Note this was initially erroneously assigned as KIP-178, which was already taken, and has been reassigned KIP-179.
Status
Current state: Under Discussion [One of "Under Discussion", "Accepted", "Rejected"]
Discussion thread: here (when initially misnumbered as KIP-178) and here (when assigned KIP-179)
JIRA: here
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).e
Motivation
- The AdminClient API currently lacks any functionality for reassigning partitions: Users have to use the
kafka-reassign-partitions.sh
tool (ReassignPartitionsCommand
) which talks directly to ZooKeeper. This prevents the tool being used in deployments where only the brokers are exposed to clients (i.e. where the zookeeper servers are intentionally not exposed). In addition, there is a general push to refactor/rewrite/replace tools which need ZooKeeper access with equivalents which use theAdminClient
API. ReassignPartitionsCommand
currently has no proper facility to report progress of a reassignment;--verify
can be used periodically to check whether the request assignments have been achieved. It would be useful if the tool could report progress better.ReassignPartitionsCommand
, when used with a replication throttle, requires a --verify invocation when the reassignment has finished in order to remove the throttled. Therefor there exists the possibility that throttles are not removed after reassignment, with negative consequences for the performance of the cluster.
Public Interfaces
The AdminClient
API will have new methods added (plus overloads for options):
alterInterbrokerThrottledRates(Map<Integer, ThrottledRate> throttledRates)
alterInterbrokerThrottledReplicas(Map<TopicPartition, ThrottledReplicas> replicas)
The options for reassignPartitions() will support setting a throttle, and a flag for its automatic removal at the end of the reassignment. Likewise the options for changing the throttled rates and replicas will include the ability to have the throttles automatically removed.
New network protocol APIs will be added to support these AdminClient APIs
AlterInterbrokerThrottledRatesRequest and
AlterInterbrokerThrottledRatesResponse
AlterInterbrokerThrottledReplicasRequest and
AlterInterbrokerThrottledReplicasResponse
The options accepted by kafka-reassign-partitions.sh
command will change:
--zookeeper
will be deprecated, with a warning message- a new
--bootstrap-server
option will be added - a new
--progress
action option will be added
When run with --bootstrap-server
it will no longer be necessary to run kafka-reassign-partitions.sh
--verify to remove a throttle: This will be done automatically.
Proposed Changes
Summary of use cases
Use case | command | AdminClient |
---|---|---|
Change replication factor | kafka-reassign-partitions --execute --reassignment-json-file J |
|
Change partition assignment | kafka-reassign-partitions --execute --reassignment-json-file J | reassignPartitions() |
Change partition assignment with throttle | kafka-reassign-partitions --execute --reassignment-json-file J --throttle R |
|
Change throttled rate | kafka-reassign-partitions --execute --reassignment-json-file J --throttle R |
|
Check progress of a reassignment | kafka-reassign-partitions --progress --reassignment-json-file J | (see KIP-113) |
Check result and clear throttle | kafka-reassign-partitions --verify --reassignment-json-file J |
|
kafka-reassign-partitions.sh
and ReassignPartitionsCommand
The --zookeeper
option will be retained and will:
- Cause a deprecation warning to be printed to standard error. The message will say that the
--zookeeper
option will be removed in a future version and that--bootstrap-server
is the replacement option. - Perform the reassignment via ZooKeeper, as currently.
A new --bootstrap-server
option will be added and will:
- Perform the reassignment via the AdminClient API (described below) using the given broker(s) as bootstrap brokers.
- When used with --execute and --throttle, the throttle will be an auto-removed one.
Using both --zookeeper
and --bootstrap-server
in the same command will produce an error message and the tool will exit without doing the intended operation.
It is anticipated that a future version of Kafka would remove support for the --zookeeper
option.
A new --progress
action option will be added. This will only be supported when used with --bootstrap-server
. If used with --zookeeper
the command will produce an error message and the tool will exit without doing the intended operation. --progress
will report on the synchronisation of each of the partitions and brokers in the reassignment given via the --reassignment-json-file
option
For example:
# If the following command is used to start a reassignment bin/kafka-reassign-partitions.sh --bootstrap-server localhost:9878 \ --reassignment-json-file expand-cluster-reassignment.json \ --execute # then the following command will print the progress of # that reassignment, then exit immediately bin/kafka-reassign-partitions.sh --bootstrap-server localhost:9878 \ --reassignment-json-file expand-cluster-reassignment.json \ --progress
That might print something like the following:
Topic Partition Broker Status ------------------------------------- my_topic 0 0 In sync my_topic 0 1 Behind: 10456 messages behind asdf 0 1 Unknown topic my_topic 42 1 Unknown partition my_topic 0 42 Unknown broker my_topic 1 0 Broker does not host this partition
The implementation of --progress
will make use of the
method from KIP-113 to find the lag of the syncing replica.describeReplicaLogDirs
()
Internally, the ReassignPartitionsCommand
will be refactored to support the above changes to the options. An interface will abstract the commands currently issued directly to zookeeper.
There will be an implementation which makes the current calls to ZooKeeper, and another implementation which uses the AdminClient
API described below.
In all other respects, the public API of ReassignPartitionsCommand
will not be changed.
AdminClient: reassignPartitions()
This API will support a number of use cases:
- Changing the partition assignment, via
kafka-reassign-partitions.sh
- Changing the replication factor, via
kafka-reassign-partitions.sh
Notes:
- This API is asynchronous in the sense that the client cannot assume that the request is complete (or the request was rejected) once they have obtained the result for the topic from the
ReassignPartitionsResult
. - The
method from KIP-113 can be used to determine progress.describeReplicaLogDirs
() A call to
reassignPartitions()
with thevalidateOnly
option can be used to determine whether a reassignment is currently running, and therefore whether the last reassignment has finished.
/** * <p>Reassign the partitions given as the key of the given <code>assignments</code> to the corresponding * list of brokers. The first broker in each list is the one which holds the "preferred replica".</p> * * <p>Inter-broker reassignment causes significant inter-broker traffic and can take a long time * in order to copy the replica data to brokers. The given options can be used impose a quota on * inter-broker traffic for the duration of the reassignment so that client-broker traffic is not * adversely affected.</p> * * <h3>Preferred replica</h3> * <p>When brokers are configured with <code>auto.leader.rebalance.enable=true</code>, the broker * with the preferred replica will be elected leader automatically. * <code>kafka-preferred-replica-election.sh</code> provides a manual trigger for this * election when <code>auto.leader.rebalance.enable=false</code>.</p> * * @param assignments The partition assignments. * @param options The options to use when reassigning the partitions * @return The ReassignPartitionsResult */ public abstract ReassignPartitionsResult reassignPartitions(Map<TopicPartition, List<Integer>> assignments, ReassignPartitionsOptions options);
Where:
public class ReassignPartitionsOptions extends AbstractOptions<ReassignPartitionsOptions> { // Note timeoutMs() inherited from AbstractOptions public boolean validateOnly() /** * Validate the request only: Do not actually trigger replica reassignment. */ public ReassignPartitionsOptions validateOnly(boolean validateOnly) public long throttle() { return throttle; } /** * <p>Set the throttle rate and throttled replicas for the reassignments. * The given throttle is in bytes/second and should be at least 1 KB/s. * Interbroker replication traffic will be throttled to approximately the given value. * Use Long.MAX_VALUE if the reassignment should not be throttled.</p> * * <p>A positive throttle is equivalent to setting:</p> * <ul> * <li>The leader and follower throttled rates to the given value given by throttle.</li> * <li>The leader throttled replicas of each topic in the request to include the existing brokers having * replicas of the partitions in the request.</li> * <li>The follower throttled replicas of each topic in the request to include the new brokers * for each partition in that topic.</li> * </ul> * * <p>The value of {@link #autoRemoveThrottle()} will determine whether these * throttles will be removed automatically when the reassignment completes.</p> * * @see AdminClient#alterInterbrokerThrottledRate(int, long, long) * @see AdminClient#alterInterbrokerThrottledReplicas(Map) */ public ReassignPartitionsOptions throttle(long throttle) { ... } public boolean autoRemoveThrottle() { ... } /** * True to automatically remove the throttle at the end of the current reassignment. */ public ReassignPartitionsOptions autoRemoveThrottle(boolean autoRemoveThrottle) { ... } } } public class ReassignPartitionsResult { public Map<TopicPartition, KafkaFuture<Void>> values(); public KafkaFuture<Void> all(); }
Network Protocol: ReassignPartitionsRequest and ReassignPartitionsResponse
A ReassignPartitionsRequest
initiates the movement of replicas between brokers, and is the basis of the AdminClient.reassignPartitions()
method
Notes:
- The request must be sent to the controller.
- The request requires the
Alter
operation on theCLUSTER
resource, since it can require significant inter-broker communication. - The request will be subject to a policy, as described in KIP-201.
ReassignPartitionsRequest => [topic_reassignments] timeout validate_only throttle remove_throttle topic_reassignments => topic [partition_reassignments] topic => STRING partition_reassignments => partition_id [broker] partition_id => INT32 broker => INT32 timeout => INT32 validate_only => BOOLEAN throttle => INT64 remove_throttle => BOOLEAN
Where
Field | Description |
---|---|
topic | the name of a topic |
partition_id | a partition of that topic |
broker | a broker id |
timeout | The maximum time to await a response in ms. |
validate_only | when true: validate the request, but don't actually reassign the partitions |
Algorithm
- The controller validates the request against configured authz, policy etc.
- The controller computes set of topics in the request, and writes this as JSON to the new /admin/throttled_replicas_removal znode
- The controller then updates the existing leader.replication.throttled.replicas and follower.replication.throttled.replicas properties of each topic config.
- The controller computes the union of 1) The brokers currently hosting replicas of the topic partitions in the request 2) the brokers assigned to host topic partitions in the request, and write this as JSON to the new /admin/throttled_rates_removal znode.
- The controller then updates the existing leader.replication.throttled.rates and follower.replication.throttled.rates properties of each broker config.
- The controller writes reassignment JSON to the
/admin/reassign_partitions
znode
A ReassignPartitionsResponse
describes which partitions in the request will be moved, and what was wrong with the request for those partitions which will not be moved.
ReassignPartitionsResponse => throttle_time_ms [reassign_partition_result] throttle_time_ms => INT32 reassign_partition_result => topic [partition_error] topic => STRING partition_error => partition_id error_code error_message partition_id => INT32 error_code => INT16 error_message => NULLABLE_STRING
Where
Field | Description |
---|---|
throttle_time_ms | duration in milliseconds for which the request was throttled |
topic | a topic name from the request |
partition_id | a partition id for that topic, from the request |
error_code | an error code for that topic partition |
error_message | more detailed information about any error for that topic |
Anticipated errors:
CLUSTER_AUTHORIZATION_FAILED
(31) Authorization failed- POLICY_VIOLATION(44) The request violated the configured policy
INVALID_TOPIC_EXCEPTION
(17) If the topic doesn't existUNKNOWN_MEMBER_ID
(25) If any broker ids in the partition_reassignments included an unknown broker idINVALID_REQUEST
(42) If duplicate topics appeared in the requestPARTITION_REASSIGNMENT_IN_PROGRESS
(new) If the reassignment cannot be started because a reassignment is currently running (i.e. the/admin/reassign_partitions
znode exists)- INVALID_THROTTLE (new) If the given throttle is <=0.
INVALID_REPLICA_ASSIGNMENT
(39) If a partition, replica or broker id in the partition_assignment doesn't exist or is incompatible with the requested num_partitions and /or replication_factor. The error_message would contain further information.NONE
(0) reassignment has started
AdminClient: alterInterbrokerThrottledRates()
/** * Change the rate at which interbroker replication is throttled, replacing existing throttled rates. * The given {@code leaderRate} is the throttled rate when the broker is acting as leader. * The given {@code followerRate} is the throttled rate when the broker is acting as follower. * For the throttled rates to take effect, the given broker must also be present in the * list of throttled replicas, which can be set by {@link #alterInterbrokerThrottledReplicas()}. * The throttled rates will be removed at the end of the current reassignment. */ AlterInterbrokerThrottledRateResult alterInterbrokerThrottledRate(Map<Integer, ThrottledRate>, AlterInterbrokerThrottledRateOptions options) // HOW do I get the current throttled rate? class AlterInterbrokerThrottledRateOptions { boolean autoRemoveThrottle() ReassignPartitionsOptions autoRemoveThrottle(boolean) }
Where:
public class ThrottledRate { public ThrottledRate(long leaderRate, long followerRate) { ... } long leaderRate() { ... } long followerRate() { ... } } public class AlterInterbrokerThrottledRateOptions extends AbstractOptions<AlterInterbrokerThrottledRateOptions> { public boolean autoRemoveThrottle() { ... } /** * True to automatically remove the throttle at the end of the current reassignment. */ public AlterInterbrokerThrottledRateOptions autoRemoveThrottle(boolean autoRemoveThrottle) { ... } } public class AlterInterbrokerThrottledRateResult { // package-access ctor public Map<Integer, KafkaFuture<Void>> values() { ... } public KafkaFuture<Void> all() { ... } }
Network API: AlterInterbrokerThrottledRatesRequest and AlterInterbrokerThrottledRatesResponse
AlterInterbrokerThrottledRatesRequest => [broker_throttles] remove_throttle timeout validate_only broker_throttles => broker_id leader_rate follower_rate broker_id => INT32 leader_rate => INT64 follower_rate => INT64 timeout => INT32 validate_only => BOOLEAN remove_throttle => BOOLEAN
Algorithm:
- The controller validates the brokers and rates in the request.
- The controller gets the current value of the /admin/throttled_rates_removal znode, forms the union of those brokers with those in the request and updates the /admin/throttled_rates_removal znode with JSON representation of this union
- The controller then subtracts the brokers in the request from the current brokers and removes the leader.replication.throttled.rates and follower.replication.throttled.rates properties from each broker config
- The controller then, for each broker in the request, adds the leader.replication.throttled.rates and follower.replication.throttled.rates properties to each broker config.
- The controller then updates /admin/throttled_rates_removal znode with JSON representation of brokers in the request.
AlterInterbrokerThrottledRatesResponse => [broker_error] broker_error => broker_id error_code error_messgae broker_id => INT32 error_code => INT16 error_message => NULLABLE_STRING
AdminClient: alterInterbrokerThrottledReplicas()
/** * Set the partitions and brokers subject to the * {@linkplain #alterInterbrokerThrottledRate(Map) * interbroker throttled rate}. * The brokers specified as the {@link ThrottledReplicas#leaders()} corresponding to a * topic partition given in {@code replicas} will be subject to the leader throttled rate * when acting as the leader for that partition. * The brokers specified as the {@link ThrottledReplicas#followers()} corresponding to a * topic partition given in {@code repicas} will be subject to the follower throttled rate * when acting as the follower for that partition. * If the given {@code replicas} is null then the throttle will apply to all topic partitions on all brokers. * TODO No! It should depend on the current reassignment * * The throttle will be automatically removed at the end of the current reassignment, * unless overridden in the given options. * * The current throttled replicas can be obtained via {@link #describeConfigs(Collection)} with a * ConfigResource with type {@link ConfigResource.Type#TOPIC TOPIC} and name "leader.replication.throttled.replicas" * or "follower.replication.throttled.replicas". */ public abstract AlterInterbrokerThrottledReplicasResult alterInterbrokerThrottledReplicas( Map<TopicPartition, ThrottledReplicas> replicas, AlterInterbrokerThrottledReplicasOptions options);
Where:
public class ThrottledReplicas { public ThrottledReplicas(Collection<Integer> leaders, Collection<Integer> followers) { ... } /** * The brokers which should be throttled when acting as leader */ public Collection<Integer> leaders() { .. } /** * The brokers which should be throttled when acting as follower */ public Collection<Integer> followers() { ... } } public class AlterInterbrokerThrottledReplicasOptions extends AbstractOptions<AlterInterbrokerThrottledReplicasOptions> { public boolean autoRemoveThrottle() { ... } /** * True to automatically remove the throttle at the end of the current reassignment. */ public AlterInterbrokerThrottledReplicasOptions autoRemoveThrottle(boolean autoRemoveThrottle) { ... } } public class AlterInterbrokerThrottledReplicasResult { // package-access ctor public Map<TopicPartition, KafkaFuture<Void>> values() { ... } public KafkaFuture<Void> all() { ... } }
Network API: AlterInterbrokerThrottledReplicasRequest and AlterInterbrokerThrottledReplicasResponse
AlterInterbrokerThrottledRatesRequest => [topic_throttles] remove_throttle timeout validate_only topic_throttles => topic [partition_throttles] topic => STRING partition_throttles => partition_id [broker_id] partition_id => INT32 broker_id => INT32 timeout => INT32 validate_only => BOOLEAN remove_throttle => BOOLEAN
Algorithm:
- The controller validates the partitions and brokers in the request.
- The controller gets the current value of the /admin/throttled_replicas_removal znode, forms the union of those topics with those in the request and updates the /admin/throttled_rates_removal znode with JSON representation of this union
- The controller then subtracts the topics in the request from the current topics and removes the leader.replication.throttled.replicas and follower.replication.throttled.replicas properties from each topic config
- The controller then, for each topic in the request, adds the leader.replication.throttled.rates and follower.replication.throttled.rates properties to each topic config.
- The controller then updates /admin/throttled_replicas_removal znode with JSON representation of topics in the request.
AlterInterbrokerThrottledReplicasResponse => [topic_errors] topic_errors => topic [partition_errors] topic => STRING partition_errors => partition_id error_code error_messgae partition_id => INT32 error_code => INT16 error_message => NULLABLE_STRING
On Controller Failover
The algorithms presented above, using the new znodes, are constructed so that should the controller fail, on election of a new controller ZooKeeper is not left in an inconsistent state where throttles which should be removed automatically are not removed at the end of the reassignment. The recovery algorithm is as follows:
- If the /admin/reassign_partitions znode exists we assume a reassignment is on-going and do nothing.
- Otherwise, if the /admin/reassign_partitions znode does not exists, we proceed to remove the throttles, as detailed in "Throttle removal" section below.
On reassignment completion
When reassignment is complete:
- The /admin/reassign_partitions znode gets removed.
- We remove the throttles, as detailed in "Throttle removal" section below.
Throttle removal
- If /admin/remove_throttled_replicas is set:
For each of the topics listed in that znode:
Remove the (leader|follower).replication.throttled.replicas propertes for that topic config.
- Remove the /admin/remove_throttled_replicas znode
The symmetric algorithm is used for /admin/remove_throttled_rates, only with broker configs.
Compatibility, Deprecation, and Migration Plan
Existing users of the kafka-reassign-partitions.sh
will receive a deprecation warning when they use the --zookeeper
option. The option will be removed in a future version of Kafka. If this KIP is introduced in version 1.0.0 the removal could happen in 2.0.0.
Implementing AdminClient.alterConfigs()
for (dynamic) broker configs was considered as a way of implementing throttle management but this would not support the auto removal feature.
Not supporting passing a throttle in the AdminClient.reassignPartitions() (and just using the APIs for altering throttles) was considered, but:
- Being able to specify a throttle at the same time at starting the reassignment is very convenient.
- Race conditions are possible if the APIs requires throttles set up before reassignment starts. What if reassignPartitions() doesn't get called, or none of the partitions in the call can be be reassigned?
Rejected Alternatives
If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.
One alternative is to do nothing: Let the ReassignPartitionsCommand continue to communicate with ZooKeeper directly.
Another alternative is to do exactly this KIP, but without the deprecation of --zookeeper
. That would have a higher long term maintenance burden, and would prevent any future plans to, for example, provide alternative cluster technologies than ZooKeeper.
An alterTopics()
AdminClient API, mirroring the existing createTopics()
API, was considered, but:
- Some calls to
alterTopics()
(such as increasing the partition count) would have been synchronous, while others (such as moving replicas between brokers) would have been long running and thus asynchronous. This made for an API which synchronousness depended on the arguments. createTopics()
allows to specify topic configs, whereasalterConfigs()
is already provided to change topic configs, so it wasn't an exact mirror
Just providing reassignPartitions()
was considered, with changes to partition count inferred from partitions present in the assignments
argument. This would require the caller to provide an assignment of partitions to brokers, whereas currently it's possible to increase the partition count without specifying an assignment. It also suffered from the synchronous/asynchronous API problem.
Similarly a alterReplicationFactors()
method, separate from reassignPartitions()
was considered, but both require a partition to brokers assignment, and both are implemented in the same way (by writing to the /admin/reassign_partitions
znode), so there didn't seem much point in making an API which distinguished them.
Algorithms making use of the ZooKeeper "multi" feature for atomic update of multiple znodes were considered. It wasn't clear that these would be better than the algorithms presented above.