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 [Change the link from the KIP proposal email archive to your own email thread]
JIRA: here
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
Describe the problems you are trying to solve.
Firstly, the ReassignPartitionsCommand
(which is used by the kafka-reassign-partitions.sh
tool) 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 the AdminClient API. Thus it is necessary to change the ReassignPartitionsCommand
so that it no longer talks to ZooKeeper directly, but via an intermediating broker.
Secondly, 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.
Public Interfaces
Briefly list any new interfaces that will be introduced as part of this proposal or any existing interfaces that will be removed or changed. The purpose of this section is to concisely call out the public contract that will come along with this feature.
A public interface is any change to the following:
Binary log format
The network protocol and api behavior
Any class in the public packages under clientsConfiguration, especially client configuration
org/apache/kafka/common/serialization
org/apache/kafka/common
org/apache/kafka/common/errors
org/apache/kafka/clients/producer
org/apache/kafka/clients/consumer (eventually, once stable)
Monitoring
Command line tools and arguments
- Anything else that will likely break existing users in some way when they upgrade
Two new network protocol APIs will be added:
PartitionAssignmentRequest
andPartitionAssignmentResponse
PartitionAssignmentStatusRequest
andPartitionAssignmentStatusResponse
The AdminClient
API will have two new methods added (plus overloads for options):
assignPartitions()
partitionAssignmentStatus()
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, with a new--request
option
Reassignment requests will be identified by a monotonic integer (the "request id"), which will be apparent in each of these public interfaces. The intention is to avoid ambiguity in the event that one reassignment is immediately followed by another. Note that this KIP will not add support for concurrent reassignments, but having a request id is compatible with disambiguating concurrent reassignments, should support for them be added at a later date.
Proposed Changes
Describe the new thing you want to do in appropriate detail. This may be fairly extensive and have large subsections of its own. Or it may be a few sentences. Use judgement based on the scope of the change.
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 given intermediating broker.
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.
New --progress
and --request
options will be added. These 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. The existing --execute
option will cause the request id to be printed to the standard output. This can be used as an argument to the --request
option to identify the request that the --progress
pertains to. If --progress
is used without a --request
option present the most recent request will be used. If an invalid --request
option is given the tool will exit with an error message. For example:
# Start the execution of a reassignment and exit immediately, # printing a reassignment request id to standard output kafka-reassign-partitions.sh --bootstrap-server localhost:1234 \ --execute ... # Assuming the above command printed 42, the following # will print the progress of that reassignment, then exit immediately kafka-reassign-partitions.sh --bootstrap-server localhost:1234 \ --progress --request 42 ...
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
The following methods will be added to AdminClient
to support the ability to reassign partitions:
public AssignPartitionsResult assignPartitions(Map<TopicPartition, List<Integer>>) public AssignPartitionsResult assignPartitions(Map<TopicPartition, List<Integer>>, AssignPartitionOptions options)
Where:
public class AssignPartitionsResult { // private constructor public Map<TopicPartition, KafkaFuture<Void>> values() public KafkaFuture<Void> all() }
The following methods will be added to AdminClient to support the progress reporting functionality:
public PartitionAssignmentStatusResult partitionAssignmentStatus(int requestId) public PartitionAssignmentStatusResult partitionAssignmentStatus(int requestId, PartitionAssignmentStatusOptions options)
Where:
public class PartitionAssignmentStatusResult { // private constructor public Map<TopicPartition, KafkaFuture<Void>> values() public KafkaFuture<Void> all() } public class PartitionAssignmentStatusResult { // private constructor public Map<PartitionReplica, Progress> all() /** The progress of moving the given partition */ public KafkaFuture<Progress> progress(PartitionReplica) /** * Whether the overall reassignment is complete, * that is, if all the values in the map returned by {@link #all()} * have complete() == true */ public KafkaFuture<Boolean> complete() } /** Represents the reassignment of a partition to a broker. */ public class PartitionReplica { public String getTopic() public int getPartition() public int getBroker() // plus override equals() and hashCode() } public class Progress { /** The number of completed units of work. */ long getNumerator() /** The total number of units of work. */ Long getDenominator() /** The error code, if the task ended erroneously. */ Errors getError() }
The Progress
class presented here can easily be reused in future APIs also dealing with progress of long running jobs. The result of getDenominator()
is nullable because it cannot be expected that all tasks can efficiently compute their total size. The isComplete()
method is provided to determine completeness since getNumerator() == getDenominator()
cannot be used in the event of a null denominator.
In general, the denominator will not be constant between successive requests since new messages will arrive from producers. Indeed it is possible for a partition to receive messages from producers at a faster rate than the syncing replica is catching up, i.e. the percentage completion implied by numerator ÷ denominator can decrease.
For partition reassignment, Progress
has sufficient information for a percent completion calculation, but not any kind of time-to-completion, since the AdminClient doesn't know the time that the initial request was made.
Authorization
With broker-mediated reassignment it becomes possible limit the authority to perform reassignment to something finer-grained than "anyone with access to zookeeper".
The reasons for reassignment are usually operational. For example, migrating partitions to new brokers when expanding the cluster, or attempting to find a more balanced assignment (according to some notion of balance). These are cluster-wide considerations and so authority should be for the reassign operation being performed on the cluster.
Given the standard form for authorization in Kafka
"Principal P is [Allowed/Denied] Operation O From Host H On Resource R"
the authorized operation will be ClusterAction
, on the CLUSTER
resource. So to authorize Alice and Bob to perform rebalancing one might need to configure an ACL like this:
bin/kafka-acls.sh --authorizer-properties zookeeper.connect=localhost:2181 \
--add --allow-principal User:Bob --allow-principal User:Alice \
--allow-host 198.51.100.0 --allow-host 198.51.100.1 \
--operation ClusterAction --cluster
Network Protocol: AssignPartitionsRequest
and AssignPartitionsResponse
An AssignPartitionsRequest
will initiate the process of partition reassignment
AssignPartitionsRequest => [PartitionAssignment] PartitionAssignment => Topic Partition Brokers Topic => string Partition =>int32 Brokers => [int32]
Where:
Topic
a topic namePartition
a partition of that topicBrokers
the broker ids which will host the partition
Possible Error Codes:
ClusterAuthorizationFailedCode
(31)PartitionReassignmentInProgress
(new)
As currently, it will not be possible to have multiple reassignments running concurrently, hence the addition of PartitionReassignmentInProgress
It is not necessary to send an AssignPartitionsRequest
to the leader for a given partition. Any broker will do.
The AssignPartitionsResponse
enumerates those topics and partitions in the request, together with any error in initiating reassignment that partition:
AssignPartitionsResponse => RequestId [PartitionAssignmentResult] RequestId => int32 PartitionAssignmentResult => Topic Partition Broker Error Topic => string Partition => int32 Broker => int32 Error => int16
Where:
RequestId
is an id which can be used to make in aPartitionAssignmentStatusRequest
Topic
is a topic namePartition
is a partition idBroker
is the brokerError
is an error code for why the initiation of assignment of the given topic to the given broker failed
The anticipated errors in the header are:
INVALID_TOPIC_EXCEPTION
(17) If the topic doesn't existINVALID_PARTITIONS
(37) If the partition doesn't existUNKNOWN_MEMBER_ID
(25) If theBrokers
in theAssignPartitionsRequest
included an unknown broker id
Network Protocol: PartitionAssignmentStatusRequest and PartitionAssignmentStatusResponse
To request progress information about the given request
PartitionAssignmentStatusRequest => RequestId RequestId => int32
Where:
RequestId
is the identifier from a previousAssignPartitionsResponse
PartitionAssignmentStatusResponse => [RequestId] [PartitionReplica Progress] PartitionReplica => Topic Partition Broker Topic => string Partition => int32 Broker => int32 Progress => Error Numerator Denominator Error => int16 Numerator => int64 Denominator => int64
Where:
RequestId
is the identifier from a previous AssignPartitionsResponseTopic
the a topic namePartition
a partition of the topicBroker
the id of the receiving brokerError
is an error code about the transfer of the partition to the given broker:PartitionReassignmentInProgress
(new) if the reassignment is still in progress,None
(0) if the reassignment completed normally,- Any other error if the reassignment completed exceptionally.
Numerator
the number of items that have been processed so farDenominator
the total number of items we expect to process for the request. Or -1 if the total is not known. Never 0.
The anticipated errors in the header are:
INVALID_REQUEST
(42) If the request id is not known
As with the AdminClient API, the Progess
Schema could be reused in other progress reporting APIs.
Implementation
- The mediating broker will write the reassignment JSON to the
/admin/reassign_partitions
znode, as present. The znode version for/admin/reassign_partitions
becomes the request_id. This means that the request id is monotonic increasing, but that request id may not be contiguous. The monotonic increasing property is only required so that when the request id is omitted it is easy to figure out the last request. - Receiving brokers are notified of the change to
/admin/reassign_partitions
(thus obtaining the request id) and start fetching replicas as necessary, as present - As the leader serves fetch requests to followers, at the same time as deciding whether to update the ISR, it will publish their progress as JSON to
/admin/reassign_partitions_status/$request_id
- To obtain progress the mediating broker will read the
/admin/reassign_partitions_status/$request_id
znodes. - When the leader adds a syncing broker to the ISR the progress JSON for that partition in
/admin/reassign_partitions_status/$request_id
will be marked as complete.
It will be necessary to delete the /admin/reassign_partitions_status/$request_id
znodes to prevent the /admin/reassign_partitions_status
subtree growing arbitrarily. At the same time the leader publishes status to the current /admin/reassign_partitions_status/$request_id
it will remove znodes which had completed a given amount of time ago. This will be configurable via the broker property reassign.partitions.status.min.retention.ms
which will default to 1 day.
Compatibility, Deprecation, and Migration Plan
- What impact (if any) will there be on existing users?
- If we are changing behavior how will we phase out the older behavior?
- If we need special migration tools, describe them here.
- When will we remove the existing behavior?
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.
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.