When developing textris gem, I've stumbled upon an issue with handling optional dependencies. For example, you may want to support features of other gem in yours, but only when gem user actually uses that other gem in his project. This is important, as all dependencies required by your project will be installed by bundler in user's project.
Let's look at my case from textris. I wanted to support delaying texter invocation with sidekiq, just like the one sidekiq provides out of the box for ActionMailer. Still, installing sidekiq for every project that just wants to send some SMSes without touching background jobs is not a good idea, to say the least. Note that sidekiq depends on redis server, so should I just advice every textris user to install redis in his system?
Unfortunately, there's no stock solution for this in bundler and
Gem::Specification builder. So what can you do?
Rescue to the rescue
Ruby community has came up with the following pattern to solve the problem:
begin require 'twilio-ruby' rescue LoadError # handle lack of library or just leave it be end
Yes, this is how most (if not all) gems out there do it. Of course, there are many things that may happen in the rescue blocks.
Let's do nothing
Sometimes you don't need to do anything in rescue block. That's the case with
twilio-ruby dependency of textris. The
twilio-ruby gem is only used in the Twilio delivery class which will never be invoked unless user explicitly configures textris to deliver via Twilio. And if that's the case, there's no reason to avoid the undefined constant exception.
But then again, why require
twilio-ruby if undefined constant error is not an issue? For example, to actually require the gem in tests and verify the collaboration of own gem with another. Another reason may be a need to require a file that is a part of another gem that doesn't get required by bundler automatically, i.e. user has something like this in her
gem 'some_gem', :require => false
Requiring optional dependencies from within the main gem file may be also the only way to have an actual place in code where you state the optional dependencies, as you cannot put them into the gemspec (which really seems like a necessary addition to bundler todo list).
Let's make some noise
Many libraries choose to print a warning message in the rescue block in order to make it clear to the developer that there's something wrong. Here's an example taken from the carrierwave_securefile gem:
begin # require aes require 'carrierwave/securefile/aes_file.rb' rescue LoadError puts "WARNING: Failed to require aes_file or openssl, AES encryption may fail!" end
Personally I don't like the idea. I think that if dependency is important enough to bother user with errors on every application load, then it should be added as runtime dependency and just install along with gem. And if it's not required then why bother the user at all? Still, I can think of some scenarios when warnings might be useful:
- if you support multiple alternative libraries for the same functionality (like supporting either sidekiq or resque for delayed jobs), you may have to react somehow if none of them is present in user's bundle
- if specific version of dependent library that is present in user's bundle has some compatibility problems with your code, you may want to encourage the user to change the library version she bundles
Sometimes you may want to offer some replacement functionality for your code when a specific gem is not available. That was the case with textris and sidekiq. I couldn't allow loading my sidekiq related classes when sidekiq is missing because I inherit from its classes, so there would be a constant missing error during textris gem initialization. Also, simply not loading these classes would result in lack of some textris documented constants at runtime which would be confusing to the user. You can fix this like I did by using one of ruby's sweet object reopening & mixing mechanics like the one below:
begin require 'sidekiq' rescue LoadError require 'textris/delay/sidekiq/missing' Textris::Delay::Sidekiq.include(Textris::Delay::Sidekiq::Missing) else require 'textris/delay/sidekiq' require 'textris/delay/sidekiq/proxy' require 'textris/delay/sidekiq/serializer' require 'textris/delay/sidekiq/worker' end
What does this do? If sidekiq is present, I load all regular classes belonging to the
Textris::Delay::Sidekiq module (the
else block). But if there's no sidekiq in the project, I load a fallback code kept in separate file and include it in the
Textris::Delay::Sidekiq module which is empty at the time because the regular classes were not loaded. This fallback code is actually very simple, it just throws an appropriate error if any of delay methods gets invoked by the user who doesn't have sidekiq in her bundle yet.
Fallback code injection allows to inform user about the missing dependency at appropriate time and in appropriate way. It also becomes handy in tests, which should describe both the "gem present" and "gem missing" scenarios. That's how you handle both scenarios:
- gem present: as gem is added via
add_development_dependencyto the gemspec and main gem file requires it, you can test this scenario without further effort
- gem missing: you can
remove_constto unload dependent gem constants and invoke the code from load rescue manually to inject the fallback code for proper testing
I hope gemspec will be extended with ability to add optional dependencies, if not for any other purpose then just to present them on the gem's RubyGems page. Honestly, I have no idea on a more elegant solution for this pattern than the rescue of dependency loading, but I'm sure there is some out there waiting to be popularized. Until then, the solution described in this article should be fine.