Apache Kafka Protocol in Haskell

This package covers the Apache Kafka protocol (version 0.8 and beyond) implementation, written in Haskell. This includes: types, parsing and serializing for the given APIs. On top of this, a client library is provided which can be used to build consumer and producer clients.

Current State

• Metadata API
  • Topic Metadata Request
  • Metadata Response
• Produce API
  • Produce Request
  • Produce Response
• Fetch API
  • Fetch Request
  • Fetch Response
• Offset API
  • Offset Request
  • Offset Response
• Offset Commit/Fetch API
  • Consumer Metadata Request
  • Consumer Metadata Response
  • Offset Commit Request
  • Offset Commit Response
  • Offset Fetch Request
  • Offset Fetch Response

Protocol Library

Types

We basically separate between three kind of types:

1. Related to request
2. Related to response
3. Related to data (for either request or response but can also be used for the Apache Kafka Log component since log files (on-disk) hold the same structure - which is awesome!)

The following table gives an overview about the naming convention used within this package as well as the different kind of data types available for requests and responses:

API	Request (Rq)	Response (Rs)
Metadata API (Md)	MetadataRequest	MetadataResponse
Produce API (Pr)	ProduceRequest	ProduceResponse
Fetch API (Ft)	FetchRequest	FetchResponse
Offset API (Of)	OffsetRequest	OffsetResponse
Offset Commit API (Ofc)	OffsetCommitRequest	OffsetCommitResponse
Offset Fetch API (Oft)	OffsetFetchRequest	OffsetFetchResponse

TODO: ConsumerMetadataRequest?!

Parser

As for binary serialization, we rely the Data.Binary library. Using the Get Monad (Data.Binary.Get) we are able to comfortablyparse ByteString in it's big endian network order and decode it to appropriate types.

Request

To determine the type of request, the Apache Kafka protocol describes an Api Key field which holds a numeric code to determine the type of request.

We thus provide a function readRequest which gives an IO RequestMessage that allows to handle the request based on the rqApiKey field:

requestMessage <- readRequest i
case rqApiKey requestMessage of
    0  -> handleProduceRequest (rqRequest requestMessage)
    1  -> handleFetchRequest (rqRequest requestMessage)
    ...

Response

TODO

Serializer

Obviously the serializer is provides the opposite functionalities as the parser and was separated just for clarity. Thus we also rely on Data.Binary and use the Put Monad (Data.Binary.Put) to construct ByteStrings.

Request

As for now, we provide build functions for every kind of request. Probably a more intuitive way to do so, is to determine the type of request based on the rqApiKey provided in the message header - basically the same way we do for parsing. See issue.

A build function takes a type RequestMessage and gives a ByteString (lazy). Here is an example with a produce request:

let msg = buildPrRqMessage $ RequestMessage ...

Response

TODO

Client Library

The client libary provides functionalities to easily implement an Apache Kafka client in Haskell. The goal of this, is to be able to setup a client without knowing too much details about the Apache Kafka protocol itself. We therefore provide separate types to construct a message very easily. In the background we then pack the input to the appropriate Request/Response Message.

Consumer example

Using the packFtRqMessage (TODO: generic functio to pack, see issues) function we generate RequestMessage (FetchRequest in this case) out of a InputFT. After all, we send this request to a socket and expact a response in the format of type FetchResponse - which will be parsed using the readFtResponse function.

...
forever $ do
    let req = packFtRqMessage $ InputFt
                (C.pack "myClient") -- client id
                (C.pack "myTopic") -- topic name
                (fromIntegral $ (read offset ::Int)) -- start offset
    sendFtRequest sock req
    forkIO $ do
      input <- SBL.recv sock 4096
      print input
      let response = readFtResponse input
      print response
    threadDelay 1000000
  ...

Producer example

Again, we first have to pack our InputMessage into a RequestMessage (in this case ProduceRequest) to further send it to a socket. In return, we expect a ProduceResponse which firt has to be parsed after we print the message to console.

...
forever $ do 
    inputMessage <- getLine
    let req = packRequest $ InputMessage 
                (C.pack clientId) -- client id
                (C.pack topicName) -- topic name
                (fromIntegral 0)-- partition number
                (C.pack inputMessage) -- payload data
    sendRequest sock req

    i <- SBL.recv sock 4096
    response <- readProduceResponse i
    print response 
  ...