A Ruby client library for Apache Kafka, a distributed log and message bus. The focus of this library will be operational simplicity, with good logging and metrics that can make debugging issues easier.
Although parts of this library work with Kafka 0.8 – specifically, the Producer API – it's being tested and developed against Kafka 0.9. The Consumer API is Kafka 0.9+ only.
- Installation
- Compatibility
- Usage
- Design
- Development
- Support and Discussion
- Roadmap
- Higher level libraries
Add this line to your application's Gemfile:
gem 'ruby-kafka'
And then execute:
$ bundle
Or install it yourself as:
$ gem install ruby-kafka
Kafka 0.8 | Kafka 0.9 | Kafka 0.10 | Kafka 0.11 | Kafka 1.0 | |
---|---|---|---|---|---|
Producer API | Full support in v0.4.x | Full support in v0.4.x | Full support in v0.5.x | Limited support | Limited support |
Consumer API | Unsupported | Full support in v0.4.x | Full support in v0.5.x | Limited support | Limited support |
This library is targeting Kafka 0.9 with the v0.4.x series and Kafka 0.10 with the v0.5.x series. There's limited support for Kafka 0.8, and things should work with Kafka 0.11, although there may be performance issues due to changes in the protocol.
- Kafka 0.8: Full support for the Producer API in ruby-kafka v0.4.x, but no support for consumer groups. Simple message fetching works.
- Kafka 0.9: Full support for the Producer and Consumer API in ruby-kafka v0.4.x.
- Kafka 0.10: Full support for the Producer and Consumer API in ruby-kafka v0.5.x. Note that you must run version 0.10.1 or higher of Kafka due to limitations in 0.10.0.
- Kafka 0.11: Everything that works with Kafka 0.10 should still work, but so far no features specific to Kafka 0.11 have been added.
- Kafka 0.11: Everything that works with Kafka 0.10 should still work, but so far no features specific to Kafka 1.0 have been added.
This library requires Ruby 2.1 or higher.
Please see the documentation site for detailed documentation on the latest release. Note that the documentation on GitHub may not match the version of the library you're using – there are still being made many changes to the API.
A client must be initialized with at least one Kafka broker, from which the entire Kafka cluster will be discovered. Each client keeps a separate pool of broker connections. Don't use the same client from more than one thread.
require "kafka"
# The first argument is a list of "seed brokers" that will be queried for the full
# cluster topology. At least one of these *must* be available. `client_id` is
# used to identify this client in logs and metrics. It's optional but recommended.
kafka = Kafka.new(["kafka1:9092", "kafka2:9092"], client_id: "my-application")
The simplest way to write a message to a Kafka topic is to call #deliver_message
:
kafka = Kafka.new(...)
kafka.deliver_message("Hello, World!", topic: "greetings")
This will write the message to a random partition in the greetings
topic. If you want to write to a specific partition, pass the partition
parameter:
# Will write to partition 42.
kafka.deliver_message("Hello, World!", topic: "greetings", partition: 42)
If you don't know exactly how many partitions are in the topic, or if you'd rather have some level of indirection, you can pass in partition_key
instead. Two messages with the same partition key will always be assigned to the same partition. This is useful if you want to make sure all messages with a given attribute are always written to the same partition, e.g. all purchase events for a given customer id.
# Partition keys assign a partition deterministically.
kafka.deliver_message("Hello, World!", topic: "greetings", partition_key: "hello")
Kafka also supports message keys. When passed, a message key can be used instead of a partition key. The message key is written alongside the message value and can be read by consumers. Message keys in Kafka can be used for interesting things such as Log Compaction. See Partitioning for more information.
# Set a message key; the key will be used for partitioning since no explicit
# `partition_key` is set.
kafka.deliver_message("Hello, World!", key: "hello", topic: "greetings")
While #deliver_message
works fine for infrequent writes, there are a number of downsides:
- Kafka is optimized for transmitting messages in batches rather than individually, so there's a significant overhead and performance penalty in using the single-message API.
- The message delivery can fail in a number of different ways, but this simplistic API does not provide automatic retries.
- The message is not buffered, so if there is an error, it is lost.
The Producer API solves all these problems and more:
# Instantiate a new producer.
producer = kafka.producer
# Add a message to the producer buffer.
producer.produce("hello1", topic: "test-messages")
# Deliver the messages to Kafka.
producer.deliver_messages
#produce
will buffer the message in the producer but will not actually send it to the Kafka cluster. Buffered messages are only delivered to the Kafka cluster once #deliver_messages
is called. Since messages may be destined for different partitions, this could involve writing to more than one Kafka broker. Note that a failure to send all buffered messages after the configured number of retries will result in Kafka::DeliveryFailed
being raised. This can be rescued and ignored; the messages will be kept in the buffer until the next attempt.
Read the docs for Kafka::Producer for more details.
A normal producer will block while #deliver_messages
is sending messages to Kafka, possibly for tens of seconds or even minutes at a time, depending on your timeout and retry settings. Furthermore, you have to call #deliver_messages
manually, with a frequency that balances batch size with message delay.
In order to avoid blocking during message deliveries you can use the asynchronous producer API. It is mostly similar to the synchronous API, with calls to #produce
and #deliver_messages
. The main difference is that rather than blocking, these calls will return immediately. The actual work will be done in a background thread, with the messages and operations being sent from the caller over a thread safe queue.
# `#async_producer` will create a new asynchronous producer.
producer = kafka.async_producer
# The `#produce` API works as normal.
producer.produce("hello", topic: "greetings")
# `#deliver_messages` will return immediately.
producer.deliver_messages
# Make sure to call `#shutdown` on the producer in order to avoid leaking
# resources. `#shutdown` will wait for any pending messages to be delivered
# before returning.
producer.shutdown
By default, the delivery policy will be the same as for a synchronous producer: only when #deliver_messages
is called will the messages be delivered. However, the asynchronous producer offers two complementary policies for automatic delivery:
- Trigger a delivery once the producer's message buffer reaches a specified threshold. This can be used to improve efficiency by increasing the batch size when sending messages to the Kafka cluster.
- Trigger a delivery at a fixed time interval. This puts an upper bound on message delays.
These policies can be used alone or in combination.
# `async_producer` will create a new asynchronous producer.
producer = kafka.async_producer(
# Trigger a delivery once 100 messages have been buffered.
delivery_threshold: 100,
# Trigger a delivery every 30 seconds.
delivery_interval: 30,
)
producer.produce("hello", topic: "greetings")
# ...
When calling #shutdown
, the producer will attempt to deliver the messages and the method call will block until that has happened. Note that there's no guarantee that the messages will be delivered.
Note: if the calling thread produces messages faster than the producer can write them to Kafka, you'll eventually run into problems. The internal queue used for sending messages from the calling thread to the background worker has a size limit; once this limit is reached, a call to #produce
will raise Kafka::BufferOverflow
.
This library is agnostic to which serialization format you prefer. Both the value and key of a message is treated as a binary string of data. This makes it easier to use whatever serialization format you want, since you don't have to do anything special to make it work with ruby-kafka. Here's an example of encoding data with JSON:
require "json"
# ...
event = {
"name" => "pageview",
"url" => "https://example.com/posts/123",
# ...
}
data = JSON.dump(event)
producer.produce(data, topic: "events")
There's also an example of encoding messages with Apache Avro.
Kafka topics are partitioned, with messages being assigned to a partition by the client. This allows a great deal of flexibility for the users. This section describes several strategies for partitioning and how they impact performance, data locality, etc.
When optimizing for efficiency, we either distribute messages as evenly as possible to all partitions, or make sure each producer always writes to a single partition. The former ensures an even load for downstream consumers; the latter ensures the highest producer performance, since message batching is done per partition.
If no explicit partition is specified, the producer will look to the partition key or the message key for a value that can be used to deterministically assign the message to a partition. If there is a big number of different keys, the resulting distribution will be pretty even. If no keys are passed, the producer will randomly assign a partition. Random partitioning can be achieved even if you use message keys by passing a random partition key, e.g. partition_key: rand(100)
.
If you wish to have the producer write all messages to a single partition, simply generate a random value and re-use that as the partition key:
partition_key = rand(100)
producer.produce(msg1, topic: "messages", partition_key: partition_key)
producer.produce(msg2, topic: "messages", partition_key: partition_key)
# ...
You can also base the partition key on some property of the producer, for example the host name.
By assigning messages to a partition based on some property of the message, e.g. making sure all events tracked in a user session are assigned to the same partition, downstream consumers can make simplifying assumptions about data locality. In this example, a consumer can keep process local state pertaining to a user session knowing that all events for the session will be read from a single partition. This is also called semantic partitioning, since the partition assignment is part of the application behavior.
Typically it's sufficient to simply pass a partition key in order to guarantee that a set of messages will be assigned to the same partition, e.g.
# All messages with the same `session_id` will be assigned to the same partition.
producer.produce(event, topic: "user-events", partition_key: session_id)
However, sometimes it's necessary to select a specific partition. When doing this, make sure that you don't pick a partition number outside the range of partitions for the topic:
partitions = kafka.partitions_for("events")
# Make sure that we don't exceed the partition count!
partition = some_number % partitions
producer.produce(event, topic: "events", partition: partition)
There's no standardized way to assign messages to partitions across different Kafka client implementations. If you have a heterogeneous set of clients producing messages to the same topics it may be important to ensure a consistent partitioning scheme. This library doesn't try to implement all schemes, so you'll have to figure out which scheme the other client is using and replicate it. An example:
partitions = kafka.partitions_for("events")
# Insert your custom partitioning scheme here:
partition = PartitioningScheme.assign(partitions, event)
producer.produce(event, topic: "events", partition: partition)
The producer is designed for resilience in the face of temporary network errors, Kafka broker failovers, and other issues that prevent the client from writing messages to the destination topics. It does this by employing local, in-memory buffers. Only when messages are acknowledged by a Kafka broker will they be removed from the buffer.
Typically, you'd configure the producer to retry failed attempts at sending messages, but sometimes all retries are exhausted. In that case, Kafka::DeliveryFailed
is raised from Kafka::Producer#deliver_messages
. If you wish to have your application be resilient to this happening (e.g. if you're logging to Kafka from a web application) you can rescue this exception. The failed messages are still retained in the buffer, so a subsequent call to #deliver_messages
will still attempt to send them.
Note that there's a maximum buffer size; by default, it's set to 1,000 messages and 10MB. It's possible to configure both these numbers:
producer = kafka.producer(
max_buffer_size: 5_000, # Allow at most 5K messages to be buffered.
max_buffer_bytesize: 100_000_000, # Allow at most 100MB to be buffered.
...
)
A final note on buffers: local buffers give resilience against broker and network failures, and allow higher throughput due to message batching, but they also trade off consistency guarantees for higher availability and resilience. If your local process dies while messages are buffered, those messages will be lost. If you require high levels of consistency, you should call #deliver_messages
immediately after #produce
.
Once the client has delivered a set of messages to a Kafka broker the broker will forward them to its replicas, thus ensuring that a single broker failure will not result in message loss. However, the client can choose when the leader acknowledges the write. At one extreme, the client can choose fire-and-forget delivery, not even bothering to check whether the messages have been acknowledged. At the other end, the client can ask the broker to wait until all its replicas have acknowledged the write before returning. This is the safest option, and the default. It's also possible to have the broker return as soon as it has written the messages to its own log but before the replicas have done so. This leaves a window of time where a failure of the leader will result in the messages being lost, although this should not be a common occurrence.
Write latency and throughput are negatively impacted by having more replicas acknowledge a write, so if you require low-latency, high throughput writes you may want to accept lower durability.
This behavior is controlled by the required_acks
option to #producer
and #async_producer
:
# This is the default: all replicas must acknowledge.
producer = kafka.producer(required_acks: :all)
# This is fire-and-forget: messages can easily be lost.
producer = kafka.producer(required_acks: 0)
# This only waits for the leader to acknowledge.
producer = kafka.producer(required_acks: 1)
Unless you absolutely need lower latency it's highly recommended to use the default setting (:all
).
There are basically two different and incompatible guarantees that can be made in a message delivery system such as Kafka:
- at-most-once delivery guarantees that a message is at most delivered to the recipient once. This is useful only if delivering the message twice carries some risk and should be avoided. Implicit is the fact that there's no guarantee that the message will be delivered at all.
- at-least-once delivery guarantees that a message is delivered, but it may be delivered more than once. If the final recipient de-duplicates messages, e.g. by checking a unique message id, then it's even possible to implement exactly-once delivery.
Of these two options, ruby-kafka implements the second one: when in doubt about whether a message has been delivered, a producer will try to deliver it again.
The guarantee is made only for the synchronous producer and boils down to this:
producer = kafka.producer
producer.produce("hello", topic: "greetings")
# If this line fails with Kafka::DeliveryFailed we *may* have succeeded in delivering
# the message to Kafka but won't know for sure.
producer.deliver_messages
# If we get to this line we can be sure that the message has been delivered to Kafka!
That is, once #deliver_messages
returns we can be sure that Kafka has received the message. Note that there are some big caveats here:
- Depending on how your cluster and topic is configured the message could still be lost by Kafka.
- If you configure the producer to not require acknowledgements from the Kafka brokers by setting
required_acks
to zero there is no guarantee that the message will ever make it to a Kafka broker. - If you use the asynchronous producer there's no guarantee that messages will have been delivered after
#deliver_messages
returns. A way of blocking until a message has been delivered with the asynchronous producer may be implemented in the future.
It's possible to improve your chances of success when calling #deliver_messages
, at the price of a longer max latency:
producer = kafka.producer(
# The number of retries when attempting to deliver messages. The default is
# 2, so 3 attempts in total, but you can configure a higher or lower number:
max_retries: 5,
# The number of seconds to wait between retries. In order to handle longer
# periods of Kafka being unavailable, increase this number. The default is
# 1 second.
retry_backoff: 5,
)
Note that these values affect the max latency of the operation; see Understanding Timeouts for an explanation of the various timeouts and latencies.
If you use the asynchronous producer you typically don't have to worry too much about this, as retries will be done in the background.
Depending on what kind of data you produce, enabling compression may yield improved bandwidth and space usage. Compression in Kafka is done on entire messages sets rather than on individual messages. This improves the compression rate and generally means that compressions works better the larger your buffers get, since the message sets will be larger by the time they're compressed.
Since many workloads have variations in throughput and distribution across partitions, it's possible to configure a threshold for when to enable compression by setting compression_threshold
. Only if the defined number of messages are buffered for a partition will the messages be compressed.
Compression is enabled by passing the compression_codec
parameter to #producer
with the name of one of the algorithms allowed by Kafka:
By default, all message sets will be compressed if you specify a compression codec. To increase the compression threshold, set compression_threshold
to an integer value higher than one.
producer = kafka.producer(
compression_codec: :snappy,
compression_threshold: 10,
)
A typical use case for Kafka is tracking events that occur in web applications. Oftentimes it's advisable to avoid having a hard dependency on Kafka being available, allowing your application to survive a Kafka outage. By using an asynchronous producer, you can avoid doing IO within the individual request/response cycles, instead pushing that to the producer's internal background thread.
In this example, a producer is configured in a Rails initializer:
# config/initializers/kafka_producer.rb
require "kafka"
# Configure the Kafka client with the broker hosts and the Rails
# logger.
$kafka = Kafka.new(["kafka1:9092", "kafka2:9092"], logger: Rails.logger)
# Set up an asynchronous producer that delivers its buffered messages
# every ten seconds:
$kafka_producer = $kafka.async_producer(
delivery_interval: 10,
)
# Make sure to shut down the producer when exiting.
at_exit { $kafka_producer.shutdown }
In your controllers, simply call the producer directly:
# app/controllers/orders_controller.rb
class OrdersController
def create
@order = Order.create!(params[:order])
event = {
order_id: @order.id,
amount: @order.amount,
timestamp: Time.now,
}
$kafka_producer.produce(event.to_json, topic: "order_events")
end
end
Note: If you're just looking to get started with Kafka consumers, you might be interested in visiting the Higher level libraries section that lists ruby-kafka based frameworks. Read on, if you're interested in either rolling your own executable consumers or if you want to learn more about how consumers work in Kafka.
Consuming messages from a Kafka topic with ruby-kafka is simple:
require "kafka"
kafka = Kafka.new(["kafka1:9092", "kafka2:9092"])
kafka.each_message(topic: "greetings") do |message|
puts message.offset, message.key, message.value
end
While this is great for extremely simple use cases, there are a number of downsides:
- You can only fetch from a single topic at a time.
- If you want to have multiple processes consume from the same topic, there's no way of coordinating which processes should fetch from which partitions.
- If the process dies, there's no way to have another process resume fetching from the point in the partition that the original process had reached.
The Consumer API solves all of the above issues, and more. It uses the Consumer Groups feature released in Kafka 0.9 to allow multiple consumer processes to coordinate access to a topic, assigning each partition to a single consumer. When a consumer fails, the partitions that were assigned to it are re-assigned to other members of the group.
Using the API is simple:
require "kafka"
kafka = Kafka.new(["kafka1:9092", "kafka2:9092"])
# Consumers with the same group id will form a Consumer Group together.
consumer = kafka.consumer(group_id: "my-consumer")
# It's possible to subscribe to multiple topics by calling `subscribe`
# repeatedly.
consumer.subscribe("greetings")
# Stop the consumer when the SIGTERM signal is sent to the process.
# It's better to shut down gracefully than to kill the process.
trap("TERM") { consumer.stop }
# This will loop indefinitely, yielding each message in turn.
consumer.each_message do |message|
puts message.topic, message.partition
puts message.offset, message.key, message.value
end
Each consumer process will be assigned one or more partitions from each topic that the group subscribes to. In order to handle more messages, simply start more processes.
In order to be able to resume processing after a consumer crashes, each consumer will periodically checkpoint its position within each partition it reads from. Since each partition has a monotonically increasing sequence of message offsets, this works by committing the offset of the last message that was processed in a given partition. Kafka handles these commits and allows another consumer in a group to resume from the last commit when a member crashes or becomes unresponsive.
By default, offsets are committed every 10 seconds. You can increase the frequency, known as the offset commit interval, to limit the duration of double-processing scenarios, at the cost of a lower throughput due to the added coordination. If you want to improve throughput, and double-processing is of less concern to you, then you can decrease the frequency. Set the commit interval to zero in order to disable the timer-based commit trigger entirely.
In addition to the time based trigger it's possible to trigger checkpointing in response to n messages having been processed, known as the offset commit threshold. This puts a bound on the number of messages that can be double-processed before the problem is detected. Setting this to 1 will cause an offset commit to take place every time a message has been processed. By default this trigger is disabled (set to zero).
It is possible to trigger an immediate offset commit by calling Consumer#commit_offsets
. This blocks the caller until the Kafka cluster has acknowledged the commit.
Stale offsets are periodically purged by the broker. The broker setting offsets.retention.minutes
controls the retention window for committed offsets, and defaults to 1 day. The length of the retention window, known as offset retention time, can be changed for the consumer.
Previously committed offsets are re-committed, to reset the retention window, at the first commit and periodically at an interval of half the offset retention time.
consumer = kafka.consumer(
group_id: "some-group",
# Increase offset commit frequency to once every 5 seconds.
offset_commit_interval: 5,
# Commit offsets when 100 messages have been processed.
offset_commit_threshold: 100,
# Increase the length of time that committed offsets are kept.
offset_retention_time: 7 * 60 * 60
)
For some use cases it may be necessary to control when messages are marked as processed. Note that since only the consumer position within each partition can be saved, marking a message as processed implies that all messages in the partition with a lower offset should also be considered as having been processed.
The method Consumer#mark_message_as_processed
marks a message (and all those that precede it in a partition) as having been processed. This is an advanced API that you should only use if you know what you're doing.
# Manually controlling checkpointing:
# Typically you want to use this API in order to buffer messages until some
# special "commit" message is received, e.g. in order to group together
# transactions consisting of several items.
buffer = []
# Messages will not be marked as processed automatically. If you shut down the
# consumer without calling `#mark_message_as_processed` first, the consumer will
# not resume where you left off!
consumer.each_message(automatically_mark_as_processed: false) do |message|
# Our messages are JSON with a `type` field and other stuff.
event = JSON.parse(message.value)
case event.fetch("type")
when "add_to_cart"
buffer << event
when "complete_purchase"
# We've received all the messages we need, time to save the transaction.
save_transaction(buffer)
# Now we can set the checkpoint by marking the last message as processed.
consumer.mark_message_as_processed(message)
# We can optionally trigger an immediate, blocking offset commit in order
# to minimize the risk of crashing before the automatic triggers have
# kicked in.
consumer.commit_offsets
# Make the buffer ready for the next transaction.
buffer.clear
end
end
For each topic subscription it's possible to decide whether to consume messages starting at the beginning of the topic or to just consume new messages that are produced to the topic. This policy is configured by setting the start_from_beginning
argument when calling #subscribe
:
# Consume messages from the very beginning of the topic. This is the default.
consumer.subscribe("users", start_from_beginning: true)
# Only consume new messages.
consumer.subscribe("notifications", start_from_beginning: false)
Once the consumer group has checkpointed its progress in the topic's partitions, the consumers will always start from the checkpointed offsets, regardless of start_from_beginning
. As such, this setting only applies when the consumer initially starts consuming from a topic.
In order to shut down a running consumer process cleanly, call #stop
on it. A common pattern is to trap a process signal and initiate the shutdown from there:
consumer = kafka.consumer(...)
# The consumer can be stopped from the command line by executing
# `kill -s TERM <process-id>`.
trap("TERM") { consumer.stop }
consumer.each_message do |message|
...
end
Sometimes it is easier to deal with messages in batches rather than individually. A batch is a sequence of one or more Kafka messages that all belong to the same topic and partition. One common reason to want to use batches is when some external system has a batch or transactional API.
# A mock search index that we'll be keeping up to date with new Kafka messages.
index = SearchIndex.new
consumer.subscribe("posts")
consumer.each_batch do |batch|
puts "Received batch: #{batch.topic}/#{batch.partition}"
transaction = index.transaction
batch.messages.each do |message|
# Let's assume that adding a document is idempotent.
transaction.add(id: message.key, body: message.value)
end
# Once this method returns, the messages have been successfully written to the
# search index. The consumer will only checkpoint a batch *after* the block
# has completed without an exception.
transaction.commit!
end
One important thing to note is that the client commits the offset of the batch's messages only after the entire batch has been processed.
There are two performance properties that can at times be at odds: throughput and latency. Throughput is the number of messages that can be processed in a given timespan; latency is the time it takes from a message is written to a topic until it has been processed.
In order to optimize for throughput, you want to make sure to fetch as many messages as possible every time you do a round trip to the Kafka cluster. This minimizes network overhead and allows processing data in big chunks.
In order to optimize for low latency, you want to process a message as soon as possible, even if that means fetching a smaller batch of messages.
There are three values that can be tuned in order to balance these two concerns.
min_bytes
is the minimum number of bytes to return from a single message fetch. By setting this to a high value you can increase the processing throughput. The default value is one byte.max_wait_time
is the maximum number of seconds to wait before returning data from a single message fetch. By setting this high you also increase the processing throughput – and by setting it low you set a bound on latency. This configuration overridesmin_bytes
, so you'll always get data back within the time specified. The default value is one second. If you want to have at most five seconds of latency, setmax_wait_time
to 5. You should make suremax_wait_time
* num brokers +heartbeat_interval
is less thansession_timeout
.max_bytes_per_partition
is the maximum amount of data a broker will return for a single partition when fetching new messages. The default is 1MB, but increasing this number may lead to better throughtput since you'll need to fetch less frequently. Setting it to a lower value is not recommended unless you have so many partitions that it's causing network and latency issues to transfer a fetch response from a broker to a client. Setting the number too high may result in instability, so be careful.
The first two settings can be passed to either #each_message
or #each_batch
, e.g.
# Waits for data for up to 5 seconds on each broker, preferring to fetch at least 5KB at a time.
# This can wait up to num brokers * 5 seconds.
consumer.each_message(min_bytes: 1024 * 5, max_wait_time: 5) do |message|
# ...
end
The last setting is configured when subscribing to a topic, and can vary between topics:
# Fetches up to 5MB per partition at a time for better throughput.
consumer.subscribe("greetings", max_bytes_per_partition: 5 * 1024 * 1024)
consumer.each_message do |message|
# ...
end
You typically don't want to share a Kafka client object between threads, since the network communication is not synchronized. Furthermore, you should avoid using threads in a consumer unless you're very careful about waiting for all work to complete before returning from the #each_message
or #each_batch
block. This is because checkpointing assumes that returning from the block means that the messages that have been yielded have been successfully processed.
You should also avoid sharing a synchronous producer between threads, as the internal buffers are not thread safe. However, the asynchronous producer should be safe to use in a multi-threaded environment. This is because producers, when instantiated, get their own copy of any non-thread-safe data such as network sockets. Furthermore, the asynchronous producer has been designed in such a way to only a single background thread operates on this data while any foreground thread with a reference to the producer object can only send messages to that background thread over a safe queue. Therefore it is safe to share an async producer object between many threads.
It's a very good idea to configure the Kafka client with a logger. All important operations and errors are logged. When instantiating your client, simply pass in a valid logger:
logger = Logger.new("log/kafka.log")
kafka = Kafka.new(logger: logger, ...)
By default, nothing is logged.
Most operations are instrumented using Active Support Notifications. In order to subscribe to notifications, make sure to require the notifications library:
require "active_support/notifications"
require "kafka"
The notifications are namespaced based on their origin, with separate namespaces for the producer and the consumer.
In order to receive notifications you can either subscribe to individual notification names or use regular expressions to subscribe to entire namespaces. This example will subscribe to all notifications sent by ruby-kafka:
ActiveSupport::Notifications.subscribe(/.*\.kafka$/) do |*args|
event = ActiveSupport::Notifications::Event.new(*args)
puts "Received notification `#{event.name}` with payload: #{event.payload.inspect}"
end
All notification events have the client_id
key in the payload, referring to the Kafka client id.
-
produce_message.producer.kafka
is sent whenever a message is produced to a buffer. It includes the following payload:value
is the message value.key
is the message key.topic
is the topic that the message was produced to.buffer_size
is the size of the producer buffer after adding the message.max_buffer_size
is the maximum size of the producer buffer.
-
deliver_messages.producer.kafka
is sent whenever a producer attempts to deliver its buffered messages to the Kafka brokers. It includes the following payload:attempts
is the number of times delivery was attempted.message_count
is the number of messages for which delivery was attempted.delivered_message_count
is the number of messages that were acknowledged by the brokers - if this number is smaller thanmessage_count
not all messages were successfully delivered.
All notifications have group_id
in the payload, referring to the Kafka consumer group id.
-
process_message.consumer.kafka
is sent whenever a message is processed by a consumer. It includes the following payload:value
is the message value.key
is the message key.topic
is the topic that the message was consumed from.partition
is the topic partition that the message was consumed from.offset
is the message's offset within the topic partition.offset_lag
is the number of messages within the topic partition that have not yet been consumed.
-
start_process_message.consumer.kafka
is sent beforeprocess_message.consumer.kafka
, and contains the same payload. It is delivered before the message is processed, rather than after. -
process_batch.consumer.kafka
is sent whenever a message batch is processed by a consumer. It includes the following payload:message_count
is the number of messages in the batch.topic
is the topic that the message batch was consumed from.partition
is the topic partition that the message batch was consumed from.highwater_mark_offset
is the message batch's highest offset within the topic partition.offset_lag
is the number of messages within the topic partition that have not yet been consumed.
-
start_process_batch.consumer.kafka
is sent beforeprocess_batch.consumer.kafka
, and contains the same payload. It is delivered before the batch is processed, rather than after. -
join_group.consumer.kafka
is sent whenever a consumer joins a consumer group. It includes the following payload:group_id
is the consumer group id.
-
sync_group.consumer.kafka
is sent whenever a consumer is assigned topic partitions within a consumer group. It includes the following payload:group_id
is the consumer group id.
-
leave_group.consumer.kafka
is sent whenever a consumer leaves a consumer group. It includes the following payload:group_id
is the consumer group id.
request.connection.kafka
is sent whenever a network request is sent to a Kafka broker. It includes the following payload:api
is the name of the API that was called, e.g.produce
orfetch
.request_size
is the number of bytes in the request.response_size
is the number of bytes in the response.
It is highly recommended that you monitor your Kafka client applications in production. Typical problems you'll see are:
- high network error rates, which may impact performance and time-to-delivery;
- producer buffer growth, which may indicate that producers are unable to deliver messages at the rate they're being produced;
- consumer processing errors, indicating exceptions are being raised in the processing code;
- frequent consumer rebalances, which may indicate unstable network conditions or consumer configurations.
You can quite easily build monitoring on top of the provided instrumentation hooks. In order to further help with monitoring, a prebuilt Statsd and Datadog reporter is included with ruby-kafka.
We recommend monitoring the following:
- Low-level Kafka API calls:
- The rate of API call errors to the total number of calls by both API and broker.
- The API call throughput by both API and broker.
- The API call latency by both API and broker.
- Producer-level metrics:
- Delivery throughput by topic.
- The latency of deliveries.
- The producer buffer fill ratios.
- The async producer queue sizes.
- Message delivery delays.
- Failed delivery attempts.
- Consumer-level metrics:
- Message processing throughput by topic.
- Processing latency by topic.
- Processing errors by topic.
- Consumer lag (how many messages are yet to be processed) by topic/partition.
- Group join/sync/leave by client host.
The Statsd reporter is automatically enabled when the kafka/statsd
library is required. You can optionally change the configuration.
require "kafka/statsd"
# Default is "ruby_kafka".
Kafka::Statsd.namespace = "custom-namespace"
# Default is "127.0.0.1".
Kafka::Statsd.host = "statsd.something.com"
# Default is 8125.
Kafka::Statsd.port = 1234
The Datadog reporter is automatically enabled when the kafka/datadog
library is required. You can optionally change the configuration.
# This enables the reporter:
require "kafka/datadog"
# Default is "ruby_kafka".
Kafka::Datadog.namespace = "custom-namespace"
# Default is "127.0.0.1".
Kafka::Datadog.host = "statsd.something.com"
# Default is 8125.
Kafka::Datadog.port = 1234
It's important to understand how timeouts work if you have a latency sensitive application. This library allows configuring timeouts on different levels:
Network timeouts apply to network connections to individual Kafka brokers. There are two config keys here, each passed to Kafka.new
:
connect_timeout
sets the number of seconds to wait while connecting to a broker for the first time. When ruby-kafka initializes, it needs to connect to at least one host inseed_brokers
in order to discover the Kafka cluster. Each host is tried until there's one that works. Usually that means the first one, but if your entire cluster is down, or there's a network partition, you could wait up ton * connect_timeout
seconds, wheren
is the number of seed brokers.socket_timeout
sets the number of seconds to wait when reading from or writing to a socket connection to a broker. After this timeout expires the connection will be killed. Note that some Kafka operations are by definition long-running, such as waiting for new messages to arrive in a partition, so don't set this value too low. When configuring timeouts relating to specific Kafka operations, make sure to make them shorter than this one.
Producer timeouts can be configured when calling #producer
on a client instance:
ack_timeout
is a timeout executed by a broker when the client is sending messages to it. It defines the number of seconds the broker should wait for replicas to acknowledge the write before responding to the client with an error. As such, it relates to therequired_acks
setting. It should be set lower thansocket_timeout
.retry_backoff
configures the number of seconds to wait after a failed attempt to send messages to a Kafka broker before retrying. Themax_retries
setting defines the maximum number of retries to attempt, and so the total duration could be up tomax_retries * retry_backoff
seconds. The timeout can be arbitrarily long, and shouldn't be too short: if a broker goes down its partitions will be handed off to another broker, and that can take tens of seconds.
When sending many messages, it's likely that the client needs to send some messages to each broker in the cluster. Given n
brokers in the cluster, the total wait time when calling Kafka::Producer#deliver_messages
can be up to
n * (connect_timeout + socket_timeout + retry_backoff) * max_retries
Make sure your application can survive being blocked for so long.
By default, communication between Kafka clients and brokers is unencrypted and unauthenticated. Kafka 0.9 added optional support for encryption and client authentication and authorization. There are two layers of security made possible by this:
By enabling SSL encryption you can have some confidence that messages can be sent to Kafka over an untrusted network without being intercepted.
In this case you just need to pass a valid CA certificate as a string when configuring your Kafka
client:
kafka = Kafka.new(["kafka1:9092"], ssl_ca_cert: File.read('my_ca_cert.pem'))
Without passing the CA certificate to the client it would be impossible to protect against man-in-the-middle attacks.
If you want to use the CA certs from your system's default certificate store, you can use:
kafka = Kafka.new(["kafka1:9092"], ssl_ca_certs_from_system: true)
This configures the store to look up CA certificates from the system default certificate store on an as needed basis. The location of the store can usually be determined by:
OpenSSL::X509::DEFAULT_CERT_FILE
In order to authenticate the client to the cluster, you need to pass in a certificate and key created for the client and trusted by the brokers.
kafka = Kafka.new(
["kafka1:9092"],
ssl_ca_cert: File.read('my_ca_cert.pem'),
ssl_client_cert: File.read('my_client_cert.pem'),
ssl_client_cert_key: File.read('my_client_cert_key.pem'),
# ...
)
Once client authentication is set up, it is possible to configure the Kafka cluster to authorize client requests.
Typically, Kafka certificates come in the JKS format, which isn't supported by ruby-kafka. There's a wiki page that describes how to generate valid X509 certificates from JKS certificates.
Kafka has support for using SASL to authenticate clients. Currently GSSAPI, SCRAM and PLAIN mechanisms are supported by ruby-kafka.
In order to authenticate using GSSAPI, set your principal and optionally your keytab when initializing the Kafka client:
kafka = Kafka.new(
["kafka1:9092"],
sasl_gssapi_principal: 'kafka/[email protected]',
sasl_gssapi_keytab: '/etc/keytabs/kafka.keytab',
# ...
)
In order to authenticate using PLAIN, you must set your username and password when initializing the Kafka client:
kafka = Kafka.new(
["kafka1:9092"],
ssl_ca_cert: File.read('/etc/openssl/cert.pem'), # Optional but highly recommended
sasl_plain_username: 'username',
sasl_plain_password: 'password'
# ...
)
NOTE: It is highly recommended that you use SSL for encryption when using SASL_PLAIN
Since 0.11 kafka supports SCRAM.
kafka = Kafka.new(
["kafka1:9092"],
sasl_scram_username: 'username',
sasl_scram_password: 'password',
sasl_scram_mechanism: 'sha256',
# ...
)
The library has been designed as a layered system, with each layer having a clear responsibility:
- The network layer handles low-level connection tasks, such as keeping open connections to each Kafka broker, reconnecting when there's an error, etc. See
Kafka::Connection
for more details. - The protocol layer is responsible for encoding and decoding the Kafka protocol's various structures. See
Kafka::Protocol
for more details. - The operational layer provides high-level operations, such as fetching messages from a topic, that may involve more than one API request to the Kafka cluster. Some complex operations are made available through
Kafka::Cluster
, which represents an entire cluster, while simpler ones are only available throughKafka::Broker
, which represents a single Kafka broker. In general,Kafka::Cluster
is the high-level API, with more polish. - The API layer provides APIs to users of the libraries. The Consumer API is implemented in
Kafka::Consumer
while the Producer API is implemented inKafka::Producer
andKafka::AsyncProducer
. - The configuration layer provides a way to set up and configure the client, as well as easy entrypoints to the various APIs.
Kafka::Client
implements the public APIs. For convenience, the methodKafka.new
can instantiate the class for you.
Note that only the API and configuration layers have any backwards compatibility guarantees – the other layers are considered internal and may change without warning. Don't use them directly.
The producer is designed with resilience and operational ease of use in mind, sometimes at the cost of raw performance. For instance, the operation is heavily instrumented, allowing operators to monitor the producer at a very granular level.
The producer has two main internal data structures: a list of pending messages and a message buffer. When the user calls Kafka::Producer#produce
, a message is appended to the pending message list, but no network communication takes place. This means that the call site does not have to handle the broad range of errors that can happen at the network or protocol level. Instead, those errors will only happen once Kafka::Producer#deliver_messages
is called. This method will go through the pending messages one by one, making sure they're assigned a partition. This may fail for some messages, as it could require knowing the current configuration for the message's topic, necessitating API calls to Kafka. Messages that cannot be assigned a partition are kept in the list, while the others are written into the message buffer. The producer then figures out which topic partitions are led by which Kafka brokers so that messages can be sent to the right place – in Kafka, it is the responsibility of the client to do this routing. A separate produce API request will be sent to each broker; the response will be inspected; and messages that were acknowledged by the broker will be removed from the message buffer. Any messages that were not acknowledged will be kept in the buffer.
If there are any messages left in either the pending message list or the message buffer after this operation, Kafka::DeliveryFailed
will be raised. This exception must be rescued and handled by the user, possibly by calling #deliver_messages
at a later time.
The synchronous producer allows the user fine-grained control over when network activity and the possible errors arising from that will take place, but it requires the user to handle the errors nonetheless. The async producer provides a more hands-off approach that trades off control for ease of use and resilience.
Instead of writing directly into the pending message list, Kafka::AsyncProducer
writes the message to an internal thread-safe queue, returning immediately. A background thread reads messages off the queue and passes them to a synchronous producer.
Rather than triggering message deliveries directly, users of the async producer will typically set up automatic triggers, such as a timer.
The Consumer API is designed for flexibility and stability. The first is accomplished by not dictating any high-level object model, instead opting for a simple loop-based approach. The second is accomplished by handling group membership, heartbeats, and checkpointing automatically. Messages are marked as processed as soon as they've been successfully yielded to the user-supplied processing block, minimizing the cost of processing errors.
After checking out the repo, run bin/setup
to install dependencies. Then, run rake spec
to run the tests. You can also run bin/console
for an interactive prompt that will allow you to experiment.
Note: the specs require a working Docker instance, but should work out of the box if you have Docker installed. Please create an issue if that's not the case.
If you would like to contribute to ruby-kafka, please join our Slack team and ask how best to do it.
If you've discovered a bug, please file a Github issue, and make sure to include all the relevant information, including the version of ruby-kafka and Kafka that you're using.
If you have other questions, or would like to discuss best practises, how to contribute to the project, or any other ruby-kafka related topic, join our Slack team!
Version 0.4 will be the last minor release with support for the Kafka 0.9 protocol. It is recommended that you pin your dependency on ruby-kafka to ~> 0.4.0
in order to receive bugfixes and security updates. New features will only target version 0.5 and up, which will be incompatible with the Kafka 0.9 protocol.
Last stable release with support for the Kafka 0.9 protocol. Bug and security fixes will be released in patch updates.
Latest stable release, with native support for the Kafka 0.10 protocol and eventually newer protocol versions. Kafka 0.9 is no longer supported by this release series.
Currently, there are three actively developed frameworks based on ruby-kafka, that provide higher level API that can be used to work with Kafka messages and two libraries for publishing messages.
-
Racecar - A simple framework that integrates with Ruby on Rails to provide a seamless way to write, test, configure, and run Kafka consumers. It comes with sensible defaults and conventions.
-
Karafka - Framework used to simplify Apache Kafka based Ruby and Rails applications development. Karafka provides higher abstraction layers, including Capistrano, Docker and Heroku support.
-
Phobos - Micro framework and library for applications dealing with Apache Kafka. It wraps common behaviors needed by consumers and producers in an easy and convenient API.
-
DeliveryBoy – A library that integrates with Ruby on Rails, making it easy to publish Kafka messages from any Rails application.
-
WaterDrop – A library for Ruby and Ruby on Rails applications, to easy publish Kafka messages in both sync and async way.
There are a few existing Kafka clients in Ruby:
- Poseidon seems to work for Kafka 0.8, but the project is unmaintained and has known issues.
- Hermann wraps the C library librdkafka and seems to be very efficient, but its API and mode of operation is too intrusive for our needs.
- jruby-kafka is a great option if you're running on JRuby.
We needed a robust client that could be used from our existing Ruby apps, allowed our Ops to monitor operation, and provided flexible error handling. There didn't exist such a client, hence this project.
Bug reports and pull requests are welcome on GitHub at https://github.com/zendesk/ruby-kafka.
Copyright 2015 Zendesk
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.