...

Rebalancing is the process where a group of consumer instances (belonging to the same group) co-ordinate to own a mutually exclusive set of partitions of topics that the group is subscribed to. At the end of a successful rebalance operation for a consumer group, every partition for all subscribed topics will be owned by a single consumer instance within the group. The way rebalancing works is as follows. Every broker is elected as the coordinator for a subset of the consumer groups. The co-ordinator broker for a group is responsible for co-ordinating orchestrating a rebalance operation on consumer group membership changes or subscribed topic partition changes for the subscribed topics. It is also responsible for communicating the resulting partition ownership configuration to all consumers of the group undergoing a rebalance operation.

...

The consumer specifies a session timeout in the JoinGroupRequest that it sends to the co-ordinator in order to join a consumer group. When the consumer has successfully joined a group, the failure detection process starts on the consumer as well as the co-ordinator. The consumer initiates periodic heartbeats (HeartbeatRequest), every session.timeout.ms/heartbeat.frequency to the co-ordinator and waits for a response. If the co-ordinator does not receive a HeartbeatRequest from a consumer for session.timeout.ms, it marks the consumer dead. Similarly, if the consumer does not receive the HeartbeatResponse within session.timeout.ms, it assumes the co-ordinator is dead and starts the co-ordinator rediscovery process. Here is the protocol in more detail -

1. After receiving a ConsumerMetadataResponse or a JoinGroupResponse, a consumer periodically sends a HeartbeatRequest to the coordinator every session.timeout.ms/heartbeat.frequency milliseconds.
2. Upon receiving the HeartbeatRequest, coordinator checks the generation number, the consumer id and the consumer group. If the consumer specifies an invalid or stale generation id, it send an IllegalGeneration error code in the HeartbeatResponse.
3. If the coordinator does not receive a HeartbeatRequest from a consumer at least once in session.timeout.ms, it marks the consumer dead and triggers a rebalance process for the group.
4. If the consumer does not receive a HeartbeatResponse from the coordinator after session.timeout.ms or finds the socket channel to the coordinator to be closed, it treats the coordinator as failed and triggers the co-ordinator re-discovery process.

Note that on coordinator failover, the consumers may discover the new coordinator before or after the new coordinator has finished the failover process including loading the consumer group metadata from ZK, etc. In the latter case, the new coordinator will just accept its ping request as normal; in the former case, the new coordinator may reject its request, causing it to re-dicover the co-ordinator and re-connect again, which is fine. Also, if the consumer connects to the new coordinator too late, the co-ordinator may have marked the consumer dead and will be treat the consumer as a new consumer, which is also fine.

State diagram

Consumer

Here is a description of the states -

Down - The consumer process is down

Startup up & discover co-ordinator - In this state, the consumer discovers the co-ordinator for it's group

The heartbeat.frequency is a consumer side config that determins the interval at which the consumer sends a heartbeat to the co-ordinator. The heartbeat.frequency puts a lower bound on the rebalance latency since the co-ordinator notifies a consumer regarding a rebalance operation through the heartbeat response. So, it is important to set it to a reasonably high value especially if the session.timeout.ms is relatively large. At the same time, care should be taken to not set the heartbeat.frequency too high since that would cause overhead due to request overload on the brokers. Here is the protocol in more detail -

1. After receiving a ConsumerMetadataResponse or a JoinGroupResponse, a consumer periodically sends a HeartbeatRequest to the coordinator every session.timeout.ms/heartbeat.frequency milliseconds.
2. Upon receiving the HeartbeatRequest, coordinator checks the generation number, the consumer id and the consumer group. If the consumer specifies an invalid or stale generation id, it send an IllegalGeneration error code in the HeartbeatResponse. If the proposed session.timeout.ms is too high, it sends the SessionTimeoutTooHigh error code in the HeartbeatResponse.
3. If the coordinator does not receive a HeartbeatRequest from a consumer at least once in session.timeout.ms, it marks the consumer dead and triggers a rebalance process for the group.
4. If the consumer does not receive a HeartbeatResponse from the coordinator after session.timeout.ms or finds the socket channel to the coordinator to be closed, it treats the coordinator as failed and triggers the co-ordinator re-discovery process.

Note that on coordinator failover, the consumers may discover the new coordinator before or after the new coordinator has finished the failover process including loading the consumer group metadata from ZK, etc. In the latter case, the new coordinator will just accept its heartbeat request as normal; in the former case, the new coordinator may reject its request, causing it to re-dicover the co-ordinator and re-connect again, which is fine. Also, if the consumer connects to the new coordinator too late, the co-ordinator may have marked the consumer dead and will be treat the consumer like a new consumer and trigger a rebalance.

State diagram

Consumer

Here is a description of the states -

Down - The consumer process is down

Startup up & discover co-ordinator - In this state, the consumer discovers the co-ordinator for it's group

Part of Part of a group - In this state, the consumer is part of a group if it receives a JoinGroupResponse with no error code, a consumer id and the generation id for it's group.

...

1. After startup, a consumer learns it's consumer id in the very first JoinGroupResponse it receives from the co-ordinator. From that point onwards, the consumer is expected to must include this consumer id in every request HeartbeatRequest and OffsetCommitRequest it sends to the co-ordinator (HeartbeatRequest, JoinGroupRequest, OffsetCommitRequest). . If the co-ordinator receives a HeartbeatRequest or an OffsetCommitRequest with a consumer id that is different from the ones in the group, it sends an UnknownConsumer error code in the corresponding responses.
   
2. The co-ordinator assigns a consumer id to a consumer on a successful rebalance and sends it in the JoinGroupResponse. The consumer should consumer can choose to include this id in every subsequent JoinGroupRequest as well until it is shutdown or dies.
   
3. The co-ordinator does consumer id assignment after it has received a JoinGroupRequest from all existing consumers in a group. At this point, it assigns a new id <group>-<consumer_host>-<sequence> to every consumer that did not have a consumer id in the JoinGroupRequest. The assumption is that such consumers are newly started up.
   
4. If a consumer fails to send the same consumer id on subsequent JoinGroupRequests, it will cause a chain of rebalance attempts and can cause the group to never finish a rebalance operation successfully. This is because the way a co-ordinator knows that a rebalance operation should be triggered due to a new consumer, is by checking the consumer id in the JoinGroupRequest. If there is no consumer id, it assumes that a new consumer wants to join the group.
   
5. If a consumer id is specified in the JoinGroupRequest but it does not match the ids in the current group membership, the co-ordinator sends an UnknownConsumer error code in the JoinGroupResponse and prevents the consumer from joining the group. This does not cause a rebalance operation for the rest of the consumers in the group, but also does not allow such a consumer to join an existing group.

Memory management

TBD

Request formats

For each consumer group, the coordinator stores the following information:

1) For each consumer group, the group metadata containing:

1. The advantage of including the consumer id in subsequent JoinGroupRequests is a lower latency on rebalance operations. When a rebalance is triggered, the co-ordinator waits for all consumers in the previous generation to send it a JoinGroupRequest. The way it identifies a consumer is by it's consumer id. If the consumer chooses to send JoinGroupRequest with empty consumer id, the co-ordinator waits a full session.timeout.ms before it proceeds with the rest of the rebalance operation. It does this since there is no way to map the incoming JoinGroupRequest with a consumer in the absence of a consumer id. This puts a lower bound on the rebalance latency (session.timeout.ms). On the other hand, if the consumer sends it's consumer id on subsequent JoinGroupRequests, the co-ordinator can immediately identify the consumers and proceed with a rebalance once all the known consumers have sent a JoinGroupRequest.
   
2. The co-ordinator does consumer id assignment after it has received a JoinGroupRequest from all existing consumers in a group. At this point, it assigns a new id <group>-<uuid> to every consumer that did not have a consumer id in the JoinGroupRequest. The assumption is that such consumers are either newly started up or choose to not send the previously assigned consumer id.
   
3. If a consumer id is specified in the JoinGroupRequest but it does not match the ids in the current group membership, the co-ordinator sends an UnknownConsumer error code in the JoinGroupResponse and prevents the consumer from joining the group. This does not cause a rebalance operation for the rest of the consumers in the group, but also does not allow such a consumer to join an existing group.

Memory management

TBD

Request formats

For each consumer group, the coordinator stores the following information:

1) For each consumer group, the group metadata containing:

• List of topics the group subscribes to
• Group configs, including session timeout, etc.
• Consumer metadata for each consumer in the group.
• Current offsets for each consumed topic/partition.
• List of topics the group subscribes to
• Group configs, including session timeout, etc.
• Consumer metadata for each consumer in the group.
• Current offsets for each consumed topic/partition.
• Partition ownership metadata, including the consumer-assigned-partitions map.

2) For each existing topic, a list of consumer groups that are currently subscribing to it.

ConsumerMetadataRequest

It is assumed that all the following requests and responses have the common header. So the fields mentioned exclude the ones in the header.

ConsumerMetadataRequest

{
  VersionGroupId                => String
}

ConsumerMetadataResponse

{int16
  CorrelationIdErrorCode              => int32int16
  ClientIdCoordinator               => StringBroker
  GroupIdCoordinatorEpoch       => int32
}

JoinGroupRequest

{
  GroupId                => String

...

ConsumerMetadataResponse

...

  ConsumerHost           => 

...

String
  

...

SessionTimeout         => int32
  Topics  

...

...

...

            => 

...

[String]
  ConsumerId     

...

        => 

...

String 
 }

...

{
  VersionErrorCode                => int16
  CorrelationIdGroupId          => int32
  GroupId   => String 
  GroupGenerationId         => Stringint32
  ConsumerId  ConsumerHost           => String
  SessionTimeoutPartitionsToOwn         => int32[TopicAndPartition]
}

HeartbeatRequest

{
  

...

GroupId     

...

...

...

        => String
  GroupGenerationId      => 

...

int32

...

  ConsumerId             => String

...

...

}

...

HeartbeatResponse

{
  

...

GroupId                => 

...

String
  

...

ConsumerId             => 

...

String

...

  ErrorCode              => int16

...

}

HeartbeatRequest

{
  Version                => int16
  CorrelationId          => int32
  GroupId                => String
  GroupGenerationId      => int32
  ConsumerId             => String
}

HeartbeatResponse

{
  CorrelationId          => int32
  GroupId                => String
  ConsumerId             => String
  ErrorCode              => int16
}

With wildcard subscription (for example, whitelist and blacklist), the consumers are responsible to discover matching topics through topic metadata request. That is, its topic metadata request will contain an empty topic list, whose response then will return the partition info of all topics, it will then filter the topics that match its wildcard expression, and then updates update the subscription list using the subscribe() API. Again, if the subscription list has changed from the previous valuesvalue, it will trigger rebalance for the consumer group by sending the JoinGroupRequest during the next poll().

Interesting scenarios to consider

...

1. On co-ordinator failover, the controller elects a new leader for a subset of the consumer groups affected due to the co-ordinator failure. As part of becoming the leader for a subset of the offset topic partitions, the co-ordinator reads metadata for each group that it is responsible for, from zookeeper. The metadata includes the group's consumer ids, the generation id and the subscribed list of topics. Until the co-ordinator has read all metadata from zookeeper, it returns the CoordinatorStartupNotComplete error code in HeartbeatResponse, ConsumerMetadataResponse. It is illegal for a consumer to send a JoinGroupRequest during this time, so the error code returned to such a consumer is different (probably IllegalProtocolState).
   
2. If a consumer sends a ConsumerMetadataRequest to a broker before the broker has received the updated group metadata through the UpdateMetadataRequest from the controlller, the ConsumerMetadataResponse will return stale information about the co-ordinator. The consumer will receive NotCoordinatorForGroup error code on the heartbeat/commit offset responses. On receiving the NotCoordinatorForGroup error code, the consumer backs off and resends the ConsumerMetadataRequest.
3. The consumer does not stop fetching data during the co-ordinator failover and re-discovery process.
   

...

1. The co-ordinator for a group detects changes to the number of partitions for the subscribed list of topics.
2. The co-ordinator then marks the group ready for rebalance and sends the IllegalGeneration error code in the HeartbeatResponse. The consumer then stops fetching data, commits offsets and sends a JoinGroupRequest to the co-ordinator.
3. The co-ordinator waits for all consumers to send it the JoinGroupRequest for the group. Once it receives all expected JoinGroupRequests, it increments the group's generation id in zookeeper, computes the new partition assignment and returns the updated assignment and the new generation id in the JoinGroupResponsein the JoinGroupResponse. Note that the generation id is incremented even if the group membership does not change.
4. On receiving the JoinGroupResponse, the consumer stores the new generation id and consumer id locally and starts fetching data for the returned set of partitions. The subsequent requests sent by the consumer to the co-ordinator will use the new generation id and consumer id returned by the last JoinGroupResponse.

...

1. Consumer periodically sends a HeartbeatRequest to the coordinator every session.timeout.ms/hearbeat.frequency milliseconds. If the consumer receives the IllegalGeneration error code in the HeartbeatResponse, it stops fetching, commits offsets and sends a JoinGroupRequest to the co-ordinator. Until the consumer receives a JoinGroupResponse, it does not send any more HearbeatRequests to the co-ordinator. 
2. A higher heartbeat.frequency ensures lower latency on a rebalance operation since the co-ordinator notifies a consumer of the need to rebalance only on a HeartbeatResponse. To protect the broker from receiving too many heartbeats, the broker will have a min.heartbeat.interval.ms config that puts a lower bound on the frequency of heartbeats sent by consumers.
3. The co-ordinator pauses failure detection for a consumer that has sent a JoinGroupRequest until a JoinGroupResponse is sent out. It restarts the hearbeat timer once the JoinGroupResponse is sent out and marks a consumer dead if it does not receive a HeartbeatRequest from that time for another session.timeout.ms milliseconds. The reason the co-ordinator stops depending on heartbeats to detect failures during a rebalance is due to a design decision on the broker's socket server - Kafka only allows the broker to read and process one outstanding request per client. This is done to make it simpler to reason about ordering. This prevents the consumer and broker from processing a heartbeat request at the same time as the join group request. Marking failures based on JoinGroupRequest The co-ordinator pauses failure detection for a consumer that has sent a JoinGroupRequest until a JoinGroupResponse is sent out. It restarts the hearbeat timer once the JoinGroupResponse is sent out and marks a consumer dead if it does not receive a HeartbeatRequest from that time for another session.timeout.ms milliseconds. This prevents the co-ordinator from marking a consumer dead during a rebalance operation. Note that this does not prevent the rebalance operation from finishing if a consumer goes through a soft failure during a rebalance operation. If the consumer pauses before it sends a JoinGroupRequest, the co-ordinator will mark it dead during the rebalance and send a JoinGroupResponse to the rest of the consumers with a RebalanceFailed error code. If a consumer pauses after it sends a JoinGroupRequest, the co-ordinator will send it the JoinGroupResponse assuming the rebalance completed successfully and will restart it's heartbeat timer. If the consumer resumes before session.timeout.ms, consumption starts normally. If the consumer pauses for session.timeout.ms after that, then it is marked dead by the co-ordinator and it will trigger a rebalance operation.
   
4. The co-ordinator returns the new generation id only in the JoinGroupResponse. Once the consumer receives a JoinGroupResponse, it sends the next HeartbeatRequest with the new generation id.
   

...