Uploaded image for project: 'Flink'
  1. Flink
  2. FLINK-21436

Speed ​​up the restore of UnionListState

Attach filesAttach ScreenshotAdd voteVotersWatch issueWatchersCreate sub-taskLinkCloneUpdate Comment AuthorReplace String in CommentUpdate Comment VisibilityDelete Comments



      1. Problem introduction and cause analysis

      Problem description: The duration of UnionListState restore under large concurrency is more than 2 minutes.

      the reason:

      2000 subtasks write 2000 files during checkpoint, and each subtask needs to read 2000 files during restore.
      2000*2000 = 4 million, so 4 million small files need to be read to hdfs during restore. HDFS has become a bottleneck, causing restore to be particularly time-consuming.

      2. Optimize ideas

      Under normal circumstances, the UnionListState state is relatively small. Typical usage scenario: Kafka offset information.
      When restoring, JM can directly read all 2000 small files, merge UnionListState into a byte array and send it to all TMs to avoid frequent access to hdfs by TMs.

      3. Benefits after optimization

      Before optimization: 2000 concurrent, Kafka offset restore takes 90~130 s.
      After optimization: 2000 concurrent, Kafka offset restore takes less than 1s.

       4. Risk points

      Too big UnionListState leads to too much pressure on JM.

      Solution 1:
      Add configuration and decide whether to enable this feature. The default is false, which means the old plan is used. When the user is set to true, JM will merge.

      Solution 2:
      The above configuration is not required, which is equivalent to enabling merge by default.
      JM detects the size of the state before merge, and if it is less than the threshold, the state is considered to be relatively small, and the state is sent to all TMs through ByteStreamStateHandle.
      If the threshold is exceeded, the state is considered to be greater. At this time, write an hdfs file, and send FileStateHandle to all TMs, and TM can read this file.


      Note: Most of the scenarios where Flink uses UnionListState are Kafka offset (small state). In theory, most jobs are risk-free.




            • Assignee:
              fanrui fanrui


              • Created:

                Issue deployment