Status

Current state:

Discussion thread: 

JIRA:

Released: 

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

Motivation

Apache Kafka is in the process of moving from storing metadata in Apache Zookeeper, to storing metadata in an internal Raft topic.  KIP-500 described the overall architecture and plan.  The purpose of this KIP is to go into detail about how the Kafka Controller will change during this transition.

Proposed Changes

Deployment

KIP-500 Mode

Once this KIP is implemented, system administrators will have the option of running in KIP-500 mode.  In this mode, we will not use ZooKeeper  The alternative mode where KIP-500 support is not enabled will be referred to as legacy mode.

KIP-500 mode must be enabled for the entire cluster, not just for specific nodes.  Initially, this mode will be considered experimental and not ready for production.  As we do more testing and gain more confidence, we will remove the experimental label.  Eventually, in a future release, KIP-500 mode will be the only supported mode.  Since dropping support for legacy mode is an incompatible change, it will need to happen in a major release, of course.

Initially, we will not support upgrading a cluster from legacy mode to KIP-500 mode.  This is in keeping with the experimental nature of KIP-500 mode.  A future KIP will describe and implement an upgrade process from legacy mode to KIP-500 mode.

Nodes

Currently, a ZooKeeper cluster must be deployed when running Kafka.  This KIP will eliminate that requirement, as well as the requirement to configure the addresses of the zookeeper nodes on each broker.

Currently, any broker node can be elected as the controller.  As part of this KIP, the active controller will instead be selected among a small pool of nodes specifically configured to act as controllers.  Typically three or five nodes in the cluster will be selected to be controllers.

System administrators will be able to choose whether to run separate controller nodes, or whether to run controller nodes which are co-located with broker nodes.  Kafka will provide support for running a controller in the same JVM as a broker, in order to save memory and enable single-process test deployments.

The addresses and ports of the controller nodes must be configured on each broker, so that the broker can contact the controller quorum when starting up.  This is similar to how we configure the ZooKeeper quorum on each node today.

Note that as long as at least one of the provided controller addresses is valid, the broker will be able to learn about the current metadata quorum and start up.  Once the broker is in contact with the metadata quorum, the quorum bootstrap addresses will not be needed.  This makes it possible to reconfigure the metadata quorum over time.  For example, if we start with a metadata quorum of host1, host2, host3, we could replace host3 with host4 without disrupting any of the brokers.  Then we could roll the brokers to apply the new metadata quorum bootstrap configuration of host1, host2, host4 on each one.

Node IDs

Just like brokers, controller nodes will have non-negative integer node IDs.  There will be a single ID space.  In other words, no controller should share the same ID as a broker.  Even when a broker and a controller are co-located in the same JVM, they must have different node IDs.

Automatic node ID assignment via ZooKeeper will no longer be supported in KIP-500 mode.  Node IDs must be set in the configuration file for brokers and controllers.

Networking

Controller processes will listen on a separate port from brokers.  This will be true even when the broker and controller are co-located in the same JVM.

In a well-run Kafka deployment, controller ports, like ZooKeeper ports, should be firewalled off from clients.  This will prevent clients from disrupting the cluster by flooding the controller ports with requests.  In the realm of ACLs, this translates to controllers requiring CLUSTERACTION on CLUSTER for all operations.

The only time when clients should contact a controller node directly is when they are debugging system issues.  This is similar to ZooKeeper, where we have things like zk-shell, but only for debugging.

The Controller Quorum

The Metadata Topic

As described in KIP-500, the controller will store its data in the internal __metadata topic.  This topic will be managed by Raft on the controller nodes.  The leader of the controller quorum will be the active controller.  The followers will function as hot standbys, ready to take over when the active leader fails or resigns.  The metadata will be stored in memory on all the active controllers.

 Metadata Persistence and Visibility

Metadata changes need to be persisted to the __metadata log before we propagate them to the other nodes in the cluster.  This means waiting for the metadata log's last stable offset to advance to the offset of the change.  After that point, we are guaranteed not to lose the change as long as we uphold the Raft invariants.

Changes that we haven't yet persisted are referred to as "uncommitted."  The controller may have several of these uncommitted changes in flight at any given time.  In essence, the controller's in-memory state is always a little bit in the future compared to the current state.  This allows the controller to continue doing things while it waits for the previous changes to be committed to the Raft log.

However, this "future state" may never be committed.  For example, the active controller might fail, truncating some of its future state.  Therefore, the controller must not make this future state "visible" to the rest of the cluster until it has been made persistent – that is, until it becomes current state.  In the case of the __metadata topic, the replication protocol itself neatly takes care of this for us.  In the case of controller RPCs like AlterIsr, the controller handles this by not sending back a response until the designated change has been persisted.

                      earlier offsets
                          +---+
     visible change ----> |   |
                          +---+       last stable offset
                          |   | <---- standby controllers'
                          +---+       in-memory state
     pending change ----> |   |
                          +---+
     pending change ----> |   |
                          +---+
     pending change ----> |   |
                          +---+
                          |   | <---- active controller's
                          +---+       in-memory state
                      latest offset

Metadata

The __metadata topic will have a single partition which contains all the metadata in the cluster.  This topic is managed by Raft, and has periodic KIP-630 snapshots.

The format of the data in the snapshots is different from the format in the log entries. The log entries are essentially deltas that must be applied to the preceding snapshot.  We call the log entries "edits."  Using edits makes the metadata much smaller than it would otherwise be.  For example, we can represent a change in the ISR of a partition without restating all the other details about the partition.

Snapshots

Clearly, as we keep creating metadata edits, the metadata partition will grow without bound, even if the size of the actual metadata stays the same.  To avoid this, we will periodically write out our metadata state into a snapshot.

The implementation of Raft snapshots is described at the protocol level in KIP-630.  For the purpose of this KIP, the important thing is to understand the format of what the controller writes out, as well as how it is generated.

Snapshots are local to each replica.  For example, replica A may have a snapshot at offset 100, and edits up to offset 150, whereas replica B may have a snapshot at 125 and edits up to offset 150.  Any snapshot must be usable as a starting point for loading the entire state of metadata.  In other words, a new controller node must be able to load the a snapshot, and then apply all the edits which follow it, and come up-to-date.

The currently active controller will monitor the offset of the latest snapshot made by all replicas, including itself.  The snapshotting state of each node is considered soft state: it is not persisted anywhere in the log, but purely communicated by heartbeats and stored in memory by the active controller.

When the controller feels that a remote node should start a snapshot, it will communicate that information in its response to the periodic heartbeat sent by that node.  When the controller feels like it itself should create a snapshot, it will first try to give up the leadership of the Raft quorum.  The active controller will not tell a node to begin snapshotting if it is aware that another node is also snapshotting.

The controller will also snapshot less frequently when too many members of the quorum have fallen behind.  Specifically, if losing a node would probably impact availability, we will use a separate set of configurations for determining when to snapshot.

Broker Heartbeats

Every distributed system needs a way of managing cluster membership.  Prior to KIP-500, Kafka brokers registered ephemeral znodes in order to register themselves as part of the cluster.  The Kafka controller passively consumed the registration information from Zookeeper.

In the post-KIP-500 world there is no ZooKeeper and no ephemeral znodes.  Instead, each broker sends a ControllerHeartbeat request to the active controller every few seconds.

This heartbeat acts as a registration.  However, the controller has a choice about whether to accept it.  It will reject brokers whose metadata is too stale, or whose IDs have been claimed by another broker.  It will also reject brokers that do not support the minimum feature level of all KIP-584 features that are enabled.

When the broker accepts the registration, it grants or renews a broker ID lease associating the broker process with its ID.  Leases are time-bounded.  A broker cannot continue using a lease indefinitely after sending a single heartbeat.  When brokers are rejected by the controller, or otherwise unable to renew their lease before it expires, they enter the "fenced" state.

Broker Fencing

Brokers that don't have a broker ID lease are said to be "fenced."  When a broker is fenced, it cannot process any client requests.  This prevents brokers which are not receiving metadata updates or that are not receiving and processing them fast enough from causing issues to clients.

Brokers start up in the fenced state, and can leave this state only by sending a heartbeat to the active controller and getting back a response that tells them they can become active.

Controlled Shutdown

In the pre-KIP-500 world, brokers triggered a controller shutdown by making an RPC to the controller.  When the controller returned a successful result from this RPC, the broker knew that it could shut down.

In the post-KIP-500 world, controller shutdown is handled by the broker heartbeat system instead.  In its periodic heartbeats, the broker asks the controller if it can transition into the SHUTDOWN state.  This motivates the controller to move all of the leaders off of that broker.  Once they are all moved, the controller responds to the heartbeat with a nextState of SHUTDOWN.

Broker ID Conflicts

Clearly, in a correctly managed cluster, there should be no broker ID conflicts.  Each broker should be configured with a unique ID.  However, we want the system to be robust against misconfigurations.  Therefore, if there are two brokers that claim the same ID, the controller will choose only one and tell the other to fence itself.

When a broker first starts up, before it has received any responses from the controller, it will always "win" broker ID conflicts.  However, once it has communicated with the controller, it may lose subsequent conflicts if its broker epoch is stale.  The reason for favoring new processes is to accommodate the common case where a process is killed with kill -9 and then restarted.  We want it to be able to reclaim its old ID quickly in this case.

Broker State Machine

This state machine ties together a few of the previous sections, such as broker registration and shutdown.

                        INITIAL
sent initial heartbeart,  | 
got back an epoch         |
                          V
                        ACTIVE -------------------------> SHUTDOWN
lease expired or revoked  |  ^                          |
(stale metadata,          |  | lease restored           | controller granted
 id conflict,             |  | with new broker epoch    | controlled shutdown
 unable to communicate,   V  |                          |
 etc.)                  FENCED -------------------------+ 

Public Interfaces

Configurations

RPCs

Deprecating the Metadata Propagation RPCs

As discussed earlier, the new controller will use FetchRequest to fetch metadata from the active controller.  The details of how Raft fetching will work are spelled out in

KIP-595: A Raft Protocol for the Metadata Quorum.

Since we propagate the metadata via Raft, we will no longer need to send out LeaderAndIsrRequest, UpdateMetadataRequest, and StopReplicaRequest.  These requests will be sent out only when we're in legacy mode, not when we're in KIP-500 mode.  Eventually we will add some support for these requests to the new controller, in order to support rolling upgrade from a pre-KIP-500 release. However, for the purpose of this KIP, the new controller will not use these requests.

Deprecating the Controlled Shutdown RPC

The broker heartbeat mechanism replaces the controlled shutdown RPC.  Therefore, we will not need to support the this RPC any more in the controller-- except for compatibility during upgrades, which will be described further in a follow-on KIP.

BrokerHeartbeat

As described earlier, the broker periodically sends out a heartbeat request to the active controller.

{
  "apiKey": 50,
  "type": "request",
  "name": "BrokerHeartbeatRequest",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "TargetState", "type": "int8", "versions": "0+",
      "about": "The state that the broker wants to reach." },
    { "name": "BrokerId", "type": "int32", "versions": "0+",
      "about": "The broker ID." },
    { "name": "BrokerEpoch", "type": "int64", "versions": "0+", "default": "-1",
      "about": "The broker epoch, or -1 if one has not yet been assigned." },
    { "name": "LeaseStartTimeMs", "type": "int64", "versions": "0+",
      "about": "The time which the broker wants the lease to start at in milliseconds." },
    { "name": "CurMetadataOffset", "type": "int64", "versions": "0+",
      "about": "The highest metadata offset which the broker has reached." },
    { "name": "Listeners", "type": "[]Listener",
      "about": "The listeners of this broker", "versions": "0+", "fields": [
        { "name": "Name", "type": "string", "versions": "0+", "mapKey": true,
          "about": "The name of the endpoint." },
        { "name": "Host", "type": "string", "versions": "0+",
          "about": "The hostname." },
        { "name": "Port", "type": "int16", "versions": "0+",
          "about": "The port." },
        { "name": "SecurityProtocol", "type": "int16", "versions": "0+",
          "about": "The security protocol." }
      ]
    }
  ]
}

{
  "apiKey": 50,
  "type": "response",
  "name": "BrokerHeartbeatResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "ErrorCode", "type": "int16", "versions": "0+",
      "about": "The error code, or 0 if there was no error." },
    { "name": "ActiveControllerId", "type": "int32", "versions": "0+",
      "about": "The ID of the active controller, or -1 if the controller doesn't know." },
    { "name": "NextState", "type": "int8", "versions": "0+",
      "about": "The state to which the broker should transition." },
    { "name": "BrokerEpoch", "type": "int64", "versions": "0+", "default": "-1",
      "about": "The broker's assigned epoch, or -1 if none was assigned." },
    { "name": "LeaseEndTimeMs", "type": "int64", "versions": "0+",
      "about": "The time in milliseconds at which the lease should end. This is based on the start time that was passed, not the controller's local clock." }
  ]
}

enum BrokerState {
    UNKNOWN(0),
    INITIAL(1),
    FENCED(2),
    ACTIVE(3),
    SHUTDOWN(4);
}

As always with enums, the UNKNOWN state is used only to translate values that our software is too old to understand.

The controller will return NOT_CONTROLLER if it is not active.  Brokers will always return NOT_CONTROLLER for these RPCs.

ControllerHeartbeat

The controllers periodically send out a heartbeat request to the active controller.

{
  "apiKey": 51,
  "type": "request",
  "name": "ControllerHeartbeatRequest",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "ControllerId", "type": "int32", "versions": "0+",
      "about": "The controller ID." },
    { "name": "LastSnapshotOffset", "type": "int64", "versions": "0+",
      "about": "The offset of the last snapshot on this controller node." }
  ]
}

{
  "apiKey": 51,
  "type": "response",
  "name": "ControllerHeartbeatResponse",
  "validVersions": "0",
  "flexibleVersions": "0+",
  "fields": [
    { "name": "ErrorCode", "type": "int16", "versions": "0+",
      "about": "The error code, or 0 if there was no error." },
    { "name": "StartSnapshot", "type": "bool", "versions": "0+",
      "about": "True if we should start creating a snapshot." }
  ]
}

The controller will return NOT_CONTROLLER if it is not active.  Brokers will always return NOT_CONTROLLER for these RPCs.

Snapshot Format

Delta Format

Metrics

Each change that the controller makes will generate a metadata edit log entry, also known as an"edit."  These edits will be persisted to the metadata partition.

The format of each edit will be:

1. a varint specifying the edit type
2. a varint specifying the edit version
3. the payload in Kafka RPC format

If the controller tries to read an entry whose type or version is unknown, it will fail and abort the process.  The types and versions which the controller uses to write will be controlled by the KIP-584 feature flags which we have enabled.

It's important to note that in many cases, it will be better to add a tagged field than to bump the version number of an edit type.  A tagged field is appropriate any time older brokers can simply ignore a field that newer brokers might want to see.

Compatibility, Deprecation, and Migration Plan

Test Plan

Rejected Alternatives

Support Automatic Broker ID Assignment

This KIP proposes to drop support for automatic broker ID assignment.  What if we decided to continue to support it?

If we were willing to take a little bit more complexity on board, it would be relatively easy to support automatic broker ID assignment.  Brokers could simply ask the active controller to assign them a new ID when starting up, just as they previously obtained one from ZooKeeper.

However, automatic controller ID assignment is a much more difficult problem.  We never actually supported automatically assigning ZooKeeper IDs, so there is no pattern to follow here.  In general, Raft assumes that nodes know their IDs before the protocol begins. We cannot rely on random assignment because the 31 bit space is not large enough. We could perhaps create a separate protocol for assigning node IDs, but it might be complex.  

In general it's not clear how useful automatic broker ID assignment really is.  Configuration management software like Puppet, Chef, or Ansible can easily create a new ID for each node's configuration file.  Therefore, it's probably best to use this compatibility break to drop support for automatic broker ID assignment.

Combined Heartbeats and Fetch Requests

The brokers are always fetching new metadata from the controller.  Why not combine these fetch requests with the heartbeat requests, so that the brokers only have to send one request rather than two?

The main reason for making them separate requests is to have better separation of concerns.  Fetching metadata is logically a bit different than sending a heartbeat, and coupling them could result in a messy design and code.  We would have to add significant extra complexity to the FetchRequest schema.  Perhaps even worse, we would need to make the timing of fetch requests line up with the timing needed for broker heartbeats.