You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 13 Next »

Status

Current state: in review

Discussion thread: here

JIRA:  Unable to render Jira issues macro, execution error.

Release:

Motivation

Currently, users initiate replica reassignment by writing directly to a ZooKeeper node named /admin/reassign_partitions.

As explained in KIP-4, ZooKeeper based APIs have many problems.  For example, there is no way to return a helpful error code if an invalid reassignment is proposed.  Adding new features over time is difficult when external tools can access and overwrite Kafka's internal metadata data structures.  Also, ZooKeeper-based APIs inherently lack security and auditability.

In addition to all the general problems of ZK-based APIs, the current reassignment interface has some problems specific to its particular structure.  There is no mechanism provided for aborting a reassignment operation once it has been initiated.  Furthermore, the API assumes that reassignment operations are launched as a batch – there is no way to incrementally add a new reassignment operation once the batch has been initiated.

We would like to provide a well-supported AdminClient API that does not suffer from these problems. Namely, it should support incremental replica reassignments and cancellation of ongoing reassignments.
This API can be used as a foundation on which to build future improvements.

Public Interfaces

AdminClient APIs

We will add two new APIs: alterPartitionAssignments, and listPartitionReassignments.  As the names imply, the alter API modifies partition reassignments, and the list API lists the ones which are ongoing.

Unlike the current ZooKeeper-based API, alterPartitionAssignments can add or remove partition reassignments without interrupting unrelated assignments that are in progress.  Partition reassignments can be added or removed by the API on a per-partition basis-- there are no global "before" or "after" snapshots.


/**
 * Change the reassignments for one or more partitions.
 *
 * @param reassignments   The reassignments to add, modify, or remove.
 * @param options         The options to use.
 * @return                The result.
 */
public AlterPartitionReassignmentsResult alterPartitionReassignments(
         Map<TopicPartition, Optional<NewPartitionReassignment>> reassignments,
         AlterPartitionReassignmentsOptions options);

/**
 * A new partition reassignment, which can be applied via {@link AdminClient#alterPartitionReassignments(Map)}.
 */
public class NewPartitionReassignment {
    private final List<Integer> targetBrokers;
    ...
}

class AlterPartitionAssignmentsResult {
  Map<TopicPartition, Future<Void>> futures; // maps partitions to the results for each partition (success / failure)
  Future<Void> all(); // Throws an exception if any reassignment was rejected
}

class AlterPartitionAssignmentsOptions extends AbstractOptions<> {
  // contains timeoutMs
}

/**
 * List some of the current partition reassignments.
 * 
 * @param options        The options to use.
 * @param partitions     The partitions to show reassignments for.  Must be non-null.
 */
ListPartitionReassignmentsResult listPartitionReassignments(
      ListPartitionReassignmentsOptions options,
      Set<TopicPartition> partitions);


/**
 * List all of the current partition reassignments for the given topics.
 * 
 * @param options        The options to use.
 * @param topics         The topics to show reassignments for.  Must be non-null.
 */
ListPartitionReassignmentsResult listPartitionReassignments(
      ListPartitionReassignmentsOptions options,
      Set<String> topics);


/**
 * List all of the current partition reassignments.
 * 
 * @param options        The options to use.
 */
ListPartitionReassignmentsResult listPartitionReassignments(
      ListPartitionReassignmentsOptions options);

class ListPartitionReassignmentsOptions extends AbstractOptions<> {
  // contains timeoutMs
}

class ListPartitionReassignmentsResult {
  private final KafkaFuture<Map<TopicPartition, PartitionReassignment>> reassignments;
}

/**
 * A partition reassignment, which has been listed via {@link AdminClient#listPartitionReassignments()}.
 */
public class PartitionReassignment {
    /**
     * The brokers which we want this partition to reside on.
     */
    private final List<Integer> targetBrokers;

    /**
     * The brokers which this partition currently resides on.
     */
    private final List<Integer> currentBrokers;
...
}

New RPCs

AlterPartitionReassignmentsRequest

This is used to create or cancel a set of partition reassignments.  It must be sent to the controller.

This operation requires ALTER on CLUSTER.

{
  "apiKey": 45,
  "type": "request",
  "name": "AlterPartitionReassignmentsRequest",
  "validVersions": "0",
  "fields": [
    { "name": "TimeoutMs", "type": "int32", "versions": "0+", "default": "60000",
      "about": "The time in ms to wait for the request to complete." },
    { "name": "Topics", "type": "[]ReassignableTopic", "versions": "0+", "nullableVersions": "0+",
      "about": "The topics to reassign, or null to cancel all the pending reassignments for all topics.", "fields": [
      { "name": "Name", "type": "string", "versions": "0+",
        "about": "The topic name." },
      { "name": "Partitions", "type": "[]ReassignablePartition", "versions": "0+", "nullableVersions": "0+",
        "about": "The partitions to reassign, or null to cancel all pending reassignments for this topic", "fields": [
        { "name": "PartitionIndex", "type": "int32", "versions": "0+",
          "about": "The partition index." },
        { "name": "Replicas", "type": "[]int32", "versions": "0+", "nullableVersions": "0+",
          "about": "The replicas to place the partitions on, or null to cancel a pending reassignment." }
      ]}
    ]}
  ]
}

AlterPartitionReassignmentsResponse

This is the response from AlterPartitionsReassignmentsRequest.

Possible errors:

  • REQUEST_TIMED_OUT: if the request timed out.
  • NOT_CONTROLLER: if the node we sent the request to was not the controller.
  • INVALID_REPLICA_ASSIGNMENT: if the specified replica assignment was not valid-- for example, if it included negative numbers, repeated numbers, or specified a broker ID that the controller was not aware of.
  • CLUSTER_AUTHORIZATION_FAILED


{
  "apiKey": 45,
  "type": "response",
  "name": "AlterPartitionReassignmentsResponse",
  "validVersions": "0",
  "fields": [
    { "name": "ThrottleTimeMs", "type": "int32", "versions": "0+",
      "about": "The duration in milliseconds for which the request was throttled due to a quota violation, or zero if the request did not violate any quota." },
    { "name": "Responses", "type": "[]ReassignablePartitionResponse", "versions": "0+",
      "about": "The responses to partitions to reassign.", "fields": [
      { "name": "ErrorCode", "type": "int16", "versions": "0+",
        "about": "The error code." },
      { "name": "ErrorString", "type": "string", "versions": "0+", "nullableVersions": "0+",
        "about": "The error string, or null if there was no error." }
      ]
    }
  ]
}

ListPartitionReassignmentsRequest

This RPC lists the currently active partition reassignments.  It must be sent to the controller.

It requires DESCRIBE on CLUSTER.

{
  "apiKey": 46,
  "type": "request",
  "name": "ListPartitionReassignmentsRequest",
  "validVersions": "0",
  "fields": [
    { "name": "TimeoutMs", "type": "int32", "versions": "0+", "default": "60000",
      "about": "The time in ms to wait for the request to complete." }
    { "name": "Topics", "type": "[]ListPartitionReassignmentsTopics", "versions": "0+", "nullableVersions": "0+",
      "about": "The topics to list partition reassignments for, or null to list everything.", "fields": [
      { "name": "PartitionIds", "type": "[]int32", "versions": "0+",
        "about": "The partitions to list partition reassignments for, or null to list everything" }
    ]}
  ]
}

ListPartitionReassignmentsResponse

Possible errors:

  • REQUEST_TIMED_OUT: if the request timed out.
  • NOT_CONTROLLER: if the node we sent the request to is not the controller.
  • CLUSTER_AUTHORIZATION_FAILED

If the top-level error code is set, no responses will be provided.


{
  "apiKey": 46,
  "type": "response",
  "name": "ListPartitionReassignmentsResponse",
  "validVersions": "0",
  "fields": [
    { "name": "ThrottleTimeMs", "type": "int32", "versions": "0+",
      "about": "The duration in milliseconds for which the request was throttled due to a quota violation, or zero if the request did not violate any quota." },
    { "name": "ErrorCode", "type": "int16", "versions": "0+",
      "about": "The top-level error code, or 0 on success." },
    { "name": "ErrorMessage", "type": "string", "versions": "0+", "nullableVersions": "0+",
      "about": "The top-level error message, or null on success." },
    { "name": "Topics", "type": "[]OngoingTopicReassignment", "versions": "0+",
      "about": "The ongoing reassignments for each topic.", "fields": [
        { "name": "Name", "type": "string", "versions": "0+",
          "about": "The topic name." },
        { "name": "Partitions", "type": "[]OngoingPartitionReassignment", "versions": "0+",
          "about": "The ongoing reassignments for each partition.", "fields": [
          { "name": "PartitionId", "type": "int32", "versions": "0+",
            "about": "The partition ID." },
          { "name": "CurrentReplicas", "type": "[]int32", "versions": "0+",
            "about": "The replicas which the partition is currently assigned to." },
          { "name": "TargetReplicas", "type": "[]int32", "versions": "0+",
            "about": "The replicas which the partition is being reassigned to." }
      ]}
    ]}
  ]
}

Modified RPCs

LeaderAndIsrRequest

We will add a new field to LeaderAndIsrRequest named TargetReplicas.  This field will contain the replicas which will exist once the partition reassignment is complete.

diff --git a/clients/src/main/resources/common/message/LeaderAndIsrRequest.json b/clients/src/main/resources/common/message/LeaderAndIsrRequest.json
index b8988351c..d45727078 100644
--- a/clients/src/main/resources/common/message/LeaderAndIsrRequest.json
+++ b/clients/src/main/resources/common/message/LeaderAndIsrRequest.json
@@ -20,6 +20,8 @@
   // Version 1 adds IsNew.
   //
   // Version 2 adds broker epoch and reorganizes the partitions by topic.
+  //
+  // Version 3 adds TargetReplicas.
   "validVersions": "0-3",
   "fields": [
     { "name": "ControllerId", "type": "int32", "versions": "0+",
@@ -48,6 +50,8 @@
           "about": "The ZooKeeper version." },
         { "name": "Replicas", "type": "[]int32", "versions": "0+",
           "about": "The replica IDs." },
+        { "name": "TargetReplicas", "type": "[]int32", "versions": "3+", "nullableVersions": "0+",
+          "about": "The replica IDs that we are reassigning the partition to, or null if the partition is not being reassigned." },
         { "name": "IsNew", "type": "bool", "versions": "1+", "default": "false", "ignorable": true, 
           "about": "Whether the replica should have existed on the broker or not." }
       ]}

Proposed Changes

Cancellation

We propose supporting the "cancellation" of an on-going rebalance. This is not a revert. To concretize the notion - a cancellation of a rebalance means stopping on-going partition movements. A partition reassignment which is half done will not have the new replicas that have entered the ISR be reverted, it will only stop the in-flight movements for replicas that are not in ISR.
Note that it is possible to have a partition's replication factor become higher than configured if a reassignment is stopped half-way. We do not propose a clear way to guard against this nor revert it – users are expected to run another reassignment

kafka-reassign-partitions.sh

kafka-reassign-partitions.sh is a script to create, inspect, or modify partition reassignments.

We will remove the --zookeeper option from this script.  Instead, all operations will be performed through the admin client.  This will be initialized by the --bootstrap-server flag, which already exists.

There will be a new flag, --list, which will list all of the current replica reassignments that are going on in the cluster.  They will be output in the same JSON format that is accepted by --reassignment-json-file.  If there are no ongoing replica reassignments, an empty JSON object will be printed.

Cancellation

The reassign-partitions tool will work in an incremental fashion – any new reassignment will be added on top of the current ones. It will continue to output a rollback JSON as part of any reassignment in order to let the user revert the changes.
Additionally, we will support a --cancel-all flag which will cancel all on-going partition reassignments. This providers users with an intuitive, quick and easy way to cancel multiple reassignments at once.

kafka-topics.sh

When describing partitions, kafka-topics.sh will show an additional field, "targetReplicas."  This will have information about the target replicas for the given partition.  kafka-topics.sh will get this information from listReassigningPartitions.

Implementation

ZooKeeper Changes

/brokers/topics/[topic]/partitions/[partition]/state

We will store the reassignment data for each partition in this zNode, as that is where we already store the current ISR.
This node will store a new structure named targetReplicas - an array of ints.  If this array is not present or null, there is no ongoing reassignment for this topic.  If it is non-null, it contains the ongoing reassignments for a certain partition to the target replicas for that partition.

partition/state zNode content
{  
    "version":1,
    "leader":1,
    "leader_epoch":1,
    "controller_epoch":1,
    "isr":[1, 2, 3],
    "targetReplicas":[4, 5, 6] # <-- NEW
}


Once the whole reassignment for a given partition is done, we will first remove the old replicas from that partition's entry in the `/brokers/topics/[topic]` zNode and then clear the targetReplicas the /state zNode. This is consistent with our current behavior and illustrated further in an example below.

Additionally, because we are not making incompatible changes to the zNode, we will not be bumping up the inter-broker protocol version.

Algorithm

# Illustrating the rebalance of a single partition.
# R is the current replicas, I is the ISR list and T is the targetReplicas -- what we want to reassign the partition to.
R: [1, 2, 3], I: [1, 2, 3], T: []
# Reassignment called
R: [1, 2, 3], I: [1, 2, 3], T: [4, 5, 6]
R: [1, 2, 3], I: [1, 2, 3, 4], T: [4, 5, 6]
R: [1, 2, 3, 4], I: [1, 2, 3, 4], T: [4, 5, 6]
R: [1, 2, 3, 4], I: [1, 2, 3, 4, 5], T: [4, 5, 6]
R: [1, 2, 3, 4, 5], I: [1, 2, 3, 4, 5], T: [4, 5, 6]
R: [1, 2, 3, 4, 5], I: [1, 2, 3, 4, 5, 6], T: [4, 5, 6]
R: [1, 2, 3, 4, 5, 6], I: [1, 2, 3, 4, 5, 6], T: [4, 5, 6]
R: [4, 5, 6], I: [1, 2, 3, 4, 5, 6], T: null
R: [4, 5, 6], I: [4, 5, 6], T: null

Note that we finish the reassignment at the very next-to-last step only. We first remove the old replicas from the '/brokers/topics/[topic]' zNode. Then, as part of the 'leader_epoch' bump we will also remove the 'targetReplicas' in the /state zNode. We send a LeaderAndIsr request to the appropriate leader with the new replicas (4, 5, 6) and the full ISR (1-6). The leader itself shrinks the ISR to only the replicaSet. This is all consistent to Kafka's current behavior, the only new addition being clearing the 'targetReplicas' collection.

Saving 'targetReplicas' in tact until the very end opens up the possibility to easily revert an on-going reassignment in the future (KIP-236 may implement this).
The alternative of iteratively shrinking the replicas and targetReplicas at each step introduces more complexity, more potential edge cases and extra ZK writes. It does not seem to yield many benefits.

Controller Changes

When the controller starts up, it will load all of the currently active reassignments from the `[partition]/state' zNodes and the `admin/reassign_partitions` zNode, with the latter taking precedence. This will not impose any additional startup overhead, because the controller needs to read these zNode anyway to start up.
The reason for having the `reassign_partitions` take precedence is 

The controller will handle ListPartitionReassignments by listing the current active reassignments. It will keep an in-memory cache of on-going reassignments that it loads from the zNode and updates on further reassignments.

The controller will handle the AlterPartitionAssignmentsResponse RPC by modifying the specified zNode in /brokers/topics/[topic].  Once the node is modified, the controller will send out LeaderAndIsr to start all the movements.
At this point, we preserve the old behavior. Namely, tracking when new replicas enter the ISR, reconciling them against the target reassignments and stopping

Tool Changes

The kafka-reassign-partitions.sh tool will use the new AdminClient APIs to submit the reassignment plan.  It will not use the old ZooKeeper APIs any more.  In order to contact the admin APIs, the tool will accept a --bootstrap-server argument.

When changing the throttle, we will use IncrementalAlterConfigs rather than directly writing to Zookeeper.

Compatibility, Deprecation, and Migration Plan

/admin/reassign_partitions znode

For compatibility purposes, we will continue to allow assignments to be submitted through the /admin/reassign_partitions node.  Just as with the current code, this will only be possible if there are no current assignments.  In other words, the znode has two states: empty and waiting for a write, and non-empty because there are assignments in progress. Once the znode is non-empty, further writes to it will be ignored.

Applications using the old API will not be able to cancel in-progress assignments.  They will also not be able to monitor the status of in-progress assignments, except by polling the znode to see when it becomes empty, which indicates that no assignments are ongoing.  To get these benefits, applications will have to be upgraded to the AdminClient API.

The deprecated ZooKeeper-based API will be removed in a future release.

There is one known edge-case with using both APIs which we will intentionally leave unaddressed:

1. zNode reassignment involving partitionA and partitionB starts
2. API reassignment involving partitionA starts
3. Controller failover. It will effectively ignore the API reassignment in step 2. and move partitionA to the replica given
reassignment by the zNode

There are ways to handle this edge-case, but we propose ignoring it due to the fact that:

  • it is very unlikely for users to use both APIs at once
  • we will be deprecating the zNode API anyway
  • it is not worth the complexity and performance hit from addressing such a rare and seemingly non-critical edge-case

Reassignments while upgrading versions

In the event of a software upgrade, we may end up with a temporarily-stuck reassignment. Imagine the following scenario:

Brokers A, B and C
Controller=BBroker A gets upgraded to 2.4 (which contains this KIP)
Broker B gets upgraded to 2.4. Controller moves to Broker A
Reassignments happens, Broker A sends a LeaderAndIsr request with TargetReplicas to C but it is ignored by C, as it does not know about the field yet.
Broker C gets upgraded to 2.4.

At this point, part of the reassignment is "stuck". Partitions with a targetReplica of C will never complete until the controller gets restarted and a new LeaderAndIsr is sent out.

Reassignments while downgrading versions

In the event of a software downgrade, the extra field in the LeaderAndIsr request will be ignored by brokers running the old code and the reassignment will stop. As soon as the controller gets downgraded, it will send a LeaderAndIsr with no targetReplicas which ensures that all brokers will stop their reassignments.


Rejected Alternatives

Store the pending replicas in a single ZNode rather than in a per-partition ZNode

We could store the pending replicas in a single ZNode per cluster, rather than putting them beside the ISR in /brokers/topics/[topic]/partitions/[partitionId]/state.  However, this ZNode might become very large, which could lead to us running into maximum znode size problems.  This is an existing scalability limitation which we would like to solve.

Another reason for storing the targetReplicas state in the ISR state ZNode is that this state belongs in the LeaderAndIsrRequest.  By making it clear which replicas are pending and which are existing, we pave the way for future improvements in reassignment.  For example, we will be able to distinguish between reassignment traffic and non-reassignment traffic for quota purposes, and provide better metrics.

Future Work

Reassignment Quotas that only throttle reassignment traffic

Currently, the quotas used for reassignment also apply to all non-ISR traffic.  This is undesirable and can lead to problems if nodes fall out of the ISR.

Since this KIP makes brokers aware of which replicas are being used for reassignment, it will be possible to implement quotas that only hit reassignment traffic, and not all non-ISR traffic.  This will also obviate the need to manually specify the list of reassigning partitions when setting the per-broker quotas.

Add reassignment metrics

We would like to provide metrics for reassignment.  For example, we would like to know how many partitions are being reassigned on a given broker, what write and read bandwidth is being used, etc.  Since this KIP clearly separates replicas that are being reassigned from those that are not, and makes this information available on each broker through the LeaderAndIsrRequest, this will be possible in the future.

Improved API for Reassignment Quotas

We would like a better API for managing reassignment quotas-- perhaps a new AdminClient function and RPC.  This is somewhat of a minor improvement, since we already have a way to change the configuration via admin client.  However, it would still be helpful.

Internal Batching

It would be nice if the controller could internally batch reassignment operations.  The controller has more information with which to make decisions than external systems.  If this were implemented, we would want to enrich the List API with more information, such as which reassignments were currently being worked on, and which the controller had decided to defer until later.

This could be added in a compatible way by making the default batch size infinite, so that users that didn't enable controller-side batching would get the current behavior of all reassignments starting immediately.

  • No labels