Hi Hector,

Thanks for your comment.

In my experience, I've managed a few 2 rack/location clusters where the RF was 
big enough to cover a site failure + 1 down node on the remaining site which 
leads to a some pretty big RF

The only way to guarantee site resilient data is to either do manual placement 
or set the RF to be number of brokers in a rack + 1

To be fair I don't think i've ever seen replica placement issues to make it 
necessary but i've managed clusters in places where they were really worried 
about data loss.

This KIP is intended to allow you to set a sensible RF and still have a 
deterministic guarantee that writes span multiple racks

Karl

________________________________
From: Hector Geraldino (BLOOMBERG/ 919 3RD A) <[email protected]>
Sent: 05 March 2026 15:36
To: [email protected] <[email protected]>
Subject: Re: [DISCUSS] KIP-1290: Rack-Aware Minimum In-Sync Replicas

Hi Karl, thanks for your KIP.

I'm curious to know if, in your experience (or anyone else's), this is 
something that actually happens.

I'd imagine it is pretty uncommon to configure topics to have RF > number of 
racks, however KIP-36 does call out this possibility. In my view, it kinda 
defeats the purpose of having racks/AZs in the first place, but maybe this is 
something that does happens and I'm just unaware of.

cheers,

From: [email protected] At: 03/05/26 08:40:12 UTC-5:00To:  
[email protected]
Subject: Re: [DISCUSS] KIP-1290: Rack-Aware Minimum In-Sync Replicas

Hi,

I haven't had any replies on this so just bumping.

I've got a patch that I've tested.

Thanks
Karl


________________________________
From: Karl Sorensen <[email protected]>
Sent: 25 February 2026 11:27
To: [email protected] <[email protected]>
Subject: [DISCUSS] KIP-1290: Rack-Aware Minimum In-Sync Replicas

Hi,

I'd like to start a discussion about KIP-1290
https://cwiki.apache.org/confluence/x/ao48G

This proposes adding a new config min.insync.racks to ensure producers write to
replicas spread across multiple racks.

Let me know if you have any questions.

Thanks
Karl


Reply via email to