In defence of Rails

I’ve been hearing a lot of sledging recently stating that design flaws in Rails cause it to scale poorly to large organisations. Highlights include that it’s slow to boot because it loads everything upfront; that Active Record models make it easy to create N+1 issues; and too much logic (including DB queries) end up in views.

When designing a framework, authors need to decide what they are optimising for. As is well described by Sid Sijbrandij in his article, Why We’re Sticking with Ruby on Rails at GitLab, DHH wanted to build something approachable but well structured.

I think it’s hard to deny that Rails is both easy to get started with and it allows you to get to market at lightning speed. Of course, being approachable comes with some compromises. Most of the “getting started” tutorials I’ve seen make a lot of choices I would never make in a production app. For example, people are generally taught to put their business logic into models. Despite the fact that some might consider me a geriatric engineer, I still remember the many, many years feeling overwhelmed with all there is to learn. Given this, it has always been my opinion that it’s best to hide as much complexity as possible until a newer developer has the space for the next concept.

This is not everyone’s opinion. Frameworks like Hanami have made the choice to drop you straight into the deep end. Their belief is that you will become a better developer if you have no choice but to do things “right”. There are also people who believe you should be forced to learn SQL very early on so that you know what your ORM is doing. While I respect their perspectives, my experience is that there is a finite amount of information that anyone can take in and trying to learn everything at once leads to someone who isn’t very good at anything; at least not for many years. You don’t become a great engineer from a few months of cramming; it’s a life of continual learning.

It’s certainly true to say that if you take the approachable version of Rails and scale it up to a large organisation, you’re going to have a bad time. The thing is, if someone has risen through to ranks to be in charge of engineering at a large organisation should know better. None of the complaints I’ve heard are enforced by Rails, they’re just approachable defaults that are very simple to change.

The following are some of the ways you can prepare Rails for scalability.

Boot time#

Eager loading#

In production, Rails will boot your whole application when it starts. This ensures your users aren’t made to wait while modules are booted during requests. The default is to disable this in development and only enable it in test when being run on CI.

Despite the defaults, I’ve seen apps with eager loading enabled for local testing which means that even running a single unit test requires booting the whole application.

It will depend on your application as to how much saving you’ll get from disabling eager loading but, in one of my applications, the difference in boot time is about 20%.

Rails helper#

If you use RSpec, you’ll know that it has a spec_helper.rb which contains the settings needed to run tests on plain old Ruby objects but there is also a rails_helper.rb which handles booting Rails and any settings that might be needed to test Rails objects.

It’s easy and convenient to place everything you need into the Rails helper, but you need to remember that everything you put in there will be loaded regardless of whether it’s required for every spec.

When I optimised one Rails helper so it contained only what was needed for every spec, it reduced boot time by 25%. The trade-off is that you will need to load the additional settings when they are required but if you can’t get test feedback in a couple seconds (at worst) the feedback loop becomes too long. You should want to run your tests frequently.

Initializers#

Even if you have eager loading disabled, it’s likely that you are booting gems, regardless of whether they are required, simply because you have an initializer setting some config. Initializers is one area I haven’t seen a clear solution for and so, while I have ideas, I’m keen to crowdsource opinions.

In one application where boot times were getting a bit long, I simply disabled any initializer that wasn’t required in development or test. In those cases, I’m using mocks and so there was no point in booting the gem when it would never be used. This reduced boot time by about a third.

I’ve been doing some work in Hanami 2 and it leans very heavily on dependency injection. Any class that is a child of a Hanami class can include Deps, passing in a list of dependencies. While it’s a different use case, it has me wondering if a similar convention for lazy loading gem initalizers could be valuable. Given the added verbosity, most likely you’d want to check the real impact on boot time per application and decide if it’s worth it. I have typically found that initializers can be a big culprit for slow boot times but other approaches might be better suited to your app.

If you have other ideas, I’d love to hear them.

Minimal#

If you’re building a small application with minimal dependencies, e.g., a microservice, you probably don’t need to bring in all of Rails. When you create a Rails app, you can skip anything you don’t need or, if you really only want Rails to provide some structure and consistent conventions, you can create a minimal build.

rails new --minimal --skip-active-record --skip-asset-pipeline --skip-test my_application_name

That command will leave you with models, views, and controllers but none of the extras you would normally expect. If you’d like to include some extras, run rails help and curate your own selection.

If you have an existing application with dependencies you don’t need, you can easily disable them in application.rb. In the case of something like Active Record, remember to also remove the database dependency from you Gemfile.

In a tiny application, you may not see huge gains from this thanks to lazy loading in test and development and the caching provided by Spring. Saying this, it can be valuable to keep a microservice clean and focused.

Code cleanliness#

ViewComponent#

Created by GitHub, ViewComponent is “a framework for creating reusable, testable & encapsulated view components, built to integrate seamlessly with Ruby on Rails.”

Essentially, it solves the issue of business logic and DB queries in the view. Components include a Ruby object for your logic, a test/spec for testing both your logic and your component’s view in isolation, a template, and optionally, component specific assets. This makes components completely self-contained and highly testable. Since you can unit test your logic without rendering the view, it makes your tests substantially faster. It doesn’t remove the need for rendered tests but it allows you to minimise them and render only that component, not the whole page.

Command objects#

A common way to organise business logic is the command/query pattern. Commands are classes that handle a single task that changes state and they should be given a name that reflects that action (e.g., CreatePost or CalculateCharge).

It’s easy enough to create your own implementation of the pattern but there are also helpful gems that can take away the boilerplate. My preference is ActiveInteraction. ActiveInteraction is an implementation of the command pattern, but they refer to commands as “interactions”. You, however, can use whatever name suits you.

I appreciate that interactions have typed inputs, but it also utilises Active Model Validations which helps maintain a consistent validation style across your application. When you have a hash input, you can choose to fully type the hash values and let ActiveInteraction strip out any unapproved keys.

The ActiveInteraction::Extras gem includes a whole range of valuable helpers including the ability to offload handling of Strong Parameters to your interactions. Handling parameter filtering in my controllers has never felt quite right to me. I like my controllers to be simple paper-pushers with all the implementation details hidden elsewhere. When I’m trying to understand some code, I want to be able to quickly skim through an overview of the process and only dig into the implementation details if it’s required.

Because ActiveInteraction embraces the Active Model conventions that you’re used to, an instance of an interaction has methods like valid? and any validation errors are added to an ActiveModel::Errors object available by calling errors, just like when you work with an Active Record model.

Every interaction has an execute method as the entry point and you commence the interaction by sending either a run or run! message. Calling run will return you the instance with success?, failure?, and result methods. run!, on the other hand, returns the result or raises an exception if there is an error. Interactions are also highly composable and, if you are using run, errors are passed back up the chain so you can deal with them gracefully.

Leaning into that composability allows you to ensure that each interaction is only ever handing a single responsibility. I’ve found the combination of the interaction interface and maintaining single responsibilities has heavily simplified my testing. In fact, if I ever think “I wonder how I can test this”, it’s almost certainly a sign that there is a missing abstraction.

Query objects#

Queries encapsulate your interactions with the database without fattening up your model. Unlike commands, queries shouldn’t have side effects and, when using Active Record, they should generally be composable.

It was great to see Thiago Araújo Silva from thoughtbot considering ways to blend the value of Active Record scopes with query objects. Whether or not you agree with their pattern, the article is great for driving the kinds of discussions that can uncover what is important to your organisation.

Strict loading#

A common complaint with Active Record is that it’s easy to create N+1 query issues. There multiple gems designed to uncover these—the most popular of which is Bullet—but the only thing better than fixing N+1 queries is not creating them in the first place. This is where strict loading comes into play.

By adding strict_loading: true to your Active Record associations, Rails will not allow you to call upon any association that hasn’t been eager loaded. For example:

class Post < ApplicationRecord
  has_many :comments, strict_loading: true
end

Post.last.comments.each
# ActiveRecord::StrictLoadingViolationError: `Post` is marked for
# strict_loading. The `Comment` association named `:comments` cannot be lazily
# loaded.

Post.includes(:comments).last.comments.each
# This is fine

I quite often use the Rails console to get a feel for the data I’m working with and find strict loading to get in the way so I have it enabled everywhere except in the console.

class Post < ApplicationRecord
  has_many :comments, strict_loading: strict_loading?
end

class ApplicationRecord < ActiveRecord::Base
  self.abstract_class = true

  def self.strict_loading?
    $PROGRAM_NAME != "rails_console"
  end
end

If there is a reason why enabling strict loading at the model level won’t work for you but you’re debugging an N+1, you can also enable it just for one query. If you don’t have the strict structures discussed above in place, it may be hard to track down where the N+1 is occurring. Enabling strict loading at the entry point should cause and exception at the N+1 which (hopefully) you can uncover with a simple run of your test suite.

post = Post.strict_loading.last
...
post.comments.each
# ActiveRecord::StrictLoadingViolationError: `Post` is marked for
# strict_loading. The `Comment` association named `:comments` cannot be lazily
# loaded.

Domains#

As an application grows, it can be difficult to navigate and understand the boundaries if all your files sit alongside each other. This becomes even more of a concern when there are many teams that each have ownership of different parts of the project.

While some organisations use tooling to map teams to files, I prefer to separate my application into clearly defined domains. This, of course, assumes that your organisation assigns team to domains.

At first, it might seem that you need to follow the Rails convention of placing your models, views, and controllers in the automatically generated directories under app, but there is no reason why you can’t have a set of those directories within every domain. For example:

  • app
    • domains
      • first_domain
        • assets
        • controllers
        • models
        • views
      • second_domain
        • assets
        • controllers
        • models
        • views

The compromise is a bit of verbosity since all your classes will be namespaced. Having been exposed to years of Java development, I don’t accept verbosity without careful consideration, but I find that this allows me to navigate an application with more clarity and it’s easy to see the boundaries of each domain. If your domains also map to teams, you know immediately what you are free to modify and what is going to require consultation.

Coupling between modules#

Even with your application grouped by domain, you might find that you are lacking clear boundaries. If domains are allowed to call any class within another domain, the domains become tightly coupled and velocity plummets as teams need to coordinate on changes. Smaller organisations can often get away with enforcing conventions but, in larger organisations, it might be necessary to lean upon tooling.

Shopify’s solution was to create the packwerk gem which enforces strict boundaries between code groups. You categorise your classes as either public or private. Classes within the group can send each other messages but classes from other groups can only interface with public classes. By creating these boundaries, you get all the benefits of a monolith without hindering productivity.

While I support certain use cases for microservices, I’m convinced that most of what has been driving the movement is organisations creating a big ball of mud and then blaming the monolith for their lack of application architecture.

Another solution I’ve heard of is splitting your monolith into multiple Rails Engines. With separate Engines, it’s impossible not to see the boundaries of responsibility and there is increased scope for teams to have differences in their styles. On the other hand, it does increase complexity when you need to communicate between domains.

Conclusion#

Even though I’m a big fan of the Rails way, I’m excited that projects like Hanami offer a different methodology for those with differing views. There is plenty of space for differing styles and there are some interesting patterns in there that we can all be inspired by.

I would hate for anyone working on a sub-optimal Rails project to believe that is just what working with Rails is like. As someone who is outcome driven, I get a lot of joy working on a well optimised Rails project. Importantly, I’ve been able to accomplish more with scarce resources than many large, well-funded teams by leaning into the simplicity but applying the appropriate guard rails.

If you’re working on a project with some of these concerns, you can do something about it. If you’re dealing with years of neglect, it might take time, but you can chip away at the problem and make your life continually better. Ruby is all about developer happiness, after all.