower remote LEO) = 0
• 把数据和当前分区HW值（依然是0）发送给follower副本

而follower副本接收到FETCH response后依次执行下列操作：

• 写入本地log（同时更新follower LEO）
• 更新follower HW——比较本地LEO和当前leader HW取小者，故follower HW = 0

此时，第一轮FETCH RPC结束，我们会发现虽然leader和follower都已经在log中保存了这条消息，但分区HW值尚未被更新。实际上，它是在第二轮FETCH RPC中被更新的，如下图所示：

KafkaC4.png

上图中，follower发来了第二轮FETCH请求，leader端接收到后仍然会依次执行下列操作：

• 读取底层log数据
• 更新remote LEO = 1（这次为什么是1了？ 因为这轮FETCH RPC携带的fetch offset是1，那么为什么这轮携带的就是1了呢，因为上一轮结束后follower LEO被更新为1了）
• 尝试更新分区HW——此时leader LEO = 1，remote LEO = 1，故分区HW值= min(leader LEO, follower remote LEO) = 1。注意分区HW值此时被更新了！！！
• 把数据（实际上没有数据）和当前分区HW值（已更新为1）发送给follower副本

同样地，follower副本接收到FETCH response后依次执行下列操作：

• 写入本地log，当然没东西可写，故follower LEO也不会变化，依然是1
• 更新follower HW——比较本地LEO和当前leader LEO取小者。由于此时两者都是1，故更新follower HW = 1

producer端发送消息后broker端完整的处理流程就讲完了。此时消息已经成功地被复制到leader和follower的log中且分区HW是1，表明consumer能够消费offset = 0的这条消息。

以上所有的东西其实就想说明一件事情：Kafka使用HW值来决定副本备份的进度，而HW值的更新通常需要额外一轮FETCH RPC才能完成，故而这种设计是有问题的。它们可能引起的问题包括：

• 备份数据丢失
• 备份数据不一致

这部分这里自己思考下就能够明白，不在赘述。

References

• 1.Kafka水位(high watermark)与leader epoch的讨论