debian-mirror-gitlab/doc/development/module_with_instance_variables.md

stage	group	info
none	unassigned	To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/engineering/ux/technical-writing/#assignments

Modules with instance variables could be considered harmful

Background

Rails somehow encourages people using modules and instance variables everywhere. For example, using instance variables in the controllers, helpers, and views. They're also encouraging the use of ActiveSupport::Concern, which further strengthens the idea of saving everything in a giant, single object, and people could access everything in that one giant object.

The problems

Of course this is convenient to develop, because we just have everything within reach. However this has a number of downsides when that chosen object is growing, it would later become out of control for the same reason.

There are just too many things in the same context, and we don't know if those things are tightly coupled or not, depending on each others or not. It's very hard to tell when the complexity grows to a point, and it makes tracking the code also extremely hard. For example, a class could be using 3 different instance variables, and all of them could be initialized and manipulated from 3 different modules. It's hard to track when those variables start giving us troubles. We don't know which module would suddenly change one of the variables. Everything could touch anything.

Similar concerns

People are saying multiple inheritance is bad. Mixing multiple modules with multiple instance variables scattering everywhere suffer from the same issue. The same applies to ActiveSupport::Concern. See: Consider replacing concerns with dedicated classes & composition

There's also a similar idea: Use decorators and interface segregation to solve overgrowing models problem

Note that included doesn't solve the whole issue. They define the dependencies, but they still allow each modules to talk implicitly via the instance variables in the final giant object, and that's where the problem is.

Solutions

We should split the giant object into multiple objects, and they communicate with each other with the API, i.e. public methods. In short, composition over inheritance. This way, each smaller objects would have their own respective limited states, i.e. instance variables. If one instance variable goes wrong, we would be very clear that it's from that single small object, because no one else could be touching it.

With clearly defined API, this would make things less coupled and much easier to debug and track, and much more extensible for other objects to use, because they communicate in a clear way, rather than implicit dependencies.

Acceptable use

However, it's not always bad to use instance variables in a module, as long as it's contained in the same module; that is, no other modules or objects are touching them, then it would be an acceptable use.

We especially allow the case where a single instance variable is used with ||= to set up the value. This would look like:

module M
  def f
    @f ||= true
  end
end

Unfortunately it's not easy to code more complex rules into the cop, so we rely on people's best judgement. If we could find another good pattern we could easily add to the cop, we should do it.

How to rewrite and avoid disabling this cop

Even if we could just disable the cop, we should avoid doing so. Some code could be easily rewritten in simple form. Consider this acceptable method:

module Gitlab
  module Emoji
    def emoji_unicode_version(name)
      @emoji_unicode_versions_by_name ||=
        JSON.parse(File.read(Rails.root.join('fixtures', 'emojis', 'emoji-unicode-version-map.json')))
      @emoji_unicode_versions_by_name[name]
    end
  end
end

This method is totally fine because it's already self-contained. No other methods should be using @emoji_unicode_versions_by_name and we're good. However it's still offending the cop because it's not just ||=, and the cop is not smart enough to judge that this is fine.

On the other hand, we could split this method into two:

module Gitlab
  module Emoji
    def emoji_unicode_version(name)
      emoji_unicode_versions_by_name[name]
    end

    private

    def emoji_unicode_versions_by_name
      @emoji_unicode_versions_by_name ||=
        JSON.parse(File.read(Rails.root.join('fixtures', 'emojis', 'emoji-unicode-version-map.json')))
    end
  end
end

Now the cop doesn't complain. Here's a bad example which we could rewrite:

module SpamCheckService
  def filter_spam_check_params
    @request            = params.delete(:request)
    @api                = params.delete(:api)
    @recaptcha_verified = params.delete(:recaptcha_verified)
    @spam_log_id        = params.delete(:spam_log_id)
  end

  def spam_check(spammable, user)
    spam_service = SpamService.new(spammable, @request)

    spam_service.when_recaptcha_verified(@recaptcha_verified, @api) do
      user.spam_logs.find_by(id: @spam_log_id)&.update!(recaptcha_verified: true)
    end
  end
end

There are several implicit dependencies here. First, params should be defined before use. Second, filter_spam_check_params should be called before spam_check. These are all implicit and the includer could be using those instance variables without awareness.

This should be rewritten like:

class SpamCheckService
  def initialize(request:, api:, recaptcha_verified:, spam_log_id:)
    @request            = request
    @api                = api
    @recaptcha_verified = recaptcha_verified
    @spam_log_id        = spam_log_id
  end

  def spam_check(spammable, user)
    spam_service = SpamService.new(spammable, @request)

    spam_service.when_recaptcha_verified(@recaptcha_verified, @api) do
      user.spam_logs.find_by(id: @spam_log_id)&.update!(recaptcha_verified: true)
    end
  end
end

And use it like:

class UpdateSnippetService < BaseService
  def execute
    # ...
    spam = SpamCheckService.new(params.slice!(:request, :api, :recaptcha_verified, :spam_log_id))

    spam.check(snippet, current_user)
    # ...
  end
end

This way, all those instance variables are isolated in SpamCheckService rather than whatever includes the module, and those modules which were also included, making it much easier to track down any issues, and reducing the chance of having name conflicts.

How to disable this cop

Put the disabling comment right after your code in the same line:

module M
  def violating_method
    @f + @g # rubocop:disable Gitlab/ModuleWithInstanceVariables
  end
end

If there are multiple lines, you could also enable and disable for a section:

module M
  # rubocop:disable Gitlab/ModuleWithInstanceVariables
  def violating_method
    @f = 0
    @g = 1
    @h = 2
  end
  # rubocop:enable Gitlab/ModuleWithInstanceVariables
end

Note that you need to enable it at some point, otherwise nothing below that point is checked.

Things we might need to ignore right now

Because of the way Rails helpers and mailers work, we might not be able to avoid the use of instance variables there. For those cases, we could ignore them at the moment. Those modules are not shared with other random objects, so they're still somewhat isolated.

Instance variables in views

They're bad because we can't easily tell who's using the instance variables (from controller's point of view) and where we set them up (from partials' point of view), making it extremely hard to track data dependency.

We're trying to use something like this instead:

= render 'projects/commits/commit', commit: commit, ref: ref, project: project

And in the partial:

- ref = local_assigns.fetch(:ref)
- commit = local_assigns.fetch(:commit)
- project = local_assigns.fetch(:project)

This way it's clearer where those values were coming from, and we gain the benefit to have typo check over using instance variables. In the future, we should also forbid the use of instance variables in partials.