...

Author: Gokul Ramanan Subramanian

Contributors: Stanislav Kozlovski, Alexandre Dupriez, Tom Bentley, Colin McCabe, Ismael Juma, Boyang Chen, Stanislav Kozlovski

Status

Current state: Voting

...

PR: https://github.com/apache/kafka/pull/8499 (currently only prototype, slightly out of date wrt KIP, but gets the idea across)

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

...

We did some performance experiments to understand the effect of increasing the number of partitions. See Appendix A1 for producer performance, and A2 for topic creation and deletion times. These consistently indicate that having a large number of partitions can lead to a malfunctioning cluster.

Topic creation policy plugins specified via the create.topic.policy.class.name configuration can partially help solve this problem by rejecting requests that result in a large number of partitions. However, these policies cannot produce a replica assignment that respects the partitions limits, instead they can only either accept or reject a request. Therefore, we need a more native solution for addressing the problem of partition limit-aware replica assignment. (See rejected alternatives for more details on why the policy approach does not work.)

We propose having two configurations (a) max.broker.In order to prevent a cluster from entering a bad state due to a large number of topic partitions, we propose having two configurations (a) max.broker.partitions to limit the number of partitions per broker, and (b) max.partitions to limit the number of partitions in the cluster overall. These can act as guard rails ensuring that the cluster is never operating with a higher number of partitions than it can handle.

...

• These limits are cluster-wide. This is obviously true for max.partitions which is meant to apply at the cluster level. However, we choose this for max.broker.partitions too, instead of supporting different values for each broker. This is in alignment with the current recommendation to run homogenous Kafka clusters where all brokers have the same specifications (CPU, RAM, disk etc.).
• If both limits max.partitions and max.broker.partitions are specified, then the more restrictive of the two apply. It is possible that a request is rejected because it causes the max.partitions limit to be hit without causing any broker to hit the max.broker.partitions limit. The vice versa is true as well.
• These limits can be changed at runtime, without restarting brokers. This provides greater flexibility. See the "Rejected alternatives" section for why we did not go with read-only configuration.
• These limits also won't apply to topics created via auto topic creation (currently possible via Metadata API requests) . By enforcing this, we disallow having a backdoor to bypass these limitsuntil KIP-590. With KIP-590, auto-topic creation will leverage the CreateTopics API, and will have same behavior as the creation of any other topic.
• These limits do not apply to internal topics (i.e. __consumer_offsets and __transaction_state), which usually are not configured with too many partitions. This ensures that any internal Kafka behaviors do not break because of partition limitspartition limits. The topic partitions corresponding to these internal topics won't also count towards the limit.
• These limits do not apply when creating topics or partitions, or reassigning partitions via the ZooKeeper-based admin tools. This is unfortunate, because it does create a backdoor to bypass these limits. However, we leave this out of scope here given that ZooKeeper will eventually be deprecated from Kafka.

...

Kafka administrators can specify these in the server.properties file. 

They can also use the following to set/modify these configurations via ZooKeeperthe kafka-config.sh admin tool.

./kafka-config.sh --zookeeperbootstrap-server $ZOOKEEPER$SERVERS --alter --add-config max.broker.partitions=4000 --entity-type brokers --entity-default
./kafka-config.sh --zookeeperbootstrap-server $ZOOKEEPER$SERVERS --alter --add-config max.partitions=200000 --entity-type brokers --entity-default

...

./kafka-config.sh --bootstrap-zookeeperserver $ZOOKEEPER$SERVERS --alter --add-config max.broker.partitions=4000 --entity-type brokers --entity-name 1
./kafka-config.sh --bootstrap-zookeeperserver $ZOOKEEPER$SERVERS --alter --add-config max.partitions=200000 --entity-type brokers --entity-name 1

...

CreateTopics, CreatePartitions, and AlterPartitionReassignments and Metadata APIs will throw the following exception APIs will throw PolicyViolationException and correspondingly the POLICY_VIOLATION(44) error code if it is not possible to satisfy the request while respecting the max.broker.partitions or max.partitions limits. This applies to Metadata requests only in case auto-topic creation is enabled .

public class WillExceedPartitionLimitsException extends ApiException { ... }

post KIP-590, which will modify the Metadata API to call CreateTopics. We will bump up the version of these APIs by one for new clients.

Corresponding to this exception, we will have the following API error code. The actual exception will contain the values of max.broker.partitions and max.partitions in order to make it easy for users to understand why their request got rejected.

WILL_EXCEED_PARTITION_LIMITS(88, "Cannot satisfy request without exceeding the partition limits", WillExceedPartitionLimitsException::new); 

Proposed Changes

The following table shows the list of methods that will need to change in order to support the max.broker.partitions and max.partitions configurations. (We skip a few internal methods for the sake of simplicity.)

...

For all the methods in the above table that are used in the context of both Kafka API request handling paths and ZooKeeper-based admin tools (`AdminUtils.assignReplicasToBrokers`, `AdminZkClient.createTopicWithAssignment`,  `AdminZkClient.createTopic` and `AdminZkClient.addPartitions`), we will pass the values for maximum number of partitions per broker, maximum number of partitions overall, and the current number of partitions for each broker as arguments.

We will modify the core algorithm for replica assignment in the `AdminUtils.assignReplicasToBrokers` method. The modified algorithm will ensure that as replicas are being assigned to brokers iteratively one partition at a time, if assigning the next partition to a broker causes the broker to exceed the max.broker.partitions limit, then the broker is skipped. If all brokers are skipped successively in a row, then the algorithm will terminate and throw WillExceedPartitionLimitsExceptionthrow PolicyViolationException. The check for max.partitions is much simpler and based purely on the total number of partitions that exist across all brokers.

When the methods are invoked in the context of a Kafka API call, we will get the values for the maximum number of partitions per broker by reading the max.broker.partitions configuration from the `KafkaConfig` object (which holds the current value after applying precedence rules on configuration supplied via server.properties and those set via ZooKeeper). Similarly, we will get the maximum number of partitions overall by reading the max.partitions configuration from the `KafkaConfig` object. We will fetch the current number of partitions for each broker from either the `AdminManager` or `KafkaControllerContext` depending on the method.

When the methods are invoked in the context of ZooKeeper-based admin tools, we will set these limits equal to the maximum int32 value int64 value that Java can represent. This is basically because it is not easy (and we don't want to make it easy) to get a reference to the broker-specific `KafkaConfig` object in this context. We will also set the object representing the current number of partitions for each broker to None, since it is not relevant when the limits are not specified.

...

This change is backwards-compatible in practice because we will set the default values for max.broker.partitions and max.partitions equal to the maximum int32 value int64 value that Java can represent, which is quite large (2^{31 63} - 1). Users will anyway run into system issues far before hitting these limits.

...