kafka-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Maysam Yabandeh (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (KAFKA-3693) Race condition between highwatermark-checkpoint thread and handleLeaderAndIsrRequest at broker start-up
Date Tue, 07 Jun 2016 20:12:20 GMT

    [ https://issues.apache.org/jira/browse/KAFKA-3693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15319273#comment-15319273
] 

Maysam Yabandeh commented on KAFKA-3693:
----------------------------------------

Thanks for the grep hints [~junrao]. According to controller logs, it did not label broker
16 as dead:
{code}
$ grep "Newly added brokers" controller.log.2016-05-10.2 
2016-05-10 05:58:30,633 INFO controller.ReplicaStateMachine$BrokerChangeListener: [BrokerChangeListener
on Controller 17]: Newly added brokers: 18, deleted brokers: , all live brokers: 14,20,21,13,17,22,18,16,19,15
2016-05-10 06:17:48,512 INFO controller.ReplicaStateMachine$BrokerChangeListener: [BrokerChangeListener
on Controller 17]: Newly added brokers: , deleted brokers: 14,21,13,22,18,19, all live brokers:
15,16
2016-05-10 06:29:57,981 INFO controller.ReplicaStateMachine$BrokerChangeListener: [BrokerChangeListener
on Controller 17]: Newly added brokers: 14,21,13,22,18,19, deleted brokers: , all live brokers:
14,21,13,22,18,16,19,15
{code}

I am still not sure what is the mechanism by which the controller is supposed to mark the
broker dead but it seems that the controller is started AFTER the broker 16 has started shutting
down, so if this mechanism depends on a message from the broker to the controller it might
have not received by the new controller:
{code}
$ grep "Controller starting up\|Controller-17-to-broker-16-send-thread" controller.log.2016-05-10.2

2016-05-10 05:50:20,645 INFO controller.KafkaController: [Controller 17]: Controller starting
up
2016-05-10 05:57:27,440 INFO controller.RequestSendThread: [Controller-17-to-broker-16-send-thread],
Starting 
2016-05-10 05:57:28,309 INFO controller.RequestSendThread: [Controller-17-to-broker-16-send-thread],
Controller 17 connected to Node(16, node16.com, 9092) for sending state change requests
2016-05-10 06:17:01,771 WARN controller.RequestSendThread: [Controller-17-to-broker-16-send-thread],
Controller 17 epoch 269 fails to send request ... to broker Node(16, node16.com, 9092). Reconnecting
to broker.
2016-05-10 06:17:02,076 WARN controller.RequestSendThread: [Controller-17-to-broker-16-send-thread],
Controller 17's connection to broker Node(16, node16.com, 9092) was unsuccessful
...
2016-05-10 06:17:33,043 WARN controller.RequestSendThread: [Controller-17-to-broker-16-send-thread],
Controller 17's connection to broker Node(16, node16.com, 9092) was unsuccessful
2016-05-10 06:17:33,344 INFO controller.RequestSendThread: [Controller-17-to-broker-16-send-thread],
Controller 17 connected to Node(16, node16.com, 9092) for sending state change requests
{code}
Regarding the slow shutdown, that is also a question for us but investigating it had less
priority compared to losing hw checkpoints.

> Race condition between highwatermark-checkpoint thread and handleLeaderAndIsrRequest
at broker start-up
> -------------------------------------------------------------------------------------------------------
>
>                 Key: KAFKA-3693
>                 URL: https://issues.apache.org/jira/browse/KAFKA-3693
>             Project: Kafka
>          Issue Type: Bug
>          Components: core
>    Affects Versions: 0.9.0.1
>            Reporter: Maysam Yabandeh
>
> Upon broker start-up, a race between highwatermark-checkpoint thread to write replication-offset-checkpoint
file and handleLeaderAndIsrRequest thread reading from it causes the highwatermark for some
partitions to be reset to 0. In the good case, this results the replica to truncate its entire
log to 0 and hence initiates fetching of terabytes of data from the lead broker, which sometimes
leads to hours of downtime. We observed the bad cases that the reset offset can propagate
to recovery-point-offset-checkpoint file, making a lead broker to truncate the file. This
seems to have the potential to lead to data loss if the truncation happens at both follower
and leader brokers.
> This is the particular faulty scenario manifested in our tests:
> # The broker restarts and receive LeaderAndIsr from the controller
> # LeaderAndIsr message however does not contain all the partitions (probably because
other brokers were churning at the same time)
> # becomeLeaderOrFollower calls getOrCreatePartition and updates the allPartitions with
the partitions included in the LeaderAndIsr message {code}
>   def getOrCreatePartition(topic: String, partitionId: Int): Partition = {
>     var partition = allPartitions.get((topic, partitionId))
>     if (partition == null) {
>       allPartitions.putIfNotExists((topic, partitionId), new Partition(topic, partitionId,
time, this))
> {code}
> # replication-offset-checkpoint jumps in taking a snapshot of (the partial) allReplicas'
high watermark into replication-offset-checkpoint file {code}  def checkpointHighWatermarks()
{
>     val replicas = allPartitions.values.map(_.getReplica(config.brokerId)).collect{case
Some(replica) => replica}{code} hence rewriting the previous highwatermarks.
> # Later becomeLeaderOrFollower calls makeLeaders and makeFollowers which read the (now
partial) file through Partition::getOrCreateReplica {code}
>           val checkpoint = replicaManager.highWatermarkCheckpoints(log.dir.getParentFile.getAbsolutePath)
>           val offsetMap = checkpoint.read
>           if (!offsetMap.contains(TopicAndPartition(topic, partitionId)))
>             info("No checkpointed highwatermark is found for partition [%s,%d]".format(topic,
partitionId))
> {code}
> We are not entirely sure whether the initial LeaderAndIsr message including a subset
of partitions is critical in making this race condition manifest or not. But it is an important
detail since it clarifies that a solution based on not letting the highwatermark-checkpoint
thread jumping in the middle of processing a LeaderAndIsr message would not suffice.
> The solution we are thinking of is to force initializing allPartitions by the partitions
listed in the replication-offset-checkpoint (and perhaps recovery-point-offset-checkpoint
file too) when a server starts.
> Thoughts?



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message