A Guide To The Kafka Protocol

Introduction

This document covers the protocol implemented in Kafka 0.8. 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 in a client. It is aimed at making it easier to implement a client. This document assumes you understand the basic design and terminology described here.

Preliminaries

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

Fixed Width Primitives

int8, int16, int32, int64, uint8, uint16, uint32, uint64 - Integers with the given precision (in bits) stored in big endian order. unitXX variants are unsigned and have double the range.

Variable Length Primitives

bytes16, bytes32, string16 - These types consist of a signed integer giving a length N followed by N bytes of content. -1 indicates null. bytes16 and string16 use a two byte int16 size and bytes32 uses a four byte int32 size. string16 is identical in format to bytes16 but the bytes should be interpreted as UTF8 encoded characters.

Arrays

It is often useful to repeat some structure multiple times. This will always be encoded as a unit size containing the length N followed by N repetitions of the structure. In the BNF grammars below we will show an array of a structure foo as [foo].

Client Requests

The Kafka protocol is extremely 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 great detail below.

Request Response Structure

There are a number of common fields shared by many or all requests. I will repeat these in each BNF, but only describe their usage here:

RequestOrResponse => Size (Request | Response)


Request => RequestId VersionId ClientId RequestMessage
Response => VersionId ResponseMessage



Size => uint32
RequestId => uint16
VersionId => uint16
ClientId => string
RequestMessage => MetadataRequest | ProduceRequest | FetchRequest | OffsetRequest
ResponseMessage => MetadataResponse | ProduceResponse | FetchResponse | OffsetResponse

The above BNF gives the grammar for all requests and responses.

Notes on reading the BNF

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 inclosed in parenthesis for grouping.

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.

MessageSet => [Offset MessageSize Message]


Offset => int64
MessageSize => int32


Message => Crc MagicByte Attributes Key Value


Crc => int32
MagicByte => int8
Attributes => int8
Key => bytes32
Value => bytes32

Metadata API

This API answers the question "who has what data and where are they?". Specifically this request will tell for each topic how many partitions it has, which brokers currently host each of these partitions, and which of these is the master. Since Kafka is a partitioned system requests need to be directed to the appropriate server--the one currently acting as the master for the partition you want to interact with. Since cluster membership in Kafka is dynamic, you can't just give all the clients a config file with all the brokers (some of them may be down, or partitions may have moved); instead you need to ask the cluster about its current state at run time. Hence the first thing a client needs to do when it connects is ask, "where is everyone?" using this metadata API.

This is the only request that can be made to any server without regard to partition ownership and all servers will give the same answer (disregarding timing differences). Fetch and produce requests always interact with particular partitions, and sending these to the wrong broker will result in an invalid metadata error. The client is expected to cache the cluster metadata locally, using it to direct requests to the correct hosts, until it gets an invalid metadata error or can't reach a particular broker, at which point it should fetch the metadata again and update its cache.

This presents a bit of a catch-22, since the only way to find out which Kafka servers exists is to ask a Kafka server, so how can a client ever connect the first time? To do this a client should take a "bootstrap urls" configuration from which it can find out the list of currently available servers. Importantly this need not contain all the servers in the cluster, maybe just two or three for redundancy. The client should try each of these until it finds one it can connect to. This will ensure that even if one of the bootstrap servers is down the client can still fetch the cluster metadata.

For deployment you may not want to hardcode such a list and may prefer to rely on dns or a VIP or something like that to find a bootstap server.

So the lifecycle of most clients looks something like this:

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.

Metadata Request And Response Format

The client sends a request for metadata. 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.

MetadataRequest => [TopicName]
MetadataResponse => [Metadata]

TopicName => string
Metadata => TopicName [PartitionMetadata] TopicErrorCode
PartitionMetadata => PartitionId Leader [Replicas] [Isr] PartitionErrorCode
Replica => Broker
Isr => Broker
Broker => NodeId CreatorId Host Port
NodeId => int32
CreatorId => string
Host => string
Port => int32
VersionId => int16
ClientId => string

Notes: Replicas should contain the broker information for the relevant brokers. The leader and isr should just be ids.

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 this field in any way it likes.

ProduceRequest => RequiredAcks Timeout TopicData

ProduceResponse => [TopicProduceResponse]


TopicResponse => TopicName [PartitionProduceResponse]
PartitionProduceResponse => Partition ErrorCode Offset


RequiredAcks => int16
Timeout => int32


TopicData => TopicName [PartitionData]
TopicName => string


PartitionData => Partition MessageSetSize MessageSet
Partition => int32
MessageSetSize => 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.

FetchRequest => ReplicaId MaxWaitTime MinBytes [TopicFetchRequest]


ReplicaId => int32
MaxWaitTime => int32
MinBytes => int32
TopicFetchRequest => TopicName [PartitionFetchRequest]
TopicName => string
PartitionFetchResponse => Partition FetchOffset MaxBytes
FetchOffset => int64


FetchResponse => [TopicData]
TopicData => Partition [PartitionData]
PartitionData => ErrorCode FetchedOffset HighwaterMarkOffset MessageSetSize MessageSet
ErrorCode => int16
FetchedOffset => uint64
HighwaterMarkOffset => int64
MessageSetSize => int32

Offset API

This API describes the valid offsets 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.

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


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

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:

Interaction With The Server

Interaction with the server uses request response pairs. 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.

The client will likely need to maintain a connection to multiple brokers, however it should not be necessary to maintain multiple connections to a single broker from a single client instance (i.e. connection pooling).

The producer is responsible for determining the partitioning of data sent, the brokers will check that they are leader for the partition of data sent to them however they will not attempt to determine if this the the "right" partition for data to be sent to as that is purely a client concern. Most clients will want to partition by the message key, but the exact nature of this partitioning is purely a client concern. In the scala client we have made it pluggable via a partitioner interface that controls the mapping of key to partition.

As indicated above the client must address its requests to the correct broker. The metadata request can be issued against any server and using the resulting cluster metadata the client should direct requests to the leader for the partition it wants to produce to or fetch from. If the client gets an error it should refresh its metadata.

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

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 different.

A final question is why we don't use a system like Protocol Buffers or Thrift to define our 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 compatibilit