[
https://issues.apache.org/jira/browse/KAFKA-19622?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18096894#comment-18096894
]
David Arthur commented on KAFKA-19622:
--------------------------------------
[~gargsha] we will look into adding some docs around this.
In general, I would not recommend running in dual-write mode for _too_ long.
The mode with the most risk is hybrid mode when you have ZK brokers but a KRaft
controller. Once the brokers have come up as KRaft (dual write mode), things
are pretty much the same as they would be in a pure-KRaft system.
I would suggest only being in hybrid mode as long as it takes to restart the
brokers as KRaft. In other words, the least amount of time as necessary. You
can stay in dual write mode for a few hours, days, or weeks. It's really up to
you. Just note that once you finalize the migration, there is no going back.
The validations that you perform are up to you. Generally, we like to see that
client traffic is unaffected and broker health is stable.
I can share an anecdote from when I was migrating a lot of clusters back in
2024. We rolled through hybrid mode as quickly as possible. After triggering
the migration, we would immediately start restarting brokers as KRaft. Once in
dual write mode, we would leave clusters there for at least one week before
doing the next restart to finalize the migration. While in dual write mode, we
would check that client throughput was normal and that we didn't have things
like sustained URPs or under min ISR partitions.
> Limitations of KRAFT Dual Write Mode for Production Support
> -----------------------------------------------------------
>
> Key: KAFKA-19622
> URL: https://issues.apache.org/jira/browse/KAFKA-19622
> Project: Kafka
> Issue Type: Task
> Components: kraft
> Affects Versions: 3.9.0
> Reporter: Sharad Garg
> Priority: Minor
>
> We are currently running Kafka version 3.9.0 and are in the process of
> migrating to KRaft. As part of the migration, we intend to operate in
> dual-write mode in production for an initial period to help identify and
> address any issues.
> Are there any known limitations or risks associated with running in
> dual-write mode? Would you recommend maintaining this mode for production
> stability, and are there best practices we should follow?
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
