具体的协议看原文，
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 persistent 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.
The broker's request processing allows only a single in-flight request per connection in order to guarantee this ordering. Note that clients can (and ideally should) use non-blocking IO to implement request pipelining and achieve higher throughput. i.e., clients can send requests even while awaiting responses for preceding requests since the outstanding requests will be buffered in the underlying OS socket buffer. All requests are initiated by the client, and result in a corresponding response message from the server except where noted.
The server has a configurable maximum limit on request size and any request that exceeds this limit will result in the socket being disconnected.
在TCP上直接实现的二进制协议
所有的Apis都是由request，reponse对来构成
client对每个需要通信的broker建立连接，但是没有必要对同一brokers建立多条链接
Broker server可以保证在同一个TCP链接中的数据是保序处理的，因为broker同时只会处理来自某个connection的一条message
broker server可以配置最大的request size，如果request超过这个limit，就会被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.
- Cycle through a list of "bootstrap" kafka urls until we find one we can connect to. Fetch cluster metadata.
- Process fetch or produce requests, directing them to the appropriate broker based on the topic/partitions they send to or fetch from.
- If we get an appropriate error, refresh the metadata and try again.
kafka是基于数据partition到多个broker上的，如果去做partition取决于client的逻辑和策略；
所以client需要知道topic和partition的metadata，其实从任意一个broker都可以得到这个信息，但是为了防止某台kafka down或者load balance，我们会配置多个broker地址
client在获取metadata后，会local cache，只有当下次读写失败，才需要去再次更新metadata
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:
- It balances data and request load over brokers
- It serves as a way to divvy 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.
partition策略，可以round robin，这样的流量比较平均，但是会需要对所有的broker建链接
当，producer数远远大于brokers的时候，也可以一个producer随机选一个partition写，这样的好处是只需要建立一条连接，但是问题是处理不好，会导致数据的倾斜；
当然，你可以定义策略，比如按照key划分partition，来满足业务需求，因为对于同一个partition是可以保序的，比如按用户id，但这样的问题也是，容易导致数据倾斜
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 implementer can choose to ignore this and send everything one at a time if they like.
支持batch读写
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.
为了支持版本演进和向后兼容，对于api，提供版本号
Retrieving Supported API versions
In order for a client to successfully talk to a broker, it must use request versions supported by the broker.
Clients may work against multiple broker versions, however to do so the clients need to know what versions of various APIs a broker supports.
Starting from 0.10.0.0, brokers provide information on various versions of APIs they support. Details of this new capability can be found here. Clients may use the supported API versions information to take appropriate actions such as propagating an unsupported API version error to application or choose an API request/response version supported by both the client and broker. The following sequence maybe used by a client to obtain supported API versions from a broker.
SASL Authentication Sequence
The following sequence is used for SASL authentication:
- Kafka
ApiVersionsRequestmay be sent by the client to obtain the version ranges of requests supported by the broker. This is optional. - Kafka
SaslHandshakeRequestcontaining the SASL mechanism for authentication is sent by the client. If the requested mechanism is not enabled in the server, the server responds with the list of supported mechanisms and closes the client connection. If the mechanism is enabled in the server, the server sends a successful response and continues with SASL authentication. - The actual SASL authentication is now performed. A series of SASL client and server tokens corresponding to the mechanism are sent as opaque packets. These packets contain a 32-bit size followed by the token as defined by the protocol for the SASL mechanism.
- If authentication succeeds, subsequent packets are handled as Kafka API requests. Otherwise, the client connection is closed.
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.
为什么不用Http？想利用TCP的特性，比如多路request，或同步poll多个connection
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.
为什么不支持多个不同的协议？对用户没有意义，因为用户往往是用现成的库
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.
为什么不用现成的XMPP, STOMP, AMQP 协议？因为不想被这些协议所束缚
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.
为什么不用Protocol Buffers或Thrift来做序列化？因为messages的类型不多，不需要因为更多依赖
