You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 2 Next »

Current stateUnder Discussion

Discussion thread:  TBD

JIRAhttps://issues.apache.org/jira/browse/KAFKA-7728

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

Motivation

The goal of this KIP continues the effort in KIP-345 to mitigate unnecessary rebalances in order to achieve better consumer performance. Today consumer group triggers rebalance on the following circumstances:

  1. A new member joins the group with UNKNOWN_MEMBER_ID
  2. A known member joins with changed metadata
  3. A leader rejoins the group
  4. A current member gets session timeout/leaves the group

We already aim to address 1, 4 through KIP-345 by applying group.instance.id to recognize members as static throughout restarts, and avoid sending LeaveGroupRequest when the member is under static membership, thus only using session timeout to kick off expired members. For circumstance 2, it is clear that if a member has metadata update such as assignment protocol change should require another rebalance to address this change as necessary. However, we have space to improve circumstance 3, because it is not a valid condition to trigger rebalance for the most time when the leader instance is just doing a restart under static membership. It is beneficial to distinguish whether leader is rejoining for the sake of rebalance, or is rejoining just due to service restart. By specifying the join reason of the request could entirely avoid rebalance during normal consumer bounces.

Furthermore, as we are promoting incremental rebalances such as KIP-415, later we hope to support stateful consumers such as KStream group to have new member only taking in standby task and give them time to replay the state when first joined. These new followers need to indicate a change of status when they have finished replaying the state. If no JoinReason is specified, brokers will not be able to distinguish the joiner's purpose: whether you are requiring an incremental rebalance, or you are just joining for restart? 

In conclusion, having JoinReason to gracefully handle the problem of rebalance necessity could simplify the implementation logic by a lot, and hide enough details to brokers' perspective on whether to move the group towards PrepareRebalance.  

Public Interfaces

We will add a new enum field to the JoinGroupRequest interface, and bump the version to v6:

JoinGroupRequest => GroupId SessionTimeout RebalanceTimeout MemberId GroupInstanceId ProtocolType GroupProtocols JoinReason
  GroupId             => String
  SessionTimeout      => int32
  RebalanceTimeout	  => int32
  MemberId            => String
  GroupInstanceId     => String
  ProtocolType        => String
  GroupProtocols      => [Protocol MemberMetadata]
  Protocol            => String
  MemberMetadata      => bytes
  JoinReason		  => Enum


JoinGroupRequest.java
public static Schema[] schemaVersions() {
    return new Schema[] {JOIN_GROUP_REQUEST_V0, JOIN_GROUP_REQUEST_V1, JOIN_GROUP_REQUEST_V2, JOIN_GROUP_REQUEST_V3, JOIN_GROUP_REQUEST_V5, JOIN_GROUP_REQUEST_V6};
}


JoinGroupResponse.java
public static Schema[] schemaVersions() {
    return new Schema[] {JOIN_GROUP_RESPONSE_V0, JOIN_GROUP_RESPONSE_V1, JOIN_GROUP_RESPONSE_V2, JOIN_GROUP_RESPONSE_V3, JOIN_GROUP_RESPONSE_V4, JOIN_GROUP_REQUEST_V5, JOIN_GROUP_REQUEST_V6};
}


For the version one, we will just have two JoinReason to handle the leader rejoin case:

public enum JoinReason {
  BLIND("blind"), // Join request from a start-up consumer 
  TOPIC_METADATA_CHANGE("topic_metadata_change"); // The topic metadata has changed (must be from the leader)
}

Proposed Changes

When leader consumer detects a change of topic metadata, it should currently be the only case when it wants to trigger a group rebalance. We will explicitly set the JoinReason to `topic_metadata_change` so that group coordinator will proceed to rebalance stage. If there is no member protocol change and JoinReason is `Blind`, then we shall not rebalance on known member rejoin. Be aware that JoinReason takes lower priority than the protocol change monitoring.

Compatibility, Deprecation, and Migration Plan

  • This change doesn't involve backward incompatible change. Broker will still be able to read existing join group request, and for request < v6 the JoinReason field will be interpreted as "unknown" which shouldn't confuse broker's judgement on whether to trigger rebalance.

Rejected Alternatives

Not applicable so far.

  • No labels