Home

Awesome

Scientist!

A Ruby library for carefully refactoring critical paths. Build Status

How do I science?

Let's pretend you're changing the way you handle permissions in a large web app. Tests can help guide your refactoring, but you really want to compare the current and refactored behaviors under load.

require "scientist"

class MyWidget
  def allows?(user)
    experiment = Scientist::Default.new "widget-permissions"
    experiment.use { model.check_user(user).valid? } # old way
    experiment.try { user.can?(:read, model) } # new way

    experiment.run
  end
end

Wrap a use block around the code's original behavior, and wrap try around the new behavior. experiment.run will always return whatever the use block returns, but it does a bunch of stuff behind the scenes:

The use block is called the control. The try block is called the candidate.

Creating an experiment is wordy, but when you include the Scientist module, the science helper will instantiate an experiment and call run for you:

require "scientist"

class MyWidget
  include Scientist

  def allows?(user)
    science "widget-permissions" do |experiment|
      experiment.use { model.check_user(user).valid? } # old way
      experiment.try { user.can?(:read, model) } # new way
    end # returns the control value
  end
end

If you don't declare any try blocks, none of the Scientist machinery is invoked and the control value is always returned.

Making science useful

The examples above will run, but they're not really doing anything. The try blocks don't run yet and none of the results get published. Replace the default experiment implementation to control execution and reporting:

require "scientist/experiment"

class MyExperiment
  include Scientist::Experiment

  attr_accessor :name

  def initialize(name)
    @name = name
  end

  def enabled?
    # see "Ramping up experiments" below
    true
  end

  def raised(operation, error)
    # see "In a Scientist callback" below
    p "Operation '#{operation}' failed with error '#{error.inspect}'"
    super # will re-raise
  end

  def publish(result)
    # see "Publishing results" below
    p result
  end
end

When Scientist::Experiment is included in a class, it automatically sets it as the default implementation via Scientist::Experiment.set_default. This set_default call is skipped if you include Scientist::Experiment in a module.

Now calls to the science helper will load instances of MyExperiment.

Controlling comparison

Scientist compares control and candidate values using ==. To override this behavior, use compare to define how to compare observed values instead:

class MyWidget
  include Scientist

  def users
    science "users" do |e|
      e.use { User.all }         # returns User instances
      e.try { UserService.list } # returns UserService::User instances

      e.compare do |control, candidate|
        control.map(&:login) == candidate.map(&:login)
      end
    end
  end
end

If either the control block or candidate block raises an error, Scientist compares the two observations' classes and messages using ==. To override this behavior, use compare_errors to define how to compare observed errors instead:

class MyWidget
  include Scientist

  def slug_from_login(login)
    science "slug_from_login" do |e|
      e.use { User.slug_from_login login }         # returns String instance or ArgumentError
      e.try { UserService.slug_from_login login }  # returns String instance or ArgumentError

      compare_error_message_and_class = -> (control, candidate) do
        control.class == candidate.class &&
        control.message == candidate.message
      end

      compare_argument_errors = -> (control, candidate) do
        control.class == ArgumentError &&
        candidate.class == ArgumentError &&
        control.message.start_with?("Input has invalid characters") &&
        candidate.message.start_with?("Invalid characters in input")
      end

      e.compare_errors do |control, candidate|
        compare_error_message_and_class.call(control, candidate) ||
        compare_argument_errors.call(control, candidate)
      end
    end
  end
end

Adding context

Results aren't very useful without some way to identify them. Use the context method to add to or retrieve the context for an experiment:

science "widget-permissions" do |e|
  e.context :user => user

  e.use { model.check_user(user).valid? }
  e.try { user.can?(:read, model) }
end

context takes a Symbol-keyed Hash of extra data. The data is available in Experiment#publish via the context method. If you're using the science helper a lot in a class, you can provide a default context:

class MyWidget
  include Scientist

  def allows?(user)
    science "widget-permissions" do |e|
      e.context :user => user

      e.use { model.check_user(user).valid? }
      e.try { user.can?(:read, model) }
    end
  end

  def destroy
    science "widget-destruction" do |e|
      e.use { old_scary_destroy }
      e.try { new_safe_destroy }
    end
  end

  def default_scientist_context
    { :widget => self }
  end
end

The widget-permissions and widget-destruction experiments will both have a :widget key in their contexts.

Expensive setup

If an experiment requires expensive setup that should only occur when the experiment is going to be run, define it with the before_run method:

# Code under test modifies this in-place. We want to copy it for the
# candidate code, but only when needed:
value_for_original_code = big_object
value_for_new_code      = nil

science "expensive-but-worthwhile" do |e|
  e.before_run do
    value_for_new_code = big_object.deep_copy
  end
  e.use { original_code(value_for_original_code) }
  e.try { new_code(value_for_new_code) }
end

Keeping it clean

Sometimes you don't want to store the full value for later analysis. For example, an experiment may return User instances, but when researching a mismatch, all you care about is the logins. You can define how to clean these values in an experiment:

class MyWidget
  include Scientist

  def users
    science "users" do |e|
      e.use { User.all }
      e.try { UserService.list }

      e.clean do |value|
        value.map(&:login).sort
      end
    end
  end
end

And this cleaned value is available in observations in the final published result:

class MyExperiment
  include Scientist::Experiment

  # ...

  def publish(result)
    result.control.value         # [<User alice>, <User bob>, <User carol>]
    result.control.cleaned_value # ["alice", "bob", "carol"]
  end
end

Note that the #clean method will discard the previous cleaner block if you call it again. If for some reason you need to access the currently configured cleaner block, Scientist::Experiment#cleaner will return the block without further ado. (This probably won't come up in normal usage, but comes in handy if you're writing, say, a custom experiment runner that provides default cleaners.)

The #clean method will not be used for comparison of the results, so in the following example it is not possible to remove the #compare method without the experiment failing:

def user_ids
  science "user_ids" do
    e.use { [1,2,3] }
    e.try { [1,3,2] }
    e.clean { |value| value.sort }
    e.compare { |a, b| a.sort == b.sort }
  end
end

Ignoring mismatches

During the early stages of an experiment, it's possible that some of your code will always generate a mismatch for reasons you know and understand but haven't yet fixed. Instead of these known cases always showing up as mismatches in your metrics or analysis, you can tell an experiment whether or not to ignore a mismatch using the ignore method. You may include more than one block if needed:

def admin?(user)
  science "widget-permissions" do |e|
    e.use { model.check_user(user).admin? }
    e.try { user.can?(:admin, model) }

    e.ignore { user.staff? } # user is staff, always an admin in the new system
    e.ignore do |control, candidate|
      # new system doesn't handle unconfirmed users yet:
      control && !candidate && !user.confirmed_email?
    end
  end
end

The ignore blocks are only called if the values don't match. Unless a compare_errors comparator is defined, two cases are considered mismatches: a) one observation raising an exception and the other not, b) observations raising exceptions with different classes or messages.

Enabling/disabling experiments

Sometimes you don't want an experiment to run. Say, disabling a new codepath for anyone who isn't staff. You can disable an experiment by setting a run_if block. If this returns false, the experiment will merely return the control value. Otherwise, it defers to the experiment's configured enabled? method.

class DashboardController
  include Scientist

  def dashboard_items
    science "dashboard-items" do |e|
      # only run this experiment for staff members
      e.run_if { current_user.staff? }
      # ...
  end
end

Ramping up experiments

As a scientist, you know it's always important to be able to turn your experiment off, lest it run amok and result in villagers with pitchforks on your doorstep. In order to control whether or not an experiment is enabled, you must include the enabled? method in your Scientist::Experiment implementation.

class MyExperiment
  include Scientist::Experiment

  attr_accessor :name, :percent_enabled

  def initialize(name)
    @name = name
    @percent_enabled = 100
  end

  def enabled?
    percent_enabled > 0 && rand(100) < percent_enabled
  end

  # ...

end

This code will be invoked for every method with an experiment every time, so be sensitive about its performance. For example, you can store an experiment in the database but wrap it in various levels of caching such as memcache or per-request thread-locals.

Publishing results

What good is science if you can't publish your results?

You must implement the publish(result) method, and can publish data however you like. For example, timing data can be sent to graphite, and mismatches can be placed in a capped collection in redis for debugging later.

The publish method is given a Scientist::Result instance with its associated Scientist::Observations:

class MyExperiment
  include Scientist::Experiment

  # ...

  def publish(result)

    # Wall time
    # Store the timing for the control value,
    $statsd.timing "science.#{name}.control", result.control.duration
    # for the candidate (only the first, see "Breaking the rules" below,
    $statsd.timing "science.#{name}.candidate", result.candidates.first.duration

    # CPU time
    # Store the timing for the control value,
    $statsd.timing "science.cpu.#{name}.control", result.control.cpu_time
    # for the candidate (only the first, see "Breaking the rules" below,
    $statsd.timing "science.cpu.#{name}.candidate", result.candidates.first.cpu_time

    # and counts for match/ignore/mismatch:
    if result.matched?
      $statsd.increment "science.#{name}.matched"
    elsif result.ignored?
      $statsd.increment "science.#{name}.ignored"
    else
      $statsd.increment "science.#{name}.mismatched"
      # Finally, store mismatches in redis so they can be retrieved and examined
      # later on, for debugging and research.
      store_mismatch_data(result)
    end
  end

  def store_mismatch_data(result)
    payload = {
      :name            => name,
      :context         => context,
      :control         => observation_payload(result.control),
      :candidate       => observation_payload(result.candidates.first),
      :execution_order => result.observations.map(&:name)
    }

    key = "science.#{name}.mismatch"
    $redis.lpush key, payload
    $redis.ltrim key, 0, 1000
  end

  def observation_payload(observation)
    if observation.raised?
      {
        :exception => observation.exception.class,
        :message   => observation.exception.message,
        :backtrace => observation.exception.backtrace
      }
    else
      {
        # see "Keeping it clean" above
        :value => observation.cleaned_value
      }
    end
  end
end

Testing

When running your test suite, it's helpful to know that the experimental results always match. To help with testing, Scientist defines a raise_on_mismatches class attribute when you include Scientist::Experiment. Only do this in your test suite!

To raise on mismatches:

class MyExperiment
  include Scientist::Experiment
  # ... implementation
end

MyExperiment.raise_on_mismatches = true

Scientist will raise a Scientist::Experiment::MismatchError exception if any observations don't match.

Custom mismatch errors

To instruct Scientist to raise a custom error instead of the default Scientist::Experiment::MismatchError:

class CustomMismatchError < Scientist::Experiment::MismatchError
  def to_s
    message = "There was a mismatch! Here's the diff:"

    diffs = result.candidates.map do |candidate|
      Diff.new(result.control, candidate)
    end.join("\n")

    "#{message}\n#{diffs}"
  end
end

science "widget-permissions" do |e|
  e.use { Report.find(id) }
  e.try { ReportService.new.fetch(id) }

  e.raise_with CustomMismatchError
end

This allows for pre-processing on mismatch error exception messages.

Handling errors

In candidate code

Scientist rescues and tracks all exceptions raised in a try or use block, including some where rescuing may cause unexpected behavior (like SystemExit or ScriptError). To rescue a more restrictive set of exceptions, modify the RESCUES list:

# default is [Exception]
Scientist::Observation::RESCUES.replace [StandardError]

Timeout ⏲️: If you're introducing a candidate that could possibly timeout, use caution. ⚠️ While Scientist rescues all exceptions that occur in the candidate block, it does not protect you from timeouts, as doing so would be complicated. It would likely require running the candidate code in a background job and tracking the time of a request. We feel the cost of this complexity would outweigh the benefit, so make sure that your code doesn't cause timeouts. This risk can be reduced by running the experiment on a low percentage so that users can (most likely) bypass the experiment by refreshing the page if they hit a timeout. See Ramping up experiments below for how details on how to set the percentage for your experiment.

In a Scientist callback

If an exception is raised within any of Scientist's internal helpers, like publish, compare, or clean, the raised method is called with the symbol name of the internal operation that failed and the exception that was raised. The default behavior of Scientist::Default is to simply re-raise the exception. Since this halts the experiment entirely, it's often a better idea to handle this error and continue so the experiment as a whole isn't canceled entirely:

class MyExperiment
  include Scientist::Experiment

  # ...

  def raised(operation, error)
    InternalErrorTracker.track! "science failure in #{name}: #{operation}", error
  end
end

The operations that may be handled here are:

Designing an experiment

Because enabled? and run_if determine when a candidate runs, it's impossible to guarantee that it will run every time. For this reason, Scientist is only safe for wrapping methods that aren't changing data.

When using Scientist, we've found it most useful to modify both the existing and new systems simultaneously anywhere writes happen, and verify the results at read time with science. raise_on_mismatches has also been useful to ensure that the correct data was written during tests, and reviewing published mismatches has helped us find any situations we overlooked with our production data at runtime. When writing to and reading from two systems, it's also useful to write some data reconciliation scripts to verify and clean up production data alongside any running experiments.

Noise and error rates

Keep in mind that Scientist's try and use blocks run sequentially in random order. As such, any data upon which your code depends may change before the second block is invoked, potentially yielding a mismatch between the candidate and control return values. To calibrate your expectations with respect to false negatives arising from systemic conditions external to your proposed changes, consider starting with an experiment in which both the try and use blocks invoke the control method. Then proceed with introducing a candidate.

Finishing an experiment

As your candidate behavior converges on the controls, you'll start thinking about removing an experiment and using the new behavior.

Breaking the rules

Sometimes scientists just gotta do weird stuff. We understand.

Ignoring results entirely

Science is useful even when all you care about is the timing data or even whether or not a new code path blew up. If you have the ability to incrementally control how often an experiment runs via your enabled? method, you can use it to silently and carefully test new code paths and ignore the results altogether. You can do this by setting ignore { true }, or for greater efficiency, compare { true }.

This will still log mismatches if any exceptions are raised, but will disregard the values entirely.

Trying more than one thing

It's not usually a good idea to try more than one alternative simultaneously. Behavior isn't guaranteed to be isolated and reporting + visualization get quite a bit harder. Still, it's sometimes useful.

To try more than one alternative at once, add names to some try blocks:

require "scientist"

class MyWidget
  include Scientist

  def allows?(user)
    science "widget-permissions" do |e|
      e.use { model.check_user(user).valid? } # old way

      e.try("api") { user.can?(:read, model) } # new service API
      e.try("raw-sql") { user.can_sql?(:read, model) } # raw query
    end
  end
end

When the experiment runs, all candidate behaviors are tested and each candidate observation is compared with the control in turn.

No control, just candidates

Define the candidates with named try blocks, omit a use, and pass a candidate name to run:

experiment = MyExperiment.new("various-ways") do |e|
  e.try("first-way")  { ... }
  e.try("second-way") { ... }
end

experiment.run("second-way")

The science helper also knows this trick:

science "various-ways", run: "first-way" do |e|
  e.try("first-way")  { ... }
  e.try("second-way") { ... }
end

Providing fake timing data

If you're writing tests that depend on specific timing values, you can provide canned durations using the fabricate_durations_for_testing_purposes method, and Scientist will report these in Scientist::Observation#duration and Scientist::Observation#cpu_time instead of the actual execution times.

science "absolutely-nothing-suspicious-happening-here" do |e|
  e.use { ... } # "control"
  e.try { ... } # "candidate"
  e.fabricate_durations_for_testing_purposes({
    "control" => { "duration" => 1.0, "cpu_time" => 0.9 },
    "candidate" => { "duration" => 0.5, "cpu_time" => 0.4 }
  })
end

fabricate_durations_for_testing_purposes takes a Hash of duration & cpu_time values, keyed by behavior names. (By default, Scientist uses "control" and "candidate", but if you override these as shown in Trying more than one thing or No control, just candidates, use matching names here.) If a name is not provided, the actual execution time will be reported instead.

We should mention these durations will be used both for the duration field and the cpu_time field.

Like Scientist::Experiment#cleaner, this probably won't come up in normal usage. It's here to make it easier to test code that extends Scientist.

Without including Scientist

If you need to use Scientist in a place where you aren't able to include the Scientist module, you can call Scientist.run:

Scientist.run "widget-permissions" do |e|
  e.use { model.check_user(user).valid? }
  e.try { user.can?(:read, model) }
end

Hacking

Be on a Unixy box. Make sure a modern Bundler is available. script/test runs the unit tests. All development dependencies are installed automatically. Scientist requires Ruby 2.3 or newer.

Wrappers

Alternatives

Maintainers

@jbarnette, @jesseplusplus, @rick, and @zerowidth