[
https://issues.apache.org/jira/browse/KAFKA-16514?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17839240#comment-17839240
]
Sal Sorrentino commented on KAFKA-16514:
----------------------------------------
I spent a little time looking at the code, and I really believe this is a
Streams implementation issue. Every member of a consumer group has some sort of
"MemberId" associated with it, regardless of whether or not it is considered
static. It must, otherwise basic functions such as rebalancing could not work.
If the static group id is not used, some randomly generated member id must be,
and the real issue is that this identifier is not be tracked in the stream task.
That being said, the current implementation uses Thread scoped variables to
store the "group.instance.id" and to really fix this, the consumer group member
id would need to be stored there as well. One quick fix would be to
auto-generate this id. It seems to me that if there the existence of a group
instance id is necessary to drive fundamental consumer group logic, that one
should be generated if not supplied by the user. The scope of that sort of
change is a bit large and not to be taken lightly.
I would recommend, that in the short term, we add a `GROUP_INSTANCE_ID` public
static variable to the StreamsConfig class and document thoroughly the logic
that this config drives.
Does that make sense?
> Kafka Streams: stream.close(CloseOptions) does not respect options.leaveGroup
> flag.
> -----------------------------------------------------------------------------------
>
> Key: KAFKA-16514
> URL: https://issues.apache.org/jira/browse/KAFKA-16514
> Project: Kafka
> Issue Type: Bug
> Components: streams
> Affects Versions: 3.7.0
> Reporter: Sal Sorrentino
> Priority: Minor
>
> Working with Kafka Streams 3.7.0, but may affect earlier versions as well.
> When attempting to shutdown a streams application and leave the associated
> consumer group, the supplied `leaveGroup` option seems to have no effect.
> Sample code:
> {code:java}
> CloseOptions options = new CloseOptions().leaveGroup(true);
> stream.close(options);{code}
> The expected behavior here is that the group member would shutdown and leave
> the group, immediately triggering a consumer group rebalance. In practice,
> the rebalance happens after the appropriate timeout configuration has expired.
> I understand the default behavior in that there is an assumption that any
> associated StateStores would be persisted to disk and that in the case of a
> rolling restart/deployment, the rebalance delay may be preferable. However,
> in our application we are using in-memory state stores and standby replicas.
> There is no benefit in delaying the rebalance in this setup and we are in
> need of a way to force a member to leave the group when shutting down.
> The workaround we found is to set an undocumented internal StreamConfig to
> enforce this behavior:
> {code:java}
> props.put("internal.leave.group.on.close", true);
> {code}
> To state the obvious, this is less than ideal.
> Additional configuration details:
> {code:java}
> Properties props = new Properties();
> props.put(StreamsConfig.APPLICATION_ID_CONFIG, "someApplicationId");
> props.put(
> StreamsConfig.BOOTSTRAP_SERVERS_CONFIG,
> "localhost:9092,localhost:9093,localhost:9094");
> props.put(StreamsConfig.REPLICATION_FACTOR_CONFIG, 3);
> props.put(StreamsConfig.NUM_STANDBY_REPLICAS_CONFIG, 1);
> props.put(StreamsConfig.NUM_STREAM_THREADS_CONFIG, numProcessors);
> props.put(StreamsConfig.PROCESSING_GUARANTEE_CONFIG,
> StreamsConfig.EXACTLY_ONCE_V2);{code}
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
