Apache Kafka is a framework for streaming data between internal systems, and Confluent offers Kafka as a managed service.
Kafka is new, but it’s getting pretty popular: managed service provider Confluent, founded by the original creators of Kafka, filed their S-1 last week.
Kafka exists to solve two fundamental problems facing almost every data infrastructure team at every company.
As storage has gotten cheaper, we’ve been collecting more and more data. Most software companies record every single website visit and click, and some go even deeper. Once you have more than few users interacting with your product, you’re talking about millions of different events per day. Storing and managing that size and velocity of data is hard.
Even if you’re a wiz at collecting and storing your data, there’s a problem: you’re going to need to move it around for it to be valuable. Where data gets initially collected and stored is rarely where it’s going to be useful.
Let’s use the internal systems we had at DigitalOcean as an example. When users interacted with our product, that would generate an event:
resource_destroyed, etc. Those events needed to be routed to multiple systems: our analytics team needed them in a warehouse for dashboarding, and our billing team needed them in the billing system to track and generate invoices.
Moving this data on a consistent basis and schedule is actually really hard, especially after the fact.
A few engineers were dealing with these two problems at LinkedIn back in 2011, and developed Kafka to solve them (among other things). Kafka is a distributed messaging system – a bunch of jargon that means that it makes it easy to collect and move data around your organization. Kafka acts as a central bus station for your data, ingesting it from different sources (your site, your app, Salesforce, etc.) and distributing it to all of the destinations it needs to get to.
Let’s dive into how Kafka works. Because that’s what you’re here for.
Kafka works through what engineers call a publish-subscribe model, often shortened to pub-sub (to save 10 milliseconds pronouncing it). It’s an architecture for sharing data and tasks across an organization through messages. To understand what that means, think about starting a newsletter. But in reality, definitely don’t think about starting a newsletter.
If you want to get people to read your newsletter, there are basically two ways you can go about distributing it.
You can wait for someone to ask you for a particular issue of your newsletter, and then send it over via email. You wait for their request (“I want to read this”) and then send back your response (“here’s the newsletter”). In software, this is called the request-response model.
Waiting for people to ask for your content is silly and inefficient. Instead, what everyone does is publish it widely on the web, and then anyone who wants can sign up to get it. You publish (“here’s the newsletter”) and then others subscribe (“I want to read this”), hence pub-sub. The order of operations is reversed.
There’s a lot more to the distinction between request-response and pub-sub, and it can be hard to understand the first time around. To keep it simple, just keep your eye on the goal of pub-sub: to make sure data gets to where it needs to go easily.
Kafka is a system for implementing pattern #2: you publish your data to a central location (Kafka), and then destinations (a data warehouse, another app, your billing system) can subscribe and consume that data as it comes in. Data only gets retained for a limited period of time: those destinations need to take that data and use it, or it gets thrown out to make room for new data coming in.
One thing worth noting: Kafka is not the only piece of software that you can use to build pub-sub systems, and it’s definitely not the first or most popular. RabbitMQ, for example, has been around for almost a decade. What makes Kafka unique is how powerful it is and how it helps shepherd data to where it needs to go.
You could also make a reasonable case that Segment is a Kafka competitor. They orient themselves towards different target customers, but the underlying tech is very similar; in fact, the Segment engineering team wrote a technical blog post about how they built what basically appears to be Kafka-lite.
Kafka is a piece of software, but it’s not something you just plop into your codebase and it solves your problem. You need to *orient your architecture around it* – it’s a helpful tool for you to implement a specific type of data infrastructure, but it’s only a piece of the puzzle.
Now that we’ve got the basics down, let’s get into how people actually use Kafka. There’s special terminology for the different types of system elements that we’ve spoken about. Because it wouldn’t be software if it wasn’t confusing.
This might be your website producing click data, or your app producing product data. Producers do the publishing, like you producing your newsletter(s).
This might be your data warehouse storing that click data, or your CRM storing customer details. Consumers subscribe, like the homies subscribing to and reading your newsletter(s).
If you produce multiple newsletters (one called Dank Memes Weekly and another called Shitposting Quarterly), this corresponds to an individual one of them. Producers and consumers can produce and subscribe to a topic.
This is the lowest level of data, and corresponds to an individual issue of an individual newsletter in our (mediocre) example.
To put it all together in a sentence: producers write messages to topics that consumers subscribe to. Phew. Part of why pub-sub is important is that multiple consumers can subscribe to a topic; that means that you only need to send data once, even if multiple places need access to it.
The last important thing to know about Kafka – and part of why it can be complex to set up – is that it’s built as a distributed system by default. In order to handle crazy throughput - data coming in very, very quickly – Kafka is usually deployed across multiple servers in a data center. Networking these servers together and making sure they’re in sync is pretty difficult, and it’s part of why managed service providers like Confluent are printing money.
(this section is usually reserved for paid subscribers, but what the hell, have fun)
Confluent provides Apache Kafka as a managed service. What does that mean, you ask? If you’re a developer, using Confluent means that you get to use Kafka without having to set it up on your own servers. Confluent hosts Kafka for you, and you pay on a usage basis. The company was founded by the team of engineers at LinkedIn who originally built and open-sourced Kafka, so they know what they’re doing.
Jay Kreps, one of the Confluent founders, wrote this lengthy post explaining logs and why he thinks software is moving in an event-driven direction. It’s technical, but highly worth a read if you can.
Confluent is a classic open source story, and fits into the same category as Elastic, Redis Labs, MongoDB, etc. It goes like this:
When I say hosting, that doesn’t necessarily mean that you’re abdicating all control of your environment to Confluent – you can also self host.
In addition to managing your Kafka deployment for you, Confluent has also built a nice series of plugins (they call it a “hub”) for getting data in and out of Kafka. They’ve also been putting muscle behind ksqlDB, a new product built on a top of Kafka that helps companies build applications that require real-time streaming (i.e. real time fraud detection).
Confluent’s pricing is predictably confusing. If you’re using their cloud product, there are 5 line items you need to calculate to understand what your monthly bill is going to be. Here’s a screenshot of their pricing calculator that I configured to use a standard multi-zone setup on AWS:
The absolute number here isn’t that important (managed services are expensive, and that’s fine) – it’s more about this pricing model optimizes for flexibility over simplicity. The first time I checked out Confluent – maybe about 2 years ago or so – they were charging simply for data in and data out. Easy. But as companies move upmarket, pricing models are destined to follow. And that’s why they need a calculator now.