ABaldwinHunter / suture

A gem that helps people refactor or reimplement legacy code

Suture

A refactoring tool for Ruby, designed to make it safe to change code you don't confidently understand. In fact, changing untrustworthy code is so fraught, Suture hopes to make it safer to completely reimplement a code path.

Suture provides help to the entire lifecycle of refactoring poorly-understood code, from local development, to a staging environment, and even in production.

Refactoring or reimplementing important code is an involved process! Instead of listing out Suture's API without sufficient exposition, here is an example that we'll take you through each stage of the lifecycle.

development

Suppose you have a really nasty worker method:

class MyWorker
  def do_work(id)
    thing = Thing.find(id)
    # … 99 lines of terribleness …
    MyMailer.send(thing.result)
  end
end

1. Identify a seam

A seam serves as an artificial entry point that sets a boundary around the code you'd like to change. A good seam is:

• easy to invoke in isolation
• takes arguments, returns a value
• eliminates (or at least minimizes) side effects

Then, to create a seam, typically we create a new unit to house the code that we excise from its original site, and then we call it. This adds a level of indirection, which gives us the flexibility we'll need later.

In this case, to create a seam, we might start with this:

class MyWorker
  def do_work(id)
    MyMailer.send(LegacyWorker.new.call(id))
  end
end

class LegacyWorker
  def call(id)
    thing = Thing.find(id)
    # … Still 99 lines. Still terrible …
    thing.result
  end
end

As you can see, the call to MyMailer.send is left at the original call site, since its effectively a void method being invoked for its side effect, and its much easier to verify return values.

Since any changes to the code while it's untested are very dangerous, it's important to minimize changes made for the sake of creating a clear seam.

2. Create our suture

Next, we introduce Suture to the call site so we can start analyzing its behavior:

class MyWorker
  def do_work(id)
    MyMailer.send(Suture.create(:worker, {
      old: LegacyWorker.new,
      args: [id]
    }))
  end
end

Where old can be anything callable with call (like the class above, a method, or a Proc/lambda) and args is an array of the args to pass to it.

At this point, running this code will result in Suture just delegating to LegacyWorker without taking any other meaningful action.

3. Record the current behavior

Next, we want to start observing how the legacy worker is actually called: what arguments are sent to it and what values does it return? By recording the calls as we use our app locally, we can later test that the old and new implementations behave the same way.

First, we tell Suture to start recording calls by setting the environment variable SUTURE_RECORD_CALLS to something truthy (e.g. SUTURE_RECORD_CALLS=true bundle exec rails s). So long as this variable is set, any calls to our suture will record the arguments passed to the legacy code path and the return value.

As you use the application (whether it's a queue system, a web app, or a CLI), the calls will be saved to a sqlite database. If the legacy code path relies on external data sources or services, keep in mind that your recorded inputs and outputs will rely on them as well. You may want to narrow the scope of your seam accordingly (e.g. to receive an object as an argument instead of a database id).

Hard to exploratory test the code locally?

If it's difficult to generate realistic usage locally, then consider running this step in production and fetching the sqlite DB after you've generated enough inputs and outputs to be confident you've covered most realistic uses. Keep in mind that this approach means your test environment will probably need access to the same data stores as the environment that made the recording, which may not be feasible or appropriate in many cases.

4. Ensure current behavior with a test

Next, we should probably write a test that will ensure our new implementation will continue to behave like the old one. We can use these recordings to help us automate some of the drudgery typically associated with writing characterization tests.

We could write a test like this:

class MyWorkerCharacterizationTest < Minitest::Test
  def setup
    # Load the test data needed to resemble the environment when recording
  end

  def test_that_it_still_works
    Suture.verify(:worker, {
      :subject => LegacyWorker.new
      :fail_fast => true
    })
  end
end

Suture.verify will fail if any of the recorded arguments don't return the expected value. It's a good idea to run this against the legacy code first, for two reasons:

• running the characterization tests against the legacy code path will ensure the test environment has the data needed to behave the same way as when it was recorded (it may be appropriate to take a snapshot of the database before you start recording and load it before you run your tests)

• by generating a code coverage report ( simplecov is a good one to start with) from running this test in isolation, we can see what LegacyWorker is actually calling, in an attempt to do two things:

  • maximize coverage for code in the LegacyWorker (and for code that's subordinate to it) to make sure our characterization test sufficiently exercises it
  • identify incidental coverage of code paths that are outside the scope of what we hope to refactor, and in turn analyzing whether LegacyWorker has side effects we didn't anticipate and should additionally write tests for

5. Specify and test a path for new code

Once our automated characterization test of our recordings is passing, then we can start work on a NewWorker. To get started, we can update our Suture configuration:

class MyWorker
  def do_work(id)
    MyMailer.send(Suture.create(:worker, {
      old: LegacyWorker.new,
      new: NewWorker.new,
      args: [id]
    }))
  end
end

class NewWorker
  def call(id)
  end
end

Next, we specify a NewWorker under the :new key. For now, Suture will start sending all of its calls to NewWorker#call.

Next, let's write a test to verify the new code path also passes the recorded interactions:

class MyWorkerCharacterizationTest < Minitest::Test
  def setup
    # Load the test data needed to resemble the environment when recording
  end

  def test_that_it_still_works
    Suture.verify(:worker, {
      subject: LegacyWorker.new,
      fail_fast: true
    })
  end

  def test_new_thing_also_works
    Suture.verify(:worker, {
      subject: NewWorker.new,
      fail_fast: false
    })
  end
end

Obviously, this should fail until NewWorker's implementation covers all the cases we recorded from LegacyWorker.

Remember, characterization tests aren't designed to be kept around forever. Once you're confident that the new implementation is sufficient, it's typically better to discard them and design focused, intention-revealing tests for the new implementation and its component parts.

6. Refactor or reimplement the legacy code.

This step is the hardest part and there's not much Suture can do to make it any easier. How you go about implementing your improvements depends on whether you intend to rewrite the legacy code path or refactor it. Some comment on each approach follows:

Reimplementing

The best time to rewrite a piece of software is when you have a better understanding of the real-world process it models than the original authors did when they first wrote it. If that's the case, it's likely you'll think of more reliable names and abstractions than they did.

As for workflow, consider writing the new implementation like you would any other new part of the system, with the added benefit of being able to run the characterization tests as a progress indicator and a backstop for any missed edge cases. The ultimate goal of this workflow should be to incrementally arrive at a clean design that completely passes the characterization test run by Suture.verify.

Refactoring

If you choose to refactor the working implementation, though, you should start by copying it (and all of its subordinate types) into the new, separate code path. The goal should be to keep the legacy code path in a working state, so that Suture can run it when needed until we're supremely confident that it can be safely discarded. (It's also nice to be able to perform side-by-side comparisons without having to check out a different git reference.)

The workflow when refactoring should be to take small, safe steps using well understood refactoring patterns and running the characterization test suite frequently to ensure nothing was accidentally broken.

Once the code is factored well enough to work with (i.e. it is clear enough to incorporate future anticipated changes), consider writing some clear and clean unit tests around new units that shook out from the activity. Having good tests for well-factored code is the best guard against seeing it slip once again into poorly-understood "legacy" code.

staging

Once you've changed the code, you still may not be confident enough to delete it entirely. It's possible (even likely) that your local exploratory testing didn't exercise every branch in the original code with the full range of potential arguments and broader state.

Suture gives users a way to experiment with risky refactors by deploying them to a staging environment and running both the original and new code paths side-by-side, raising an error in the event they don't return the same value. This is governed by the :run_both to true:

class MyWorker
  def do_work(id)
    MyMailer.send(Suture.create(:worker, {
      old: LegacyWorker.new,
      new: NewWorker.new,
      args: [id],
      run_both: true
    }))
  end
end

With this setting, the seam will call through to both legacy and refactored implementations, and will raise an error if they don't return the same value. Obviously, this setting is only helpful if the paths don't trigger major or destructive side effects.

production

You're almost ready to delete the old code path and switch production over to the new one, but fear lingers: maybe there's an edge case your testing to this point hasn't caught.

Suture was written to minimize the inhibition to moving forward with changing code, so it provides a couple features designed to be run in production when you're yet unsure that your refactor or reimplementation is complete.

Logging errors

While your application's logs aren't affected by Suture, it may be helpful for Suture to maintain a separate log file for any errors that are raised by the refactored code path.

Suture has a handful of process-wide logging settings that can be set at any point as your app starts up (if you're using Rails, then your environment-specific (e.g. config/environments/production.rb) config file is a good choice).

Suture.config({
  :log_level => "WARN", #<-- defaults to "INFO"
  :log_stdout => false, #<-- defaults to true
  :log_file => "log/suture.log" #<-- defaults to nil
})

When your new code path raises an error with the above settings, it will propogate and log the error to the specified file.

Custom error handlers

Additionally, you may have some idea of what you want to do (i.e. phone home to a reporting service) in the event that your new code path fails. To add custom error handling before, set the :on_error option to a callable.

class MyWorker
  def do_work(id)
    MyMailer.send(Suture.create(:worker, {
      old: LegacyWorker.new,
      new: NewWorker.new,
      args: [id],
      on_error: -> (name, args) { PhonesHome.new.phone(name, args) }
    }))
  end
end

Retrying failures

Since the legacy code path hasn't been deleted yet, there's no reason to leave users hanging if the new code path explodes. By setting the :fallback_to_old entry to true, Suture will rescue any errors raised from the new code path and attempt to invoke the legacy code path instead.

class MyWorker
  def do_work(id)
    MyMailer.send(Suture.create(:worker, {
      old: LegacyWorker.new,
      new: NewWorker.new,
      args: [id],
      fallback_to_old: true
    }))
  end
end

Since this approach rescues errors, it's possible that errors in the new code path will go unnoticed, so it's best used in conjunction with Suture's logging feature. Before ultimately deciding to finally delete the legacy code path, double-check that the logs aren't full of rescued errors!

Configuration

Legacy code is, necessarily, complex and hard-to-wrangle. That's why Suture comes with a bunch of configuration options to modify its behavior, particularly for hard-to-compare objects.

Setting configuration options

In general, most configuration options can be set in several places:

• Globally, via an environment variable. The flag record_calls will translate to an expected ENV var named SUTURE_RECORD_CALLS and can be set from the command line like so: SUTURE_RECORD_CALLS=true bundle exec rails server, to tell Suture to record all your interactions with your seams without touching the source code.

• Globally, via the top-level Suture.config method. Most variables can be set via this top-level configuration, like Suture.config(:database_path => 'my.db'). Once set, this will apply to all your interactions with Suture for the life of the process until you call Suture.reset!.

• At a Suture.create or Suture.verify call-site as part of its options hash. If you have several seams, you'll probably want to set most options locally where you call Suture, like Suture.create(:foo, { :comparator => my_thing })

Supported options

Suture.create

TODO

Suture.verify

TODO

Creating a custom comparator

Out-of-the-box, Suture will do its best to compare your recorded & actual results to ensure that things are equivalent to one another, but reality is often less tidy than a gem can predict up-front. When the built-in equivalency comparator fails you, you can define a custom one—globally or at each Suture.create or Suture.verify call-site.

Extending the built-in comparator class

If you have a bunch of value types that require special equivalency checks, it makes sense to invest the time to extend built-in one:

class MyComparator < Suture::Comparator
  def call(recorded, actual)
    if recorded.kind_of?(MyType)
      recorded.data_stuff == actual.data_stuff
    else
      super
    end
  end
end

So long as you return super for non-special cases, it should be safe to set an instance of your custom comparator globally for the life of the process with:

Suture.config({
  :comparator => MyComparator.new
})

Creating a one-off comparator

If a particular seam requires a custom comparator and will always return sufficiently homogeneous types, it may be good enough to set a custom comparator inline at the Suture.create or Suture.verify call-site, like so:

Suture.create(:my_type, {
  :old => method(:old_method),
  :args => [42],
  :comparator => ->(recorded, actual){ recorded.data_thing == actual.data_thing }
})

Just be sure to set it the same way if you want Suture.verify to be able to test your recorded values!

Suture.verify(:my_type, {
  :subject => method(:old_method),
  :comparator => ->(recorded, actual){ recorded.data_thing == actual.data_thing }
})

Troubleshooting

Some ideas if you can't get a particular verification to work or if you keep seeing false negatives:

• There may be a side effect in your code that you haven't found, extracted, replicated, or controlled for. Consider contributing to this milestone, which specifies a side-effect detector to be paired with Suture to make it easier to see when observable database, network, and in-memory changes are made during a Suture operation
• Consider writing a custom comparator with a relaxed conception of equivalence between the recorded and observed results
• If a recording was made in error, you can always delete it, either by dropping Suture's database (which is, by default, stored in db/suture.sqlite3) or by observing the ID of the recording from an error message and invoking Suture.delete(42)