Status

Current state: Under Discussion

Discussion thread: here

JIRA: here

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

Note that this is joint work by Unknown User (lucasbru) and Unknown User (cadonna) .

Motivation

Kafka Streams clients use a protocol to distribute stream processing tasks among Streams clients. Since the inception of Streams, this has worked by piggybacking custom byte payloads on the pre-KIP-848 consumer rebalance protocol. A custom partition assignor on one of the Streams clients would perform task assignment and various other centralized tasks, such as internal topic initialization. While this approach has served its purpose well and the streams partition assignor has achieved impressive flexibility and complexity over the years, it has also had a number of deficiencies:

• A large fraction of the production incidents in Kafka Streams are caused by bugs in the rebalancing logic or misconfiguration thereof. With the rebalance logic being run on the processing node, the implementer of the business logic is also typically responsible for configuring, upgrading and debugging the rebalance logic. 

  • When rebalance problems occur, the timeline of events and information that caused the rebalance and influenced the assignment is spread across all clients, making it difficult to debug.

  • Since streams versions on the client are notoriously slow to be picked up by teams building the processing logic, bugs in the rebalance logic are prevalent for years after being fixed on trunk.

  • Since assignment logic is controlled by the streams application, tuning the parameters of the assignment logic requires redeployment of the application, which in itself triggers rebalances (and reassignments), creating a vicious cycle.

• With the rebalance logic being based on the metadata on one of the Streams clients, metadata inconsistencies among clients have caused a variety of issues.

• Particularly, the use of raw byte arrays in the protocol means that brokers cannot understand or validate the payload that is propagated among Streams clients.

• In the implementation, piggybacking the task assignment logic on the consumer partition assignment and injecting the required Streams metadata and functionality into the general-purpose KafkaConsumer interface has required a number of workarounds, increasing code complexity and hurting the maintainability of the code. This has affected both the streams code, which had to work around the restrictions of the consumer interface, and the consumer, whose interface had to be padded to account for Kafka Streams requirements.

• Because the protocol is not Kafka Streams native, advanced features like “smooth scale out” or “application shutdown on error” and others require workarounds like probing rebalances, which are disruptive for the running application as well as often hard to reason about by users.

KIP-848 introduced the next generation of the consumer rebalance protocol, which also requires a redefinition of the streams rebalance protocol. We propose to use this change in the protocol to solve or alleviate the above problems by introducing a dedicated protocol for Kafka Streams, which uses dedicated RPCs for task rebalancing and where centralized logic such as topic initialization and task assignment are performed centrally by the brokers’ group coordinator.

Design Goals

We propose to introduce new RPCs designed to make Streams initialization and task assignment first-class citizens of the Kafka protocol. The proposed RPCs are built on top of the following design goals:

• We want to inherit all the advantages of the new consumer rebalance protocol introduced in KIP-848, such as greatly simplified clients, task assignment on the broker by default, truly incremental and cooperative rebalancing without global synchronization, and the ability of each instance to initiate a metadata update to the rebalance logic.

• Move central initialization tasks, such as the creation of internal topics, to the group coordinator.

• Make all data required for initialization and assignment explicit in the protocol and deserializable by the broker, with corresponding compatibility guarantees.

• Establish a single source of truth for stream application metadata such as the internal topic configuration, the topology and the assignment configuration to detect inconsistencies early by persisting them on the broker.

• Make the assignment logic parameters centrally tunable, without the need to redeploy the clients.

• Make active tasks, standby tasks and warm-up tasks first-class citizens in assignment and reconciliation

Proposed Changes

Overview

We introduce the concept of a Streams group in parallel to a consumer group. Streams clients use a dedicated heartbeat RPC, StreamsGroupHeartbeat to join a group, leave a group, and update the group coordinator about its currently owned tasks and its client-specific metadata. The group coordinator manages a Streams group similarly to a consumer group, continuously updating the group member metadata via heartbeat responses and running assignment logic when changes are detected. 

To assign tasks among Streams clients, the group coordinator uses topology metadata that is initialized using a dedicated StreamsGroupInitialize RPC and persisted in the consumer offsets topic. To trigger topology initialization and/or detect topology inconsistencies, each heartbeat contains a unique ID of the topology (which can be derived automatically from the topology via hashing), and the group coordinator can ask a Streams client to trigger an initialize request in the heartbeat response. 

Streams groups

We will introduce a new group type called streams to the group coordinator. For each Streams group, we define new record key and value types for the group metadata, topology metadata and group member metadata. The records are persisted in the __consumer_offsets topic. A group can either be a streams group, a share group or a consumer group, defined by the first heartbeat request using the corresponding groupId. 

Joining, leaving and staying in a Streams group

During the complete lifetime of a Streams client, it continuously sends a StreamsGroupHeartbeatRequest to the group coordinator. The request contains the following information:

• All non-assignment metadata as defined in the ConsumerGroupHeartbeatRequest such as the group ID, member ID, member epoch, instance ID, rack ID and rebalance timeout.

• The currently assigned active, standby and warm-up tasks. Each task is identified by a subtopology ID (which can be an arbitrary string in protocol for future extendability, even though subtopologies are currently numbered) and a partition number.

• The process ID to identify multiple Streams clients running in the same process.

• The user-defined endpoint to be used for running interactive queries.

• A unique ID of the processing topology. The ID is defined by the client, and will be hash of the topology representation by default.

• For each task, the current sum of task changelog offsets and end offsets. These are used to determine whether a warm-up task is below acceptable.recovery.lag and can be used to optimize the assignment in the assignor. For more details on these offsets see "Cumulative task changelog offsets and task changelog end offsets" section below

• A flag to request shutdown of the whole application (only if the member epoch is the group leave epoch).

Group ID, member ID, member epoch are sent with each heartbeat request. Any other information that has not changed since the last heartbeat can be omitted.

Heartbeat handling & response

If the heartbeat is valid and any data in the heartbeat has changed, the group coordinator will update its internal representation of the member correspondingly and persist the updated member metadata in a record on the __consumer_offsets topic. It then potentially updates the group assignment and reconciles member assignments, which is described in the section "Rebalance Process".

The group coordinator responds with a StreamsGroupHeartbeatResponse, which contains 

• All non-assignment metadata as defined in the ConsumerGroupHeartbeatResponse such as the member ID (when joining as a new member), new member epoch and heartbeat interval ms.

• The current target assignment for this member, as sets of active, standby and warm-up task IDs.

• For each client in the group, the user-defined endpoint and the assigned tasks for IQ

• A flag to indicate whether this client should initiate a StreamsGroupInitialize request to (re-)initialize the topology metadata

Again, fields that don’t change will be omitted from the heartbeat response.

Initializing and updating topology metadata on the broker

Whenever a Streams client receives a heartbeat response indicating that it should initialize or update the topology metadata on the broker, it initiates a StreamsGroupInitializeRequest on the group coordinator. The request describes the topology as a set of subtopologies, each identified by a unique string identifier and containing the following metadata that is relevant for creation of internal topics and assignment:

• The set of source topics. Here, the source topics can be defined through a Google RE2/J regular expression.

• The set of repartition sink topics. We do not include non-repartition sink topics or output topics that the topology produces in using dynamic routing.

• The set of changelog topics, replication factors and other topic-level configurations.

• The set of repartitions topics, the number of partitions for each of them (optional), replication factors and other topic-level configurations.

The data sent to the broker via the StreamsGroupInitialize RPC is persisted as a new record type in the __consumer_offsets topic and can be retrieved in subsequent requests. If one or more internal topics do not exist, the group coordinator invokes an asynchronous initialization. Only after the topology metadata is initialized and the internal topics are created, the group coordinator starts assigning tasks to the members of the group. The StreamsGroupInitializeResponse only indicates whether an error occurred or not.

If internal topics exist, but no consistent set of topics can be created without changing the number of partitions of the existing topics, an error is returned. The topic-level configurations are not validated and only used for initialization.

Topology updates

Topology updates can be initiated by any member upon joining. Whenever a streams member joins the group (whether a new member, or a static member after a restart) with a topology ID that is different from the current topology of the streams group, the group coordinator will ask one of the new members to reinitialize the broker-side topology. If the topology ID of an existing member does not correspond to the current topology ID of the group, this will be indicated in the heartbeat response sent to that member (for informational purposes, e.g., logging a warning).

The topology ID of the group and the topology ID of each member are provided to the assignor, which will make sure that members with a topology ID that does not correspond to the group's topology ID will not get any new tasks assigned, avoiding problems caused by topology incompatibilities.

It is up to the client implementation how topology IDs are defined. The protocol and the broker only requires that two clients with the same topology ID should have the same topology. In the initial implementation, the topology ID will be derived from the topology by deterministic hashing. However, it would also be possible to derive an implementation where topology IDs are explicitely defined through a configuration. This would, for example, allow the topologies to be manually versioned or topology IDs be derived from a git commit hash.

Adding a strict mode, where topology updates are requested explicitly, any implicit updates of the topology are rejected and members with incompatible topologies are fenced out of the group can be built on top of this protocol, but will be postponed to a follow-up KIP.

Describing and listing streams groups

Streams groups will be returned in ListGroupsResponse with GroupType equal to the string streams. The group ID of a streams group can be used in OffsetFetchRequest and OffsetCommitRequest as usual. Sending a ConsumerGroupHeartbeatRequest to a streams group will return an GROUP_ID_NOT_FOUND error. Sending a StreamsGroupHeartbeatRequest to a consumer group will similarly return an GROUP_ID_NOT_FOUND error.

There will be a new RPC DescribeStreamsGroup that returns, given the group ID, all metadata related to the streams group, such as

• The topology metadata that was provided through the StreamsGroupInitialize API.

• The latest member metadata that each member provided through the StreamsGroupHeartbeat API.

• The current target assignment generated by the assignor.

Rebalance Process

Similar as in KIP-848 consumer groups, the rebalance process is entirely driven by the group coordinator, and based on the same three epochs: group epoch, assignment epoch, and member epoch. The main difference is the assigned resource - instead of assigning partitions to consumers, we assign tasks consisting of a subtopology ID and a partition number, and each task can be assigned to a client in three roles: as an active task, as a standby task or as a warm-up task.

We only explain the main differences in the rebalance process in regular consumer groups here.

• The group epoch will be bumped every time the topology is updated through the StreamsGroupInitialize API.

• The partition metadata for the group, which is tracked by the group coordinator, is the partition metadata of all input topics (i.e., user source topics and internal repartition topics we read from).

• The group epoch is bumped every time a member:

  • When a client with an assigned warm-up task reports a task offset and task end offset whose difference is less that acceptable.recovery.lag.

  • Updates its topology ID, rack ID, client tags or process ID. Note: Typically, these do not change within the lifetime of a Streams client, so this only happens when a member with static membership rejoins with an updated configuration.

Assignment

Every time the group epoch is bumped, a new task assignment is computed by calling the task assignor, which consists of assigning tasks as active, standby or warm-up tasks to members of the group. The task assignor is configured on the broker side both through a global static configuration that defined the default assingor for all groups, and a dynamic group-level configuration that sets the assignor for a specific group.

The AK implementation will provide the following assignors

• highly_available - Like the current HighAvailabilityTaskAssignor

• sticky - Like the current StickyTaskAssignor

Which assignors are provided by the group coordinator is not part of the protocol and is specific to each implementation. In this KIP, we do not introduce a pluggable interface for a task assignors, but we leave it open to do so in the future.

If no task assignor is configured globally or on the group-level, the broker picks the assignor. The broker is free to make a dynamic decision here - the AK implementation will choose highly_available for stateful topologies and sticky for stateless topologies.

Each task assignor must make sure to fulfill the following invariants:

• Each warm-up task that has reached acceptable.recovery.lag must be turned into an active task or a standby task, or be unassigned from the member.

• A stateful task (in the various roles) cannot be assigned to two clients with the same processId.

Cumulative task changelog offsets and task changelog end offsets

Whenever a stateful task is added or removed to/from a Streams client, or when warm-up task reaches acceptable.recovery.lag and in regular intervals, defined by a broker-side configuration task.offset.interval.ms, each client reports two sets of offsets, the sums of task changelog offsets and the sums of task changelog end offsets. These sets can be used by the assignors, in particular the highly_available assignor, to determine which tasks are caught-up to acceptable.recovery.lag and optimize the assignment when multiple clients with a partial copy of the state exist.

Cumulative task changelog offsets

The cumulative changelog offset of a task is the sum of all positions in the changelog topics of the local state stores belonging to the task. A streams client reports the cumulative changelog offsets for all tasks with local state. That is:

• For each processing (active) task, the sum of the offsets last fully replicated to the broker in the changelog topic of each local state. Under at-least-once guarantees, this is going to correspond to the high watermark, under exactly-once semantics, it’s going to be the last stable offset.

• For each restoring (active, standby or warm-up), task the sum of the positions in the changelog topic and checkpointed to the local state directory.

• For each dormant task (task which is not owned, but which has state locally on disk) the sum of the checkpointed positions in the changelog topic present in its state directory

Cumulative task end offsets

Similarly, the client reports the sum of the end offsets for the currently owned tasks, if available. This simplifies the broker-side task assignment, since the broker doesn’t need to fetch the current end-offsets. Specifically:

• For active processing tasks, the cumulative task end offset is the same as the cumulative task offset

• For all restoring tasks (active, standby or warm-up), the cumulative end-offset is the sum of the last end offsets cached by the restore consumer. If an end offset of a topic partition is unknown, no end-offset is reported.

• For dormant standby tasks, no end-offset is reported.

Using cumulative task offsets and cumulative task end offsets

By reporting cumulative task offsets and cumulative task end offsets instead of the cumulative task lag, we can determine most task lags in the group coordinator, without the group coordinator or the streams client having to do additional requests. Even if the end offsets of the topic partition is not known on that client, as long as it is known on another client in the group. However, it may happen that no end offset is known in the group coordinator - for example, for tasks that are dormant on at least one client, and no client currently owns that task as active, standby or warm-up tasks. In these cases, the task with unknown end-offsets are never considered to be caught-up, however, the cumulative task offset can still be used when deciding on where to place a new active, standby or warm-up task - typically, by selecting the client with the maximal cumulative task offset that fulfils all other requirements.

Reconciliation of the group

Once a new target assignment is installed, each member will independently reconcile their current assignment with their new target assignment. Ultimately, each member will converge to their target epoch and assignment. The reconciliation process handles active tasks and standby/warm-up tasks differently:

• Active tasks are reconciled like topic partitions in the KIP-848 consumer group protocol, that is, for reconciling the target assignment, the reconciliation follows three phases:

  • The group coordinator first asks the member to revoke any active tasks that are not assigned to that member in the target assignment any more, by removing those active tasks from the set of tasks sent to the member in the heartbeat response

  • The member must first confirm revocation of these active tasks by removing them from the set of active tasks sent in the heartbeat request

  • Now, active tasks will be incrementally assigned to the member. An active task is assigned as soon as no other client owns it anymore, that is, once the previous owner (if any) has confirmed the revocation of the corresponding active task.

• Standby and warm-up tasks are reconciled in parallel with active tasks, but following slightly different logic:

  • As for active tasks, standby and warm-up tasks removed from the members target assignment are removed from the assignment sent in the heartbeat response immediately, that is, with the next heartbeat. The member confirms the revocation of the tasks as soon as the state directory (or any other resource related to the task) is closed.

  • Newly assigned standby and warm-up tasks await that any member with the same processId owning that task (as active or standby) confirms revocation of the task by removing it from the corresponding set in a heartbeat response. That is, a task that is not an active task or a standby task on any other member with the same processId can be assigned immediately. Otherwise, it is assigned as soon as the blocking task is revoked. Target assignments that assign a task to two members with the same processId are invalid.

Streams Group States

The possible states of the streams group are EMPTY, ASSIGNING, RECONCILING , STABLE, DEAD as for consumer groups, and additionally UNINITIALIZED for streams groups that have members but are pending initialization with a topology through the StreamsGroupInitialize RPC.

Global & Dynamic Group Configuration

We make core assignment options configurable centrally on the broker, without relying on each clients configuration. This allows tuning a streams group without redeploying the streams application. The three core assignment options will be introduced on the broker-side: acceptable.recovery.lag, max.warmup.replicas and num.standby.replicas. They can be configured both globally on the broker, and dynamically for specific streams groups through the IncrementalAlterConfigs and DescribeConfigs RPCs.

Public Interfaces

This section lists the changes impacting the public interfaces.

KRPC

New Errors

The conditions in which these errors are returned are stated further down.

• STREAMS_INVALID_TOPOLOGY - The supplied topology is invalid. Returned if the client sends a topology that does not fulfill the expected invariants, see below in the sections "Request Validation".

• STREAMS_MISSING_SOURCE_TOPICS - There are source topics missing for a topology that is supposed to be initialized. Also returned if the source topic regular expression matched no topics.

• STREAMS_INCONSISTENT_INTERNAL_TOPICS - There are internal topics present on the broker that are not consistent with the internal topic requirements of the provided topology.

StreamsGroupInitialize

The StreamsGroupInitialize API is called to initialize the topology information on the broker side, that will be reused by all members of the group. The StreamsGroupInitialize RPC can be called only by a single application to explicitly initialize the group, after the broker requested the initialization from that member.

Request Schema

{
  "apiKey": TBD,
  "type": "request",
  "listeners": ["broker"],
  "name": "StreamsGroupInitializeRequest",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "0+", "entityType": "groupId",
      "about": "The group identifier." },
    { "name":  "Topology", "type": "[]Subtopology", "versions": "0+",
      "about": "The sub-topologies of the streams application.",
      "fields": [
        { "name": "Subtopology", "type": "string", "versions": "0+",
          "about": "String to uniquely identify the sub-topology. Deterministically generated from the topology" },
        { "name": "SourceTopics", "type": "[]string", "versions": "0+",
          "about": "The topics the topology reads from." },
        { "name": "SourceTopicRegex", "type": "string", "versions": "0+",
          "about": "The regular expressions identifying topics the sub-topology reads from." },
        { "name": "StateChangelogTopics", "type": "[]TopicInfo", "versions": "0+",
          "about": "The set of state changelog topics associated with this sub-topology. Created automatically." },
        { "name": "RepartitionSinkTopics", "type": "[]string", "versions": "0+",
          "about": "The repartition topics the sub-topology writes to." }, 
        { "name": "RepartitionSourceTopics", "type": "[]TopicInfo", "versions": "0+",
          "about": "The set of source topics that are internally created repartition topics. Created automatically." }
      ]
    }
  ],
  "commonStructs": [
    { "name": "TopicConfig", "versions": "0+", "fields": [
        { "name": "key", "type": "string", "versions": "0+",
          "about": "The key of the topic-level configuration." },
        { "name": "value", "type": "string", "versions": "0+",
          "about": "The value of the topic-level configuration," }
      ]
    },
    { "name": "TopicInfo", "versions": "0+", "fields": [
      { "name": "Name", "type": "string", "versions": "0+",
        "about": "The name of the topic." },
      { "name": "Partitions", "type": "int32", "versions": "0+",
        "about": "The number of partitions in the topic. Can be 0 if no specific number of partitions is enforced. Always 0 for changelog topics." },
      { "name": "TopicConfigs", "type": "[]TopicConfig", "versions": "0+", "nullableVersions": "0+", "default": "null",
        "about": "Topic-level configurations as key-value pairs."
      }
    ]}
  ]
}

Required ACL

• READ on the group

• CREATE on cluster resource

• CREATE on all topics in StateChangelogTopics and RepartitionSourceTopics

• DESCRIBE_CONFIGS on all topics included in the message

Request Validation

INVALID_REQUEST is returned should the request not obey to the following invariants:

• GroupId must be non-empty.

STREAMS_INVALID_TOPOLOGY is returned should the request not obey the following invariants:

• A stateChangelogTopic topics must not have a defined partition number.

• A repartitionSourceTopic cannot be in sourceTopics or stateChangelogTopics of any subtopology.

• A stateChangelogTopic cannot be in sourceTopics or repartitionSinkTopic or repartitionSourceTopics of any subtopology.

• A repartitionSourceTopic of one subtopology must be a repartitionSinkTopic of at least one other subtopology.

Request Handling

When the group coordinator handles a StreamsGroupInitialize request:

• Looks up the group.

• It performs request validation.

• Checks if the provided group exists. GROUP_ID_NOT_FOUND is returned if the group does not exist or if the group ID is associated with a group type that is not streams or classic (the latter will be allowed for migration).

• Creates all required internal topics.

  • STREAMS_MISSING_SOURCE_TOPICS is returned if there are source topics missing for a topology that is supposed to be initialized. Also returned if the source topic regular expression matched no topics.

  • STREAMS_INCONSISTENT_INTERNAL_TOPICS is returned if there are internal topics present on the broker that are not consistent with the internal topic requirements of the provided topology.

• Writes the topology, keyed with the groupId to the consumer offset topic. Existing records will be overwritten.

Response Schema

{
  "apiKey": TBD,
  "type": "response",
  "name": "StreamsGroupInitializeResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  // Supported errors:
  // - GROUP_AUTHORIZATION_FAILED (version 0+)
  // - GROUP_ID_NOT_FOUND (version 0+)
  // - NOT_COORDINATOR (version 0+)
  // - COORDINATOR_NOT_AVAILABLE (version 0+)
  // - COORDINATOR_LOAD_IN_PROGRESS (version 0+)
  // - INVALID_REQUEST (version 0+)
  // - STREAMS_INVALID_TOPOLOGY (version 0+)
  // - STREAMS_MISSING_SOURCE_TOPICS (version 0+)
  // - STREAMS_INCONSISTENT_INTERNAL_TOPICS (version 0+)
  "fields": [
    { "name": "ThrottleTimeMs", "type": "int32", "versions": "0+",
      "about": "The duration in milliseconds for which the request was throttled due to a quota violation, or zero if the request did not violate any quota." },
    { "name": "ErrorCode", "type": "int16", "versions": "0+",
      "about": "The top-level error code, or 0 if there was no error" },
    { "name": "ErrorMessage", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "The top-level error message, or null if there was no error." }
  ]
}

StreamsGroupHeartbeat

The StreamsGroupHeartbeat API is the new core API used by streams application to form a group. The API allows members to advertise their state, and their owned tasks. The group coordinator uses it to assign/revoke tasks to/from members. This API is also used as a liveness check.

Request Schema

The member must set all the (top level) fields with the exception of RackId and InstanceId when it joins for the first time or when an error occurs (e.g. request timed out). Otherwise, it is expected to only fill in the fields which have changed since the last heartbeat.

{
  "apiKey": TBD,  
  "type": "request",
  "listeners": ["broker"],
  "name": "StreamsGroupHeartbeatRequest",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "0+", "entityType": "groupId",
      "about": "The group identifier." },
    { "name": "MemberId", "type": "string", "versions": "0+",
      "about": "The member id generated by the coordinator. The member id must be kept during the entire lifetime of the member." },
    { "name": "MemberEpoch", "type": "int32", "versions": "0+",
      "about": "The current member epoch; 0 to join the group; -1 to leave the group; -2 to indicate that the static member will rejoin." },
    { "name": "InstanceId", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "null if not provided or if it didn't change since the last heartbeat; the instance Id otherwise." },
    { "name": "RackId", "type": "string", "versions": "0+",  "nullableVersions": "0+", "default": "null",
      "about": "null if not provided or if it didn't change since the last heartbeat; the rack ID of consumer otherwise." },
    { "name": "RebalanceTimeoutMs", "type": "int32", "versions": "0+", "default": -1,
      "about": "-1 if it didn't change since the last heartbeat; the maximum time in milliseconds that the coordinator will wait on the member to revoke its partitions otherwise." },

    { "name": "TopologyID", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": null,
      "about": "The ID of the topology. Only sent when memberEpoch = 0, must be non-empty. Null otherwise. Used to check if topology corresponds to server-topology. " },

    { "name": "ActiveTasks", "type": "[]TaskIds", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Currently owned active tasks for this client. Null if unchanged since last heartbeat." },
    { "name": "StandbyTasks", "type": "[]TaskIds", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Currently owned standby tasks for this client. Null if unchanged since last heartbeat." },
    { "name": "WarmupTasks", "type": "[]TaskIds", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Currently owned warm-up tasks for this client. Null if unchanged since last heartbeat." },

    { "name": "ProcessId", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Identity of the streams instance that may have multiple consumers. Null if unchanged since last heartbeat." },
    { "name": "UserEndpoint", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "User-defined endpoint for Interactive Queries. Null if unchanged since last heartbeat." },
    { "name": "ClientTags", "type": "[]string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Used for rack-aware assignment algorithm. Null if unchanged since last heartbeat." },

    { "name": "TaskOffsets", "type": "[]TaskOffset", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Cumulative offsets for tasks. Only updated when a warm up task has caught up, and according to the task offset interval. Null if unchanged since last heartbeat." },
    { "name": "TaskEndOffsets", "type": "[]TaskOffset", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Cumulative end-offsets for tasks. Only updated when a warm up task has caught up, and according to the task offset interval. Null if unchanged since last heartbeat." },
    { "name": "ShutdownApplication", "type": "bool", "versions": "0+", "default": false,
      "about": "Whether all Streams clients in the group should shut down. Can only be defined if memberEpoch = -1" }
  ],
  "commonStructs": [
    { "name": "TaskOffset", "versions": "0+", "fields": [
      { "name": "Subtopology", "type": "string", "versions": "0+",
        "about": "The sub-topology identifier." },
      { "name": "Partition", "type": "int32", "versions": "0+",
        "about": "The partition." },
      { "name": "Offset", "type": "int64", "versions": "0+",
        "about": "The offset." }
    ]},
    { "name": "TaskIds", "versions": "0+", "fields": [
      { "name": "Subtopology", "type": "string", "versions": "0+",
        "about": "The sub-topology identifier." },
      { "name": "Partitions", "type": "[]int32", "versions": "0+",
        "about": "The partitions of the input topics processed by this member." }
    ]}
  ]
}

Required ACL

• Read Group

Request Validation

INVALID_REQUEST is returned should the request not obey to the following invariants:

• GroupId must be non-empty.

• Either MemberId is non-empty or MemberEpoch is 0.

• MemberEpoch must be >= -2.

• InstanceId, if not null, must be non-empty.

• RebalanceTimeoutMs must be larger than zero in the first heartbeat request.

• ActiveTasks, StandbyTasks and WarmupTasks have to be disjoint sets

• Each element of ActiveTasks, StandbyTasks and WarmupTasks has to be a valid task ID in the topology initialized for the groupID.

• ActiveTasks, StandbyTasks and WarmupTasks have to be non-null and empty when joining (MemberEpoch is 0)

Request Handling

When the group coordinator handles a StreamsGroupHeartbeat request:

1. Looks up the group. GROUP_ID_NOT_FOUND is returned if the group ID does not exist, or the group ID is associated with a group type that is not streams or classic (the latter will be allowed for migration).

2. Creates the member should the member epoch be zero, otherwise checks whether it exists. If it does not exist, UNKNOWN_MEMBER_ID is returned.

3. Checks whether the member epoch matches the member epoch in its current assignment. FENCED_MEMBER_EPOCH is returned otherwise. The member is also removed from the group.

   • There is an edge case here. When the group coordinator transitions a member to its target epoch, the heartbeat response with the new member epoch may be lost. In this case, the member will retry with the member epoch that he knows about and its request will be rejected with a FENCED_MEMBER_EPOCH. This will be handled as in KIP-848.

4. Updates the members informations if any. The group epoch is incremented if there is any change.

5. Reconcile the member assignments as explained earlier in this document. 

Reponse Schema

The group coordinator will only set the ActiveTasks, StandbyTasks and WarmupTasks fields until the member acknowledges that it has converged to the desired assignment. This is done to ensure that the members converge to the target assignment.

{
  "apiKey": TBD,
  "type": "response",
  "name": "StreamsGroupHeartbeatResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  // Supported errors:
  // - GROUP_AUTHORIZATION_FAILED (version 0+)
  // - NOT_COORDINATOR (version 0+)
  // - COORDINATOR_NOT_AVAILABLE (version 0+)
  // - COORDINATOR_LOAD_IN_PROGRESS (version 0+)
  // - INVALID_REQUEST (version 0+)
  // - UNKNOWN_MEMBER_ID (version 0+)
  // - FENCED_MEMBER_EPOCH (version 0+)
  // - UNRELEASED_INSTANCE_ID (version 0+)
  // - GROUP_MAX_SIZE_REACHED (version 0+)
  "fields": [
    // Same as consumer group heart beat
    { "name": "ThrottleTimeMs", "type": "int32", "versions": "0+",
      "about": "The duration in milliseconds for which the request was throttled due to a quota violation, or zero if the request did not violate any quota." },
    { "name": "ErrorCode", "type": "int16", "versions": "0+",
      "about": "The top-level error code, or 0 if there was no error" },
    { "name": "ErrorMessage", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "The top-level error message, or null if there was no error." },
    { "name": "MemberId", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "The member id generated by the coordinator. Only provided when the member joins with MemberEpoch == 0." },
    { "name": "MemberEpoch", "type": "int32", "versions": "0+",
      "about": "The member epoch." },
    { "name": "HeartbeatIntervalMs", "type": "int32", "versions": "0+",
      "about": "The heartbeat interval in milliseconds." },
    { "name": "AcceptableRecoveryLag", "type": "int32", "versions": "0+",
      "about": "The maximal lag a warm-up task can have to be considered caught-up." },
    { "name": "TaskOffsetIntervalMs", "type": "int32", "versions": "0+",
      "about": "The interval in which the changelog task offsets on a client are updated on the broker. The offsets are sent with the next heartbeat after this time has passed." },

    // Topology updating
    { "name": "ShouldInitializeTopology", "type":  "bool", "versions":  "0+", "default": false,
      "about": "true if this streams application should initialize the topology on the broker" },
    { "name": "GroupTopologyID", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": null,
      "about": "The current ID of the topology for the group. Null if unchanged since last heartbeat, empty if group is unitialized." },

    { "name": "Status", "type":  "[]Status", "versions":  "0+",  "nullableVersions": "0+", "default": "null", 
      "about": "Indicate zero or more status for the group.  Null if unchanged since last heartbeat." },

    // The streams app knows which partitions to fetch from given this information
    { "name": "ActiveTasks", "type": "[]TaskIds", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Assigned active tasks for this client. Null if unchanged since last heartbeat." },
    { "name": "StandbyTasks", "type": "[]TaskIds", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Assigned standby tasks for this client. Null if unchanged since last heartbeat." },
    { "name": "WarmupTasks", "type": "[]TaskIds", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Assigned warm-up tasks for this client. Null if unchanged since last heartbeat." },

    // IQ-related information
    { "name": "PartitionsByUserEndpoint", "type": "[]EndpointInfo", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "Global assignment information used for IQ. Null if unchanged since last heartbeat." ,
      "fields": [
        { "name": "UserEndpoint", "type": "string", "versions": "0+",
          "about": "User-defined endpoint to connect to the node" },
        { "name": "Partitions", "type": "[]TopicPartition", "versions": "0+",
          "about": "All partitions available on the node" }
      ]
    }
  ],
  "commonStructs": [
    { "name": "Status", "versions": "0+", "fields": [
      // Possible status codes
      //  0 - UNITIALIZED_TOPOLOGY    - The group is not initialized.
      //  1 - INCONSISTENT_TOPOLOGY   - The topology ID supplied is inconsistent with the topology for this streams group.
      //  2 - MISSING_SOURCE_TOPICS   - One or more source topics do not exist. Missing topics are indicated in the StatusDetail.
      //  3 - MISSING_INTERNAL_TOPICS - One or more internal topics do not exist. Missing topics are indicated in the StatusDetail.
      //  4 - SHUTDOWN_APPLICATION    - A client requested the shutdown of the whole application.
      { "name": "StatusCode", "type": "int8", "versions": "0+",
        "about": "A code to indicate that a particular status is active for the group membership" },
      { "name": "StatusDetail", "type": "string", "versions": "0+",
        "about": "A string representation of the status." }
    ]},
    { "name": "TopicPartition", "versions": "0+", "fields": [
      { "name": "Topic", "type": "string", "versions": "0+",
        "about": "topic name" },
      { "name": "Partitions", "type": "[]int32", "versions": "0+",
        "about": "partitions" }
    ]},
    { "name": "TaskIds", "versions": "0+", "fields": [
      { "name": "Subtopology", "type": "string", "versions": "0+",
        "about": "The subtopology identifier." },
      { "name": "Partitions", "type": "[]int32", "versions": "0+",
        "about": "The partitions of the input topics processed by this member." }
    ]}
  ]
}

StreamsGroupDescribe API

Request Schema

{
  "apiKey": TBD,
  "type": "request",
  "listeners": ["broker"],
  "name": "StreamsGroupDescribeRequest",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "GroupIds", "type": "[]string", "versions": "0+", "entityType": "groupId",
      "about": "The ids of the groups to describe" },
    { "name": "IncludeAuthorizedOperations", "type": "bool", "versions": "0+",
      "about": "Whether to include authorized operations." }
  ]
}

Required ACL

• Read Group

Request Validation

INVALID_REQUEST is returned should the request not obey to the following invariants:

• GroupIds must be non-empty.

Request Handling

When the group coordinator handle a StreamsGroupDescribeRequest request:

• Checks whether the group ids exists. If it does not, GROUP_ID_NOT_FOUND is returned. Similarly, it is returned if the group ID is associated with a group type that is not streams .

• Looks up the groups and returns the response.

Response Schema

{
  "apiKey": TBD,
  "type": "response",
  "name": "StreamsGroupDescribeResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  // Supported errors:
  // - GROUP_AUTHORIZATION_FAILED (version 0+)
  // - NOT_COORDINATOR (version 0+)
  // - COORDINATOR_NOT_AVAILABLE (version 0+)
  // - COORDINATOR_LOAD_IN_PROGRESS (version 0+)
  // - INVALID_REQUEST (version 0+)
  // - INVALID_GROUP_ID (version 0+)
  // - GROUP_ID_NOT_FOUND (version 0+)
  "fields": [
    { "name": "ThrottleTimeMs", "type": "int32", "versions": "0+",
      "about": "The duration in milliseconds for which the request was throttled due to a quota violation, or zero if the request did not violate any quota." },
    { "name": "Groups", "type": "[]DescribedGroup", "versions": "0+",
      "about": "Each described group.",
      "fields": [
        { "name": "ErrorCode", "type": "int16", "versions": "0+",
          "about": "The describe error, or 0 if there was no error." },
        { "name": "ErrorMessage", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
          "about": "The top-level error message, or null if there was no error." },
        { "name": "GroupId", "type": "string", "versions": "0+", "entityType": "groupId",
          "about": "The group ID string." },
        { "name": "GroupState", "type": "string", "versions": "0+",
          "about": "The group state string, or the empty string." },
        { "name": "GroupEpoch", "type": "int32", "versions": "0+",
          "about": "The group epoch." },
        { "name": "AssignmentEpoch", "type": "int32", "versions": "0+",
          "about": "The assignment epoch." },
        { "name":  "Topology", "type": "[]Subtopology", "versions": "0+",
          "about": "The sub-topologies of the streams application.",
          "fields": [
            { "name": "Subtopology", "type": "string", "versions": "0+",
              "about": "String to uniquely identify the subtopology." },
            { "name": "SourceTopics", "type": "[]string", "versions": "0+",
              "about": "The topics the topology reads from." },
            { "name": "SourceTopicRegex", "type": "string", "versions": "0+",
              "about": "The regular expressions identifying topics the topology reads from." },
            { "name": "SinkTopics", "type": "[]string", "versions": "0+",
              "about": "The topics the topology writes to." },
            { "name": "StateChangelogTopics", "type": "[]TopicInfo", "versions": "0+",
              "about": "The set of state changelog topics associated with this subtopology. Created automatically." },
            { "name": "RepartitionSourceTopics", "type": "[]TopicInfo", "versions": "0+",
              "about": "The set of source topics that are internally created repartition topics. Created automatically." }
          ]
        },
        { "name": "Members", "type": "[]Member", "versions": "0+",
          "about": "The members.",
          "fields": [
            { "name": "MemberId", "type": "string", "versions": "0+",
              "about": "The member ID." },
            { "name": "MemberEpoch", "type": "int32", "versions": "0+",
              "about": "The member epoch." },
            { "name": "InstanceId", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
              "about": "The member instance ID." },
            { "name": "RackId", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
              "about": "The rack ID." },

            { "name": "ClientId", "type": "string", "versions": "0+",
              "about": "The client ID." },
            { "name": "ClientHost", "type": "string", "versions": "0+",
              "about": "The client host." },

            { "name": "TopologyID", "type": "string", "versions": "0+",
              "about": "The ID of the topology. Must be non-empty." },

            { "name": "ProcessId", "type": "uuid", "versions": "0+",
              "about": "Identity of the streams instance that may have multiple clients. " },
            { "name": "ClientTags", "type": "[]KeyValue", "versions": "0+",
              "about": "Used for rack-aware assignment algorithm." },
            { "name": "TaskOffsets", "type": "[]TaskOffset", "versions": "0+",
              "about": "Cumulative offsets for tasks." },
            { "name": "TaskEndOffsets", "type": "[]TaskOffset", "versions": "0+",
              "about": "Cumulative end offsets for tasks." },

            { "name": "Assignment", "type": "Assignment", "versions": "0+",
              "about": "The current assignment." },
            { "name": "TargetAssignment", "type": "Assignment", "versions": "0+",
              "about": "The target assignment." }
          ]},
        { "name": "AuthorizedOperations", "type": "int32", "versions": "0+", "default": "-2147483648",
          "about": "32-bit bitfield to represent authorized operations for this group." }
      ]
    }
  ],
  "commonStructs": [
    { "name": "TaskOffset", "versions": "0+", "fields": [
      { "name": "Subtopology", "type": "string", "versions": "0+",
        "about": "The subtopology identifier." },
      { "name": "Partition", "type": "int32", "versions": "0+",
        "about": "The partition." },
      { "name": "Offset", "type": "int64", "versions": "0+",
        "about": "The offset." }
    ]},
    { "name": "TopicPartitions", "versions": "0+", "fields": [
      { "name": "TopicId", "type": "uuid", "versions": "0+",
        "about": "The topic ID." },
      { "name": "TopicName", "type": "string", "versions": "0+", "entityType": "topicName",
        "about": "The topic name." },
      { "name": "Partitions", "type": "[]int32", "versions": "0+",
        "about": "The partitions." }
    ]},
    { "name": "Assignment", "versions": "0+", "fields": [
      { "name": "ActiveTasks", "type": "[]TaskIds", "versions": "0+",
        "about": "Active tasks for this client." },
      { "name": "StandbyTasks", "type": "[]TaskIds", "versions": "0+",
        "about": "Standby tasks for this client." },
      { "name": "WarmupTasks", "type": "[]TaskIds", "versions": "0+",
        "about": "Warm-up tasks for this client. " }
    ]},
    { "name": "TaskIds", "versions": "0+", "fields": [
      { "name": "Subtopology", "type": "string", "versions": "0+",
        "about": "The subtopology identifier." },
      { "name": "Partitions", "type": "[]int32", "versions": "0+",
        "about": "The partitions of the input topics processed by this member." }
    ]},
    { "name": "KeyValue", "versions": "0+", "fields": [
      { "name": "Key", "type": "string", "versions": "0+",
        "about": "key of the config" },
      { "name": "Value", "type": "string", "versions": "0+",
        "about": "value of the config" }
    ]},
    { "name": "TopicInfo", "versions": "0+", "fields": [
      { "name": "Name", "type": "string", "versions": "0+",
        "about": "The name of the topic." },
      { "name": "Partitions", "type": "int32", "versions": "0+",
        "about": "The number of partitions in the topic. Can be 0 if no specific number of partitions is enforced. Always 0 for changelog topics." },
      { "name": "TopicConfigs", "type": "[]KeyValue", "versions": "0+", "nullableVersions": "0+", "default": "null",
        "about": "Topic-level configurations as key-value pairs."
      }
    ]}
  ]
}

Response Handling

Nothing particular.

Records

Similar to KIP-848, we define new record types that persist all data related to groups, their members and their assignments in the __consumer_offsets compacted topic. Each record has a dedicated key type that uses the version to differentiate the records from the records introduced in KIP-848, and those used to persist the committed offsets.

Compared to consumer groups, we introduce one extra record type, that stores information about the topology of the group.

The size of the persisted member metadata in the __consumer_offset topic will have to respect the max message size. But those records are persisted separately for each member, so it should be less of a concern. The same problem affects plain consumers in KIP-848 for very large consumer groups. Eventually, we may implement a way to split the records into multiple messages, and there are currently plans to add transaction support to the group coordinator to make sure the records can be updated atomically. For the first version of the protocol we will limit those records by the max message size.

Group Metadata

StreamsGroupCurrentMemberAssignmentKey

{
  "type": "data",
  "name": "StreamsGroupCurrentMemberAssignmentKey",
  "validVersions": "14",
  "flexibleVersions": "none",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "8",
      "about": "The group id." },
    { "name": "MemberId", "type": "string", "versions": "8",
      "about": "The member id." }
  ]
}

StreamsGroupCurrentMemberAssignmentValue

{
  "type": "data",
  "name": "StreamsGroupCurrentMemberAssignmentValue",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "MemberEpoch", "versions": "0+", "type": "int32",
      "about": "The current member epoch that is expected from the member in the heartbeat request." },
    { "name": "PreviousMemberEpoch", "versions": "0+", "type": "int32",
      "about": "If the last epoch bump is lost before reaching the member, the member will retry with the previous epoch." },
    { "name": "State", "versions": "0+", "type": "int8",
      "about": "The member state. See StreamsGroupMember.MemberState for the possible values." },
    { "name": "ActiveTasks", "versions": "0+", "type": "[]TaskIds",
      "about": "Currently assigned active tasks for this streams client." },
    { "name": "StandbyTasks", "versions": "0+", "type": "[]TaskIds",
      "about": "Currently assigned standby tasks for this streams client." },
    { "name": "WarmupTasks", "versions": "0+", "type": "[]TaskIds",
      "about": "Currently assigned warm-up tasks for this streams client." },
    { "name": "ActiveTasksPendingRevocation", "versions": "0+", "type": "[]TaskIds",
      "about": "The active tasks that must be revoked by this member." }
  ],
  "commonStructs": [
    { "name": "TaskIds", "versions": "0+", "fields": [
      { "name": "Subtopology", "type": "string", "versions": "0+",
        "about": "The subtopology identifier." },
      { "name": "Partitions", "type": "[]int32", "versions": "0+",
        "about": "The partitions of the input topics processed by this member." }
    ]}
  ]
}

StreamsGroupMemberMetadataKey

{
  "type": "data",
  "name": "StreamsGroupMemberMetadataKey",
  "validVersions": "11",
  "flexibleVersions": "none",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "5",
      "about": "The group id." },
    { "name": "MemberId", "type": "string", "versions": "5",
      "about": "The member id." }
  ]
}

StreamsGroupMemberMetadataValue

{
  "type": "data",
  "name": "StreamsGroupMemberMetadataValue",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "InstanceId", "versions": "0+", "nullableVersions": "0+", "type": "string",
      "about": "The (optional) instance id." },
    { "name": "RackId", "versions": "0+", "nullableVersions": "0+", "type": "string",
      "about": "The (optional) rack id." },
    { "name": "ClientId", "versions": "0+", "type": "string",
      "about": "The client id." },
    { "name": "ClientHost", "versions": "0+", "type": "string",
      "about": "The client host." },
    { "name": "RebalanceTimeoutMs", "type": "int32", "versions": "0+", "default": -1,
      "about": "The rebalance timeout." },

    { "name": "TopologyID", "type": "string", "versions": "0+",
      "about": "The ID of the topology. Must be non-empty." },

    { "name": "ProcessId", "type": "string", "versions": "0+",
      "about": "Identity of the streams instance that may have multiple consumers." },
    { "name": "UserEndpoint", "type": "string", "versions": "0+",
      "about": "User-defined endpoint for running interactive queries on this instance." },
    { "name": "ClientTags", "type": "[]KeyValue", "versions": "0+",
      "about": "Used for rack-aware assignment algorithm." }
  ],
  "commonStructs": [
    { "name": "KeyValue", "versions": "0+",
      "fields": [
        { "name": "Key", "type": "string", "versions": "0+",
          "about": "key of the config" },
        { "name": "Value", "type": "string", "versions": "0+",
          "about": "value of the config" }
      ]
    }
  ]
}

StreamsGroupMetadataKey

{
  "type": "data",
  "name": "StreamsGroupMetadataKey",
  "validVersions": "9",
  "flexibleVersions": "none",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "3",
      "about": "The group id." }
  ]
}

StreamsGroupMetadataValue

{
  "type": "data",
  "name": "StreamsGroupMetadataValue",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "Epoch", "versions": "0+", "type": "int32",
      "about": "The group epoch." }
  ]
}

StreamsGroupPartitionMetadataKey

{
  "type": "data",
  "name": "StreamsGroupPartitionMetadataKey",
  "validVersions": "10",
  "flexibleVersions": "none",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "4",
      "about": "The group id." }
  ]
}

StreamsGroupPartitionMetadataValue

{
  "type": "data",
  "name": "StreamsGroupPartitionMetadataValue",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "Topics", "versions": "0+", "type": "[]TopicMetadata",
      "about": "The list of topic metadata.", "fields": [
      { "name": "TopicId", "versions": "0+", "type": "uuid",
        "about": "The topic id." },
      { "name": "TopicName", "versions": "0+", "type": "string",
        "about": "The topic name." },
      { "name": "NumPartitions", "versions": "0+", "type": "int32",
        "about": "The number of partitions of the topic." },
      { "name": "PartitionMetadata", "versions": "0+", "type": "[]PartitionMetadata",
        "about": "Partitions mapped to a set of racks. If the rack information is unavailable for all the partitions, an empty list is stored", "fields": [
          { "name": "Partition", "versions": "0+", "type": "int32",
            "about": "The partition number." },
          { "name": "Racks", "versions": "0+", "type": "[]string",
            "about": "The set of racks that the partition is mapped to." }
      ]}
    ]}
  ]
}

StreamsGroupTargetAssignmentMemberKey

{
  "type": "data",
  "name": "StreamsGroupTargetAssignmentMemberKey",
  "validVersions": "13",
  "flexibleVersions": "none",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "7",
      "about": "The group id." },
    { "name": "MemberId", "type": "string", "versions": "7",
      "about": "The member id." }
  ]
}

StreamsGroupTargetAssignmentMemberValue

{
  "type": "data",
  "name": "StreamsGroupTargetAssignmentMemberValue",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "ActiveTasks", "versions": "0+", "type": "[]TaskIds",
      "about": "Currently assigned active tasks for this streams client." },
    { "name": "StandbyTasks", "versions": "0+", "type": "[]TaskIds",
      "about": "Currently assigned standby tasks for this streams client." },
    { "name": "WarmupTasks", "versions": "0+", "type": "[]TaskIds",
      "about": "Currently assigned warm-up tasks for this streams client." }
  ],
  "commonStructs": [
    { "name": "TaskIds", "versions": "0+", "fields": [
      { "name": "Subtopology", "type": "string", "versions": "0+",
        "about": "The subtopology identifier." },
      { "name": "Partitions", "type": "[]int32", "versions": "0+",
        "about": "The partitions of the input topics processed by this member." }
    ]}
  ]
}

StreamsGroupTargetAssignmentMetadataKey

{
  "type": "data",
  "name": "StreamsGroupTargetAssignmentMetadataKey",
  "validVersions": "12",
  "flexibleVersions": "none",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "6",
      "about": "The group id." }
  ]
}

StreamsGroupTargetAssignmentMetadataValue

{
  "type": "data",
  "name": "StreamsGroupTargetAssignmentMetadataValue",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "AssignmentEpoch", "versions": "0+", "type": "int32",
      "about": "The assignment epoch." }
  ]
}

StreamsGroupTopologyKey

{
  "type": "data",
  "name": "StreamsGroupTopologyKey",
  "validVersions": "15",
  "flexibleVersions": "none",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "3",
      "about": "The group id." }
  ]
}

StreamsGroupTopologyValue

{
  "type": "data",
  "name": "StreamsGroupTopologyValue",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "TopologyID", "type": "string", "versions": "0+",
      "about": "The ID of the topology. Must be non-empty. " },
    { "name":  "Topology", "type": "[]Subtopology", "versions": "0+",
      "about": "The sub-topologies of the streams application.",
      "fields": [
        { "name": "Subtopology", "type": "string", "versions": "0+",
          "about": "String to uniquely identify the subtopology." },
        { "name": "SourceTopics", "type": "[]string", "versions": "0+",
          "about": "The topics the topology reads from." },
        { "name": "SourceTopicRegex", "type": "string", "versions": "0+",
          "about": "The regular expressions identifying topics the topology reads from." },
        { "name": "SinkTopics", "type": "[]string", "versions": "0+",
          "about": "The topics the topology writes to." },
        { "name": "StateChangelogTopics", "type": "[]TopicInfo", "versions": "0+",
          "about": "The set of state changelog topics associated with this subtopology. " },
        { "name": "RepartitionSourceTopics", "type": "[]TopicInfo", "versions": "0+",
          "about": "The set of source topics that are internally created repartition topics. " }
      ]
    }
  ],
  "commonStructs": [
    { "name": "TopicConfig", "versions": "0+", "fields": [
      { "name": "key", "type": "string", "versions": "0+",
        "about": "The key of the topic-level configuration." },
      { "name": "value", "type": "string", "versions": "0+",
        "about": "The value of the topic-level configuration," }
    ]
    },
    { "name": "TopicInfo", "versions": "0+", "fields": [
      { "name": "Name", "type": "string", "versions": "0+",
        "about": "The name of the topic." },
      { "name": "Partitions", "type": "int32", "versions": "0+",
        "about": "The number of partitions in the topic. Can be 0 if no specific number of partitions is enforced. Always 0 for changelog topics." },
      { "name": "TopicConfigs", "type": "[]TopicConfig", "versions": "0+", "nullableVersions": "0+", "default": "null",
        "about": "Topic-level configurations as key-value pairs."
      }
    ]}
  ]
}

Broker Metrics

The existing group metrics are extended to differentiate between streams groups and consumer groups and account for streams group states.

• Number of groups based on type of the protocol, where the list of protocols is extended by the protocol=streams variation.

kafka.server:type=group-coordinator-metrics,name=group-count,protocol={consumer|classic|streams}

• Number of streams groups based on state

kafka.server:type=group-coordinator-metrics,name=streams-group-count,state={empty|uninitialized|assigning|reconciling|stable|dead}

• Streams group rebalances sensor

kafka.server:type=group-coordinator-metrics,name=streams-group-rebalance-rate

kafka.server:type=group-coordinator-metrics,name=streams-group-rebalance-count

Broker Configurations

The new group protocol would be guarded by enabling a new version of the group.version feature flag on the broker side.

Updated properties

New properties

Group Configurations

We will add new configurations for the resource type GROUP in DescribeConfigs and IncrementalAlterConfigs to override the default broker configurations dynamically for specific groups.

Streams Interfaces

In a future major version, when the classic group protocol will be deprecated, the interfaces introduced in KIP-924 would be deprecated, unless an integration with this KIP is proposed in a follow-up KIP.

Streams Configurations

New configurations

Deprecations

In a future major version, when the classic group protocol will be deprecated, the following configuration options will be deprecated:

acceptable.recovery.lag

max.warmup.replicas

num.standby.replicas

probing.rebalance.interval.ms

rack.aware.assignment.tags

rack.aware.assignment.strategy

rack.aware.assignment.traffic_cost

rack.aware.assignment.non_overlap_cost

task.assignor.class

Note that both the rack-aware task assignor and customizable client-side task assignment may be introduced into the new protocol in follow-up KIPs (see Future Work section). In this case, the latter 5 configuration options would not be deprecated in the future major version.

kafka-streams-application-reset

The streams application reset tool will have to be updated to use the streams rebalance protocol to reset offsets and force existing members out of the group. It can detect the type of group using a ListGroups RPC, so be able to handle both kinds of groups, so no extension of the public API are required.

Compatibility, Deprecation and Migration Plan

The new broker-side group management will require more information than the current UserData to compute the assignment, so existing Streams versions cannot join a Streams group. 

To enable rolling upgrades from the old protocol to the new protocol, we will have to extend the current UserData with the missing (topology-derived) data. We can then introduce a compatibility layer in the new group coordinator that translates RPCs from the legacy rebalance protocol into requests in the streams rebalance protocol by deserializing the UserData payload. The migration path for the user is then to update to a new version of Streams, before switching from the legacy consumer-based protocol to the streams rebalance protocol in a second step. The first time a consumer sends a StreamsGroupHeartbeat with the group ID of an existing legacy consumer group, the group will be converted into a Streams group, similar to how the migration is handled for legacy consumer groups in KIP-848.

Named topologies will not be supported in the new protocol. A Streams application using named topologies will throw an error on start-up when configured with the new protocol. While noteworthy, this is not strictly a compatibility concern, since the configuration to enable named topologies is not part of the public API.

Test Plan

Our primary method of validation will be the execution of the existing test infrastructure for Kafka streams on the new protocol, including porting all integration tests and system tests. Furthermore, we will add new integration tests for offline & online migration from the classic protocol to the streams protocol.

Rejected Alternatives

• KIP-848 originally envisioned that Kafka Streams continues to using a byte-array metadata protocol on top of regular consumer group RPCs, and use a client-side assignor, and provided a rough proposal how to do so. The described approach was never implemented. In this KIP, we propose to use broker-side assignment by default for several reasons (mostly already mentioned in the motivation section):

  • Client-side assignment makes it hard to debug and tune the assignment logic, where a simple parameter change requires redeploying of all streams clients.

  • Using opaque byte-arrays to exchange the required metadata makes it hard to understand the state of a streams group on the protocol level. One has to discover the information from logs spread across all client machines. Central mechanisms of the Kafka protocol, such as versioning of RPCs, needs to be reimplemented on the level of RPCs.

  • One could argue that the Kafka protocol is simpler without Kafka Streams specific RPCs, but we think the opposite is true. The KIP-848 consumer group protocol was extended with client-side assignment, byte-array metadata and metadata versioning just for Kafka Streams. Having the concerns of Kafka Streams dealt with in separate RPCs mainly introduces a clearer separation and gives the Kafka community the ability to simplify the upcoming consumer protocol.

• Instead of using a separate initialize RPC, we considered just using the heartbeat with memberEpoch=0 to initialize the Streams group and initiate the creation of all topics. This would have a number of downsides, such as that every member needs to send the initialization data every time they start (and the initialization data may be large already in the current schema and may grow larger in the future), and the heartbeat requiring the ACL for creating topics. Furthermore, since the initialization may take a certain amount of time, we’d likely still have to respond to the initial heartbeat without having fully initialized the group, leading to temporary “uninitialized” state for the group that we also get in this proposal.

• An earlier draft of this KIP proposed introducing optional client-side task assignment, that would allow customization of the task assignment similar to KIP-924. We decided to remove it from this KIP, to limit the scope of the KIP and not introduce features in the protocol whose use we cannot fully envision yet. The RPCs defined in this KIP are, however, defined in a way that they could easily be extended with client-side assignment in a follow-up KIP, possibly together with a public interface for adding custom task assignors on the client side.

Future Work

The classic streams rebalance protocol will remain fully functional with the introduction of the streams protocol, however the aim is to eventually deprecate. There are some use cases no yet covered by this KIP, that would have to be added to the protocol in follow-up KIPs, if there is enough interest.

• Customizable client-side task assignment in the style of KIP-924
• Optimization of the assignment in cases were not every rack has a replica of the partition, as in KIP-925: Rack aware task assignment in Kafka Streams
• Injection of a customized implementation of a streams rebalance protocol using the KafkaClientSupplier.

There are also several improvements of long-standing Kafka Streams issues that can be built on top of this protocol, in particular:

• Topology updates

  • Validate topology compatibility broker-side

  • Introduce manually bumped topology versioning in addition to automatically managed topology IDs

  • Automatically drain repartition topics on a topology update

• In addition to the existing shutdown-flag, a global pause/resume functionality could be built on top of the protocol

• There is an opportunity to improve EOSv2 in the new protocol. One may add the current transaction.id to the heartbeat request and a list of closing transactions.id that the client should fence in the heartbeat response to improve EOSv2 and avoid the need to rely on a low transaction.timeout.ms config.