Status

Current state: Under Discussion

Discussion thread: TBD

JIRA:  Unable to render Jira issues macro, execution error.

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

Motivation

Rebalance during scaling up is always painful. Every newly joined member will keep the group at rebalancing stage until all of new instances finished bootstrapping. There could be multiple shuffling of active tasks around existing and new instances, thus decreasing the entire system availability. This negative impact has been mitigated after we introduced KIP-345. Under static membership, user could provide a list of hard-coded `group.instance.id`s to pre-register their identities on broker if the new host info is known, so that broker coordinator could respond to scaling operations more intelligently. For example when we scale up the fleet by defining 4 new client instance ids, the server shall wait until all 4 new members to join the group before kicking out only one rebalance, instead of four in the worst case.

Proposed Changes

This change requires us to change JoinGroup protocol to batch mode in order to easily scale multiple members at once.

Public Interfaces

We will bump JoinGroup request/response version to support batch adding members.

JoinGroupRequest => GroupId SessionTimeout RebalanceTimeout MemberId GroupInstanceId ProtocolType GroupProtocols
  GroupId             => String
  SessionTimeout      => int32
  RebalanceTimeout    => int32
  MemberId            => String // removed
  GroupInstanceId     => String // removed
  ProtocolType        => String 
  GroupProtocols      => [Protocol MemberMetadata]
  Protocol            => String // removed
  MemberMetadata      => bytes // removed
  JoinGroupMembers    => []JoinGroupRequestMember // new
 						   MemberId            => String // new
						   GroupInstanceId     => String // new
						   Protocol            => String // new

JoinGroupResponse => ThrottleTime ErrorCode GenerationId ProtocolName LeaderId MemberId Members
  ThrottleTime           => int16
  ErrorCode              => int16  // removed
  GenerationId           => int32
  ProtocolName           => String
  LeaderId               => String
  MemberId               => String // removed
  Members                => []JoinGroupResponseMember 	
							  MemberId         => String
                              GroupInstanceId  => String
                              Metadata         => bytes
  MemberJoinResponseList => []JoinGroupResult  // new
							  MemberInfo  => JoinGroupResponseMember	
							  ErrorCode   => int16

A new admin request shall be created for user to supply a list of `group.instance.id` to batch join the group:

public static AddMemberResult addMemberToGroup(String groupId, list<String> groupInstanceIdsToAdd, AddMemberToGroupOptions options);

In the meantime, for better visibility for static members, we are also going to bump DescribeGroup request/response protocol to include `group.instance.id`:

DescribeGroupRequest => ThrottleTime Groups
  ThrottleTime           => int16
  Groups                 => []DescribeGroups
							  ErrorCode        => int16
							  GroupId          => String
							  GroupState       => String
							  ProtocolType     => String
							  ProtocolData     => int16
							  Members          => []DescribedGroupMember
									    			MemberId   => String
										    		GroupInstanceId  => String // new
											    	ClientId         => String							
										    		ClientHost       => String
											    	MemberMetadata   => bytes
    												MemberAssignment => bytes

Compatibility, Deprecation, and Migration Plan

• User needs to upgrade broker to latest version to be able to use this new feature.
• Since we are only introducing new admin API, the change should be backward compatible.

Rejected Alternatives

We could trigger multiple join group requests at the same time without changing JoinGroup protocol. However, considering our change in LeaveGroupRequest, it's hard to handle multiple responses within single admin client request. Changing the protocol to adapt to this change shall be more consistent.