kafka-pixy

Kafka-Pixy (gRPC/REST Proxy for Kafka)

Kafka-Pixy is a dual API (gRPC and REST) proxy for Kafka with automatic consumer group control. It is designed to hide the complexity of the Kafka client protocol and provide a stupid simple API that is trivial to implement in any language.

Kafka-Pixy is tested against Kafka versions 1.1.1 and 2.3.0, but is likely to work with any version starting from 0.8.2.2. It uses the Kafka Offset Commit/Fetch API to keep track of consumer offsets. However Group Membership API is not yet implemented, therefore it needs to talk to Zookeeper directly to manage consumer group membership.

Warning: Kafka-Pixy does not support wildcard subscriptions and therefore cannot coexist in a consumer group with clients using them. It should be possible to use other clients in the same consumer group as kafka-pixy instance if they subscribe to topics by their full names, but that has never been tested so do that at your own risk.

If you are anxious to get started then install Kafka-Pixy and proceed with a quick start guide for your weapon of choice: Curl, Python, or Golang. If you want to use some other language, then you still can use either of the guides for inspiration, but you would need to generate gRPC client stubs from kafkapixy.proto yourself (please refer to gRPC documentation for details).

Key Features:

• Automatic Consumer Group Management: Unlike in Kafka REST Proxy by Confluent clients do not need to explicitly create a consumer instance. When Kafka-Pixy gets a consume request for a group-topic pair for the first time, it automatically joins the group and subscribes to the topic. When requests stop coming for longer than the subscription timeout it cancels the subscription;
• At Least Once Guarantee: The main feature of Kafka-Pixy is that it guarantees at-least-once message delivery. The guarantee is achieved via combination of synchronous production and explicit acknowledgement of consumed messages;
• Dual API: Kafka-Pixy provides two types of API:
  • gRPC (Protocol Buffers over HTTP/2) recommended to produce/consume messages;
  • REST (JSON over HTTP) intended for testing and operations purposes, although you can use it to produce/consume messages too;
• Multi-Cluster Support: One Kafka-Pixy instance can proxy to several Kafka clusters. You just need to define them in the config file and then address clusters by name given in the config file in your API requests.
• Aggregation: Kafka works best when messages are read/written in batches, but from an application's standpoint it is easier to deal with individual message read/writes. Kafka-Pixy provides a message based API to clients, but internally it aggregates requests and sends them to Kafka in batches.
• Locality: Kafka-Pixy is intended to run on the same host as the applications using it. Remember that it only provides a message based API - no batching, therefore using it over network is suboptimal.

gRPC API

gRPC is an open source framework that is using Protocol Buffers as interface definition language and HTTP/2 as transport protocol. Kafka-Pixy API is defined in kafkapixy.proto. Client stubs for Golang and Python are generated and provided in this repository, but you can easily generate stubs for a bunch of other languages. Please refer to the gRPC documentation for information on the language of your choice.

REST API

It is highly recommended to use the gRPC API for production/consumption. The HTTP API is only provided for quick tests and operational purposes.

Each API endpoint has two variants which differ by /clusters/ prefix. The one with the proxy prefix is to be used when multiple clusters are configured. The one without the prefix operates on the default cluster (the one that is mentioned first in the YAML configuration file).

Produce

POST /topics//messages
POST /clusters//topics//messages

Writes a message to a topic on a particular cluster. If the request content type is either text/plain or application/json then a message should be sent as the body of a request. If content type is x-www-form-urlencoded then a message should be passed as the msg form parameter.

If Kafka-Pixy is configured to use a version of the Kafka protocol (via the kafka.version proxy setting) that is 0.11.0.0 or later, it is also possible to add record headers to a message by adding HTTP headers to your message. Any HTTP header with the prefix "X-Kafka-" will have that prefix stripped and the header will be used as a record header. Since the values of Kafka headers can be arbitrary byte strings, the value of the HTTP header must be Base 64-encoded.

By default the message is written to Kafka asynchronously, that is the HTTP request completes as soon as Kafka-Pixy reads the request from the wire, and production to Kafka is performed in the background. Therefore it is not guaranteed that the message will ever get into Kafka.

If you need a guarantee that a message is written to Kafka, then pass the sync flag with your request. In that case when Kafka-Pixy returns a response is governed by producer.required_acks parameter in the YAML config. It can be one of:

• no_response: the response is returned as soon as a produce request is delivered to a partition leader Kafka broker (no disk writes performed yet).
• wait_for_local: the response is returned as soon as data is written to the disk by a partition leader Kafka broker.
• wait_for_all: the response is returned after all in-sync replicas have data committed to disk.

E.g. if a Kafka-Pixy process has been started with the --tcpAddr=0.0.0.0:8080 argument, then you can test it using curl as follows:

curl -X POST localhost:8080/topics/foo/messages?key=bar&sync \
  -H 'Content-Type: text/plain' \
  -d 'Good news everyone!'

If the message is submitted asynchronously then the response will be an empty json object {}.

If the message is submitted synchronously then in case of success (HTTP status 200) the response will be like:

{
  "partition": ,
  "offset": 
}

In case of failure (HTTP statuses 404 and 500) the response will be:

{
  "error": 
}

Consume

GET /topics//messages
GET /clusters//topics//messages

Consumes a message from a topic of a particular cluster as a member of a particular consumer group. A message previously consumed from the same topic can be optionally acknowledged.

If noAck is defined in a request then no message is acknowledged by the request. If a request defines both ackPartition and ackOffset parameters then a message previously consumed from the same topic from the specified partition with the specified offset is acknowledged by the request. If none of the ack related parameters is specified then the request will acknowledge the message consumed in this requests if any. It is called auto-ack mode.

When a message is consumed as a member of a consume group for the first time, Kafka-Pixy joins the consumer group and subscribes to the topic. All Kafka-Pixy instances that are currently members of that group and subscribed to that topic distribute partitions between themselves, so that each Kafka-Pixy instance gets a subset of partitions for exclusive consumption (Read more about Kafka consumer groups here).

If a Kafka-Pixy instance has not received consume requests for a topic for the duration of the subscription timeout, then it unsubscribes from the topic, and the topic partitions are redistributed among Kafka-Pixy instances that are still consuming from it.

If there are no unread messages in the topic the request will block waiting for the duration of the long polling timeout. If there are no messages produced during this long poll waiting then the request will return 408 Request Timeout error, otherwise the response will be a JSON document of the following structure:

{
  "key": ,
  "value": ,
  "partition": ,
  "offset": ,
  "headers": [
    {
      "key": ,
      "value": 
    }
  ]
}

e.g.:

{
  "key": "0JzQsNGA0YPRgdGP",
  "value": "0JzQvtGPINC70Y7QsdC40LzQsNGPINC00L7Rh9C10L3RjNC60LA=",
  "partition": 0,
  "offset": 13,
  "headers": [
    {
      "key": "foo",
      "value": "YmFy"
    }
  ]
}

Note that headers are only supported if the Kafka protocol version (set via the kafka.version configuration flag) is set to 0.11.0.0 or later.

Acknowledge

POST /topics//acks
POST /clusters//topics//acks

Acknowledges a previously consumed message.