Awesome
Scientist
Table of contents
- API Documentation
- How it works
- Getting started
- Errors in behaviors
- Asynchronous behaviors
- Customizing your experiment
- Side effects
- Enabling and skipping
- Why CoffeeScript?
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 resultsasync
: Turn async mode onskipWhen
: Determine whether the experiment should be skippedmap
: Change values for more simple comparison and loggingignore
: Throw away certain observationscompare
: Decide whether two observations matchclean
: 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:
- Block is called and the value or error is saved as an observation
map()
is applied to the value- Promises are settled if
async
was set totrue
- The
Result
object is instantiated and observations are passed tocompare()
- The consumer may call
inspect()
on an observation, which appliesclean()
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.