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 this is a joint worked proposed by David Jacot, guozhang Wang and Jason Gustafson.

Motivation

It has been about 8 years since we introduced the so-called new consumer which does group membership and rebalancing through Kafka. Although it was a huge improvement over the old Zookeeper based consumer, it has still been a major pain point from an operation perspective. There are multiple reasons for this, let’s dive into them:

• The protocol relies on thick clients. Thick clients are annoying in many ways:
  • We have had many bugs in the rebalance protocol over the last years and the majority of them required client side bug fixes. In the cloud area, this is really annoying because we effectively depend on the adoption of the clients in order to fix the issues in production. Unfortunately, the adoption is usually rather slow in the Kafka community;
  • It is almost impossible to debug issues in the protocol without having access to the client logs. In the cloud area, it is a bit annoying to have to request client logs to troubleshoot your system;
  • The clients specify the so-called embedded protocol in the rebalance protocol. While this allows the core protocol to be reused for different purposes, for instance it is used by both the consumer and connect in Apache Kafka, this makes inspecting the state on the broker side hard because the brokers get a bunch of raw bytes. The compatibility of the embedded protocols has also been a challenge; and
  • The clients are responsible for monitoring the metadata and for triggering rebalances. This has caused all sorts of issues in the past because clients of a given group might have a different view of the metadata at a given point in time.
• The protocol relies on a group-wide synchronization barrier. This means that a single misbehaving consumer can take down or disturb the whole group because a rebalance of the whole group is required whenever a consumer joins, leaves or fails. This also limits its scalability as the cost of a rebalance increases with the number of members in the group. Even the cooperative rebalancing protocols depend on the barrier. Specifically, one of the deficiencies of the cooperative protocol is that offsets cannot be committed while the consumer is waiting on the rebalance to complete. So even though a consumer can keep fetching while the rebalance is in progress, it still tends to get stuck behind the barrier.
• The protocol has gotten too complex over the years. We started with a rather simple protocol and we have extended it a few times over the years. For instance, we introduced KIP-429: Kafka Consumer Incremental Rebalance Protocol, KIP-345: Introduce static membership protocol to reduce consumer rebalances, and a few others. All the incremental changes that we have made have increased the complexity of the protocol.
• The group protocol has been used for general state propagation between members. This is especially the case for power users such as Kafka Streams. While the state propagation is not an issue in itself, the protocol only propagates the state during a rebalance so we have introduced fake or dummy rebalance with the sole goal to propagate some new state to the leader of the group or to all the members of the group. This has been very confusing for our users both on the client side and the broker side. This also makes the interpretation of rebalances through metrics or logs more difficult.

Design Goals

We propose to introduce a new group membership and rebalance protocol for the Kafka Consumer and, by extensions, Kafka Streams. The proposed protocol is built on top of the following design goals.

• The protocol should be truly incremental and cooperative and should not rely on a global synchronization barrier anymore. Ideally, a consumer should not be impacted at all by a rebalance if its assignment is not changed.
  
• The complexity should move away from the consumer to the group coordinator. We want to be able to troubleshoot issues without requiring client logs and we want to fix issues without having to wait on consumer adoption.
  
• The protocol should still allow power users such as Kafka Streams to run assignment logic on the client. This is important for Kafka Streams to remain independent from the broker. However, we want this process to be driven and controlled by the group coordinator.
  
• The protocol should provide the same guarantee as the current protocol that is at-least-once in the worst case scenario and exactly-once when the hand off between members is clean.
  
• The protocol should support upgrading the consumers without downtime.

Note that Kafka Connect is not supported by this new protocol. We discuss how Kafka Connect could evolve by using a similar protocol in the future in the future work section.

Proposed Changes

Rebalance Protocol in a Nutshell

The proposed rebalance protocol is based on the concept of a declarative assignment for the group and the use of reconciliation loops to drive members toward their desired assignment. Members can independently converge and the group coordinator takes care of resolving the dependencies - e.g. revoking a partition before it can be assigned - between the members if any.

The desired (or target) assignment is either directly computed by the group coordinator using a server side assignor or computed by one of the group members if a client side assignor is specified. The former is the new default for consumers while the latter allows power users such as Kafka Streams to continue using purpose-built assignors. It is important to note that the entire rebalance process is driven by the group coordinator with this new protocol.

Unlike the current protocol which keeps the heartbeat mechanism as lightweight as possible, the new protocol piggybacks on it to let the group coordinator assign/revoke partitions to/from group members while allowing group members to propagate their current state to the group coordinator. The ConsumerGroupHeartbeat API is introduced for this purpose.

When a client side assignor is used, the group coordinator requests the assignment from one group member by notifying him via the heartbeat protocol. The chosen member uses the ConsumerGroupPrepareAssignment API and the ConsumerGroupInstallAssignment API to respectively get the current state of the group and to install the computed assignment. Thanks to this, the input of the client side assignor is entirely driven by the group coordinator. The consumer is no longer responsible for maintaining any state besides its assigned partitions.

The new protocol’s RPCs are specified in the details in the public interfaces section of this document while the details of the rebalance logic is described in the next chapter.

Group Coordinator

Implementing the new rebalance protocol in the current group coordinator is not appropriate in our opinion because its requires many changes anyway in the current protocol to make it interoperable. Therefore, this KIP proposes to rewrite the group coordinator from scratch in Java. The new group coordinator will have a state machine per __consumer_offsets partitions, where each state machine is modelled as an event loop. Those state machines will be executed in N threads.

Consumer Groups

The group coordinator already supports the so called consumer groups. Those groups are groups which implement the consumer embedded protocol type. With the introduction of the new consumer rebalance protocol, we need a way to differentiate the existing groups from the new consumer groups. This is important because the existing group relies on a specific set of APIs whereas the new consumer group will use a different set of APIs.

Therefore, we propose to introduce the notion of types within the group coordinator. This will allow us to support different types of groups in the future. We propose to call the current group generic as it represents a generic implementation of the membership protocol which is specialized by a protocol type and to call the new consumer group consumer. Effectively, we would have consumer groups and generic groups using the consumer embedded protocol, the old one and the new one proposed in this document.

The ListGroups API will be extended to support both filtering on the group types and returning the group types of the queried groups.

Data Model

Before diving into the details of the new rebalance process, let’s define the data model of the group as the group coordinator will bookkeep it. Note that this data model is a logical one. The detailed records are described in the Public Interfaces section of this document.

Consumer Group & Member

The group and the members represents the current state of a consumer group.

Target Assignment

The target (or desired) assignment of the group. This represents the assignment that all the members will eventually converge to. It is a declarative assignment which is generated by the assignor based on the group state.

Current Assignment

The Current Assignment represents the current epoch and assignment of a member. Note that members of a given group could be at a different epoch but they will all eventually converge to the target assignment.

Rebalance Process

The rebalance process is entirely driven by the group coordinator and revolves around three kinds of epoch: the group epoch, the assignment epoch and the member epoch. The process and the epochs are explained in the following chapters.

Group Epoch - Trigger a rebalance

The group coordinator is responsible for triggering a rebalance of the group when the metadata of the group changes. The metadata of the group is used as the input of the assignment function. For tracking this, we introduce the group epoch which represents the generation (or the version) of the group metadata. The group epoch is incremented whenever the group metadata is updated. There are a couple of cases to consider:

• A member joins or leaves the group.
• A member updates its subscriptions.
• A member updates its assignors.
• A member updates its assignors' reason or metadata.
• A member is fenced or removed from the group by the group coordinator.
• The partition metadata is updated. For instance when a new partition is added or a new topic matching the subscribed topics is created.

In all these cases, a new version of the group metadata is persisted by the group coordinator with an incremented group epoch. This also signals that a new assignment is required for the group.

Assignment Epoch - Compute the group assignment

Whenever the group epoch is larger than the target assignment epoch, the group coordinator will trigger the computation of a new target assignment based on the latest group metadata. When the new assignment is computed, the group coordinator persists it. The assignment epoch becomes the group epoch of the group metadata used to compute the assignment.

The group coordinator either directly computes the new target assignment for the group based on its default server-side assignor or requests a new assignment from one of the members in the group. The entire delegation logic for the latter is detailed later in the document.

Member Epoch - 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 requires three phases:

1. The group coordinator revokes the partitions which are no longer in the target assignment of the member. It does so by providing the intersection of the Current Partitions and the Target Partitions in the heartbeat response until the member acknowledges the revocation in the heartbeat response. The group coordinator will give the rebalance timeout to the member for the revocation process to complete or kick it out from the group otherwise.
2. When the group coordinator receives the acknowledgement of the revocation, it updates the member current assignment to its target assignment (and target epoch) and durably persist it.
3. The group coordinator assigns the new partitions to the member. It does so by providing the Target Partitions to the member while ensuring that partitions which are not revoked by other members yet are removed from this set. In other words, new partitions are incrementally assigned to the member when they are revoked by the other members.

The rebalance timeout is provided by the member when it joins the group. It is basically the max poll interval configured on the client side. The timer starts ticking when the heartbeat response is sent out by the group coordinator.

Assignment Process

Whenever the group epoch is larger than the assignment epoch, the group coordinator must compute a new target assignment for the group. The group coordinator will either directly compute a new assignment with its server side assignor or delegate the assignment to a member of the group if a client-side assignor must be used.

The new target assignment for the group is basically a function of the current group metadata and the current target assignment. One important aspect to note here is that the assignment is declarative now instead of being incremental like it is in the current implementation. In other words, the assignor defines the desired state for the group and let the group coordinator converge to it.

Assignor Selection

The group coordinator has to determine which assignment strategy must be used for the group. The group's members may not have exactly the same assignors at any given point in time - e.g. they may migrate from an assignor to another one for instance. The group coordinator will chose the assignor as follow:

• The server side assignor is used if any member specified one. If multiple server side assignors are specified in the group, the group coordinator uses the most common one.
• The client side assignor is used otherwise. The group coordinator will use the assignor which is supported by all the members in the group and, if multiple are, it will respect the precedence defined by the members when they advertise their supposed assignors.

Server Side Mode

The server side assignor is pluggable and the client can choose the one that it wants to use by providing its name in the heartbeat request. If the selected assignor does exist, the group coordinator will reject the heartbeat with an UNSUPPORTED_ASSIGNOR error. The list of supported assignors will be configured in the broker configuration.

We will support two assignors out of the box for Apache Kafka:

• range - An assignor which co-partitions topics.
• uniform - An assignor which uniformly assign partitions amongst the members. This is somewhat similar to the existing "sticky" assignor.

Note that in both cases, assignors are sticky. The goal is to minimise partition movements.

Client Side Mode

The client side assignment is executed by the consumer. The overall process has the following phases:

• The group coordinator selects a member to run the assignment logic. The selection is explained later in this chapter.
• The group coordinator notifies the member to compute the new assignment by returning the COMPUTE_ASSIGNMENT error in its next heartbeat response.
• When the member receives this error, it is expected to call the ConsumerGroupPrepareAssignment API to get the current group metadata and the current target assignment.
• The member computes the new assignment with the relevant assignor.
• The member calls the ConsumerGroupInstallAssignment API to install the new assignment. The group coordinator validates it and persists it.

Note that the group coordinator always installs any new valid assignment, even if the group epoch has changed in the mean time, to ensure that the group can always make progress. We want to avoid the situation where a faulty member could prevent the whole group to move forward. The group coordinator only allows one inflight assignment at the time.

The chosen member is expected to complete the assignment process within the rebalance timeout. The time on the coordinator side starts ticking when the member is notified.

Metadata Version Handling (KIP-268)

Managing the compatibility of the metadata used by the client side assignors has been a challenge for our powerful users such as Kafka Streams. The metadata is usually versioned but there is not guarantee in the current protocol which ensures that the elected leader is able to handle all the versions used in the group. Kafka Streams introduces the so-called version probing (KIP-268) to mitigate this issue. This mechanism basically allows the leader to downgrade the version used by the other members in the group.

We propose to make the version a first class citizen concept in the new protocol. Every member will advertise the version used to encode their metadata, usually the highest that they support, and the minimum and the maximum version that they can handle. This allow the group coordinator to reason about the versions and to pick the member to run the assignment wisely.

The group coordinator will also ensure that any member joining with a non-overlapping version range is rejected with the UNSUPPORTED_ASSIGNOR error.

Member Selection

The group coordinator can generally pick any members to run the assignment. However, when the members support different version ranges, the group coordinator must select a member which is able to handle all the supported versions. For instance, if we have three members: A [1-5], B [3-4], C [2-4]. Member A must be selected because it supports all the other versions in the group.

Assignment Validation

Before installing any new assignment, the group coordinator will ensure that the following invariants are met:

• All partitions are assigned.
• A partition is assigned only once.
• All members exists.

Note that this validation is made with regarding to the metadata used to compute the assignment. The group may have already advanced to a newer group epoch - e.g. a member could have left during the assignment computation.

The installation will be rejected with an INVALID_ASSIGNMENT error if the invariants are not held.

Assignment Error

There could be cases where the the client side assignor can not compute a new assignment. For instance, in the context of Kafka Streams, the members may have a different topology. In this case, the client side assignor can return an error to the group coordinator. In this case, the group coordinator automatically keeps the current target assignment for group.

Member ID

Every member is uniquely identified by a UUID. This is is called the Member ID. This UUID is generated on the server side and given to the member when it joins the group. It is used in all the communication with the group coordinator and must be kept during the entirely life span of the member (e.g. the consumer). In that sense, it is similar to an incarnation ID.

Heartbeat & Session

The member uses the ConsumerGroupHeartbeat API to establish a session between him and the group coordinator. The member is expected to heartbeat every group.consumer.heartbeat.interval.ms in order to keep its session opened. If it does not heartbeat at least once within the group.consumer.session.timeout.ms, the group coordinator will kick him out from the group. group.consumer.heartbeat.interval.ms is defined on the server side and the member is told about it in the heartbeat response. The group.consumer.session.timeout.ms is also defined on the server side.

Joining & Leaving

The member joins the group by sending an heartbeat with no Member ID and a member epoch equals to 0. He can rejoins the group with a member epoch equals to 0. He can leaves the group by using a member epoch equals to -1.

Fencing

The group coordinator ensures that requests comes from a known Member ID. Any request is rejected with the UNKNOWN_MEMBER_ID error otherwise. It also ensures that the Member Epoch matches the expected member epoch. If not, the request is rejected with the FENCED_MEMBER_EPOCH error. Details for every API are given in the Public Interfaces section.

Static Membership (KIP-345)

Static membership, introduced in KIP-345, is still supported by this new rebalance protocol. When a member rejoins with the same Instance ID, the group coordinator replaces the old member with the new member. The new member can continue from where it left off.

Consumer Group States

EMPTY

When a consumer group is created or when the last member leaves the group, the consumer group is EMPTY.

ASSIGNING

When the group epoch is larger than the assignment epoch, the consumer group is ASSIGNING.

Consumer groups relying on the server-side assignor (e.g. regular consumers) are not expected to be in this state because the assignment is computed directly by the Group Coordinator.

RECONCILING

Until all the members have converged to the group epoch, the consumer group is RECONCILING.

STABLE

Once the reconciliation process is completed, the consumer group moves to the STABLE state.

DEAD

Like today, when the group remains EMPTY for a configured period, the group coordinator transitions it to DEAD to delete it.

Dynamic Group Configuration

The new rebalance protocol relies on server side configurations such as group.consumer.heartbeat.interval.ms and group.consumer.session.timeout.ms. Our goal is to give administrator the ability to use and tweak those settings for their entire consumers fleet. However, it may not always be possible to have values fitting all workloads. Therefore, we propose to extend the IncrementalAlterConfigs and the DescribeConfigs API to support a new resource type called GROUP. This allows users to override the default defined by the administrators. The dynamic group configurations are described in the Public Interfaces section.

The group coordinator is responsible for storing those group configurations in order to keep their lifecycle tight to their group. When a group is deleted, we want the configuration to be deleted as well. This assumes that IncrementalAlterConfigs and the DescribeConfigs API will be routed to the group coordinator owning the group they are acting upon.

Regex Based Subscription

The group coordinator is responsible of the regex based subscriptions. We will use Google RE2/J to compile and to execute the regular expressions on the server side. This means that all clients, regardless of their language, will use this common regular expression syntax. This should not be an issue for any common regular expressions but may require changes if specifics from the language are used. 

Feature Flag / MetadataVersion / IBP

We introduce a new feature flag named “group.new.coordinator.enable” instead of relying on the IBP to enable the new group coordinator and the consumer group protocol. This flag is particularly useful during the development phase as it allows us to enable the new group coordinator while keeping the other for production deployment. This also allows operators to explicitly enable it in the future.

When the feature will be marked as production ready, we will replace the feature flag by using the IBP / Metadata Version.

Persistence

We will introduce a new set of records to persist the new consumer group type in the existing __consumer_offsets topic. The records are detailed in the public interfaces section of this document.

Consumer

The semantics of the consumer will remain unchanged after this proposal is implemented. The goal is to swap the implementation of the group membership/assignment protocol by the new one.

Feature Flag

A new configuration setting will be used to determine whether the new protocol should be used or not. The feature flag allows the user to control when he starts using or migrating to the new protocol for its application. This is also required by our migration path as we will require to have the software on a specific version which is compatible with the new protocol.

In the case where a consumer would try to use the new protocol against a cluster which does not support it, either because the software is too old or because the feature is not enabled, the consumer would fail starting with a fatal exception.

In the beginning, the new protocol will be disabled by default. We envision enabling it by default in a future major release of Kafka.

Rebalance Process

The rebalance process in the consumer is basically the opposite of the process that was described earlier in this document. The consumer will know at any point in time its current epoch and the list of partitions that it owns. There are a few cases to consider:

• If the member is fenced by the group coordinator, it will immediately abandon all its partitions and call ConsumerRebalanceListener#onPartitionsLost. It will rejoin the group as a new member afterwards.
• Otherwise, the member will compute the difference between its currently owned partitions and the assigned partitions, as defined in the heartbeat response.
• If there are any revoked partitions, it will revoke them, commits their offsets and call ConsumerRebalanceListener#onPartitionsRevoked.
• If there are any newly assigned partitions, it will start processing them and call ConsumerRebalanceListener#onPartitionsAssigned.

Note that the process is dependent on Consumer#poll being called like with the current protocol. There is a parallel effort to this design to redesign the threading model of the consumers. This will very likely change how/when those callbacks are called.

Client-Side Assignor

By default, the consumer will entirely rely on the group coordinator but it will allow specifying a customer assignor on the client-side as already explained in this document. For this purpose, we propose to introduce a new and optional assignor interface in the Consumer called PartitionAssignor. The interface is specified in the public interfaces section of this document. The current assignor interface is strongly tied to the current group membership/assignment protocol so reusing it is not appropriate for two reasons:

• The new protocol does not really fit in the current interface and its semantic is different; and
• It seems preferable to let us evolve the current protocol independently if the need arises.

Deprecate Enforcing Rebalances

Consumer#enforceRebalance will be deprecated and will throw an IllegalStateException if used when the new protocol is enabled. Enforcing a rebalance with the new protocol does not make any sense. Instead, power users will have the ability to trigger a reassignment by either providing a non-zero reason or by updating the assignor metadata.

Streams

Kafka Streams remains a power user of the consumer so it will continue extending the consumer by providing an implementation of the new assignor interface. Streams will also rely on a feature flag to enable the new rebalance protocol.

Member Metadata & Assignment Metadata

Member Metadata refers to the metadata provided by a given member from its assignor. Assignment Metadata refers to the metadata computed by the assignor for the member. The Version, MinimumVersion, MaximumVersion, Reason and Error fields are now first class citizen in the rebalance protocol so Stream does not have the specify them in the metadata anymore. The schemas for respectively the assignor metadata and the assignment metadata are detailed in the Public Interfaces section.

Note that Streams may take this opportunity to do further changes to its metadata. We may extend this KIP or do a follow-up KIP in the future for this.

Assignor Behavior

The assignor behavior remains similar to the existing assignor. The major difference is that the assignor must serialize the assignment metadata of each member with the correct version used by the member. Another difference is that the new assignor must be able to handle the old metadata format as well during the upgrade from the old to the new protocol. This upgrade path is detailed in the upgrade section of this document.

Member Behavior

Upon receiving the assignment, each member would respectively create, close, or recycle tasks as indicated and update the global assignment information, like today. We explained earlier that partitions are incrementally assigned to the member when they are revoked by the others. This means that the assignment metadata may already reference partitions which are not assigned to the member yet. The Streams assignor must consider the assigned partitions as the source of truth in this case.

Each member encodes the lag of its standby tasks in its metadata. We can not update the lag in every heartbeat request because that would constantly trigger reassignment in the group. Instead, when a) the task lag has been reduced within the acceptable.recovery.lag threshold or b) the task lag is consistently increasing for some time, the member should consider triggering a rebalance by sending its next heartbeat with the appropriate encoded reason and the updated task lags. 

Upgrading/Downgrading to/from the New Rebalance Protocol

Upgrading to the new protocol or downgrading from is possible by rolling the consumers, assuming that the new protocol is enabled on the server side, with the correct group.protocol. When the first consumer using the new rebalance protocol joins the group, the group is converted from a generic group to a consumer group and all the JoinGroup, SyncGroup and Heartbeat calls from the non-upgrades consumers are translated to the new protocol.

Protocols Interoperability

The new rebalance protocol relies mainly on one API, the ConsumerGroupHeartbeat API, whereas the old protocol relies on three APIs:

• JoinGroup API - The consumer sends a JoinGroup request to join or re-join a group. The requests contains subscriptions, the owned partitions and the current generation id are provided (assuming KIP-792 is implemented here) serialized in the inner consumer protocol. The responses are sent back to the consumers only when all the members have (re-)joined the group. One of the member is elected as the leader of the group and is responsible for computing the assignment.
• SyncGroup API - The consumer sends a SyncGroup request when it gets its JoinGroup response back in order to get its assignment.
• Heartbeat API - The consumer heartbeats in order to maintain its session opened.

The old protocol has two modes:

• Eager Mode - In the eager mode, the consumer revokes all its partitions before re-joining the group during a rebalance.
• Cooperative Mode - In the cooperative mode, the consumer only revokes the partitions that it does not own anymore before rejoining the group. So two rebalances are required to move a partition from member A to member B. One rebalance to revoke the partition from A and another rebalance to assign the partition to B.

Note that the upgrade path will only work from the consumer protocol version 3 (as described in KIP-792).

Heartbeat Handling

The Heartbeat request updates the member session and is used to trigger a rebalance by returning REBALANCE_IN_PROGRESS if the generation id is smaller than the target assignment epoch.

JoinGroup Handling

The JoinGroup request updates the group and the member state. For this, the group coordinator parses the consumer inner protocol and maps the information provided to the new data model. The generation id is considered as the current member epoch, the subscribed topics as the subscribed topics, and the user data as the assignor data with version -1.

In the eager mode, the consumer revokes all its partitions before rejoining the group so when all the newly assigned partitions are free to be assigned, the group coordinator updates the current assignment of the member to the target assignment and then sends the JoinGroup response to the consumer with the generation id set to the new member epoch. The consumer will then collects its assignment by calling the SyncGroup API.

In the cooperative mode, the consumer rejoins without revoking any partitions. If the member must revoke partition, the group coordinator replies immediately with the current member epoch. The consumer will then collects its assignment by calling the SyncGroup API. If the member does not have to revoke partitions, either because it just did or because it does not have too, the group coordinator waits until all the newly assigned partitions are free. Then it updates the current assignment of the member to the target assignment and sends the JoinGroup response to the consumer with the generation id set to the new member epoch. The consumer will then collects its assignment by calling the SyncGroup API.

SyncGroup Handling

The SyncGroup request collects the assignment of the member. The current assignment is defined as the intersection of the Current Partitions and Target Partitions. This removes the partitions to be revokes from the set. The group coordinator has to map the data model of the new protocol to the consumer protocol schema.

Assignors Interoperability

When upgrading from a client side assignor to a server side assignor, the server sides takes over immediately. The client side assignor is never used any more.

When upgrading from a client side assignor with the old protocol to a client side assignor with the new protocol, the new assignor must be able to understand the metadata serialised by the old assignor. Those metadata will have their version set to -1. This indicates that the assignor must get the version from the metadata bytes. The new assignor has to serialise the metadata using the same version. The group coordinator will refuse a member using the new protocol to join the group if its client side assignor does not support version -1 but only if the old assignor uses metadata. 

Public Interfaces

This section lists the changes impacting the public interfaces.

KRPC

New Errors

• FENCED_MEMBER_EPOCH - The member epoch does not correspond to the member epoch expected by the coordinator.
• COMPUTE_ASSIGNMENT - The member has been selected by the coordinator to compute the new target assignment of the group.
• UNSUPPORTED_ASSIGNOR - The assignor used by the member or its version range are not supported by the group.

ConsumerGroupHeartbeat API

The ConsumerGroupHeartbeat API is the new core API used by consumers to form a group. The API allows members to advertise their subscriptions, their state, their assignors, and their owned partitions. The group coordinator uses it to assign/revoke partitions to/from members. This API is also used as a liveness check.

Request Schema

The member must set all the (top level) fields 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": ["zkBroker", "broker"],
  "name": "ConsumerGroupHeartbeatRequest",
  "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 server. The member id must be kept during the entire lifetime of the member." },
    { "name": "MemberEpoch", "type": "int32", "versions": "0+", "default": "-1",
      "about": "The current member epoch; 0 to join the group; -1 to leave the group." },
    { "name": "InstanceId", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "null it not provided or if it didn't change since the last heartbeat; the instance Id otherwise." },
    { "name": "RebalanceTimeoutMs", "type": "int32", "versions": "0+", "default": -1,
      "about": "-1 if it didn't chance since the last heartbeat; the maximum time in milliseconds that the coordinator will wait on the member to revoke its partitions otherwise." },
    { "name": "SubscribedTopicNames", "type": "[]string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "null if it didn't change since the last heartbeat; the subscribed topic names otherwise." },
    { "name": "SubscribedTopicRegex", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "null if it didn't change since the last heartbeat; the subscribed topic regex otherwise" },
    { "name": "ServerAssignor", "type": "string", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "null if not used or if it didn't change since the last heartbeat; the server side assignor to use otherwise." }, 
    { "name": "ClientAssignors", "type": "[]Assignor", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "null if not used or if it didn't change since the last heartbeat; the list of client-side assignors otherwise.",
      "fields": [
        { "name": "Name", "type": "string", "versions": "0+",
          "about": "The name of the assignor." },
        { "name": "MinimumVersion", "type": "int16", "versions": "0+",
          "about": "The minimum supported version for the metadata." },
        { "name": "MaximumVersion", "type": "int16", "versions": "0+",
          "about": "The maximum supported version for the metadata." },
        { "name": "Reason", "type": "int8", "versions": "0+",
          "about": "The reason of the metadata update." }, 
        { "name": "Version", "type": "int16", "versions": "0+",
          "about": "The version of the metadata." },
        { "name": "Metadata", "type": "bytes", "versions": "0+",
          "about": "The metadata." }
      ]},
    { "name": "TopicPartitions", "type": "[]TopicPartition", "versions": "0+", "nullableVersions": "0+", "default": "null",
      "about": "null if it didn't change since the last heartbeat; the partitions owned by the member.",
      "fields": [
        { "name": "TopicId", "type": "uuid", "versions": "0+",
          "about": "The topic ID." },
        { "name": "Partitions", "type": "[]int32", "versions": "0+",
          "about": "The partitions." }
      ]}
  ]
}

Required ACL

• Read Group

Request Validation

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

• GroupId must be non-empty.
• MemberId must be non-empty.
• MemberEpoch must be >= -1.
• InstanceId, if provided, must be non-empty.
• RebalanceTimeoutMs must be larger than zero in the first heartbeat request.
• SubscribedTopicNames and SubscribedTopicRegex cannot be used together.
• SubscribedTopicNames or SubscribedTopicRegex must be in the first heartbeat request.
• SubscribedTopicRegex must be a valid regular expression.
• ServerAssignor and ClientAssignors cannot be used together.
• Assignor.Name must be non-empty.
• Assignor.MinimumVersion must be >= -1.
• Assignor.MaximumVersion must be >= 0 and >= Assignor.MinimumVersion.
• Assignor.Version must be in the >= Assignor.MinimumVersion and <= Assignor.MaximumVersion.

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

• ServerAssignor must be supported by the server.
• ClientAssignors' version range must overlap with the other members in the group.

Request Handling

When the group coordinator handle a ConsumerGroupHeartbeat request:

1. Lookups the group or creates it.
2. Creates the member should the member epoch be zero or checks whether it exists. If it does not exist, UNKNOWN_MEMBER_ID is returned.
3. Checks wether the member epoch matches the member epoch if 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 his request will be rejected with a FENCED_MEMBER_EPOCH. This is not optimal. Instead, the group coordinator could accept the request if the partitions owned by the members are a subset of the target partitions. If it is the case, it is safe to transition the member to its target epoch again.
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. 

Response Schema 

The group coordinator will only set the Assignment field when the member epoch is smaller than the target assignment epoch. This is done to ensure that the members converge to the target assignment.

{
  "apiKey": TBD,
  "type": "response",
  "name": "ConsumerGroupHeartbeatResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  // Supported errors:
  // - GROUP_AUTHORIZATION_FAILED
  // - NOT_COORDINATOR
  // - COORDINATOR_NOT_AVAILABLE
  // - COORDINATOR_LOAD_IN_PROGRESS
  // - INVALID_REQUEST
  // - UNKNOWN_MEMBER_ID
  // - FENCED_MEMBER_EPOCH
  // - UNSUPPORTED_ASSIGNOR
  // - COMPUTE_ASSIGNMENT
  "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." },
    { "name": "MemberEpoch", "type": "int32", "versions": "0+",
      "about": "The member epoch." },
    { "name": "HeartbeatIntervalMs", "type": "int32", "versions": "0+",
      "about": "The heartbeat interval in milliseconds." }, 
    { "name": "Assignment", "type": "Assignment", "versions": "0+", "nullableVersions": "0+", "default": "null",
	  "about": "null if not provided; the assignment otherwise."
      "fields": [
     	{ "name": "Error", "type": "int8", "versions": "0+",
          "about": "The assigned error." },
        { "name": "TopicPartitions", "type": "[]TopicPartition", "versions": "0+",
          "about": "The assigned topic-partitions to the member.",
          "fields": [
        	{ "name": "TopicId", "type": "uuid", "versions": "0+",
              "about": "The topic ID." },
	        { "name": "Partitions", "type": "[]int32", "versions": "0+",
              "about": "The partitions." }
      	]},
        { "name": "Version", "type": "int16", "versions": "0+",
          "about": "The version of the metadata." },
        { "name": "Metadata", "type": "bytes", "versions": "0+",
          "about": "The assigned metadata." }
	]
  ]
}

Response Handling

If the response contains no error, the member will reconcile its current assignment towards its new assignment. It does the following:

1. It updates its member epoch.
2. It computes the difference between the old and the new assignment to determine the revoked partitions and the newly assignment partitions. There should be either revoked partitions or newly assignment partitions. The protocol never does both together.
   1. It revokes the partitions, commit all the offsets, and calls ConsumerRebalanceListener#onPartitionsRevoked.
   2. It assigns the new partitions, calls PartitionAssignor#onAssignment if one is defined and calls ConsumerRebalanceListener#onPartitionsAssigned.
3. After a revocation, It sends the next heartbeat immediately to acknowledge it. 

Upon receiving the COMPUTE_ASSIGNMENT error, the consumer starts the assignment process.

Upon receiving the UNKNOWN_MEMBER_ID or FENCED_MEMBER_EPOCH error, the consumer abandon all its partitions and rejoins with the same member id and the epoch 0.

ConsumerGroupPrepareAssignment API

The ConsumerGroupPrepareAssignment API will be used by the consumer to get the information to feed its client-side assignor.

Request Schema

{
  "apiKey": TBD,
  "type": "request",
  "listeners": ["zkBroker", "broker"],
  "name": "ConsumerGroupPrepareAssignmentRequest",
  "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 assigned by the group coordinator." },
    { "name": "MemberEpoch", "type": "int32", "versions": "0+",
      "about": "The member epoch." }
  ]
}

Required ACL

• Read Group

Request Validation

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

• GroupId must be non-empty.
• MemberId must be non-empty.
• MemberEpoch must be >= 0.

Request Handling

When the group coordinator handle a ConsumerGroupPrepareAssignmentRequest request:

1. Checks wether the group exists. If it does not, GROUP_ID_NOT_FOUND is returned.
2. Checks wether the member exists. If it does not, UNKNOWN_MEMBER_ID is returned.
3. Checks wether the member epoch matches the current member epoch. If it does not, FENCED_MEMBER_EPOCH is returned.
4. Checks wether the member is the chosen one to compute the assignment. If it does not, UNKNOWN_MEMBER_ID is returned.
5. Returns the group state of the group.

Response Schema

{
  "apiKey": TBD,
  "type": "response",
  "name": "ConsumerGroupPrepareAssignmentResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  // Supported errors: 
  // - GROUP_AUTHORIZATION_FAILED
  // - NOT_COORDINATOR
  // - COORDINATOR_NOT_AVAILABLE
  // - COORDINATOR_LOAD_IN_PROGRESS
  // - INVALID_REQUEST
  // - INVALID_GROUP_ID
  // - GROUP_ID_NOT_FOUND
  // - UNKNOWN_MEMBER_ID
  // - FENCED_MEMBER_EPOCH
  "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." },
    { "name": "GroupEpoch", "type": "int32", "versions": "0+",
      "about": "The group epoch." },
    { "name": "AssignorName", "type": "string", "versions": "0+",
      "about": "The selected assignor." },
    { "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+",
        "about": "The member instance ID." },
      { "name": "SubscribedTopicIds", "type": "[]uuid", "versions": "0+",
        "about": "The subscribed topic IDs." },
      { "name": "Assignor", "type": "Assignor", "versions": "0+",
        "about": "The information of the selected assignor",
        "fields": [ 
        { "name": "Version", "type": "int16", "versions": "0+",
          "about": "The version of the metadata." },
        { "name": "Reason", "type": "int8", "versions": "0+",
          "about": "The reason of the metadata update." }, 
        { "name": "Metadata", "type": "bytes", "versions": "0+",
          "about": "The assignor metadata." }
      ]},
      { "name": "TopicPartitions", "type": "[]TopicPartition", "versions": "0+",
        "about": "The target topic-partitions of the member.",
        "fields": [
          { "name": "TopicId", "type": "uuid", "versions": "0+",
            "about": "The topic ID." },
          { "name": "Partitions", "type": "[]int32", "versions": "0+",
            "about": "The partitions." }
      ]}
    ]},
    { "name": "Topics", "type": "[]TopicMetadata", "versions": "0+",
      "about": "The topic-partition metadata.",
      "fields": [
        { "name": "TopicId", "type": "uuid", "versions": "0+",
          "about": "The topic ID." },
        { "name": "NumPartitions", "type": "int32", "versions": "0+",
          "about": "The number of partitions." }
    ]}    
  ]
}

Response Handling

If the response contains no error, the member calls the client side assignor with the group state.

Upon receiving the UNKNOWN_MEMBER_ID error, the consumer abandon the process.

Upon receiving the FENCED_MEMBER_EPOCH error, the consumer retries when receiving its next heartbeat response with its member epoch.

ConsumerGroupInstallAssignment API

The ConsumerGroupInstallAssignment API will be used by the consumer to install a new assignment for the group. The new assignment is the result of the client-side assignor.

Request Schema

{
  "apiKey": TBD,
  "type": "request",
  "listeners": ["zkBroker", "broker"],
  "name": "ConsumerGroupInstallAssignmentRequest",
  "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 assigned by the group coordinator." },
    { "name": "MemberEpoch", "type": "int32", "versions": "0+",
      "about": "The member epoch." },
    { "name": "GroupEpoch", "type": "int32", "versions": "0+",
      "about": "The group epoch." },
    { "name": "Error", "type": "int8", "versions": "0+",
      "about": "The assignment error; or zero if the assignment is successful." },
    { "name": "Members", "type": "[]Member", "versions": "0+",
      "about": "The members.", "fields": [
      { "name": "MemberId", "type": "string", "versions": "0+",
        "about": "The member ID." },
      { "name": "Partitions", "type": "[]TopicPartition", "versions": "0+",
        "about": "The assigned topic-partitions to the member.",
        "fields": [
          { "name": "TopicId", "type": "uuid", "versions": "0+",
            "about": "The topic ID." },
          { "name": "Partitions", "type": "[]int32", "versions": "0+",
            "about": "The partitions." }
        ]},
      { "name": "Version", "type": "int32", "versions": "0+",
        "about": "The metadata version." }
      { "name": "Metadata", "type": "bytes", "versions": "0+",
        "about": "The metadata bytes." }
    ]}
  ]
}

Required ACL

• Read Group

Request Validation

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

• GroupId must be non-empty.
• MemberId must be non-empty.
• MemberEpoch must be >= 0.

Request Handling

When the group coordinator handle a ConsumerGroupInstallAssignmentRequest request:

1. Checks wether the group exists. If it does not, GROUP_ID_NOT_FOUND is returned.
2. Checks wether the member exists. If it does not, UNKNOWN_MEMBER_ID is returned.
3. Checks wether the member epoch matches the current member epoch. If it does not, FENCED_MEMBER_EPOCH is returned.
4. Checks wether the member is the chosen one to compute the assignment. If it does not, UNKNOWN_MEMBER_ID is returned.
5. Validates the assignment based on the information used to compute it. If it is not valid, INVALID_ASSIGNMENT is returned.
6. Installs the new target assignment.

Response Schema

{
  "apiKey": TBD,
  "type": "response",
  "name": "ConsumerGroupInstallAssignmentResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  // Supported errors: 
  // - GROUP_AUTHORIZATION_FAILED
  // - NOT_COORDINATOR
  // - COORDINATOR_NOT_AVAILABLE
  // - COORDINATOR_LOAD_IN_PROGRESS
  // - INVALID_REQUEST
  // - INVALID_GROUP_ID
  // - GROUP_ID_NOT_FOUND
  // - UNKNOWN_MEMBER_ID
  // - FENCED_MEMBER_EPOCH  
  // - INVALID_ASSIGNMENT  
  "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." }
  ]
}

Response Handling

If the response contains no error, the member is done.

Upon receiving the FENCED_MEMBER_EPOCH error, the consumer retries when receiving its next heartbeat response with its member epoch.

Upon receiving any other errors, the consumer abandon the process.

ConsumerGroupDescribe API

Request Schema

{
  "apiKey": TBD,
  "type": "request",
  "listeners": ["zkBroker", "broker"],
  "name": "ConsumerGroupDescribe",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "GroupIds", "type": "[]string", "versions": "0+", "entityType": "groupId",
      "about": "The names 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 ConsumerGroupPrepareAssignmentRequest request:

• Checks wether the group ids exists. If it does not, GROUP_ID_NOT_FOUND is returned.
• Looks up the groups and returns the response.

Response Schema

{
  "apiKey": 71,
  "type": "response",
  "name": "ConsumerGroupDescribeResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  // Supported errors: 
  // - GROUP_AUTHORIZATION_FAILED
  // - NOT_COORDINATOR
  // - COORDINATOR_NOT_AVAILABLE
  // - COORDINATOR_LOAD_IN_PROGRESS
  // - INVALID_REQUEST
  // - INVALID_GROUP_ID
  // - GROUP_ID_NOT_FOUND
  "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": "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": "AssignorName", "type": "string", "versions": "0+",
          "about": "The selected assignor." },
        { "name": "Members", "type": "[]Member", "versions": "0+",
          "about": "The members.",
          "fields": [
          { "name": "MemberId", "type": "uuid", "versions": "0+",
            "about": "The member ID." },
          { "name": "InstanceId", "type": "string", "versions": "0+",
            "about": "The member instance ID." },
    	  { "name": "MemberEpoch", "type": "int32", "versions": "0+",
            "about": "The current member epoch." },
          { "name": "ClientId", "type": "string", "versions": "0+",
            "about": "The client ID." },
          { "name": "ClientHost", "type": "string", "versions": "0+",
            "about": "The client host." },
          { "name": "Subscriptions", "type": "[]uuid", "versions": "0+",
            "about": "The subscribed topic IDs." },
          { "name": "Assignment", "type": "Assignment", "versions": "0+",
            "about": "The current assignment.",
            "fields": [
            { "name": "Partitions", "type": "[]TopicPartition", "versions": "0+",
              "about": "The assigned topic-partitions to the member.",
              "fields": [
                { "name": "TopicId", "type": "uuid", "versions": "0+",
                  "about": "The topic ID." },
                { "name": "Partitions", "type": "[]int32", "versions": "0+",
                  "about": "The partitions." }
              ]},
            { "name": "Version", "type": "int32", "versions": "0+",
              "about": "The assignor metadata version." }
            { "name": "Metadata", "type": "bytes", "versions": "0+",
              "about": "The assignor metadata bytes." }
          ]},
          { "name": "TargetAssignment", "type": "Assignment", "versions": "0+",
            "about": "The target assignment.",
            "fields": [
            { "name": "Partitions", "type": "[]TopicPartition", "versions": "0+",
              "about": "The assigned topic-partitions to the member.",
              "fields": [
                { "name": "TopicId", "type": "uuid", "versions": "0+",
                  "about": "The topic ID." },
                { "name": "Partitions", "type": "[]int32", "versions": "0+",
                  "about": "The partitions." }
              ]},
            { "name": "Version", "type": "int32", "versions": "0+",
              "about": "The assignor metadata version." }
            { "name": "Metadata", "type": "bytes", "versions": "0+",
              "about": "The assignor metadata bytes." }
          ]},
      { "name": "AuthorizedOperations", "type": "int32", "versions": "3+", "default": "-2147483648",
        "about": "32-bit bitfield to represent authorized operations for this group." }
    ]}
  ]
}

Response Handling

Nothing particular.

ListGroups API

The existing ListGroups API will be extended to support the notion of group types and to support the new group states.

Request Schema

The TypesFilter field is introduced. It allows listing groups of certain types.

{
  "apiKey": 16,
  "type": "request",
  "listeners": ["zkBroker", "broker"],
  "name": "ListGroupsRequest",
  // Version 1 and 2 are the same as version 0.
  //
  // Version 3 is the first flexible version.
  //
  // Version 4 adds the StatesFilter field (KIP-518).
  //
  // Version 5 adds the TypesFilter field (KIP-848).
  "validVersions": "0-5",
  "flexibleVersions": "3+",
  "fields": [
    { "name": "StatesFilter", "type": "[]string", "versions": "4+",
      "about": "The states of the groups we want to list. If empty all groups are returned with their state." },
    { "name": "TypesFilter", "type": "[]string", "versions": "5+",
      "about": "The types of the groups we want to list. If empty all groups are returned" }
  ]
}

Required ACL

• Read Group

Response Schema

The GroupType field is introduced. It represents the type of the group.

{
  "apiKey": 16,
  "type": "response",
  "name": "ListGroupsResponse",
  // Version 1 adds the throttle time.
  //
  // Starting in version 2, on quota violation, brokers send out
  // responses before throttling.
  //
  // Version 3 is the first flexible version.
  //
  // Version 4 adds the GroupState field (KIP-518).
  //
  // Version 5 adds the GroupType field (KIP-848).
  "validVersions": "0-5",
  "flexibleVersions": "3+",
  "fields": [
    { "name": "ThrottleTimeMs", "type": "int32", "versions": "1+", 
      "ignorable": true,
      "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 error code, or 0 if there was no error." },
    { "name": "Groups", "type": "[]ListedGroup", "versions": "0+",
      "about": "Each group in the response.", "fields": [
      { "name": "GroupId", "type": "string", "versions": "0+",
        "entityType": "groupId",
        "about": "The group ID." },
      { "name": "ProtocolType", "type": "string", "versions": "0+",
        "about": "The group protocol type." },
      { "name": "GroupState", "type": "string", "versions": "4+", "ignorable": true,
        "about": "The group state name." },
      { "name": "GroupType", "type": "string", "versions": "5+", "ignorable": true,
        "about": "The group state name." }
    ]}
  ]
}

OffsetCommit API

The version of the API is bumped to 9.

Request Schema

We propose to rename GenerationId to GenerationIdOrMemberEpoch.

Request Handling

When the group id corresponds to a consumer group using the new rebalance protocol, the provided member epoch must match the expected member epoch.

Response Schema

The response can return FENCED_MEMBER_EPOCH.

OffsetFetch API

The version of the API is bumped to 9.

Request Schema

{
  "apiKey": 9,
  "type": "request",
  "listeners": ["zkBroker", "broker"],
  "name": "OffsetFetchRequest",
  // In version 0, the request read offsets from ZK.
  //
  // Starting in version 1, the broker supports fetching offsets from the internal __consumer_offsets topic.
  //
  // Starting in version 2, the request can contain a null topics array to indicate that offsets
  // for all topics should be fetched. It also returns a top level error code
  // for group or coordinator level errors.
  //
  // Version 3, 4, and 5 are the same as version 2.
  //
  // Version 6 is the first flexible version.
  //
  // Version 7 is adding the require stable flag.
  //
  // Version 8 is adding support for fetching offsets for multiple groups at a time
  //
  // Version 9 adds GenerationIdOrMemberEpoch and MemberId fields (KIP-848).
  "validVersions": "0-9",
  "flexibleVersions": "6+",
  "fields": [
    { "name": "GroupId", "type": "string", "versions": "0-7", "entityType": "groupId",
      "about": "The group to fetch offsets for." },
    // New fields.
    { "name": "GenerationIdOrMemberEpoch", "type": "int32", "versions": "9+", "default": "-1", "ignorable": true,
      "about": "The generation of the group." },
    { "name": "MemberId", "type": "string", "versions": "9+", "ignorable": true,
      "about": "The member ID assigned by the group coordinator." },
    // End of new fields.
    { "name": "Topics", "type": "[]OffsetFetchRequestTopic", "versions": "0-7", "nullableVersions": "2-7",
      "about": "Each topic we would like to fetch offsets for, or null to fetch offsets for all topics.", "fields": [
      { "name": "Name", "type": "string", "versions": "0-7", "entityType": "topicName",
        "about": "The topic name."},
      { "name": "PartitionIndexes", "type": "[]int32", "versions": "0-7",
        "about": "The partition indexes we would like to fetch offsets for." }
    ]},
    { "name": "Groups", "type": "[]OffsetFetchRequestGroup", "versions": "8+",
      "about": "Each group we would like to fetch offsets for", "fields": [
      { "name": "groupId", "type": "string", "versions": "8+", "entityType": "groupId",
        "about": "The group ID."},
      { "name": "Topics", "type": "[]OffsetFetchRequestTopics", "versions": "8+", "nullableVersions": "8+",
        "about": "Each topic we would like to fetch offsets for, or null to fetch offsets for all topics.", "fields": [
        { "name": "Name", "type": "string", "versions": "8+", "entityType": "topicName",
          "about": "The topic name."},
        { "name": "PartitionIndexes", "type": "[]int32", "versions": "8+",
          "about": "The partition indexes we would like to fetch offsets for." }
      ]}
    ]},
    {"name": "RequireStable", "type": "bool", "versions": "7+", "default": "false",
      "about": "Whether broker should hold on returning unstable offsets but set a retriable error code for the partitions."}
  ]
}

Request Handling

The MemberId and the GenerationIdOrMemberEpoch are verified. FENCED_MEMBER_EPOCH or UNKNOWN_MEMBER_ID is returned.

Response Schema

The response is the same. Only new FENCED_MEMBER_EPOCH or UNKNOWN_MEMBER_ID errors could be returned.

Response Handling

Upon receiving the FENCED_MEMBER_EPOCH error, the consumer retries when receiving its next heartbeat response with its member epoch.

DescribeConfigs API

The API is the same but supports a new resource type: GROUP (16). When GROUP is used, the resource name corresponds to the group id. 

AlterIncrementalConfigs API

The API is the same but supports a new resource type: GROUP (16). When GROUP is used, the resource name corresponds to the group id.

Records

This section describes the new record types required for the new protocol. The storage layout is based on the data model described earlier in this document.

As explained earlier, they will be persisted in the __consumer_offsets compacted topic. The compacted topic based storage requires a dedicated key type per record type in order for the compaction to work. The current protocol already uses versions from 0 to 2 (included) for the keys.

Group Metadata

Groups can be rather large so we propose to use several records to store a group in order to not be limited by the maximum batch size (1MB by default). Therefore we propose to store group metadata with two records types: the ConsumerGroupMetadata and the ConsumerGroupMemberMetadata.

A group with X members will be stored with X+2 records. One ConsumerGroupMemberMetadata per member, one ConsumerGroupPartitionMetadata, and one ConsumerGroupMetadata for the group at the end. Atomicity is not a concern here. All the records can be applied independently.

Moreover, the whole group does not necessarily have to be written for every epoch. Members who have not changed could be omitted as the compacted topic will retain their previous state anyway.

When a member is deleted, a tombstone for him is written to the partition.

ConsumerGroupMetadataKey

{
    "type": "data",
    "name": "ConsumerGroupMetadataKey",
    "validVersions": "3",
    "flexibleVersions": "none",
    "fields": [
      	{ "name": "GroupId", "type": "string", "versions": "3" }
    ]
}

ConsumerGroupMetadataValue

{
    "type": "data",
    "name": "ConsumerGroupMetadataValue",
    "validVersions": "0",
    "flexibleVersions": "0+",
    "fields": [
        { "name": "Epoch", "versions": "0+", "type": "int32" }
    ], 
}

ConsumerGroupPartitionMetadataKey

{
    "type": "data",
    "name": "ConsumerGroupPartitionMetadataKey",
    "validVersions": "4",
    "flexibleVersions": "none",
    "fields": [
      	{ "name": "GroupId", "type": "string", "versions": "4" }
    ]
}

ConsumerGroupPartitionMetadataValue

{
    "type": "data",
    "name": "ConsumerGroupPartitionMetadataValue",
    "validVersions": "0",
    "flexibleVersions": "0+",
    "fields": [
        { "name": "Epoch", "versions": "0+", "type": "int32" },
        { "name": "TopicPartitionMetadata", "versions": "0+",
          "type": "[]TopicPartition", "fields": [
            { "name": "TopicId", "versions": "0+", "type": "uuid" },
            { "name": "NumPartitions", "versions": "0+", "type": "int32" }
          ]}
    ], 
}

ConsumerGroupMemberMetadataKey

{
    "type": "data",
    "name": "ConsumerGroupMemberMetadataKey",
    "validVersions": "5",
    "flexibleVersions": "none",
    "fields": [
        { "name": "GroupId", "type": "string", "versions": "5" },
        { "name": "MemberId", "type": "string", "versions": "5" }
    ]
}

ConsumerGroupMemberMetadataValue

{
    "type": "data",
    "name": "ConsumerGroupMemberMetadataValue",
    "validVersions": "0",
    "flexibleVersions": "0+",
    "fields": [
        { "name": "GroupEpoch", "versions": "0+", "type": "int32" },
        { "name": "InstanceId", "versions": "0+", "type": "string" },
        { "name": "ClientId", "versions": "0+", "type": "string" },
        { "name": "ClientHost", "versions": "0+", "type": "string" },
        { "name": "SubscribedTopicNames", "versions": "0+", "type": "[]string" },
        { "name": "SubscribedTopicRegex", "versions": "0+", "type": "string" },
        { "name": "Assignors", "versions": "0+",
          "type": "[]Assignor", "fields": [
            { "name": "Name", "versions": "0+", "type": "string" },
            { "name": "MinimumVersion", "versions": "0+", "type": "int16" },
            { "name": "MaximumVersion", "versions": "0+", "type": "int16" },
            { "name": "Reason", "versions": "0+", "type": "int8" },
			{ "name": "Version", "versions": "0+", "type": "int16" },
            { "name": "Metadata", "versions": "0+", "type": "bytes" }
          ]}
    ], 
}

Target Assignment

The target assignment is stored in a single record.

ConsumerGroupTargetAssignmentKey

{
    "type": "data",
    "name": "ConsumerGroupTargetAssignmentKey",
    "validVersions": "6",
    "flexibleVersions": "none",
    "fields": [
      	{ "name": "GroupId", "type": "string", "versions": "5" }
    ]
}

ConsumerGroupTargetAssignmentValue

{
    "type": "data",
    "name": "ConsumerGroupTargetAssignmentValue",
    "validVersions": "0",
    "flexibleVersions": "0+",
    "fields": [
        { "name": "AssignmentEpoch", "versions": "0+", "type": "int32" },
        { "name": "Members", "versions": "0+", "type": "[]Member", "fields": [
        	{ "name": "MemberId", "versions": "0+", "type": "string" },
            { "name": "Error", "versions": "0+", "type": "int8" },
            { "name": "TopicPartitions", "versions": "0+",
          	  "type": "[]TopicPartition", "fields": [
            	{ "name": "TopicId", "versions": "0+", "type": "uuid" },
            	{ "name": "Partitions", "versions": "0+", "type": "[]int32" }
        	]},
        	{ "name": "Version", "versions": "0+", "type": "int16" },
        	{ "name": "Metadata", "versions": "0+", "type": "bytes" }
        ]
    ]
}

Current Member Assignment

The current member assignment represents, as the name suggests, the current assignment of a given member.

When a member is deleted from the group, a tombstone for him is written to the partition.

ConsumerGroupCurrentMemberAssignmentKey

{
    "type": "data",
    "name": "ConsumerGroupCurrentMemberAssignmentKey",
    "validVersions": "7",
    "flexibleVersions": "none",
    "fields": [
      	{ "name": "GroupId", "type": "string", "versions": "7" },
      	{ "name": "MemberId", "type": "string", "versions": "7" },
    ]
}

ConsumerGroupCurrentMemberAssignmentValue

{
    "type": "data",
    "name": "ConsumerGroupCurrentMemberAssignmentValue",
    "validVersions": "0",
    "flexibleVersions": "0+",
    "fields": [
        { "name": "MemberEpoch", "versions": "0+", "type": "int32" },
		{ "name": "Error", "versions": "0+", "type": "int8" },
        { "name": "TopicPartitions", "versions": "0+",
          "type": "[]TopicPartition", "fields": [
            { "name": "TopicId", "versions": "0+", "type": "uuid" },
            { "name": "Partitions", "versions": "0+", "type": "[]int32" }
        ]},
        { "name": "Version", "versions": "0+", "type": "int16" },
        { "name": "Metadata", "versions": "0+", "type": "bytes" }
    ], 
}

GroupConfigurationKey

{
    "type": "data",
    "name": "GroupConfigurationKey",
    "validVersions": "8",
    "flexibleVersions": "none",
    "fields": [
     	{ "name": "GroupId", "type": "string", "versions": "8" }
    ]
}

GroupConfigurationValue

{
    "type": "data",
    "name": "GroupConfigurationValue",
    "validVersions": "0",
    "flexibleVersions": "0+",
    "fields": [
        { "name": "Configurations", "versions": "0+", "type": "[]Configuration",
          "fields": [
		     { "name": "Name", "type": "string", "versions": "0+",
      		   "about": "The name of the configuration key." },
    		 { "name": "Value", "type": "string", "versions": "0+",
      		   "about": "The value of the configuration." }
		]}
    ] 
}

Broker API

The new PartitionAssignor interface will be introduced on the server side. Two implementations will be provided out of the box: RangeAssignor (range) and UniformAssignor (uniform).

package org.apache.kafka.server.group.consumer;

public interface PartitionAssignor {

    class Group {
        /**
         * The members.
         */
        List<GroupMember> members;

        /**
         * The topics' metadata.
         */
        List<TopicMetadata> topics;
    }

    class GroupMember {
        /**
         * The member ID.
         */
        String memberId;

        /**
         * The instance ID if provided.
         */
        Optional<String> instanceId;

        /**
         * The set of topic IDs that the member is subscribed to.
         */
        List<Uuid> subscribedTopicIds;

        /**
         * The partitions owned by the member at the current epoch.
         */
        List<TopicIdPartition> ownedPartitions;
    }

     class TopicMetadata {
      	/**
		 * The topic ID.
		 */
		Uuid topicId;

        /**
		 * The number of partitions.
		 */
		int numPartitions; 
    } 

    class Assignment {
        /**
         * The member assignment.
         */
        List<MemberAssignment> members;
    }

    class MemberAssignment {
        /**
         * The member ID.
         */
        String memberId;

        /**
         * The assigned partitions.
         */
        List<TopicIdPartition> partitions;
    }

    /**
     * Unique name for this assignor.
     */
    String name();

    /**
     * Perform the group assignment given the current members and
     * topic metadata.
     *
     * @param group The group state.
     * @return The new assignment for the group.
     */
    Assignment assign(Group group);
}

Broker Metrics

The set of new metrics is not clear at the moment. We plan to amend the KIP later on when progress on the implementation would have been made.

• Group count by type
• Group count by state
• Rebalance Rate

Broker Configurations

New properties in the broker configuration.

Group Configurations

New dynamic group properties.

Consumer API

New PartitionAssignor interface

The new PartitionAssignor interface will be introduced to replace the ConsumerPartitionAssignor interface. The interface is defined as follow:

package org.apache.kafka.clients.consumer;

public interface PartitionAssignor {

    class Group {
        /**
         * The members.
         */
        List<GroupMember> members;

        /**
         * The topics' metadata.
         */
        List<TopicMetadata> topics;
    }

    class GroupMember {
        /**
         * The member ID.
         */
        String memberId;

        /**
         * The instance ID if provided.
         */
        Optional<String> instanceId;

        /**
         * The set of topic IDs that the member is subscribed to.
         */
        List<Uuid> subscribedTopicIds;

   		/**
		 * The reason reported by the member.
		 */
		byte reason;  

		/**
		 * The version of the metadata encoded in {{@link GroupMember#metadata()}}.
		 */
		int version;

        /**
         * The custom metadata provided by the member as defined
         * by {{@link PartitionAssignor#metadata()}}.
         */
        ByteBuffer metadata;

        /**
         * The partitions owned by the member at the current epoch.
         */
        List<TopicIdPartition> ownedPartitions;
    }

    class TopicMetadata {
      	/**
		 * The topic ID.
		 */
		Uuid topicId;

        /**
		 * The number of partitions.
		 */
		int numPartitions; 
    }

    class Assignment {
        /**
         * The assignment error.
         */
		byte error;

        /**
         * The member assignment.
         */
        List<MemberAssignment> members;
    }

    class MemberAssignment {
        /**
         * The member ID.
         */
        String memberId;

        /**
         * The assigned partitions.
         */
        List<TopicIdPartition> partitions;

  		/**
		 * The error reported by the assignor.
		 */
		byte error; 

 		/**
		 * The version of the metadata encoded in {{@link GroupMember#metadata()}}.
		 */
		int version;

        /**
         * The custom metadata provided by the assignor.
         */
        ByteBuffer metadata;
    }

	class Metadata {
   		/**
		 * The reason reported by the assignor.
		 */
		byte reason; 

 		/**
		 * The version of the metadata encoded in {{@link Metadata#metadata()}}.
		 */
		int version;

        /**
         * The custom metadata provided by the assignor.
         */
        ByteBuffer metadata; 
    }

    /**
     * Unique name for this assignor.
     */
    String name();

    /**
     * The minimum version.
     */
    int minimumVersion();

    /**
     * The maximum version.
     */
    int maximumVersion();

    /**
     * Return serialized data that will be sent to the assignor.
     */
    Metadata metadata();

    /**
     * Perform the group assignment given the current members and
     * topic metadata.
     *
     * @param group The group state.
     * @return The new assignment for the group.
     */
    Assignment assign(Group group);

    /**
     * Callback which is invoked when the member received a new
     * assignment from the assignor/group coordinator.
     */
    void onAssignment(MemberAssignment assignment);
}

Deprecate Consumer#enforceRebalance and Consumer#enforceRebalance(String)

The enforceRebalance methods are no longer necessary and will be deprecated in a future release.

Deprecate ConsumerPartitionAssignor interface.

The ConsumerPartitionAssignor interface will be deprecated in a future (major) release.

Consumer Configurations

Streams Member Metadata and Assignment Metadata

The changes here are mainly informative at this stage. They show how we could structure the Streams' metadata. We may decide to leverage this change to do more changes.

Member Metadata Schema

This is the schema of the metadata advertised by each member.

Member Metadata Reasons

• None (0)
• Shutdown (1)
• WarmUpReady (2)
• WarmUpFailed (3)
• TopologyChanged (4)

Assignment Metadata Schema

Assignment Metadata Errors

• None (0)
• Shutdown (1)
• AssignmentError (2)
• InconsistentTopology (3)

Streams Configurations

Admin API

Admin#listConsumerGroups

The Admin#listConsumerGroups will be extended to support querying group types and retrieving/querying the new group states.

public class ListConsumerGroupsOptions extends AbstractOptions<ListConsumerGroupsOptions> {

    /**
     * If types is set, only groups with these types will be returned.
     */
    public ListConsumerGroupsOptions withTypes(Set<String> types) {
        this.types = types;
    }

    /**
     * Returns the list of Types that are requested or empty if no types
     * have been specified.
     */
    public Set<String> types() {
        return types;
    }
}

public class ConsumerGroupListing {

    /**
     * Consumer Group type, generic by default.
     */
    public String type() {
        return type;
    }
}

public enum ConsumerGroupState {
    UNKNOWN("Unknown"),
    PREPARING_REBALANCE("PreparingRebalance"),
    COMPLETING_REBALANCE("CompletingRebalance"),
    STABLE("Stable"),
    DEAD("Dead"),
    EMPTY("Empty"),
    ASSIGNING("Assigning"),
    RECONCILING("Reconciling");
}

Admin#describeConsumerGroups

The Admin#describeConsumerGroups will be extended to expose the new information related to the new protocol.

public class ConsumerGroupDescription {
    public String type() {
      return type;
    }
}

public class MemberDescription {
    // Current Assignment
    public MemberAssignment assignment() {}

    // Target Assignment
    public MemberAssignment targetAssignment() {}
}

public class MemberAssignment {
	/**
	 * The reason reported by the assignor.
	 */
	byte error; 

	/**
	 * The version of the metadata encoded in {{@link Metadata#metadata()}}.
	 */
	int version;

	/**
	 * The custom metadata provided by the assignor.
	 */
	ByteBuffer metadata;

	/**
     * The partitions owned by the member at the current epoch.
     */
    List<TopicIdPartition> ownedPartitions;
}

Admin#incrementalAlterConfigs and Admin#describeConfigs

The GROUP resource type is added.

public final class ConfigResource {
    /**
     * Type of resource.
     */
    public enum Type {
         GROUP((byte) 16), BROKER_LOGGER((byte) 8), BROKER((byte) 4), TOPIC((byte) 2), UNKNOWN((byte) 0);
    }
}

kafka-consumer-groups

The kafka-consumer-group command line tool will be extended to support the –type filter which allows to list or to describe groups implementing a specific type.

kafka-consumer-groups.sh -–bootstrap-server localhost:9092 -–list -–type <comma separated list of types>

kafka-consumer-groups.sh -–bootstrap-server localhost:9092 -–describe -–type <comma separated list of types>

Case Studies

Basic

Let’s look at a few examples to illustrate the rebalance logic. Let’s assume that the group is subscribed to the topic foo which has 3 partitions.

Let’s start with an empty group:

• Group (epoch=0)
• Empty
• Target Assignment (epoch=0)
• Empty
• Member Assignment
• Empty

Member A joins the group. The coordinator bumps the group epoch to 1, adds A to the group, and creates an empty member assignment.

• Group (epoch=1)
• A
• Target Assignment (epoch=0)
• Empty
• Member Assignment
• A - epoch=0, partitions=[]

The coordinator computes and installs the new target assignment. All the partitions are assigned to A.

• Group (epoch=1)
• A
• Target Assignment (epoch=1)
• A - partitions=[foo-0, foo-1, foo-2]
• Member Assignment
• A - epoch=0, partitions=[]

When A heartbeats, the group coordinator transitions him to its target epoch/assignment because it does not have any partitions to revoke. The group coordinator updates the member assignment and replies with the new epoch 1 and all the partitions.

• Group (epoch=1)
• A
• Target Assignment (epoch=1)
• A - partitions=[foo-0, foo-1, foo-2]
• Member Assignment
• A - epoch=1, partitions=[foo-0, foo-1, foo-2]

Member B joins the group. The coordinator adds the member to the group and bumps the group epoch to 2.

• Group (epoch=2)
• A
• B
• Target Assignment (epoch=1)
• A - partitions=[foo-0, foo-1, foo-2]
• Member Assignment
• A - epoch=1, partitions=[foo-0, foo-1, foo-2]
• B - epoch=0, partitions=[]

The coordinator computes and installs the new target assignment.

• Group (epoch=2)
• A
• B
• Target Assignment (epoch=2)
• A - partitions=[foo-0, foo-1]
• B - partitions=[foo-2]
• Member Assignment
• A - epoch=1, partitions=[foo-0, foo-1, foo-2]
• B - epoch=1, partitions=[foo-2]

At this point B can transitions to epoch 1 but cannot get foo-2 until A revokes it.

When A heartbeats, the group coordinator instructs him to revoke foo-2.

When A heartbeats again and acknowledges the revocation, the group coordinator transitions him to epoch 2.

• Group (epoch=2)
• A
• B
• Target Assignment (epoch=2)
• A - partitions=[foo-0, foo-1]
• B - partitions=[foo-2]
• Member Assignment
• A - epoch=2, partitions=[foo-0, foo-1]
• B - epoch=1, partitions=[foo-2]

When B heartbeats, he can now gets foo-2.

• Group (epoch=2)
• A
• B
• Target Assignment (epoch=2)
• A - partitions=[foo-0, foo-1]
• B - partitions=[foo-2]
• Member Assignment
• A - epoch=2, partitions=[foo-0, foo-1]
• B - epoch=2, partitions=[foo-2]

Member C joins the group. The coordinator adds the member to the group and bumps the group epoch to 3.

• Group (epoch=3)
• A
• B
• C
• Target Assignment (epoch=2)
• A - partitions=[foo-0, foo-1]
• B - partitions=[foo-2]
• Member Assignment
• A - epoch=2, partitions=[foo-0, foo-1]
• B - epoch=2, partitions=[foo-2]
• C - epoch=0, partitions=[]

The coordinator computes and installs the new target assignment.

• Group (epoch=3)
• A
• B
• C
• Target Assignment (epoch=3)
• A - partitions=[foo-0]
• B - partitions=[foo-2]
• C - partitions=[foo-1]
• Member Assignment
• A - epoch=2, partitions=[foo-0, foo-1]
• B - epoch=2, partitions=[foo-2]
• C - epoch=0, partitions=[]

When B heartbeats, the group coordinator transitions him to epoch 3 because B has no partitions to revoke. It persists the change and reply.

• Group (epoch=3)
• A
• B
• C
• Target Assignment (epoch=3)
• A - partitions=[foo-0]
• B - partitions=[foo-2]
• C - partitions=[foo-1]
• Member Assignment
• A - epoch=2, partitions=[foo-0, foo-1]
• B - epoch=3, partitions=[foo-2]
• C - epoch=3, partitions=[foo-1]

When C heartbeats, it transitions to epoch 3 but cannot get foo-1 yet.

When A heartbeats, the group coordinator instructs him to revoke foo-1.

When A heartbeats again and acknowledges the revocation, the group coordinator transitions him to epoch 2.

When C heartbeats, the group coordinator transitions him to epoch 3, persists the change, and reply.

• Group (epoch=3)
• A
• B
• C
• Target Assignment (epoch=3)
• A - partitions=[foo-0]
• B - partitions=[foo-2]
• C - partitions=[foo-1]
• Member Assignment
• A - epoch=2, partitions=[foo-0]
• B - epoch=3, partitions=[foo-2]
• C - epoch=3, partitions=[foo-1]

All the members have eventually advanced to the group epoch (3).

Incremental Revocation & Assignment

Let's imagine a group with two members and six partitions.

• Group (epoch=21)
• A
• B
• Target Assignment (epoch=21)
• A - partitions=[foo-0, foo-1, foo-2]
• B - partitions=[foo-3, foo-4, foo-5]
• Member Assignment
• A - epoch=21, partitions=[foo-0, foo-1, foo-2]
• B - epoch=21, partitions=[foo-3, foo-4, foo-5]

C joins the group. The group coordinator adds him, bumps the group epoch, create the member assignment, and computes the target assignment.

• Group (epoch=22)
• A
• B
• C
• Target Assignment (epoch=22)
• A - partitions=[foo-0, foo-1]
• B - partitions=[foo-3, foo-4]
• C - partitions=[foo-2, foo-5]
• Member Assignment
• A - epoch=21, partitions=[foo-0, foo-1, foo-2]
• B - epoch=21, partitions=[foo-3, foo-4, foo-5]
• C - epoch=0, partitions=[]

C heartbeats, the group coordinator transitions him to epoch 22 but does not yet give him any partitions because they are not revoked yet.

• Group (epoch=22)
• A
• B
• C
• Target Assignment (epoch=22)
• A - partitions=[foo-0, foo-1]
• B - partitions=[foo-3, foo-4]
• C - partitions=[foo-2, foo-5]
• Member Assignment
• A - epoch=21, partitions=[foo-0, foo-1, foo-2]
• B - epoch=21, partitions=[foo-3, foo-4, foo-5]
• C - epoch=22, partitions=[foo-2, foo-5]

A heartbeats, the group coordinator instructs him to revoke foo-2.

B heartbeats, the group coordinator instructs him to revoke foo-5.

C heartbeats, no changes for him.

A heartbeats and acknowledges the revocation, the group coordinator transitions him to epoch 22, persists and reply.

• Group (epoch=22)
• A
• B
• C
• Target Assignment (epoch=22)
• A - partitions=[foo-0, foo-1]
• B - partitions=[foo-3, foo-4]
• C - partitions=[foo-2, foo-5]
• Member Assignment
• A - epoch=22, partitions=[foo-0, foo-1]
• B - epoch=21, partitions=[foo-3, foo-4, foo-5]
• C - epoch=22, partitions=[foo-2, foo-5]

C heartbeats, the group coordinator gives him foo-2 which is now free but hold foo-5.

B heartbeats and acknowledges the revocation, the group coordinator transitions him to epoch 22, persists and reply.

• Group (epoch=22)
• A
• B
• C
• Target Assignment (epoch=22)
• A - partitions=[foo-0, foo-1]
• B - partitions=[foo-3, foo-4]
• C - partitions=[foo-2, foo-5]
• Member Assignment
• A - epoch=22, partitions=[foo-0, foo-1]
• B - epoch=22, partitions=[foo-3, foo-4]
• C - epoch=22, partitions=[foo-2, foo-5]

C heartbeats, the group coordinator gives him foo-2 and foo-5.

Member Failure

Let's start with a group with three members and six partitions.

• Group (epoch=22)
• A
• B
• C
• Target Assignment (epoch=22)
• A - partitions=[foo-0, foo-1]
• B - partitions=[foo-3, foo-4]
• C - partitions=[foo-2, foo-5]
• Member Assignment
• A - epoch=22, partitions=[foo-0, foo-1]
• B - epoch=22, partitions=[foo-3, foo-4]
• C - epoch=22, partitions=[foo-2, foo-5]

A fails to heartbeat. The group coordinator removes him after the session timeout expires and bump the group epoch.

• Group (epoch=23)
  • B
  • C
• Target Assignment (epoch=22)
  • A - partitions=[foo-0, foo-1]
  • B - partitions=[foo-3, foo-4]
  • C - partitions=[foo-2, foo-5]
• Member Assignment
  
  • B - epoch=22, partitions=[foo-3, foo-4]
  • C - epoch=22, partitions=[foo-2, foo-5]

The group coordinator computes the new target assignment.

• Group (epoch=23)
  • B
  • C
• Target Assignment (epoch=23)
  
  • B - partitions=[foo-3, foo-4, foo-0]
  • C - partitions=[foo-2, foo-5, foo-1]
• Member Assignment
  
  • B - epoch=22, partitions=[foo-3, foo-4]
  • C - epoch=22, partitions=[foo-2, foo-5]

B and C heartbeat and transition to epoch 23.

• Group (epoch=23)
  • B
  • C
• Target Assignment (epoch=23)
  
  • B - partitions=[foo-3, foo-4, foo-0]
  • C - partitions=[foo-2, foo-5, foo-1]
• Member Assignment
  
  • B - epoch=23, partitions=[foo-3, foo-4, foo-0]
  • C - epoch=23, partitions=[foo-2, foo-5, foo-1]

Partition Added

Let's start with a group with two members and one partition.

• Group (epoch=22)
• A
• B
• Target Assignment (epoch=22)
• A - partitions=[foo-0]
• B - partitions=[]
• Member Assignment
• A - epoch=22, partitions=[foo-0]
• B - epoch=22, partitions=[]

A new partition foo-1 is created. The group coordinator detects it. It updates the group and bump the group epoch.

• Group (epoch=23)
• A
• B
• Target Assignment (epoch=22)
• A - partitions=[foo-0]
• B - partitions=[]
• Member Assignment
• A - epoch=22, partitions=[foo-0]
• B - epoch=22, partitions=[]

The group coordinator computes a new target assignment.

• Group (epoch=23)
• A
• B
• Target Assignment (epoch=23)
• A - partitions=[foo-0]
• B - partitions=[foo-1]
• Member Assignment
• A - epoch=22, partitions=[foo-0]
• B - epoch=22, partitions=[]

B and C heartbeat and transition to epoch 23.

• Group (epoch=23)
• A
• B
• Target Assignment (epoch=23)
• A - partitions=[foo-0]
• B - partitions=[foo-1]
• Member Assignment
• A - epoch=23, partitions=[foo-0]
• B - epoch=23, partitions=[foo-1]

Compatibility, Deprecation, and Migration Plan

The change is mainly backward compatible. The current protocol is still supported without any changes. The consumers and streams should be able to upgrade to the new protocol without any issues. The only part that may require adaptations is the regex based subscriptions. At the moment, those are validated on the client side based and that is language specific. We plan to use Google RE2/J on the server side so clients may have to adapt their regex based subscriptions when they migrate to the new protocol. We believe that the transition should be frictionless for most of the users.

The migration requires to first upgrades the servers to the new software and enable the new group coordinator. A roll is required for this. Then, the consumers must be upgraded to the new software as well and the new protocol must be enabled. This can be done with on roll if the consumer version is supported by the upgrade path. If not, two rolls are required.

We plan to release the feature in a Kafka 4.x release, possibly 4.0. The feature will be opt-in to start. We plan to make it the default in Kafka 5.0.

We don't plan to deprecate the current rebalance protocol anytime soon because Connect and others still rely on it. However, we will deprecate the consumer embedded protocol used in the current protocol in Kafka 5.0 and to remove it in Kafka 6.0. Consumers and Streams won't be able to use it any longer after this.

Test Plan

Our primary method for testing the implementation will be through Discrete Event Simulation (DES). DES allows us to test a large number of deterministically generated random scenarios which include various kinds of faults (such as network partitions). It allows us to define system invariants programmatically which are then checked after each step in the simulation. The protocol will be formally verified with a TLA+ model as well. Other than that, we will use the typical suite of unit/integration/system tests. System tests will be parameterised to run with both protocols.

Rejected Alternatives

An epoch per partition

We started this design by using an epoch per partition instead of relying on an epoch per member. Moving a partition from A to B would have require the following step: 1) revoke the partition from A; 2) bump the partition epoch; and 3) assign the partition to B. While this was very appealing at first, it was unpractical in the end for two reasons: 1) migrating from the current protocol is much more difficult without a member id; and 2) the metadata associated to the assignment (e.g. Streams metadata) is not tight to a particular partitions. We ended up using a member epoch with an incremental reassignment algorithm which is pretty close to this.

Not reusing the current coordinator

We considered not reusing the current group coordinator. Instead, the idea was to implement a brand new consumer group coordinator dedicated for the new rebalance protocol. The main benefits of this is that we could have moved away from the __consumer_offsets storage and use something more appropriate, perhaps closer to the KRaft metadata topic. This was rejected because migrating from a generic group to a consumer group would have been much more difficult.

No more client-side assignors, even for Kafka Streams

We considered removing the client-side assignor feature. From a consumer perspective, this is rarely used nowadays. Kafka Streams is its primary user so we thought about using a server side assignor in this case as well. We abandoned this for two reasons: 1) the Streams' assignor needs to know the entire Streams's topology so each member would have had to send it out to the server. The topology could be rather big (in MB) so this is not very practical; and 2) That would have introduced a strong dependency between the server version and the Streams version. Using new features in Streams would not be possible without upgrading the servers first. 

Future Work

Eventually, we aim at deprecating the current membership/rebalance API. In order to get to this point, we would need to first move all the use cases away from it.

Connect Group/Rebalance Protocol

Kafka Connect is the second protocol type which is currently supported by Apache Kafka. We propose to use a similar approach that the one used by the current proposal for Connect in the future. We would introduce a new connect group type and introduce a new set of APIs for Connect. The rebalance protocol is very similar to the consumer rebalance protocol but works with different resource types.

Membership/Leader Election API

The group membership protocol is also used outside of Apache Kafka. For instance, the Confluent Schema Registry uses it for leader election. It is not clear whether we really want to suppose such cases in the future. If we do, we could also define a new set of APIs for it. That would be much cleaner in the long run.