Rebalance your kafka topics, partitions, replicas across your cluster
kafkabalancer
allows you to compute the set of rebalancing operations yielding a minimally-unbalanced kafka cluster, given a set of constraints:
- set of allowed brokers (globally, or per partition)
- number of desired replicas (per partition)
- current distribution of replicas (per partition)
- leader reassignment enabled/disabled (globally)
- partition weight (per partition)
The goal is to minimize the workload difference between brokers in the cluster, where the workload of a broker is measured by the sum of the weights of each partition having a replica on that broker. Additionally leaders have to do more work (producers and consumers only operate on the leader, followers fetch from the leaders as well) so the weight applied to leader partitions is assumed to be proportional to the sum of the number of replicas and consumer groups.
The tool is designed to be used iteratively: at each iteration only a single reassignment operation is returned by kafkabalancer. This is useful in a automated framework like the following:
forever:
if !kafka_cluster_is_nominal:
continue
state = get_current_state()
change = kafkabalancer(state)
if change:
apply(change)
go get github.com/cafxx/kafkabalancer
Run kafkabalancer -help
for usage instructions.
Usage of ./kafkabalancer:
-allow-leader
Consider the partition leader eligible for rebalancing
-broker-ids string
Comma-separated list of broker IDs (default "auto")
-from-zk string
Zookeeper connection string (can not be used with -input)
-full-output
Output the full partition list: by default only the changes are printed
-help
Display usage
-input string
Name of the file to read (if no file is specified read from stdin, can not be used with -from-zk)
-input-json
Parse the input as JSON
-max-reassign int
Maximum number of reassignments to generate (default 1)
-min-replicas int
Minimum number of replicas for a partition to be eligible for rebalancing (default 2)
-min-unbalance float
Minimum unbalance value required to perform rebalancing (default 1e-05)
-pprof
Enable CPU profiling
To emit the first suggested rebalancing operation for your Kafka cluster ($ZK
is the comma-separated list of your zookeeper brokers):
kafkabalancer -from-zk $ZK > reassignment.json
To perform the suggested change, run the following command:
kafka-reassign-partitions.sh --zookeeper $ZK --reassignment-json-file reassignment.json --execute
If you want to generate/run more than a single rebalancing operation, specify a value greater than 1
for -max-reassign
.
First dump the list of partitions from your Kafka broker ($ZK
is the comma-separated list of your zookeeper brokers):
kafka-topics.sh --zookeeper $ZK --describe > kafka-topics.txt
Next run kafkabalancer
on the list (note: this assumes that all partitions have the same weight and no consumers; this is functionally OK but could lead to suboptimal load distribution). kafkabalancer
will analyze the list and suggest one or more reassignments:
kafkabalancer -input kafka-topics.txt > reassignment.json
To perform the suggested change(s), run the following command:
kafka-reassign-partitions.sh --zookeeper $ZK --reassignment-json-file reassignment.json --execute
If you want to generate/run more than a single rebalancing operation, specify a value greater than 1
for -max-reassign
.
- parse the output of kafka-topic.sh --describe or the Kafka cluster state in Zookeeper
- parse the reassignment JSON format
- output the reassignment JSON format
- minimize leader unbalance (maximize global throughput)
- parse the output of kafka-offset.sh to get the per-partiton weights (number of messages)
- fetch elsewhere additional metrics to refine the weights (e.g. number of consumers, size of messages)
- minimize same-broker colocation of partitions of the same topic (maximize per-topic throughput)
- proactively minimize unbalance caused by broker failure (i.e. minimize unbalance caused by one or more brokers failing) keeping into considerations how followers are elected to leaders when the leader fails
- consider N-way rebalancing plans (e.g. swap two replicas) to avoid local minima
- prefer to relocate "small" partitions to minimize the additional load due to moving data between brokers
- use something like https://github.com/wvanbergen/kazoo-go to query state directly
- use something like https://github.com/wvanbergen/kazoo-go to apply changes directly
This section lists some examples of how kafkabalancer
operates.
Setting broker-ids=1,2,3
will move partition 1 from broker 2 to broker 3 to equalize the load:
Part | Original | Output |
---|---|---|
1 | 1,2 | 1,3 |
2 | 2,1 | 2,1 |
Setting broker-ids=1,2,3,4
will move partition 1 from brokers 1,2 to brokers 4,3 to equalize the load:
Part | Original | Output |
---|---|---|
1 | 1,2 | 4,3 |
2 | 2,1 | 2,1 |
Setting broker-ids=1,2
will move partition 3 from broker 3 to broker 2 to equalize the load:
Part | Original | Output |
---|---|---|
1 | 1,2 | 1,2 |
2 | 1 | 1 |
3 | 3 | 2 |
Setting broker-ids=1
will return error because the partition 1 requires 2 replicas.
Setting NumReplicas=2
for partition 3 will add a replica on broker 2 to equalize the load.
Part | Original | Output |
---|---|---|
1 | 1,2 | 1,2 |
2 | 1,3 | 1,3 |
3 | 3 | 2,3 |
Setting NumReplicas=1
for partition 1 will remove the replica from broker 1 to equalize the load.
Part | Original | Output |
---|---|---|
1 | 1,2 | 2 |
2 | 1 | 1 |
If no changes need to be made, kafkabalancer
will simply seek to equalize the load between brokers:
Part | Original | Output |
---|---|---|
1 | 1,2,3 | 1,4,3 |
2 | 1,2,4 | 1,2,4 |
3 | 1,2,3 | 1,2,3 |
If leader moving is enabled, also the leaders are eligible for rebalancing:
Part | Original | Output |
---|---|---|
1 | 1,2,3 | 4,2,1 |
2 | 1,2,4 | 3,2,4 |
3 | 1,2,3 | 1,2,3 |
kafkabalancer
rebalancing capabilities are split in a series of step executed in order. The order of steps is chosen to prioritize constraints application first and then performance optimization.
The steps are defined in steps.go
and the ordering of the steps in balancer.go
.
When the steps need to identify the relative load of the cluster nodes, they use each partition weight as a relative measure of the workload the cluster has to sustain for that particular partition. The partition weight is then scaled by a multiplier and added to the total load of each node; the multiplier depends on the role of the node for that partition and is defined as:
Leader | Follower |
---|---|
(Replicas)+(Consumers) |
1 |
Where (Replicas)
and (Consumers)
are, respectively, the number of replicas and consumers of the partition.
These steps simply validate that the input data is consistent and they fill in any default value that is not explicitely defined.
These steps deal with any changes in the desired number of replicas by either removing replicas from the highest-loaded cluster nodes or by adding replicas to the lowest-loaded cluster nodes.
This step detects if any replica is currently on a broker that is not in the list of allowed brokers and, if so, it moves those replicas to the lowest-loaded allowed brokers.
These steps attempt to redistribute replicas to minimize the load difference between brokers (see the section above to understand the metric used to measure load on each broker).
MoveLeaders
is a no-op if you don't specify -allow-leader
. Leaders are moved before followers because the weight of leader partitions is normally greater than the one of follower partitions.
Carlo Alberto Ferraris (@cafxx)