Master KIP
KIP-500: Replace ZooKeeper with a Self-Managed Metadata Quorum (Accepted)
Status
Current state: in reviewaccepted
Discussion thread: here
JIRA:
Jira | ||||||
---|---|---|---|---|---|---|
|
Release: controller-side changes in 2.4, command line changes in 2.6
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 partition 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.
...
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 (revert) of ongoing reassignments.
This API can be used as a foundation on which to build future improvements.
...
AdminClient APIs
We will add two new admin APIs: alterPartitionAssignments, and listPartitionReassignments. As the names imply, the alter API modifies partition reassignments, and the list API lists the ones which are ongoing.
...
Code Block | ||
---|---|---|
| ||
/** * Change the reassignments for one or more partitions. * * @param reassignments The reassignments to add, modify, or remove. * @param Providing an empty Optional (e.g via {@link Optional#empty()}) will <bold>cancel</bold> the reassignment for the associated partition. * * @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> targetBrokerstargetReplicas; ... } class AlterPartitionAssignmentsResult { Map<TopicPartition, Future<Void>> futures; // maps partitions to the results for each partition (success / failure) Future<Void> all(); /** * @throws IllegalArgumentException if no replicas are supplied */ public NewPartitionReassignment(List<Integer> targetReplicas) public List<Integer> targetReplicas() } class AlterPartitionAssignmentsResult { Map<TopicPartition, KafkaFuture<Void>> values() 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 optionspartitions The partitions The options to useto show reassignments for. Must be non-null. * @param partitionsoptions The partitions to showThe reassignmentsoptions for. Must be non-nullto use. */ public ListPartitionReassignmentsResult listPartitionReassignments( ListPartitionReassignmentsOptionsSet<TopicPartition> optionspartitions, Set<TopicPartition>ListPartitionReassignmentsOptions partitionsoptions); /** * List all of the current partition reassignments for the given topics. . * * @param options The options to use. */ public @paramListPartitionReassignmentsResult topicslistPartitionReassignments( ListPartitionReassignmentsOptions options); abstract ListPartitionReassignmentsResult The topics to show reassignments for. Must be non-null. */ ListPartitionReassignmentsResult listPartitionReassignments( listPartitionReassignments(Optional<Set<TopicPartition>> partitions, 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 tocurrently resideresides on. */ private final List<Integer> targetBrokersreplicas; /** * The brokers which that we are adding this partition currently resides on to as part of a reassignment. */ private final List<Integer> currentBrokersaddingReplicas; ... } |
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.
/**
* The brokers that we are removing this partition from as part of a reassignment.
*/
private final List<Integer> removingReplicas;
} |
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.
Note that even though we have TimeoutMs, this field will be ignored in the implementation for the time being. Its implementation is tracked in
Jira | ||||||
---|---|---|---|---|---|---|
|
Code Block | ||
---|---|---|
| ||
{
"apiKey": 45,
"type": "request",
| ||
Code Block | ||
| ||
{ "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": "TopicsAlterPartitionReassignmentsRequest", "typevalidVersions": "[]ReassignableTopic0", "fields": [ { "name": "TimeoutMs", "type": "int32", "versions": "0+", "nullableVersionsdefault": "0+60000", "about": "The topicstime to reassign, or nullin ms to cancelwait allfor the pendingrequest reassignments for all topicsto complete.", "fields": [ }, { "name": "NameTopics", "type": "string[]ReassignableTopic", "versions": "0+", "about": "The topictopics name." },to reassign.", "fields": [ { "name": "PartitionsName", "type": "[]ReassignablePartitionstring", "versions": "0+", "nullableVersions "about": "0+"The topic name." }, { "name": "Partitions", "abouttype": "[]ReassignablePartition", "The partitions to reassign, or null to cancel all pending reassignments for this topicversions": "0+", "about": "The partitions to reassign.", "fields": [ { "name": "PartitionIndex", "type": "int32", "versions": "0+", "about": "The partition index." }, { "name": "Replicas", "type": "[]int32", "versions": "0+", "nullableVersions": "0+", "default": "null", "about": "The replicas to place the partitions on, or null to cancel a pending reassignment for this partition." } ]} ]} ] } |
...
This is the response from AlterPartitionsReassignmentsRequest.
Possible top-level errors:
- REQUEST_TIMED_OUT: if the request timed out.
- NOT_CONTROLLER: if the node we sent the request to was not the controller.
- CLUSTER_AUTHORIZATION_FAILED: If we didn't have sufficient permission to perform the alteration.
Possible partition-level errors:
- INVALID_REPLICA_ASSIGNMENT : if - when the specified replica assignment was is not valid -- for example, if it included negative numbers, repeated numbers, or specified a broker ID that the controller was not aware of.(brokers are not part of the cluster, negative replica ids, empty assignment)
- NO_REASSIGNMENT_IN_PROGRESS : If - when a cancellation was called on a topic/partition which was not in the middle of a reassignment
- CLUSTERUNKNOWN_AUTHORIZATION_FAILEDTOPIC_OR_PARTITION - when a topic/partition doesn't exist or a topic deletion is in progress
Code Block | ||
---|---|---|
| ||
{ "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": "ErrorCode", "type": "int16", "versions": "0+", "about": "The error code." }, { "name": "ErrorStringErrorMessage", "type": "string", "versions": "0+", "nullableVersions": "0+", "about": "The error stringmessage, or null if there was no error." } { "name": "Responses", "type": "[]ReassignableTopicResponse", "versions": "0+", "nullableVersions": "0+", "about": "The responses to topics to reassign.", "fields": [ { "name": "ErrorCodeName", "type": "int16string", "versions": "0+", "entityType": "topicName", "about": "The errortopic code.name" }, { "name": "ErrorStringPartitions", "type": "string[]ReassignablePartitionResponse", "versions": "0+", "nullableVersions": "0+", "about": "The errorresponses string,to orpartitions nullto if there was no error." } reassign", "fields": [ { "name": "PartitionsPartitionIndex", "type": "[]ReassignablePartitionResponseint32", "versions": "0+", "nullableVersions": "0+", "about": "The responsespartition to partitions to reassign", "fields": [index." }, { "name": "PartitionIndexErrorCode", "type": "int32int16", "versions": "0+", "about": "The partitionerror indexcode." }, { "name": "ErrorCodeErrorMessage", "type": "int16string", "versions": "0+", "nullableVersions": "0+", "about": "The error code." }message, or null if there was no { "name": "ErrorString", "type": "string", "versions": "0+", "nullableVersions": "0+", "about": "The error string, or null if there was no error." error." } ] ] } ] } |
ListPartitionReassignmentsRequest
...
It requires DESCRIBE on CLUSTER.
Note that even though we have TimeoutMs, this field will be ignored in the implementation for the time being. Its implementation is tracked in
Jira | ||||||
---|---|---|---|---|---|---|
|
Code Blockcode | ||
---|---|---|
| ||
{ "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": "PartitionIndexesName", "type": "[]int32string", "versions": "0+", "nullableVersionsentityType": "0+topicName", "about": "The partitionstopic to list partition reassignments for, or null to list everything" }name" }, { "name": "PartitionIndexes", "type": "[]int32", "versions": "0+", "about": "The partitions to list partition reassignments for" } ]} ] } |
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 we didn't have sufficient permissions.
If the top-level error code is set, no responses will be provided.
If a partition is not going through a reassignment, its AddingReplicas and RemovingReplicas fields will simply be empty.
If a partition doesn't exist, no response will be returned for it.
Code Block | ||
---|---|---|
| ||
{ | ||
Code Block | ||
| ||
{ "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": "PartitionIndex", "type": "int32", "versions": "0+", "about": "The index of the partition." }, { "name": "CurrentReplicasReplicas", "type": "[]int32", "versions": "0+", "about": "The replicascurrent which the partition is currently assigned toreplica set." }, { "name": "TargetReplicasAddingReplicas", "type": "[]int32", "versions": "0+", "about": "The replicasset whichof thereplicas partitionwe isare beingcurrently reassigned toadding." }, ]} ]} ] }{ "name": "RemovingReplicas", "type": "[]int32", "versions": "0+", "about": "The set of replicas we are currently removing." } ]} ]} ] } |
Modified RPCs
LeaderAndIsrRequest
We will add a two new field fields to LeaderAndIsrRequest named TargetReplicas. This field will AddingReplicas and RemovingReplicas. These fields will, respectively, contain the replicas which will exist once the partition reassignment is completeare being added and removed from a partition's assignment.
These fields allow the controller to propagate reassignment information to the leader/follower brokers, paving the way for some of the improvements outlined in the Future Work section of this KIP. We will not make use of these fields as of this KIP.
Code Block | ||
---|---|---|
| ||
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 AddingReplicas and RetiringReplicasRemovingReplicas. "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": "AddingReplicas", "type": "[]int32", "versions": "3+", "nullableVersions": "0+", + "about": "The replica IDs that we are adding this partition to." }, + or null if the partition is not being added to any new replicas." }, + { "name"{ "name": "RetiringReplicasRemovingReplicas", "type": "[]int32", "versions": "3+", "nullableVersions": "0+", + "about": "The replica IDs that we are removing this partition from, or null if the partition is not being from any replicas." }, { "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 in all types of granularity - at the partition level, topic level and global (stop all reassignments).
Cancellation is not the same as 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 with the tool
replica should have existed on the broker or not." }
]}
|
Other Interfaces
NoReassignmentInProgressException
This exception is thrown when a user tries to cancel a reassignment which doesn't exist
Code Block | ||
---|---|---|
| ||
package org.apache.kafka.common.errors;
/**
* Thrown if a request cannot cancel a reassignment because one is not in progress.
*/
public class NoReassignmentInProgressException extends ApiException {
public NoReassignmentInProgressException(String msg) {
super(msg);
}
public NoReassignmentInProgressException(String msg, Throwable cause) {
super(msg, cause);
}
} |
Proposed Changes
Cancellation
We propose supporting the "cancellation" of an on-going partition rebalance. It is essentially a rollback of a reassignment.
To concretize the notion - a cancellation of a reassignment means stopping any on-going partition movements and restoring the replica set to the one it was prior to the reassignment being initiated.
There is no guarantee that the reverted replica set will be in the same order as it was previously.
kafka-reassign-partitions.sh
kafka-reassign-partitions.sh is a script to create, inspect, or modify partition reassignments.
Deprecation of --zookeeper
We will deprecate the --zookeeper option in this script. New operations introduced by this KIP (cancellation, list) will not be supported by the ZK flag.
All old operations and new will be supported through the admin client. This will be initialized by the --bootstrap-server flag, which already exists.
The --list flag
There will be a new flag, --list, which will list all of the current replica reassignments that are going on in the cluster. For each reassignment, we will list the replica set, the adding replicas, and the removing replicas. If there are no ongoing replica reassignments, "No partition reassignments found." will be printed.
The --additional flag
Like the underlying KIP-455 API, the reassign-partitions tool will now work in an incremental fashion. Any new reassignment will be added on top of the current ones.
However, to avoid confusing users, the tool will decline to create a new reassignment if there is already one in progress. Users can choose to create additional reassignments even if there is one in progress by specifying the new --additional flag. This flag is only supported when --bootstrap-server is provided.
There is also one slight behavior change to --execute that is worth noting here. Previously, when --execute declined to create a new reassignment because one was already in progress, it would still apply the provided throttle, if any. In a world of concurrent reassignments, this behavior no longer makes sense, and has been removed. (However, users can still modify the throttle by simply supplying the --additional flag)
Cancellation with the tool
Additionally, we will support a --cancel flag. This flag will create a JSON file that, if applied, will cancel all on-going partition reassignments.
Not that there is a slight race condition here when using the tool. Someone may create a new reassignment in between creating the cancellation plan and applying it. This new reassignment would not be cancelled by the original cancellation plan. In general, we expect system administrators to stop any process or person that is creating new reassignments before cancelling all pending reassignments, so we don't expect this to be a problem in practice. If such processes are not stopped, the effect of cancelling all pending reassignments will be negated anyway, by the creation of new reassignments.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.
The protocol also supports cancellation at the topic-level but we will not be leveraging that in this tool because the JSON schema used is not suitable for coarser-grained cancellations. Any further improvements to the tool and the JSON format it uses can be slated for future work
kafka-topics.sh
When describing partitions, kafka-topics.sh will show an additional field, "targetReplicastwo additional fields, "addingReplicas," and "removingReplicas." This These will have information about the target adding and removing replicas for the given partition. kafka-topics.sh will get this information from listReassigningPartitions.
...
We will store the reassignment data for each partition in this znode, as that is where we already store the current replicaSet for every partition.
This node will store two new maps, named retiringReplicas addingReplicas and addingReplicas removingReplicas. Each key is the partition and the value is an array of integers denoting the replicas we should retire add or add remove for that partition to, respectively.
If both maps are empty, there is no on-going reassignment for that topic.
Code Block | ||||
---|---|---|---|---|
| ||||
{ "version": 12, "partitions": {"0": [1, 4, 2, 3] }, "retiringReplicasaddingReplicas": {"0": [14] }, # <----- NEW "addingReplicasremovingReplicas": {"0": [41] }, # <----- NEW } |
Because we are not making incompatible changes to the znode, we will not be bumping up the inter-broker protocol versionThe znode will now be at version 2 rather than 1. This change is backwards-compatible-- older brokers can read the new ZNode.
Controller Changes
When the controller starts up, it will load all of the currently active reassignments from the `topics/[topic]' 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 znodes anyway to already reads these znodes during start up.
The reason for having the `reassign_partitions` take precedence is that we cannot know for sure whether the Controller has taken action on that znode, so we allow it to overwrite the API-assigned reassignments in the event of a failover.
The controller will handle ListPartitionReassignments by listing the current active reassignments. It will keep an It can supply this information directly from its in-memory cache. Note that this in-memory cache of on-going reassignments that it loads from ZooKeeper and updates on further reassignmentsall ongoing reassignments already existed prior to this KIP.
The controller will handle the AlterPartitionAssignmentsResponse RPC by modifying the /brokers/topics/[topic] znode. Once the node is modified, the controller will send out LeaderAndIsr LeaderAndIsrRequest to start all the movements.
Once the controller is notified that a new replica has entered the ISR for a particular partition, it will remove it from the "addingReplicas" field in `/topics/[topic]`. If "addingReplicas" becomes empty, the controller will send a new wave of LeaderAndIsr requests to retire all the old replicas and, if applicable, elect new leaders.
Algorithm
We will always empty out add all the replicas in the "addingReplicas" field before starting to act on the "retiringReplicasremovingReplicas" field.
Code Block |
---|
# Illustrating the rebalance of a single partition. # R is the current replicas, I is the ISR list, RRAR is retiringReplicasaddingReplicas and ARRR is addingReplicasremovingReplicas R: [1, 2, 3], I: [1, 2, 3], AR: [], RR: [] # Reassignment called via API with targetReplicas=[4,3,2] R: [1, 4, 3, 2], I: [1, 2, 3], AR: [4], RR: [1] # Controller sends LeaderAndIsr requests # (We ignore RR until AR is empty) R: [1, 4, 3, 2], I: [1, 2, 3, 4], AR: [4], RR: [1] # Leader 1 adds 4 to ISR # The Controllercontroller picksrealizes upthat onall the change and removes 4 from AR. # (in-memory) - R: [1, 4, 3, 2], I: [1, 2, 3, 4], AR: [], RR: [1] # Controller realizes AR is emptyreplicas in AR have been added and starts work on RR. Removes all of RR from R and from the ISR, and sends StopReplica/LeaderAndIsr requests R: [4, 3, 2], I: [1, 2, 3, 4], AR: [], RR: [] # at this point the reassignment is considered complete # Leaders change, 1 retires and 4 starts leading. 4 shrinks the ISR R: [4, 3, 2], I: [2, 3, 4], AR: [], RR: [] considered complete |
When all of the replicas in AR have been added, we remove RR from R and empty out AR/We iteratively shrink the AR collection. When AR is empty, we empty out RR in one step.
If a new reassignment is issued during an on-going one, we cancel the current one by emptying out both AR and RR, constructing them , send StopReplica requests to the replicas that were part of the old reassignment but not part of the new one, construct AR/RR from (the updated from the last-reassignment) R and TR, and starting start anew.
In general this algorithm is consistent with the current Kafka behavior - other brokers still get the full replica set consisting of both the original replicas and the new ones.set consisting of both the original replicas and the new ones.
Note that because of iteratively shrinking the AR collection, the server will not know how to create a "revert" of a reassignment, as it is always seeing it in motion. To keep the server logic simpler, we defer to clients for potential reverts. Here is an example of a cancellation:
Code Block |
---|
# Illustrating the rebalance of a single partition. # R is the current replicas, I is the ISR list, RR is retiringReplicasremovingReplicas and AR is addingReplicas R: [1, 2, 3], I: [1, 2, 3], AR: [], RR: [] # Reassignment called via API with targetReplicas=[3, 4, 5] R: [1, 2, 3, 4, 5], I: [1, 2, 3], AR: [4, 5], RR: [1, 2] R: [1, 2, 3, 4, 5], I: [1, 2, 3, 4], AR: [4, 5], RR: [1, 2] # Cancellation called R: [1, 2, 3, 4], I: [1, 2, 3, 4], AR: [], RR: [] |
Essentially, once a cancellation is called we subtract AR from R, empty out both AR and RR, and send StopReplica/LeaderAndIsr requests to cancel the replica movements that have not yet completed.
...
We could store the pending replicas in a single Znode per cluster, rather than putting them beside the ISR replicas 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 reassigning replicas state in the ISR state /brokers/topics/[topic] Znode is that this state belongs in the LeaderAndIsrRequestthere, next to the replicaSet. 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.
...