This page is meant as a template for writing a KIP. To create a KIP choose Tools->Copy on this page and modify with your content and replace the heading with the next KIP number and a description of your issue. Replace anything in italics with your own description.

Status

Current state: DraftDiscarded

Discussion thread: here [Change the link from the KIP proposal email archive to your own email thread]

...

When creating topics or partitions, the Kafka controller has to pick brokers to host the new partitions. The current placement logic is based on a round robin algorithm and supports rack awareness. While this works relatively well in many scenarios, in a few cases the placements it generates are not optimal because it's not aware of the state of the clusters. Many cluster administrators then rely on tools like Cruise Control to move partitions to better brokers. This process is expensive as often data has to be copied between brokers.

It would be desirable to allow custom logic for the placer to leverage the state of the cluster to enable administrators to build assignment rules for their clusters and minimize the number of partition reassignments necessary. It would enable administrators to build assignment goals (similar to Cruise Control goals) for their clusters.

Some scenarios that could benefit greatly from this feature:

• When adding brokers to a cluster, Kafka currently does not necessarily place new partitions on new brokers
• When removing administrators want to remove brokers from a cluster, as Kafka currently will keep there is no way to prevent Kafka from placing partitions on all existing brokersthem
• When some brokers are near their storage/throughput limit, the assignor Kafka could avoid putting new partitions on them

...

package org.apache.kafka.server.placer;

/**
 * The interface which a Kafka replica placement policy must implement.
 */
@InterfaceStability.Evolving
public
interface ReplicaPlacer extends Configurable, Closeable {
    /**
     * Create a new replica placement.
     *
     * @param startPartition        The partition ID to start with.
     * @param numPartitions         The number of partitions to create placements for.
     * @param numReplicas           The number of replicas to create for each partition.
     * @param iterator              An iterator that yields all the usable brokers.
     *
     * @return                      A list of replica lists.
     *
     * @throws InvalidReplicationFactorException    If too many replicas were requested.
     * @throws ReplicaPlacementException            If a new replica placement can't be created
     */
    List<List<Integer>> place(int startPartition,
                              int numPartitions,
                              short numReplicas,
                              Iterator<UsableBroker> iterator)
        throws InvalidReplicationFactorException, ReplicaPlacementException;
}

...

The proposal is to expose the ReplicaPlacer API which is currently internal as public APIinterface. It will move from the org.apache.kafka.controller package in the metadata project to the org.apache.kafka.server.placer package in the clients project. Similarly the existing UsableBroker class will move from org.apache.kafka.metadata package in the metadata project to the org.apache.kafka.server.placer  placer package in the clients project.
This feature will only be . Sanity checks about the replication factor that are currently performed in StripedReplicaPlacer will be performed in ControlClusterManager as they are common to all ReplicaPlacer implementations.

To address the use cases identified in the motivation section, some knowledge about the current state of the cluster is necessary. Details whether a new broker has just been added or is being decommissioned are not part of the cluster metadata. Therefore such knowledge has to be provided via an external mean to the ReplicaPlacer, for example via the configuration. Apart from configure(), this KIP does not provide a mechanism for defining the behavior of ReplicaPlacer implementation and cluster operators wanting to use this feature will have to build such a mechanism for their specific environments.

The logic assigning replicas to partition differs so much between ZooKeeper and KRaft that I propose making this feature only available in KRaft mode.

Compatibility, Deprecation, and Migration Plan

The new exception/error code changes some of responses clients can get if DefaultReplicaAssignor encounters an error. For example, if a cluster contains brokers with and without "broker.rack", the error will be REPLICA_ASSIGNOR_FAILED instead of Older clients that don't have the new error code will interpret it as UNKNOWN_SERVER_ERROR but they will receive the generated error message indicating the reason for the failure.

Rejected Alternatives

• Computing assignments replica placement for the whole batchcreate topics/partitions request: Instead of computing assignment for each topic in the CreateTopics/CreatePartitions request one at a time, we I looked at computing assignment for all of them in a single call. We rejected this approach for the following reasons:
  • All logic (validation, policies, creation in ZK) in AdminManager current logic works on a single topic at a time. Grouping the replica assignment computation created very complicated logic
  • It's not clear if having all topics at once would significantly improve computed assignments. This is especially true for the 4 scenarios listed in the Motivation section
• Providing more details about the cluster to the placer: Instead of only passing usable brokers, I considered passing a data structure with more details about the cluster, such as Cluster. While this could allow some additional advanced use cases, this would potentially not scale well if we expect Kafka to be able to support very large number of topics with KRaft.