[ 
https://issues.apache.org/jira/browse/KAFKA-6699?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16606059#comment-16606059
 ] 

Matthias J. Sax commented on KAFKA-6699:
----------------------------------------

It's recommended to run with at least 3 broker, use replication factor 3, isr = 
2, and acks=all to guarantee consistency if one broker goes down.

With your setting, ie, ack=1, a write is acknowledge if the leader received it, 
but before is was replicated. Thus, if the leader broker fails, before 
replication happens, you might loose the write (even if it was acked to the 
producer client).

Also, with ISR=2, if one broker goes down, you are not able to write any 
longer. This is the reason why your Streams application starts to fail – the 
cluster cannot provide 2 ISRs, and thus rejects writes. You can either add one 
more broker, or reduce ISR to 1, to allow writes if one broker goes down.

Overall, this ticket seems to be a config issue, rather than a bug. In general, 
please use the mailing list and open tickets if you know it's a bug. If you can 
verify that changing the config/setup resolves your issue, please report back 
so we can close this ticket.

> When one of two Kafka nodes are dead, streaming API cannot handle messaging
> ---------------------------------------------------------------------------
>
>                 Key: KAFKA-6699
>                 URL: https://issues.apache.org/jira/browse/KAFKA-6699
>             Project: Kafka
>          Issue Type: Bug
>          Components: streams
>    Affects Versions: 0.11.0.2
>            Reporter: Seweryn Habdank-Wojewodzki
>            Priority: Major
>
> Dears,
> I am observing quite often, when Kafka Broker is partly dead(*), then 
> application, which uses streaming API are doing nothing.
> (*) Partly dead in my case it means that one of two Kafka nodes are out of 
> order. 
> Especially when disk is full on one machine, then Broker is going in some 
> strange state, where streaming API goes vacations. It seems like regular 
> producer/consumer API has no problem in such a case.
> Can you have a look on that matter?



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to