When open source has less benefits
Here at Shopify, I work as a Production Engineer on the Jobs team. Our mission is to maintain and improve the background job infrastructure for Shopify Core, one of the world’s most significant Ruby on Rails applications.
Let’s start with some context about Background Jobs, especially in the context of ruby applications. When talking about jobs, we refer to those units of work that are important for the application to function but would take too long to be processed within the lifetime of a web request. The usual way of dealing with those units is to offload them to the background process, aka “background jobs,” and for the background process to pick them and execute them.
In Ruby, there are many popular options for dealing with jobs. The most commonly used one is Sidekiq and we have others like Resque, Delayed job, Sucker punch, etc.
This post will not focus on how those libraries work or their differences, but rather why, at Shopify, we decided to build our internal library for dealing with jobs.
To provide a little more context, at Shopify, in the beginning, we were using the Delayed job, which Tobias Lütke, the current CEO of Shopify, created. The backend used to store jobs was MySQL, and at some point, MySQL couldn’t keep up with the amount of MySQL writes. Later Github released Resque in November of 2009, which uses Redis as the backend to store jobs. At that time, it was the only option that could support the requirements of Shopify. Since then, we have added functionality on top of Resque with external gems like resque-scheduler or resque-pool.
Of course, the application’s requirements constantly change, and the library you are using might not evolve in a way that would satisfy your new needs.
Throughout the years of Shopify, the complexity of our monolith kept increasing as our needs kept evolving. As a result, we introduced sharding and new sophisticated job primitives. For many years,
resque and the other gems helped to solve these problems, but those libraries could no longer provide all the functionality that Shopify needed. We had to regularly hack or patch the library to keep providing the required functionality.
For the maintainers of this system, this was not ideal. We worked with at least three
resque-* gems and dozens of custom modules that provided extra functionality. This situation was the cause of frustration, production errors, and a complex enough system that was difficult to grasp in one’s head. Most critically, adding new features took time and effort.
At the beginning of this year, my team decided that we would aggregate all the resque code and all the custom functionality we have built over the years and place everything under one library.
After four months of work, we are currently running all jobs from the monolith with our library, and we have ported all previous patches into our library at the time this post is published.
This decision has allowed the team to work on major refactors to the internals of the job infrastructure and adapt to new requirements that appeared throughout the year.
Thanks to that decision, adding new features was smooth, and the team’s confidence in the code base has increased substantially; more importantly, we are in a much better place to keep improving the job infrastructure.
Here at Shopify, we value open-source projects and contribute to them as much as possible. But even though that is the mentality at Shopify, there are occasions where owning all of the code has more benefits in the long term.
The conclusion I wanted to leave here is that if you or your team find yourselves often working around library functionality, finding hacky ways to extend it or changing the core functionality of those libraries, sometimes it’s better to rethink the approach and question what you need from them. If you need only 20% and end up hacking the rest of 80%, you might benefit from owning all of the code.
If you have any thoughts or questions, please share them, and I will be happy to answer in the comments.