Current state: Under Discussion

Discussion thread: here


Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).


The current tool imposes the limitation that only a single batch of partition reassignments can be in-flight, and it is not possible to cancel a reassignment that is in-flight cleanly, safely in a timely fashion (e.g. as reported KAFKA-6304,  the current way of reassignment cancellation requires a lot of manual steps). This has a number of consequences:

  1. Reassignments especially for large topic/partition is costly.  In some case, the performance of the Kafka cluster can be severely impacted when reassignments are kicked off.   There should be a fast, clean, safe way to cancel and rollback the pending reassignments.   e.g.  original replicas [1,2,3],  new replicas [4,5,6],   causing performance impact on Leader 1,  the reassignment should be able to get cancelled immediately and reverted back to original replicas [1,2,3],  and dropping the new replicas. 
  2. If users need to do a large-scale reassignment they end up having to do the reassignment in smaller batches, so they can abort the overall reassignment sooner, if operationally necessary.   
  3. Each batch of reassignments takes as long as the slowest partition; this slowest partition prevents other reassignments from happening.   This can be happening even in the case submitting the reassignments by grouping similar size topic/partitions into each batch. How to optimally group reassignments into one batch for faster execution and less impact to the cluster is beyond the discussion in this KIP. 
  4. The ZooKeeper-imposed limit of 1MB on znode size places an upper limit on the number of reassignments that can be done at a given time. Note that in real Production environment, it's better to do reassignments in batches with reasonable reassignments in each batch.  Large number reassignments tends to cause higher Producer latency. Between batches,  proper staggering, throttling is recommended.  

This change would enable 

Public Interfaces

Strictly speaking this is not a change that would affect any public interfaces (since ZooKeeper is not considered a public interface, and it can be made in a backward compatible way), however since some users are known to operate on the /admin/reassign_partitions znode directly,  this can break in new version of Kafka  (e.g. as reported in KAFKA-7854).   

For the existing /admin/reassign_partitions znode,  adding "original_replicas" to support rollback to its original state of the topic partition assigned replicas.   How "original_replicas" gets populated will be discussed in detail later. 

 "partitions":[{"topic": "foo1",
                "partition": 0,
		        "replicas": [1,2,4],
                "original_replicas": [1,2,3]
               {"topic": "foo2",
                "partition": 1,
		        "replicas": [5,6,8],
                "original_replicas": [7,9,8]

Proposed Changes

The main idea is to move away from using the single /admin/reassign_partitions znode to control reassignment.

Instead a protocol similar to that used for config changes will be used. To start a reassignment (of a single partition, or a collection of partitions) a persistent sequenced znode will be created in /admin/reassignment_requests. The controller will be informed of this new request via a ZooKeeper watch. It will initiate the reassignment, create a znode /admin/reassignments/$topic/$partition and then delete the /admin/reassignment_requests/request_NNNN znode.

A client can subsequently create more requests involving the same partitions (by creating a /admin/reassignment_requests/request_MMMM znode), and the controller would re-initiate reassignment as the assignment changes. This means it is possible to "cancel" the reassignment of a single partition by changing the reassignment to the old assigned replicas. Note, however, that the controller itself doesn't remember the old assigned replicas – it is up to the client to record the old state prior to starting a reassignment if that client needs to support cancellation.

When the ISR includes all the newly assigned replicas (given as the contents of the reassignment znode), the controller would remove the relevant child of /admin/reassignments (i.e. at the same time the controller currently updates/removes the /admin/reassign_partitions znode).

In this way the child nodes of /admin/reassignments are correspond precisely to the replicas currently being reassigned. Moreover the ZooKeeper czxid of a child node of /admin/reassignments identifies a reassignment uniquely. It is then possible to determine whether that exact reassignment is still on-going, or (by comparing the czxid with the mzxid) has been changed.

The following sections describe, in detail, the algorithms used to mange the various znodes

The supporting the legacy znode protocol (/admin/reassign_partitions)

  1. Some ZK client creates /admin/reassign_partitions with the same JSON content as currently
  2. ZK triggers watch on /admin/reassign_partitions, controller notified
  3. Controller creates /admin/reassignment_requests/request_xxx (for xxx the znode sequence number)

The JSON in step 3 would look like:

        "my-topic": {"0":[1,2,3]}
        "your-topic": {"3":[4,5,6]}

The "legacy" key is used to track that this request originates from /admin/reassign_partitions.

The new znode protocol

  1. Some ZK client creates /admin/reassignment_requests/request_xxx (without "legacy" key mentioned above, unless it was created via the legacy protocol described above)
  2. ZK triggers watch on /admin/reassignment_requests/request_xxx, controller notified

  3. For each /admin/reassignment_requests/request_xxx the controller:

    1. Reads the znode to find the partitions being reassigned and the newly assigned brokers.

    2. Checks whether a reassignment for the same partition exists:
      1. If it does, and one is a legacy request and the other is not, then it logs a warning that the request is being ignored.
      2. Otherwise:

        1. The controller initiates reassignment

        2. The controller creates or updates  /admin/reassignments/$topic/$partition corresponding to those partitions
    3. The controller deletes /admin/reassignment_requests/request_xxx

The check in 3.b.i prevents the same partition being reassigned via different routes.

Revised protocol when a partition "catches up

  1. If the reassignment was via the legacy path:
    1. The controller removes partition from /admin/reassign_partitions, or if this is the last partition being reassigned by a change to /admin/reassign_partitions, then /admin/reassign_partitions is deleted.
  2. The controller removes /admin/reassignments/$topic/$partition

On controller failover

When failing over, the new controller needs to cope with three cases:

  1. Creation of /admin/reassign_partitions while there was no active controller
  2. Creation of childen of /admin/reassignment_requests while there was no active controller
  3. Restarting reassignment of all the in-flight reassignments in /admin/reassignments

The controller context's partitionsBeingReassigned map can be updated as follows:

  1. Add a ReassignedPartitionsContext for each reassignment in /admin/reassignments
  2. Add ReassignedPartitionsContexts for each child of /admin/reassignment_requests, but only if the context being added has the same legacy flag as any current context for the partition.
  3. Add ReassignedPartitionsContexts for partition in /admin/reassign_partitions, but only if the context being added has the same legacy flag as any current context for the partition.

Change to the KafkaController.onPartitionReassignment() algorithm

The existing algorithm in onPartitionReassignment() is suboptimal for changing reassignments in-flight. As an example, consider an initial assignment of [1,2], reassigned to [2,3] and then changed to [2,4]. With the current algorithm broker 3 will remain in the assigned replicas until broker 4 is in sync, even though 3 wasn't one of the original assigned replicas nor one of the new assigned replicas after the second reassignment. Broker 3 will try to get and stay in-sync during this time, wasting network bandwidth and CPU on both broker 3 and the leader.

More complex scenarios exist where such "intermediate replicas" become the leader.

To solve this problem it is necessary to compute a set of replicas to be removed from the assigned replicas when a reassignment changes.

The set drop of replicas to be removed will be computed as follows:

  1. At the point a new reassignment is created we compute the value top = len(original_assignment)

    1. We check top >= min.insync.replicas, logging a warning and ignore the reassignment if the check fails.

    2. Otherwise we store this in the /admin/reassignments/$topic/$partition znode.
  2. Then, initially, and each time the reassignment changes, (i.e. in KafkaController.onPartitionReassignment()):
    1. Let assigned be the currently assigned replicas, ie. assigned = ControllerContext.partitionRepliacAssignment(tp)
    2. Define a sorted list of out-of-sync replicas: osr = assigned -- isr
    3. Create a list of current replicas existing = List(leader) ++ isr ++ osr
    4. Remove duplicates from existing (keeping the first occurrence in the list)

    5. Then, the set of replicas which can be dropped is: drop = assigned -- existing[0:top-1]


 In order to actually drop those replicas:

  1. Make the transitions in the RSM: -> Offline -> ReplicaDeletionStarted -> ReplicaDeletionSuccessful -> NonExistentReplica

  2. Update the assigned replicas in memory and in ZK

Compatibility, Deprecation, and Migration Plan

As described above, compatibility with /admin/reassign_partitions is maintained, so existing software will continue working and the only difference to a client that operates on /admin/reassign_partitions would observe would be a slight increase in latency due to the round trips needed to create the new znodes.

This compatibility behaviour could be dropped in some future version of Kafka, if that was desirable.

Rejected Alternatives

  1. A similar protocol based on just /admin/reassignments without the /admin/reassignment_requests was initially considered, but that required a ZK watch per reassignment, which would not scale well. This proposal requires only 1 more watch than the current version of the broker.