[KAFKA-1382] Update zkVersion on partition state update failures

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.8.1.2, 0.8.2.0
• Assignee: Harsha
• Reporter: Joel Jacob Koshy
• Created: 10/Apr/14 01:00
• Updated: 27/Apr/16 23:25
• Resolved: 10/Jun/14 21:56

Description

Our updateIsr code is currently:

private def updateIsr(newIsr: Set[Replica]) {
debug("Updated ISR for partition [%s,%d] to %s".format(topic, partitionId, newIsr.mkString(",")))
val newLeaderAndIsr = new LeaderAndIsr(localBrokerId, leaderEpoch, newIsr.map(r => r.brokerId).toList, zkVersion)
// use the epoch of the controller that made the leadership decision, instead of the current controller epoch
val (updateSucceeded, newVersion) = ZkUtils.conditionalUpdatePersistentPath(zkClient,
ZkUtils.getTopicPartitionLeaderAndIsrPath(topic, partitionId),
ZkUtils.leaderAndIsrZkData(newLeaderAndIsr, controllerEpoch), zkVersion)
if (updateSucceeded)

{ inSyncReplicas = newIsr zkVersion = newVersion trace("ISR updated to [%s] and zkVersion updated to [%d]".format(newIsr.mkString(","), zkVersion)) }

else

{ info("Cached zkVersion [%d] not equal to that in zookeeper, skip updating ISR".format(zkVersion)) }

We encountered an interesting scenario recently when a large producer fully
saturated the broker's NIC for over an hour. The large volume of data led to
a number of ISR shrinks (and subsequent expands). The NIC saturation
affected the zookeeper client heartbeats and led to a session timeout. The
timeline was roughly as follows:

• Attempt to expand ISR
• Expansion written to zookeeper (confirmed in zookeeper transaction logs)
• Session timeout after around 13 seconds (the configured timeout is 20
  seconds) so that lines up.
• zkclient reconnects to zookeeper (with the same session ID) and retries
  the write - but uses the old zkVersion. This fails because the zkVersion
  has already been updated (above).
• The ISR expand keeps failing after that and the only way to get out of it
  is to bounce the broker.

In the above code, if the zkVersion is different we should probably update
the cached version and even retry the expansion until it succeeds.

Issue Links

• relates to: KAFKA-3042 updateIsr should stop after failed several times due to zkVersion issue