Thank you for your reply The proposed load balancer is designed to distribute partitions uniformly across all brokers, assuming homogeneous hardware and similar partition workloads (i.e., CPU, network, and disk I/O demands are consistent). Under these conditions, it guarantees globally balanced partition distribution.
This aligns with the goals of Kafka’s current built-in load balancer, but the newly proposed rebalancing algorithm introduces key improvements described in the KIP. 在 2025年4月23日星期三，Hector Geraldino (BLOOMBERG/ 919 3RD A) < hgerald...@bloomberg.net> 写道： > Thanks for this KIP (and the companion PR) > > This is an interesting proposal. I have to admit that I've only used this > tool when I want to force a specific topic/ReplicaSet distribution. > > I wonder what criteria you used when deciding the set of goals and their > order. Reason I ask is because we use MirrorMaker extensively to balance > our clusters, and not only it offers lots of different goals, but also > allows operators to decide which ones to optimize for. > > My worry is that having a fixed/opinionated criteria on how replicas are > distributed can be limiting. > > From: dev@kafka.apache.org At: 03/28/25 23:04:50 UTC-4:00To: > dev@kafka.apache.org > Subject: [DISCUSS] KIP-1151: Minimal Movement Replica Balancing Algorithm > for Reassignment > > Hi Kafka Community, > > I would like to start a discussion on a new KIP proposal: KIP-1151: Minimal > movement replica balancing algorithm for reassignment > > Motivation > > Kafka clusters require rebalancing of topic replicas after horizontal > scaling to evenly distribute the load. However, the current approach does > not consider the existing replica distribution, often leading to > unnecessary movements. This KIP proposes an optimized strategy that > minimizes movements while achieving an even distribution. > > Proposal > > The proposal focuses on: > > 1. Minimal Movement: Minimize the number of replica relocations during > rebalancing. > 2. Replica Balancing: Ensure that replicas are evenly distributed across > brokers. > 3. Anti-Affinity Support: Support rack-aware allocation when enabled. > 4. Leader Balancing: Distribute leader replicas evenly across brokers. > 5.ISR Order Optimization: Optimize adjacency relationships to prevent > failover traffic concentration in case of broker failures. > > For a detailed description, please refer to the KIP document: > > KIP-1151: > https://cwiki.apache.org/confluence/display/KAFKA/KIP- > 1151%3A+Minimal+movement+r > eplica+balancing+algorithm+for+reassignment > > > I would appreciate any feedback, comments, or suggestions from the > community. > > Best > > >
