A Node.js library for carefully refactoring critical paths in production
CoffeeScript JavaScript
Switch branches/tags
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
docs
examples
src
test
.coffeelint.json
.gitignore
.travis.yml
CHANGELOG.md
LICENSE
README.md
console.js
index.js
package.json

README.md

Scientist

npm Build Status Coverage Status

Table of contents

How it works

So you just refactored a swath of code and all tests pass. You feel completely confident that this can go to production. Right? In reality, not so much. Be it poor test coverage or just that the refactored code is very critical, sometimes you need more reassurance.

Scientist lets you run your refactored code alongside the actual code, comparing the outputs and logging when it did not return as expected. It's heavily based on GitHub's Scientist gem. Let's walk through an example. Start with this code:

const sumList = (arr) => {
  let sum = 0;
  for (var i of arr) {
    sum += i;
  }
  return sum;
};

And let's refactor it as so:

const sumList = (arr) => {
  return _.reduce(arr, (sum, i) => sum + i);
};

To do science, all you need to do is replace the original function with a science wrapper that uses both functions:

const sumList = (arr) => {
  return science('sum-list', (experiment) => {
    experiment.use(() => sumListOld(arr));
    experiment.try(() => sumListNew(arr));
  });
};

And that's it. The science function takes a string to identify the experiment by and passes an experiment object to a function that you can use to set up your experiment. We call use to define what our control behavior is -- that's also the value that is returned from the original science call, which makes this a drop-in replacement. The try function can be used to define one or more candidates to compare. So what happens if we do this:

sumList([1, 2, 3]);
// -> 6
// Experiment candidate matched the control

But there's also a bug in our refactored code. Science logs that as appropriate, but still returns the old value that we know works.

sumList([]);
// -> 0
// Experiment candidate did not match the control
//   expected value: 0
//   received value: undefined

You can find this implemented in examples/basic.js.

Getting started

Above we just used a simple science() function to run an experiment. If you're just looking to play around, you can get the same function with require('scientist/console'). If you examine console.js, you'll notice that this is a very simple implementation of the Scientist class, which is exposed through a normal require('scientist') call.

The recommended usage is to create a file specific to your application and export the science method bound to a fully set-up Scientist instance.

const Scientist = require('scientist');

const scientist = new Scientist();

scientist.on('skip', function (experiment) { /* ... */ });
scientist.on('result', function (result) { /* ... */ });
scientist.on('error', function (err) { /* ... */ });

module.exports = scientist.science.bind(scientist);

Then you can rely on your own internal logging and metrics tools to do science.

Errors in behaviors

Scientist has built-in support for handling errors thrown by any of your behaviors.

science('throwing errors', (experiment) => {
  experiment.use(() => {
    throw Error(msg)
  });
  experiment.try("with-new", () => {
    throw new Error(msg)
  });
  experiment.try("as-type-error", () => {
    throw TypeError(msg)
  });
});

error("An error occured!");
// Experiment candidate matched the control
// Experiment candidate did not match the control
//   expected: error: [Error] 'An error occured!'
//   received: error: [TypeError] 'An error occured!'

In this case, the call to science() is actually throwing the same error that the control function threw, but after testing the other functions and readying the logging. The criteria for matching errors is based on the constructor and message.

You can find this full example at examples/errors.js.

Asynchronous behaviors

See docs/async.md.

Customizing your experiment

There are several functions you can use to configure science:

  • context: Record information to give context to results
  • async: Turn async mode on
  • skipWhen: Determine whether the experiment should be skipped
  • map: Change values for more simple comparison and logging
  • ignore: Throw away certain observations
  • compare: Decide whether two observations match
  • clean: Prepare data for logging

Because of the first-class promise support, the compare and clean functions will take values after they are settled. map happens synchronously and may also return a promise, which could be resolved.

If you want to think about the flow of data in a pipeline, it looks like this:

  1. Block is called and the value or error is saved as an observation
  2. map() is applied to the value
  3. Promises are settled if async was set to true
  4. The Result object is instantiated and observations are passed to compare()
  5. The consumer may call inspect() on an observation, which applies clean()

You can see a fairly full example at examples/complex.js

Side effects

So all of these examples were simple because they were either pure functions or functions that produced no observable side-effects. What if we want to test something more complicated? We definitely cannot let our candidate function change the state of the world permanently, such as updating an entry in the database. However, we can still use science to observe functions that change the state of some object.

science('user middleware', (experiment) => {
  experiment.use(() => {
    findUser(req);
    return req;
  });
  experiment.try(() => {
    let clone = _.clone(req);
    findUserById(clone);
    findUserByName(clone);
    return clone;
  });
});

Enabling and skipping

Often you don't want to run science on every single function call. Since we're testing under production load and running the functionality at least twice, you can imagine that some parts may get out of control. Scientist provides a solution to let you sample a test so that you can slowly ramp it up in production and stop when you have a comfortable amount of data. You can configure this with the Scientist#sample() function.

const scienceConfig = require('./science-config.json');
const scientist = new Scientist();

scientist.sample((experimentName) => {
  if (experimentName in scienceConfig) {
    // Configuration maps a name to a percentage
    return Math.random() < scienceConfig[experimentName];
  } else {
    // Default to not running for safety
    return false;
  }
});

Note that the sampling function is provided the experiment name and must be synchronous.

If you want to skip experiments based on more information, you can configure this at the experiment level with skipWhen().

science('parse headers', (experiment) => {
  experiment.skipWhen(() => 'x-internal' in headers);
  // ...
});

Why CoffeeScript?

This project started out internally at Trello and only later was spun off into a separate module. As such, it was written using the language, dependencies, and style of the Trello codebase. The code is hopefully simple enough to grok such that the language choice does not deter contributors.