Kaffe

An opinionated, highly specific, Elixir wrapper around Brod: the Erlang Kafka client. :coffee:

NOTE: Although we're using this in production at Spreedly it is still under active development. The API may change and there may be serious bugs we've yet to encounter.

Installation

1. Add kaffe to your list of dependencies in mix.exs:

   def deps do
     [{:kaffe, "~> 1.0"}]
   end

2. Ensure kaffe is started with your application:

   def application do
     [applications: [:logger, :kaffe]]
   end

3. Configure a Kaffe Consumer and/or Producer

Kaffe Consumer Usage

Consumers receive a list of messages and work as part of the :brod_group_member behavior. This has a few important benefits:

1. Group members assign a "subscriber" to each partition in the topic. Because Kafka topics scale with partitions, having a worker per partition usually increases throughput.
2. Group members correctly handle partition assignments across multiple clients in a consumer group. This means that this mode of operation will scale horizontally (e.g., multiple dynos on Heroku).
3. Downstream processing that benefits from batching (like writing to another Kafka topic) is more easily supported.

There is also legacy support for single message consumers, which process one message at a time using the :brod_group_subscriber behavior. This was the original mode of operation for Kaffe but is slow and does not scale. For this reason it is considered deprecated.

Kaffe GroupMember - Batch Message Consumer

1. Define a handle_messages/1 function in the provided module.

   handle_messages/1 This function (note the pluralization) will be called with a list of messages, with each message as a map. Each message map will include the topic and partition in addition to the normal Kafka message metadata.

   The module's handle_messages/1 function must return :ok or Kaffe will throw an error. The Kaffe consumer will block until your handle_messages/1 function returns :ok.

   defmodule MessageProcessor do
     def handle_messages(messages) do
       for %{key: key, value: value} = message <- messages do
         IO.inspect message
         IO.puts "#{key}: #{value}"
       end
       :ok # Important!
     end
   end

2. The configuration options for the GroupMember consumer are a superset of those for Kaffe.Consumer, except for :async_message_ack, which is not supported. The additional options are:

   :rebalance_delay_ms which is the time to allow for rebalancing among workers. The default is 10,000, which should give the consumers time to rebalance when scaling.

   :max_bytes limits the number of message bytes received from Kafka for a particular topic subscriber. The default is 1MB. This parameter might need tuning depending on the number of partitions in the topics being read (there is one subscriber per topic per partition). For example, if you are reading from two topics, each with 32 partitions, there is the potential of 64MB in buffered messages at any one time.

   :min_bytes Sets a minimum threshold for the number of bytes to fetch for a batch of messages. The default is 0MB.

   :max_wait_time Sets the maximum number of milliseconds that the broker is allowed to collect min_bytes of messages in a batch of messages

   :offset_reset_policy controls how the subscriber handles an expired offset. See the Kafka consumer option, auto.offset.reset. Valid values for this option are:

   • :reset_to_earliest - reset to the earliest available offset
   • :reset_to_latest - reset to the latest offset
   • :reset_by_subscriber - The subscriber receives the OffsetOutOfRange error

   More information in the Brod consumer.

   :worker_allocation_strategy controls how workers are allocated with respect to consumed topics and partitions.

   • :worker_per_partition - this is the default (for backward compatibilty) and allocates a single worker per partition across topics. This is useful for managing concurrent processing of messages that may be received from any consumed topic.
   • :worker_per_topic_partition - this strategy allocates a worker per topic partition. This means there will be a worker for every topic partition consumed. Unless you need to control concurrency across topics, you should use this strategy.

   config :kaffe,
     consumer: [
       endpoints: [kafka: 9092],
       topics: ["interesting-topic"],
       consumer_group: "your-app-consumer-group",
       message_handler: MessageProcessor,
       offset_reset_policy: :reset_to_latest,
       max_bytes: 500_000,
       worker_allocation_strategy: :worker_per_topic_partition,
   
       #optional
       sasl: %{
         mechanism: :plain,
         login: System.get_env("KAFFE_PRODUCER_USER"),
         password: System.get_env("KAFFE_PRODUCER_PASSWORD")
       }
     ],

3. Add Kaffe.GroupMemberSupervisor as a supervisor in your supervision tree

   defmodule MyApp.Application do
     use Application
   
     def start(_type, _args) do
       children = [
         %{
           id: Kaffe.GroupMemberSupervisor,
           start: {Kaffe.GroupMemberSupervisor, :start_link, []},
           type: :supervisor
         }
       ]
   
       opts = [strategy: :one_for_one, name: MyApp.Application.Supervisor]
       Supervisor.start_link(children, opts)
     end
   end

Managing how offsets are committed

In some cases you may not want to commit back the most recent offset after processing a list of messages. For example, if you're batching messages to be sent elsewhere and want to ensure that a batch can be rebuilt should there be an error further downstream. In that example you might want to keep the offset of the first message in your batch so your consumer can restart back at that point to reprocess and rebatch the messages. Your message handler can respond in the following ways to manage how offsets are committed back:

:ok - commit back the most recent offset and request more messages {:ok, :no_commit} - do not commit back the most recent offset and request more message from the offset of the last message {:ok, offset} - commit back at the offset specified and request messages from that point forward

Example:

defmodule MessageProcessor do
  def handle_messages(messages) do
    for %{key: key, value: value} = message <- messages do
      IO.inspect message
      IO.puts "#{key}: #{value}"
    end
    {:ok, :no_commit}
  end
end

Kaffe Consumer - Single Message Consumer (Deprecated)

For backward compatiblitly only, you should use Kaffe.GroupMemberSupervisor instead!

1. Add a handle_message/1 function to a local module (e.g. MessageProcessor). This function will be called with each Kafka message as a map. Each message map will include the topic and partition in addition to the normal Kafka message metadata.

   The module's handle_message/1 function must return :ok or Kaffe will throw an error. In normal (synchronous consumer) operation the Kaffe consumer will block until your handle_message/1 function returns :ok.

   Example

   defmodule MessageProcessor do
     def handle_message(%{key: key, value: value} = message) do
       IO.inspect message
       IO.puts "#{key}: #{value}"
       :ok # The handle_message function MUST return :ok
     end
   end

   Message Structure

   %{
     attributes: 0,
     crc: 1914336469,
     key: "kafka message key",
     magic_byte: 0,
     offset: 41,
     partition: 17,
     topic: "some-kafka-topic",
     value: "the actual kafka message value is here",
     ts: 1234567890123, # timestamp in milliseconds
     ts_type: :append  # timestamp type: :undefined | :create | :append
   }

2. Configure your Kaffe Consumer in your mix config

   config :kaffe,
     consumer: [
       endpoints: [kafka: 9092], # that's [hostname: kafka_port]
       topics: ["interesting-topic"], # the topic(s) that will be consumed
       consumer_group: "your-app-consumer-group", # the consumer group for tracking offsets in Kafka
       message_handler: MessageProcessor, # the module from Step 1 that will process messages
   
       # optional
       async_message_ack: false, # see "async message acknowledgement" below
       start_with_earliest_message: true # default false
     ],

   The start_with_earliest_message field controls where your consumer group starts when it starts for the very first time. Once offsets have been committed to Kafka then they will supercede this option. If omitted then your consumer group will start processing from the most recent messages in the topic instead of consuming all available messages.

   Heroku Configuration

   To configure a Kaffe Consumer for a Heroku Kafka compatible environment including SSL omit the endpoint and instead set heroku_kafka_env: true

   config :kaffe,
     consumer: [
       heroku_kafka_env: true,
       topics: ["interesting-topic"],
       consumer_group: "your-app-consumer-group",
       message_handler: MessageProcessor
     ]

   With that setting in place Kaffe will automatically pull required info from the following ENV variables:

   • KAFKA_URL
   • KAFKA_CLIENT_CERT
   • KAFKA_CLIENT_CERT_KEY
   • KAFKA_TRUSTED_CERT (not used yet)

3. Add Kaffe.Consumer as a worker in your supervision tree

   worker(Kaffe.Consumer, [])

async message acknowledgement

If you need asynchronous message consumption:

1. Add a handle_message/2 function to your processing module. This function will be called with the Consumer pid and the Kafka message. When your processing is complete you will need to call Kaffe.Consumer.ack(pid, message) to acknowledge the offset.

2. Set async to true when you start the Kaffe.Consumer

   consumer_group = "demo-commitlog-consumer"
   topic = "commitlog"
   message_handler = MessageProcessor
   async = true
   
   worker(Kaffe.Consumer, [consumer_group, topics, message_handler, async])
   
   # … in your message handler module
   
   def handle_message(pid, message) do
     spawn_message_processing_worker(pid, message)
     :ok # MUST return :ok
   end
   
   # … somewhere in your system when the worker is finished processing
   
   Kaffe.Consumer.ack(pid, message)

NOTE: Asynchronous consumption means your system will no longer provide any backpressure to the Kaffe.Consumer. You will also need to add robust measures to your system to ensure that no messages are lost in processing. IE if you spawn 5 workers processing a series of asynchronous messages from Kafka and 1 of them crashes without acknowledgement then it's possible and likely that the message will be skipped entirely.

Kafka only tracks a single numeric offset, not individual messages. If a message fails and a later offset is committed then the failed message will not be sent again.

It's possible that your topic and system are entirely ok with losing some messages (i.e. frequent metrics that aren't individually important).

Kaffe Producer Usage

Kaffe.Producer handles producing messages to Kafka and will automatically select the topic partitions per message or can be given a function to call to determine the partition per message. Kaffe automatically inserts a Kafka timestamp with each message.

Configure your Kaffe Producer in your mix config

config :kaffe,
  producer: [
    endpoints: [kafka: 9092], # [hostname: port]
    topics: ["kafka-topic"],

    # optional
    partition_strategy: :md5,
    ssl: true,
    sasl: %{
      mechanism: :plain,
      login: System.get_env("KAFFE_PRODUCER_USER"),
      password: System.get_env("KAFFE_PRODUCER_PASSWORD")
    }
  ]

The partition_strategy setting can be one of:

• :md5: (default) provides even and deterministic distrbution of the messages over the available partitions based on an MD5 hash of the key
• :random: select a random partition for each message
• function: a given function to call to determine the correct partition

You can also set any of the Brod producer configuration options in the producer section - see the Brod sources for a list of keys and their meaning.

If the Kafka broker is configured with SASL_PLAINTEXT auth, the sasl option can be added.

If using Confluent Hosted Kafka, also add ssl: true as shown above.

Heroku Configuration

To configure a Kaffe Producer for a Heroku Kafka compatible environment including SSL omit the endpoint and instead set heroku_kafka_env: true

config :kaffe,
  producer: [
    heroku_kafka_env: true,
    topics: ["kafka-topic"],

    # optional
    partition_strategy: :md5
  ]

With that setting in place Kaffe will automatically pull required info from the following ENV variables:

• KAFKA_URL
• KAFKA_CLIENT_CERT
• KAFKA_CLIENT_CERT_KEY
• KAFKA_TRUSTED_CERT

Producing to Kafka

Currently only synchronous message production is supported.

There are several ways to produce:

• topic/message_list - Produce each message in the list to the given topic. The messages are produced to the correct partition based on the configured partitioning strategy.

  Each item in the list is a tuple of the key and value: {key, value}.

  Kaffe.Producer.produce_sync("topic", [{"key1", "value1"}, {"key2", "value2"}])

• topic/partition/message_list - Produce each message in the list to the given topic/partition.

  Each item in the list is a tuple of the key and value: {key, value}.

  Kaffe.Producer.produce_sync("topic", 2, [{"key1", "value1"}, {"key2", "value2"}])

• key/value - The key/value will be produced to the first topic given to the producer when it was started. The partition will be selected with the chosen strategy or given function.

  Kaffe.Producer.produce_sync("key", "value")

• topic/key/value - The key/value will be produced to the given topic.

  Kaffe.Producer.produce_sync("whitelist", "key", "value")

• topic/partition/key/value - The key/value will be produced to the given topic/partition.

  Kaffe.Producer.produce_sync("whitelist", 2, "key", "value")

  NOTE: With this approach Kaffe will not calculate the next partition since it assumes you're taking over that job by giving it a specific partition.

Testing

Setup

In order to run the end to end tests, a Kafka topic is required. It must:

• be named kaffe-test
• have 32 partitions

If using the kafka-topics.sh script that comes with the Kafka distribution, you may use something like:

kafka-topics.sh --zookeeper localhost:2181 --create --partitions 32 --replication-factor 1 --topic kaffe-test

Running

# unit tests
mix test
# end to end test
mix test --only e2e