Ruby

Alternative Service Communication Using Pub/Sub

The HTTP protocol was designed for synchronous communication between two entities — for instance, a browser requesting a stylesheet or a server charging with a payment processor. Those are synchronous operations where nothing can proceed without an immediate response.

Often communication can be asynchronous, like when queueing up work to be performed in the background. It is possible to use HTTP for these lengthy requests via long polling or server push, but it belies the strengths of HTTP. Long-held direct HTTP requests don’t scale beyond single server-to-server communication, and they require greater resources as service scales.

There doesn’t always need to be direct communication between two services. Message queues and distributed logging are two ways to accomplish fully asynchronous, many-to-many communication, but there is a more fundamental pattern available. The Publish/Subscribe pattern, also known as Pub/Sub, is the most basic flavor of event-based messaging. It allows messages to be broadcast out to other services which may or may not be listening. Those services are free to pick up communication events they know how to respond to and may respond with asynchronous communications of their own.

Getting to Know Pub/Sub

[Services] should be able to register to receive only the events they need, and never be sent the events they don’t need. We don’t want to spam our [services]!

The Pragmatic Programmer—Dave Thomas & Andy Hunt

A Pub/Sub provider brokers messages between publishers and subscribers. Published messages are characterized into channels, which can be as broad or narrow as necessary. Subscribers register to listen to one or more channels and then receive published messages only for those channels. Subscribers have no knowledge of publishers, and publishers aren’t aware of subscribers. This decoupling enables a dynamic network topology that can scale beyond processes and applications, well into the realm of multiple services.

Pub/Sub is most fitting when there are no limits on the number of subscribers. In this situation, an unknown number of of subscribers may respond to any messages. This makes it unsuitable for queue-like operations, where an event must be popped off and handled exactly once. However, it is perfectly suited to synchronization operations, where a central publisher wishes to replicate data in parallel to multiple subscribers.

Pub/Sub as a means of distributed communication is simple to understand and simple to implement. It just so happens that Redis has an outstanding, albeit simple, Pub/Sub implementation. Chances are, Redis is already in your stack, so there isn’t a barrier to using it. No additional services are required and the overhead is low.

Warming Up with the Redis CLI

Let’s make all this abstract talk of Pub/Sub more concrete. We’ll replicate simple interprocess communication over Pub/Sub using the redis-cli. Assuming you have Redis installed, open two separate terminal sessions. In the first session use the SUBSCRIBE command to bind to the foo channel:

$ redis-cli SUBSCRIBE foo
Reading messages... (press Ctrl-C to quit)
1) "subscribe"
2) "foo"
3) (integer) 1

The subscribe command is blocking, so the process will wait for any published events. In the other session use the PUBLISH to publish an event on the foo channel:

$ redis-cli PUBLISH foo "bar"
(integer) 1

Back in the original process you’ll see the message was passed through:

1) "message"
2) "foo"
3) "bar"

The message comes through with three arguments: command type, channel name, and the payload. Any processes that subscribe to the same channel will receive identical messages. This simple message passes scales beyond processes and into disparate servers, enabling basic distributed messaging.

Now we’ll move publishers and subscribers into Ruby to synchronize data between a central source and multiple subscribers.

!New Call-to-action

Publishers and Subscribers in Ruby

Simulating publish and subscribe behavior from within the same Ruby process is a tricky business, requiring threads and passing behavior. Rather than go down that rabbit hole we’ll use two separate processes for demonstration. First, a wrapper around channel subscription:

# subscriber.rb
require 'redis'

class Subscriber 
  def subscribe(channel) 
    Redis.current.subscribe(channel) do |on| 
    on.subscribe do |channel, _| 
    puts "subscribed to #{channel}" 
    end

    on.message do |channel, message|
    puts "#{channel} received #{message}"
    end
    end
  end 
end

The Subscriber class in this example is only a thin wrapper around Redis#subscribe, which binds to the specified channel and then writes out published messages. The wrapper around Redis#publish is even simpler:

# publisher.rb
require 'redis'

class Publisher 
  def publish(channel, message) 
    Redis.current.publish(channel, message) 
  end 
end

Just as we used one terminal session with redis-cli before, now we’ll start a subscribe process in one session and publish to it from another.

$ ruby -rSubscriber -e "Subscriber.new.subscribe('custom/channel')"
> subscribed to custom/channel

The process will block while it waits for messages to arrive. Send a message to the “custom/channel”:

$ ruby -rPublisher -e "Publisher.new.publish('custom/channel', 'hello there')"

It is written out in the first session:

> custom/channel received hello there

And that’s all there is to Pub/Sub between processes in Ruby. To make use of Pub/Sub for data synchronization we must layer predictable channels and structured messages on top of simple messaging.

Structuring Channels and Messages

The previous toy examples leave out all of the details necessary to actually fold distributed messaging into an application. What you need, at minimum, are conventions for setting predictable channel names. Channel names that compose to become more specific are ideal. For example, in an application where readers can leave comments on a post, you may have a channel named posts/1 and another channel for comments named posts/1/comments. Updates to the post itself are broadcast to the posts/1 channel, while new comments are sent to the posts/1/comments channel.

Message events also need to be structured in a predictable format. Any transport that would be used for HTTP is fitting for messages, but JSON is the obvious choice. Continuing the comment synchronization example above, here is an example of a “comment updated” payload:

json
{
  action: "updated",
  channel: "posts/1/comments",
  meta: {},
  model: {
    "id": 100,
    "body": "This is my new comment"
  }
)

Just like any API, or the canonical JSON-API, payload consistency is key. When designing communication you must be aware that, unlike HTTP, there aren’t any headers or versions, the only means of filtering messages is by channel name.

Subscribing To Events Between Languages

Pub/Sub is a bare-bones protocol that is entirely language agnostic. Every major language has a usable Redis library, so the possibilities are wide open. As with any multi-service architecture, there are opportunities to lean on other languages and frameworks where your primary platform may fall short.

For example, Ruby is notoriously bad at maintaining multiple concurrent connections. That’s a perfect opportunity to write a small service in a platform like Node.js, which can hold long-standing connections with browsers. The primary application then broadcasts events out to each service, which can then relay them to the browser.

Even a framework like Phoenix, built in the highly asynchronous Elixir language, relies on Redis’ Pub/Sub to expand websocket support beyond a single web host.

The Final Countdown

Pub/Sub makes one-to-many and many-to-many communication between services possible where HTTP falls flat. Allowing applications to broadcast events to arbitrary services is a path to highly scalable, fault tolerant, and loosely coupled services.

It must be noted that Redis’ implementation has notable caveats. Unlike some other implementations, such as ZeroMQ, Redis has no guarantees on message delivery, no acknowledgements, and no persistence in the event of service or network failure. Its triumph is simplicity and ubiquity, but you’ll want to look further into the world of Pub/Sub if persistence and reliability are paramount.

Reference: Alternative Service Communication Using Pub/Sub from our WCG partner Florian Motlik at the Codeship Blog blog.

Parker Selbert

Parker is one half of Soren, a partnership that helps businesses make their production Rails apps better.
Subscribe
Notify of
guest

This site uses Akismet to reduce spam. Learn how your comment data is processed.

0 Comments
Oldest
Newest Most Voted
Inline Feedbacks
View all comments
Back to top button