[KAFKA-6870] Concurrency conflicts in SampledStat

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 1.1.1, 2.0.0
• Assignee: Chia-Ping Tsai
• Reporter: Chia-Ping Tsai
• Reviewer: Rajini Sivaram
• Created: 06/May/18 05:48
• Updated: 01/Jun/18 19:05
• Resolved: 10/May/18 13:29

Description

The samples stored in SampledStat is not thread-safe. However, ReplicaFetcherThreads used to handle replica to specified brokers may update (when the samples is empty, we will add a new sample to it) and iterate the samples concurrently, and then cause the ConcurrentModificationException.

[2018-05-03 13:50:56,087] ERROR [ReplicaFetcher replicaId=106, leaderId=100, fetcherId=0] Error due to (kafka.server.ReplicaFetcherThread:76)
java.util.ConcurrentModificationException
        at java.util.ArrayList$Itr.checkForComodification(ArrayList.java:909)
        at java.util.ArrayList$Itr.next(ArrayList.java:859)
        at org.apache.kafka.common.metrics.stats.Rate$SampledTotal.combine(Rate.java:132)
        at org.apache.kafka.common.metrics.stats.SampledStat.measure(SampledStat.java:78)
        at org.apache.kafka.common.metrics.stats.Rate.measure(Rate.java:66)
        at org.apache.kafka.common.metrics.KafkaMetric.measurableValue(KafkaMetric.java:85)
        at org.apache.kafka.common.metrics.Sensor.checkQuotas(Sensor.java:201)
        at org.apache.kafka.common.metrics.Sensor.checkQuotas(Sensor.java:192)
        at kafka.server.ReplicationQuotaManager.isQuotaExceeded(ReplicationQuotaManager.scala:104)
        at kafka.server.ReplicaFetcherThread.kafka$server$ReplicaFetcherThread$$shouldFollowerThrottle(ReplicaFetcherThread.scala:384)
        at kafka.server.ReplicaFetcherThread$$anonfun$buildFetchRequest$1.apply(ReplicaFetcherThread.scala:263)
        at kafka.server.ReplicaFetcherThread$$anonfun$buildFetchRequest$1.apply(ReplicaFetcherThread.scala:261)
        at scala.collection.mutable.ResizableArray$class.foreach(ResizableArray.scala:59)
        at scala.collection.mutable.ArrayBuffer.foreach(ArrayBuffer.scala:48)
        at kafka.server.ReplicaFetcherThread.buildFetchRequest(ReplicaFetcherThread.scala:261)
        at kafka.server.AbstractFetcherThread$$anonfun$2.apply(AbstractFetcherThread.scala:102)
        at kafka.server.AbstractFetcherThread$$anonfun$2.apply(AbstractFetcherThread.scala:101)
        at kafka.utils.CoreUtils$.inLock(CoreUtils.scala:250)
        at kafka.server.AbstractFetcherThread.doWork(AbstractFetcherThread.scala:101)
        at kafka.utils.ShutdownableThread.run(ShutdownableThread.scala:82)

Before https://github.com/apache/kafka/commit/d734f4e56d276f84b8c52b602edd67d41cbb6c35 the ConcurrentModificationException doesn't exist since all changes to samples is "add" currently. Using the get(index) is able to avoid the ConcurrentModificationException.

In short, we can just make samples thread-safe. Or just replace the foreach loop by get(index) if we have concerns about the performance of thread-safe list...

 

 

 

 

Issue Links

• links to: GitHub Pull Request #4985