Kafka GroupCoordinator机制(三)：GroupMetadataManager之groupsCache管理和offsetsCache管理

记录groupMetadata信息的消息和记录消费offset位置的消息都是通过partitionFor()选择合适的分区，他们的会分到同一个__CONSUMER_OFFSETS中。
GroupMetadataManager提供了对于groupsCache集合的管理方法，这里特别介绍removeGroup。
把GroupMetadata对象删除，向__CONSUMER_OFFSET中写入一个value为空的消息作为删除标志。

  def removeGroup(group: GroupMetadata) {
    // guard this removal in case of concurrent access (e.g. if a delayed join completes with no members
    // while the group is being removed due to coordinator emigration)
    // 删除groupCache中的GroupMetadata
    if (groupsCache.remove(group.groupId, group)) {
      // Append the tombstone messages to the partition. It is okay if the replicas don't receive these (say,
      // if we crash or leaders move) since the new leaders will still expire the consumers with heartbeat and
      // retry removing this group.
      // 获取consumerGroup在__CONSUMER_OFFSET中对应的分区ID
      val groupPartition = partitionFor(group.groupId)
      val (magicValue, timestamp) = getMessageFormatVersionAndTimestamp(groupPartition)
      // 产生删除标志消息，注意value是null，key由groupID封装而来
      val tombstone = new Message(bytes = null, key = GroupMetadataManager.groupMetadataKey(group.groupId),
        timestamp = timestamp, magicValue = magicValue)
        // 获取offsets Topic对应的Partition对象
      val partitionOpt = replicaManager.getPartition(TopicConstants.GROUP_METADATA_TOPIC_NAME, groupPartition)
      partitionOpt.foreach { partition =>
        val appendPartition = TopicAndPartition(TopicConstants.GROUP_METADATA_TOPIC_NAME, groupPartition)

        trace("Marking group %s as deleted.".format(group.groupId))

        try {
          // do not need to require acks since even if the tombstone is lost,
          // it will be appended again by the new leader
          // TODO KAFKA-2720: periodic purging instead of immediate removal of groups
           //写入消息
          partition.appendMessagesToLeader(new ByteBufferMessageSet(config.offsetsTopicCompressionCodec, tombstone))
        } catch {
          case t: Throwable =>
            error("Failed to mark group %s as deleted in %s.".format(group.groupId, appendPartition), t)
          // ignore and continue
        }
      }
    }
  }

GroupMetadataManager在初始化时会创建并且启动一个KafkaScheduler对象，周期性地调用的deleteExpiredOffsets()方法
deleteExpiredOffsets除了删除offsetsCache集合中对应的OffsetMetadata对象，还需要向__CONSUMER_OFFSET追加删除标志信息。

  private def deleteExpiredOffsets() {
    debug("Collecting expired offsets.")
    val startMs = time.milliseconds()
    // 过滤得到所有的过期的offsetAndMetadata
    val numExpiredOffsetsRemoved = inWriteLock(offsetExpireLock) {
      val expiredOffsets = offsetsCache.filter { case (groupTopicPartition, offsetAndMetadata) =>
        offsetAndMetadata.expireTimestamp < startMs
      }

      debug("Found %d expired offsets.".format(expiredOffsets.size))

      // delete the expired offsets from the table and generate tombstone messages to remove them from the log
      val tombstonesForPartition = expiredOffsets.map { case (groupTopicAndPartition, offsetAndMetadata) =>
        // 找到该groupID对应的储存offset消息的offsettopic分区的id
        val offsetsPartition = partitionFor(groupTopicAndPartition.group)
        trace("Removing expired offset and metadata for %s: %s".format(groupTopicAndPartition, offsetAndMetadata))
        // 删除对应的offsetAndMetadata对象
        offsetsCache.remove(groupTopicAndPartition)
        // 获取消息的key
        val commitKey = GroupMetadataManager.offsetCommitKey(groupTopicAndPartition.group,
          groupTopicAndPartition.topicPartition.topic, groupTopicAndPartition.topicPartition.partition)
        // 获取分区使用的magic和时间戳
        val (magicValue, timestamp) = getMessageFormatVersionAndTimestamp(offsetsPartition)
        // 返回的是对应partition的ID和一个删除标志消息
        (offsetsPartition, new Message(bytes = null, key = commitKey, timestamp = timestamp, magicValue = magicValue))
      }.groupBy { case (partition, tombstone) => partition }

      // Append the tombstone messages to the offset partitions. It is okay if the replicas don't receive these (say,
      // if we crash or leaders move) since the new leaders will get rid of expired offsets during their own purge cycles.
      // __CONSUMER_OFFSET的一个分区可能记录多个group的offset信息，为了便于追加，按照offsets Topic的分区id进行一次分组
      tombstonesForPartition.flatMap { case (offsetsPartition, tombstones) =>
        val partitionOpt = replicaManager.getPartition(TopicConstants.GROUP_METADATA_TOPIC_NAME, offsetsPartition)
        partitionOpt.map { partition =>
          val appendPartition = TopicAndPartition(TopicConstants.GROUP_METADATA_TOPIC_NAME, offsetsPartition)
          // 获取要追加到此offsettopic分区中的删除标志消息集合
          val messages = tombstones.map(_._2).toSeq

          trace("Marked %d offsets in %s for deletion.".format(messages.size, appendPartition))

          try {
            // do not need to require acks since even if the tombstone is lost,
            // it will be appended again in the next purge cycle
            // 追加删除标志信息
            partition.appendMessagesToLeader(new ByteBufferMessageSet(config.offsetsTopicCompressionCodec, messages: _*))
            tombstones.size
          }
          catch {
            case t: Throwable =>
              error("Failed to mark %d expired offsets for deletion in %s.".format(messages.size, appendPartition), t)
              // ignore and continue
              0
          }
        }
      }.sum
    }