A Guide To The Kafka Protocol

This is a draft document. We hope to turn it into a useful guide to implementing a Kafka client, so any feedback on how to make it more useful would be much appreciated. The 0.8 release has not yet happened so there may yet be changes in the protocol described, though we hope to minimize them. We feel this is stable enough that people can build against it without major disruption.

Introduction

This document covers the protocol implemented in Kafka 0.8 and beyond. It is meant to give a readable guide to the protocol that covers the available requests, their binary format, and the proper way to make use of them to implement a client. This document assumes you understand the basic design and terminology described here.

The protocol used in 0.7 and earlier is similar to this, but we chose to make a one time (we hope) break in compatibility to be able to clean up cruft and generalize things.

Overview

The Kafka protocol is fairly simple, there are only four client requests APIs.

1. Metadata - Describes the currently available brokers, their host and port information, and gives information about which broker hosts which partitions.
2. Send - Send messages to a broker
3. Fetch - Fetch messages from a broker, one which fetches data, one which gets cluster metadata, and one which gets offset information about a topic.
4. Offsets - Get information about the available offsets for a given topic partition.

Each of these will be described in detail below.

Preliminaries

Network

Kafka uses a binary protocol over TCP. The protocol defines all apis as request response message pairs. All messages are size delimited and are made up of the following primitive types.

The client initiates a socket connection and then writes a sequence of request messages and reads back the corresponding response message. No handshake is required on connection or disconnection. TCP is happier if you maintain persistant connections used for many requests to amortize the cost of the TCP handshake, but beyond this penalty connecting is pretty cheap.

The client will likely need to maintain a connection to multiple brokers, as data is partitioned and the clients will need to talk to the server that has their data. However it should not generally be necessary to maintain multiple connections to a single broker from a single client instance (i.e. connection pooling).

The server guarantees that on a single TCP connection, requests will be processed in the order they are sent and responses will return in that order as well. All requests are initiated by the client and result in a corresponding response message from the server.

The server has a configurable maximum limit on request size and any request that exceeds this limit will result in the socket being disconnected.

Partitioning and bootstrapping

Kafka is a partitioned system so not all servers have the complete data set. Instead recall that topics are split into a pre-defined number of partitions, P, and each partition is replicated with some replication factor, N. Topic partitions themselves are just ordered "commit logs" numbered 0, 1, ..., P.

All systems of this nature have the question of how a particular piece of data is assigned to a particular partition. Kafka clients directly control this assignment, the brokers themselves enforce no particular semantics of which messages should be published to a particular partition. Rather, to publish messages the client directly addresses messages to a particular partition, and when fetching messages, fetches from a particular partition. If two clients want to use the same partitioning scheme they must use the same method to compute the mapping of key to partition.

These requests to publish or fetch data must be sent to the broker that is currently acting as the leader for a given partition. This condition is enforced by the broker, so a request for a particular partition to the wrong broker will result in an the NotLeaderForPartition error code (described below). 

How can the client find out which topics exist, what partitions they have, and which brokers currently host those partitions so that it can direct its requests to the right hosts? This information is dynamic, so you can't just configure each client with some static mapping file. Instead all Kafka brokers can answer a metadata request that describes the current state of the cluster: what topics there are, which partitions those topics have, which broker is the leader for those partitions, and the host and port information for these brokers.

In other words, the client needs to somehow find one broker and that broker will tell the client about all the other brokers that exist and what partitions they host. This first broker may itself go down so the best practice for a client implementation is to take a list of two or three urls to bootstrap from. The user can then choose to use a load balancer or just statically configure two or three of their kafka hosts in the clients.

The client does not need to keep polling to see if the cluster has changed; it can fetch metadata once when it is instantiated cache that metadata until it receives an error indicating that the metadata is out of date. This error can come in two forms: (1) a socket error indicating the client cannot communicate with a particular broker, (2) an error code in the response to a request indicating that this broker no longer hosts the partition for which data was requested.

1. Cycle through a list of "bootstrap" kafka urls until we find one we can connect to. Fetch cluster metadata.
2. Process fetch or produce requests, directing them to the appropriate broker based on the topic/partitions they send to or fetch from.
3. If we get an appropriate error, refresh the metadata and try again.

Partitioning Strategies

As mentioned above the assignment of messages to partitions is something the producing client controls. That said, how should this functionality be exposed to the end-user?

Partitioning really serves two purposes in Kafka:

1. It balances data and request load over brokers
2. It serves as a way to divy up processing among consumer processes while allowing local state and preserving order within the partition. We call this semantic partitioning.

For a given use case you may care about only one of these or both.

To accomplish simple load balancing a simple approach would be for the client to just round robin requests over all brokers. Another alternative, in an environment where there are many more producers than brokers, would be to have each client chose a single partition at random and publish to that. This later strategy will result in far fewer TCP connections.

Semantic partitioning means using some key in the message to assign messages to partitions. For example if you were processing a click message stream you might want to partition the stream by the user id so that all data for a particular user would go to a single consumer. To accomplish this the client can take a key associated with the message and use some hash of this key to choose the partition to which to deliver the message.

Batching

Our apis encourage batching small things together for efficiency. We have found this is a very significant performance win. Both our API to send messages and our API to fetch messages always work with a sequence of messages not a single message to encourage this. A clever client can make use of this and support an "asynchronous" mode in which it batches together messages sent individually and sends them in larger clumps. We go even further with this and allow the batching across multiple topics and partitions, so a produce request may contain data to append to many partitions and a fetch request may pull data from many partitions all at once.

The client implementor can choose to ignore this and send everything one at a time if they like.

Versioning and Compatibility

The protocol is designed to enable incremental evolution in a backward compatible fashion. Our versioning is on a per-api basis, each version consisting of a request and response pair. Each request contains an API key that identifies the API being invoked and a version number that indicates the format of the request and the expected format of the response.

The intention is that clients would implement a particular version of the protocol, and indicate this version in their requests. Our goal is primarily to allow API evolution in an environment where downtime is not allowed and clients and servers cannot all be changed at once.

The server will reject requests with a version it does not support, and will always respond to the client with exactly the protocol format it expects based on the version it included in its request. The intended upgrade path is that new features would first be rolled out on the server (with the older clients not making use of them) and then as newer clients are deployed these new features would gradually be taken advantage of.

Currently all versions are baselined at 0, as we evolve these APIs we will indicate the format for each version individually.

The Protocol

Protocol Primitive Types

The protocol is built out of the following primitive types.

Fixed Width Primitives

int8, int16, int32, int64 - Signed integers with the given precision (in bits) stored in big endian order.

Variable Length Primitives

bytes, string - These types consist of a signed integer giving a length N followed by N bytes of content. A length of -1 indicates null. string uses an int16 for its size, and bytes uses an int32.

Arrays

This is a notation for handling repeated structures. These will always be encoded as an int32 size containing the length N followed by N repetitions of the structure which can itself be made up of other primitive types. In the BNF grammars below we will show an array of a structure foo as [foo].

Notes on reading the request format grammars

The BNFs below give an exact context free grammar for the request and response binary format. For each API I will give the request and response together followed by all the sub-definitions. The BNF is intentionally not compact in order to give human-readable name (for example I define a production for ErrorCode even though it is just an int16 in order to give it a symbolic name). As always in a BNF a sequence of productions indicates concatenation, so the MetadataRequest given below would be a sequence of bytes containing first a VersionId, then a ClientId, and then an array of TopicNames (each of which has its own definition). Productions are always given in camel case and primitive types in lower case. When there are multiple possible productions these are separated with '|' and may be enclosed in parenthesis for grouping. The top-level definition is always given first and subsequent sub-parts are indented.

Common Request and Response Structure

All requests and responses originate from the following grammar which will be incrementally describe through the rest of this document:

RequestOrResponse => Size (RequestMessage | ResponseMessage)
  Size => int32

Requests

Requests all have the following format:

RequestMessage => ApiKey ApiVersion CorrelationId ClientId RequestMessage
  ApiKey => int16
  ApiVersion => int16
  CorrelationId => int32
  ClientId => string
  RequestMessage => MetadataRequest | ProduceRequest | FetchRequest | OffsetRequest

The various request and response messages will be described below.

Responses

Response => CorrelationId ResponseMessage
CorrelationId => int32
ResponseMessage => MetadataResponse | ProduceResponse | FetchResponse | OffsetResponse

The response will always match the paired request (e.g. we will send a MetadataResponse in return to a MetadataRequest).

Message sets

One structure common to both the produce and fetch requests is the message set format. A message in kafka is a key-value pair with a small amount of associated metadata. A message set is just a sequence of messages with offset and size information. This format happens to be used both for the on-disk storage on the broker and the on-the-wire format.

A message set is also the unit of compression in Kafka, and we allow messages to recursively contain compressed message sets to allow batch compression.

N.B., MessageSets are not preceded by an int32 like other array elements in the protocol.

MessageSet => [Offset MessageSize Message]
  Offset => int64
  MessageSize => int32

Message format

Message => Crc MagicByte Attributes Key Value
  Crc => int32
  MagicByte => int8
  Attributes => int8
  Key => bytes
  Value => bytes

Compression

Kafka supports batch compression of a message set. Batch compression is important as individual messages may not have sufficient redundancy to enable good compression rations. Batch compression is supported by allowing the value field of a message to hold a compressed message set in the format defined above with the attributes field of the message set to indicate the compression codec used.

Kafka currently supports two compression codecs for message sets with the following codec numbers:

The APIs

This section gives details on each of the individual APIs, their usage, their binary format, and the meaning of their fields.

Metadata API

This API answers the following questions:

• What topics exist?
• How many partitions does each topic have?
• Which broker is currently the leader for each partition?
• What is the host and port for each of these brokers

This is the only request that can be addressed to any broker in the cluster.

Since there may be many topics the client can give an optional list of topic names in order to only return metadata for a subset of topics.

The metdata returned is at the partition level, but grouped together by topic for convenience and to avoid redundancy. For each partition the metadata contains the information for the leader as well as for all the replicas and the list of replicas that are currently in-sync.

Metadata Request

MetadataRequest => [TopicName]
  TopicName => string

Metadata Response

The response contains metadata for each partition, with partitions grouped together by topic. This metadata refers to brokers by their broker id. The brokers each have a host and port.

MetadataResponse => [Broker][TopicMetadata]
  Broker => NodeId Host Port
  NodeId => int32
  Host => string
  Port => int32
  TopicMetadata => TopicErrorCode TopicName [PartitionMetadata]
  PartitionMetadata => PartitionErrorCode PartitionId Leader Replicas Isr
  PartitionErrorCode => int16
  PartitionId => int32
  Leader => int32
  Replicas => [int32]
  Isr => [int32]

Produce API

The produce API is used to send message sets to the server. For efficiency it allows sending message sets intended for many topic partitions in a single request.

The produce API uses the generic message set format, but since no offset has been assigned to the messages at the time of the send the producer is free to fill in that field in any way it likes.

Produce Request

ProduceRequest => RequiredAcks Timeout [TopicName [Partition MessageSetSize MessageSet]]
  RequiredAcks => int16
  Timeout => int32
  Partition => int32
  MessageSetSize => int32

Produce Response

ProduceResponse => [TopicName [Partition ErrorCode Offset]]
  TopicName => string
  Partition => int32
  ErrorCode => int16
  Offset => int64

Fetch API

The fetch API is used to fetch a chunk of one or more logs for some topic-partitions. Logically one specifies the topics, partitions, and starting offset at which to begin the fetch and gets back a chunk of messages.

Fetch requests follow a long poll model so they can be made to block for a period of time if sufficient data is not immediately available.

One thing to note is that the fetch API requires specifying the partition to consume from. The question is how should a consumer know what partitions to consume from? In particular how can you balance the partitions over a set of consumers acting as a group so that each consumer gets a subset of partitions. We have done this assignment dynamically using zookeeper for the scala and java client. The downside of this approach is that it requires a fairly fat client and a zookeeper connection. We haven't yet created a Kafka API to allow this functionality to be moved to the server side and accessed more conveniently. A simple consumer client can be implemented by simply requiring that the partitions be specified in config, though this will not allow dynamic reassignment of partitions should that consumer fail. We hope to address this gap in the next major release.

Fetch Request

FetchRequest => ReplicaId MaxWaitTime MinBytes [TopicName [Partition FetchOffset MaxBytes]]
  ReplicaId => int32
  MaxWaitTime => int32
  MinBytes => int32
  TopicName => string
  Partition => int32
  FetchOffset => int64
  MaxBytes => int32

Fetch Response

FetchResponse => [TopicName [Partition ErrorCode FetchedOffset HighwaterMarkOffset MessageSetSize MessageSet]]
  TopicName => string
  Partition => int32
  ErrorCode => int16
  HighwaterMarkOffset => int64
  MessageSetSize => int32

Offset API

This API describes the valid offset range available for a set of topic-partitions. As with the produce and fetch APIs requests must be directed to the broker that is currently the leader for the partitions in question. This can be determined using the metadata API.

The response contains the starting offset of each segment for the requested partition as well as the "log end offset" i.e. the offset of the next message that would be appended to the given partition.

We agree that this API is slightly funky.

Offset Request

OffsetRequest => ReplicaId [TopicName [Partition Time MaxNumberOfOffsets]]
  ReplicaId => int32
  TopicName => string
  Partition => int32
  Time => int64
  MaxNumberOfOffsets => int32

Offset Response

OffsetResponse => [TopicName [PartitionOffsets]]
  PartitionOffsets => Partition ErrorCode [Offset]
  Partition => int32
  ErrorCode => int16
  Offset => int64

Constants

Api Keys

The following are the numeric codes that the ApiKey in the request can take for each of the above request types.

Error Codes

We use numeric codes to indicate what problem occurred on the server. These can be translated by the client into exceptions or whatever the appropriate error handling mechanism in the client language. Here is a table of the error codes currently in use:

Some Common Philosophical Questions

Some people have asked why we don't use HTTP. There are a number of reasons, the best is that client implementors can make use of some of the more advanced TCP features--the ability to multiplex requests, the ability to simultaneously poll many connections, etc. We have also found HTTP libraries in many languages to be surprisingly shabby.

Others have asked if maybe we shouldn't support many different protocols. Prior experience with this was that it makes it very hard to add and test new features if they have to be ported across many protocol implementations. Our feeling is that most users don't really see multiple protocols as a feature, they just want a good reliable client in the language of their choice.

Another question is why we don't adopt XMPP, STOMP, AMQP or an existing protocol. The answer to this varies by protocol, but in general the problem is that the protocol does determine large parts of the implementation and we couldn't do what we are doing if we didn't have control over the protocol. Our belief is that it is possible to do better than existing messaging systems have in providing a truly distributed messaging system, and to do this we need to build something that works differently.

A final question is why we don't use a system like Protocol Buffers or Thrift to define our request messages. These packages excel at helping you to managing lots and lots of serialized messages. However we have only a few messages. Support across languages is somewhat spotty (depending on the package). Finally the mapping between binary log format and wire protocol is something we manage somewhat carefully and this would not be possible with these systems. Finally we prefer the style of versioning APIs explicitly and checking this to inferring new values as nulls as it allows more nuanced control of compatibility.