check for stale cluster info before retrying

[sarcilav]

So we keep stale data from the cluster even when we try to reread the leader.

so basically #get_leader keeps getting the same error over on NotLeaderForPartition, when leader rebalance happens and we are sending a batch, and it wasn't triggering a reread.

[ptrykov]

could we try this approach instead https://github.com/zendesk/ruby-kafka/pull/901/files (back port from main repo) and see if this will solve an issue. (Just to be closer to the main repo)

[sarcilav]

I think we will need both, but yeah, good call

[ptrykov]

Hey @fcbr, before updating Catbus gem we want to monitor #3 to check if that’s is enough to fix the rebalancing issue.

[dhruvparanjape]

@ptrykov & @fcbr is this something we should work on ? could you perhaps share more details on what's happening and how to reproduce the issue ?

[sarcilav]

@dhruvCW the issue is more evident if you spawn more than one consumer process, as you start dealing with more changes in the state of the cluster, which will create more chances for the cluster data in the clients to go stale, and basically this change tries to refresh the cluster metadata before retrying after an error.