You are viewing a plain text version of this content. The canonical link for it is here.
Posted to users@kafka.apache.org by Mohit Kathuria <mk...@sprinklr.com> on 2014/11/03 14:37:29 UTC

Fwd: Rebalance not happening even after increasing max retries causing conflict in ZK

Neha,

Looks like an issue with the consumer rebalance not able to complete
successfully. We were able to reproduce the issue on topic with 30
partitions,  3 consumer processes(p1,p2 and p3), properties -  40
rebalance.max.retries and 10000(10s) rebalance.backoff.ms.

Before the process p3 was started, partition ownership was as expected:

partitions 0-14 owned by p1
partitions 15-29 -> owner p2

As the process p3 started, rebalance was triggered. Process p3 was
successfully able to acquire partition ownership for partitions 20-29 as
expected as per the rebalance algorithm. However, process p2 while trying
to acquire ownership of partitions 10-19 saw rebalance failure after 40
retries.

Attaching the logs from process p2 and process p1. It says that p2 was
attempting to rebalance, it was trying to acquire ownership of partitions
10-14 which were owned by process p1. However, at the same time process p1
did not get any event for giving up the partition ownership for partitions
1-14.
We were expecting a rebalance to have triggered in p1 - but it didn't and
hence not giving up ownership. Is our assumption correct/incorrect?
And if the rebalance gets triggered in p1 - how to figure out apart from
logs as the logs on p1 did not have anything.

*2014-11-03 06:57:36 k.c.ZookeeperConsumerConnector [INFO]
[topic_consumerIdString], waiting for the partition ownership to be
deleted: 11*

During and after the rebalance failed on process p2, Partition Ownership
was as below:
0-14 -> owner p1
15-19 -> none
20-29 -> owner p3

This left the consumers in inconsistent state as 5 partitions were never
consumer from and neither was the partitions ownership balanced.

However, there was no conflict in creating the ephemeral node which was the
case last time. Just to note that the ephemeral node conflict which we were
seeing earlier also appeared after rebalance failed. My hunch is that
fixing the rebalance failure will fix that issue as well.

-Thanks,
Mohit