Status

Current state: Under discussion

Discussion thread: here 

JIRA: here

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

Summary

Today, kafka-reassign-partitions.sh --execute submits all partition reassignments from the JSON in one AdminClient#alterPartitionReassignments call (unless the user manually splits work). For large clusters this can create large bursts of replication traffic and controller work.

This KIP proposes tool-only pacing controls:

  1. --reassignment-batch-size — caps how many topic partitions are submitted per step (semantics differ slightly depending on mode; see below). 0 preserves legacy behavior (single request for the entire plan).
  2. --incremental — optional mode used with --reassignment-batch-size > 0 to keep at most N partition reassignments in flight for this execution, submitting the next partition from a deterministic queue when a slot frees up.

Ordering for both modes is (topic name, partition index), not the order of entries in the JSON file.

No broker protocol, controller, or metadata changes are required; pacing is implemented entirely in the reassignment tool using existing Admin APIs (alterPartitionReassignments, listPartitionReassignments, metadata reads).

Motivation

Problem

  1. Operational risk: Submitting hundreds or thousands of partition reassignments in one RPC can stress replication links, disk, and the controller in ways that are hard to predict during maintenance windows.
  2. Limited operator control: Throttling (--throttle) limits bandwidth but does not limit how many partitions are simultaneously moving; operators often want serialised waves or bounded concurrency without hand-splitting JSON files.
  3. Manual workarounds: Teams split reassignment JSON by hand or wrap the tool in scripts that call alterPartitionReassignments in chunks — error-prone and inconsistent across deployments.

Goals

Non-goals

Public interfaces

CLI (kafka-reassign-partitions.sh)

Option

Applies to

Semantics

--reassignment-batch-size <int>

--execute only

  • Default 0: legacy — entire plan in one alterPartitionReassignments request; no extra waits in the tool.
  • > 0 without --incremental: split the plan into contiguous batches of at most N partitions (sorted by topic, then partition id). After each batch except the last, the tool blocks until every partition in that batch reports complete (current ISR matches target and reassignment not active) before submitting the next batch.
  • > 0 with --incremental: N is the maximum number of partition reassignments from this JSON that may be active at once; when one completes, the tool submits the next partition from the sorted queue.

--incremental

--execute only

Requires --reassignment-batch-size > 0. Mutually exclusive interpretation of batch size as in-flight cap (see above).


Validation and compatibility

Programmatic API

ReassignPartitionsCommand.executeAssignment gains parameters:

Existing callers that omit pacing pass 0 and false to preserve legacy behaviour.

Detailed design

Partition counting

One row in the JSON (one TopicPartition) = one partition reassignment for pacing purposes, regardless of how many brokers join or leave the replica set in that single move.

Batch construction (non-incremental, reassignmentBatchSize > 0)

  1. Build the map of TopicPartition → target replicas from the JSON.
  2. Split into batches: sort keys with compareTopicPartitions, then chunk in groups of N partitions.
  3. For each batch except the last:
    alterPartitionReassignments(batch) → wait until all partitions in that batch are complete → next batch.
  4. For the last batch:
    alterPartitionReassignments only — the tool does not wait for completion before printing success (same family as legacy “started” messaging). Operators should run --verify for full completion.

Incremental mode (--incremental)

  1. Sort partitions deterministically (compareTopicPartitions).
  2. Maintain a pending deque and an in-flight map (submissions for this execute only).
  3. Loop until pending is empty: remove completed partitions from in-flight (using the same completion predicate as non-incremental wait paths), then submit new partitions up to the N in-flight cap.
  4. Poll interval between iterations when work remains: INCREMENTAL_REASSIGNMENT_POLL_INTERVAL_MS (500 ms in the default implementation).

Semantics: The tool returns after all partitions in this JSON have been successfully submitted to alterPartitionReassignments, not necessarily after all replication has finished (consistent with legacy execute for “completion” of submission).

Non-incremental wait between batches

Interaction with --additional

--additional only bypasses the “existing reassignment on cluster” guard. It does not merge pacing across concurrent executes: --reassignment-batch-size applies per process invocation to that JSON. Multiple overlapping executes can each contribute up to N in-flight partition reassignments from their respective plans.

Interaction with throttles

Existing --throttle / --replica-alter-log-dirs-throttle behaviour is unchanged; pacing is orthogonal and can be combined.

Compatibility

Limitations and future work

  1. No timeout on batch-completion waits in the reference implementation; stuck reassignments can poll indefinitely until operator intervention (same class of risk as long-running admin operations without deadlines).
  2. Incremental + --additional: total cluster in-flight work can exceed N when multiple tool processes run overlapping plans; documentation / optional warnings are recommended.

Rejected alternatives

  1. Only documentation: “Split your JSON manually” — does not scale and yields inconsistent operations.
  2. New broker-side “max concurrent reassignments” quota: much larger scope; tool-side pacing addresses the common case without protocol work.
  3. Order = JSON file order only: rejected in favour of deterministic sorted order so behaviour is reproducible and independent of file editing.

Test plan

Documentation