Uploaded image for project: 'Kafka'
  1. Kafka
  2. KAFKA-4959

remove controller concurrent access to non-threadsafe NetworkClient, Selector, and SSLEngine

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


    • Bug
    • Status: Resolved
    • Major
    • Resolution: Fixed
    • None
    • None
    • None


      This brought down a cluster by causing continuous controller moves.

      ZkClient's ZkEventThread and a RequestSendThread can concurrently use objects that aren't thread-safe:

      • Selector
      • NetworkClient
      • SSLEngine (this was the big one for us. We turn on SSL for interbroker communication).

      As per the "Concurrency Notes" section from the SSLEngine javadoc:

      two threads must not attempt to call the same method (either wrap() or unwrap()) concurrently

      SSLEngine.wrap gets called in:

      • SslTransportLayer.write
      • SslTransportLayer.handshake
      • SslTransportLayer.close

      It turns out that the ZkEventThread and RequestSendThread can concurrently call SSLEngine.wrap:

      • ZkEventThread calls SslTransportLayer.close from ControllerChannelManager.removeExistingBroker
      • RequestSendThread can call SslTransportLayer.write or SslTransportLayer.handshake from NetworkClient.poll

      Suppose the controller moves for whatever reason. The former controller could have had a RequestSendThread who was in the middle of sending out messages to the cluster while the ZkEventThread began executing KafkaController.onControllerResignation, which calls ControllerChannelManager.shutdown, which sequentially cleans up the controller-to-broker queue and connection for every broker in the cluster. This cleanup includes the call to ControllerChannelManager.removeExistingBroker as mentioned earlier, causing the concurrent call to SSLEngine.wrap. This concurrent call throws a BufferOverflowException which ControllerChannelManager.removeExistingBroker catches so the ControllerChannelManager.shutdown moves onto cleaning up the next controller-to-broker queue and connection, skipping the cleanup steps such as clearing the queue, stopping the RequestSendThread, and removing the entry from its brokerStateInfo map.

      By failing out of the Selector.close, the sensors corresponding to the broker connection has not been cleaned up. Any later attempt at initializing an identical Selector will result in a sensor collision and therefore cause Selector initialization to throw an exception. In other words, any later attempts by this broker to become controller again will fail on initialization. When controller initialization fails, the controller deletes the /controller znode and lets another broker take over.

      Now suppose the controller moves enough times such that every broker hits the BufferOverflowException concurrency issue. We're now guaranteed to fail controller initialization due to the sensor collision on every controller transition, so the controller will move across brokers continuously.


        Issue Links


          This comment will be Viewable by All Users Viewable by All Users


            onurkaraman Onur Karaman
            onurkaraman Onur Karaman
            0 Vote for this issue
            2 Start watching this issue




                Issue deployment