[
https://issues.apache.org/jira/browse/KAFKA-14374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17642135#comment-17642135
]
Matthias J. Sax commented on KAFKA-14374:
-----------------------------------------
Thanks for the details. – It would be very concerning if it turn out to be a
bug.
Did you check the content of RocksDB (using interactive queries) to see if data
is missing there? If it's missing in RocksDB too, flushing to the changelog
should not be the issue, as the data was never put into the store to begin with.
I also see in the topology visualization you added to the ticket, that there is
two filters before repartitioning – just wondering what they do and if they
drop records, how can the content of the input topic be the same as the content
of the result topic?
Last, after data from the repartition topic was processed downstream, we would
issue a "purge" request and delete records. How do you ensure that the
repartition topic is really complete?
Are there any deletes happening? Note that the repartition topic is configured
with "log retention", while the changelog topic is configured with "log
compaction" – thus, the repartition topic could contain two record `<k1,v1>`
and `<k1,null>` (ie k1 exists physically but is actually logically deleted),
while the changelog topic might have been compacted already and the key k1 was
purged from it.
> Kafka streams losing messages in State Store during first launch of app
> -----------------------------------------------------------------------
>
> Key: KAFKA-14374
> URL: https://issues.apache.org/jira/browse/KAFKA-14374
> Project: Kafka
> Issue Type: Bug
> Components: streams
> Affects Versions: 2.8.0, 3.2.0
> Reporter: Youcef Sebiat
> Priority: Major
> Attachments: Screenshot 2022-11-09 at 14.56.00.png
>
>
> We have been using Kafka Streams to implement a CDC based app. Attached is
> the sub topology of interest.
> `table2` topic is created by Debezium who is connected to a SQL DB. It
> contains 26K lines. We take `table2` and create a key which is only a
> conversion of the key from `string` to `int`. This means that we should
> expect that #table2=#repartition-topic=#state-store; which actually is not
> verified. What we end up with is the following #table2=#repartition-topic,
> but #repartition-topic>#state-store. We actually lose messages and thus
> corrupt the state store, which makes the app live in incorrect state. (Please
> note that there is no insertion in `table2` as we paused the connector to
> verify the cardinality.)
> The above happens only during the first launch, i.e. the app has never been
> launched before, so internal topics do not exist yet. Restarts of
> pre-existing apps do not yield any problems.
> We have:
> 1. Broker on Kafka 3.2.
> 2. Client app on 2.8|3.2 (we tried both and we faced the same issue).
> 2. All parameters are by default except `CACHE_MAX_BYTES_BUFFERING_CONFIG`
> set to `0` and `NUM_STREAM_THREADS_CONFIG` set to `>1`.
>
> *What actually worked*
> 1. Use a monothread at first launch: using one thread solves the problem. The
> #table2=#repartition-topic=#state-store is verified.
> 2. Pre-creating kafka internal topics: we noticed that whenever there is
> rebalance during the first launch of Kafka Streams app, the state stores
> ended up missing values. This also happens when you launch multiple pods in
> K8s for example. When we read through the logs, we noticed that there is a
> rebalance that is triggered when we first launch the app. This comes from the
> fact that the internal topics get created and assigned, thus the rebalance.
> So by creating the internal topics before, we avoid the rebalance and we end
> up by #table2=#repartition-topic=#state-store.
> *What we noticed from the logs*
> On multi-thread mode, we noticed that it is the partitions that are assigned
> to the thread chosen by the Coordinator to be the Leader of consumers that
> suffer the data loss. What we think is happening is the following:
> 1. Consumers threads are launched and inform the coordinator.
> 2. Coordinator assign topics and choses the Leader among the threads.
> 3. App create internal topics.
> 4. Consumers/producers process data. Specifically the Consumer leader
> consumes from the repartition topic, which triggers the delete of those
> messages without flushing them to changelog topic.
> 5. Leader notified of new assignment with internal topics. Triggers rebalance.
> 6. Leader pauses partitions.
> 7. Rebalance finished. The leader resumes partitions.
> 8. Leader fetches the oldest offset of repartition partitions he got
> assigned. He will not start from zero, but instead from where he got
> interrupted in 4. The chunk of early messages are thus lost.
> Please note, that on mono-thread mode, there is no data loss which is weird
> since the leader is actually the unique thread.
> So my questions are:
> 1. Are we understanding wrongly what's happening?
> 2. What can be the origin of this problem?
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
