« Scott Nonnenberg

A functional distinction

2016 May 10

I had a moment of testing/architecture clarity recently while working on a new Node.js module which works with PostgreSQL. Previous leanings and intuitions became concrete. Maybe you’re wondering how best to structure Node.js code involving callbacks? Or how to test it? I’ve got some ideas for you!

How to test this?

First, take a look at this code. How you would you unit test it?

Grator.prototype.list = function(cb) {
  const _this = this;

  this._init(function(err) {
    if (breadcrumbs.add(err, cb)) {

    _this._getMigrations(function(err, scripts) {
      if (breadcrumbs.add(err, cb)) {

      _this._getMigrationsStatus(scripts, null, function(err, migrations) {
        if (breadcrumbs.add(err, cb)) {

        const list = _this._mergeFilesAndDB(scripts, migrations);

        return cb(null, list);

Note: this code uses breadcrumbs functionality from my thehelp-core library to make it easier to debug async errors. It’s just a better way to write if (err) return cb(err).

Go ahead, I’ll wait…

cookie monster is waiting

So what do you think? Maybe bring in Sinon.js to stub out the three calls on this so we could cover the _mergeFilesAndDB() call without going to the database? Refactor list() to use Async.js primitives? Refactor it to reduce the callback depth?

My answer? None of the above.

A heuristic

First, let’s talk about a heuristic I’ve had in the back of my mind for a while now:

If your tests must use Sinon.js, it’s a code smell.

Put another way: If you can’t test certain code any other way than Sinon.js stubs or mocks, you might have some refactoring to do. Or you have an opportunity to be a little bit more creative with your test approach. Interesting logic buried inside a nest of callbacks? You could pull that logic into a separate method.

You can see where I’ve already done that refactoring in the code above. _mergeFilesAndDB() contains non-trivial logic which was previously inline, deep in list(). I couldn’t unit test it without using Sinon.js. When I pulled it out into its own method I could test it very fast, covering lots of scenarios, with no stubs required.

And, with list() newly simplified, I had a new approach available to me.

The role of unit tests

Unit tests are valuable because they run very quickly and focus like a laser on your core logic. But that strength becomes a weakness with what I like to call ‘workflow’ logic. In list() above, it calls three methods in order, and then sends the results of the last two methods to another method, finally returning the result.

What value would a unit test provide in its coverage of list()? It could perhaps could verify that values are properly flowed from one step to the next. But it doesn’t validate that the contracts between each of these component match up with each other. A unit test could still pass even if the entire application is broken.

Another approach to validate the kind of code shown in list() is a integration test. Because this type of test involves real dependencies like the database and filesystem, it will require more setup and run more slowly. But that tradeoff allows you to validate real user scenarios and real customer value.

So my initial question was a bit of a trick.

I’m not covering list() with unit tests at all. Just integration tests.

A clear distinction

I like to think of my apps as a combination of two very different types of functions.

Data transformation functions are simply described: data of a certain structure in, and data of a different structure out. The are deterministic, always returning the same result for the same inputs, and pure, with no side effects (no modification of data in the application or outside of it). This makes them very easily unit tested. These functions might be analyzing or summarizing data, or generating/modifying the data which will be used to render a user interface.

Workflow functions tie my applications together, flowing data through the app. Like list() above they might do some async operations, or they might call a number of data transformation functions in a row, assembling from them the final result. Integration tests ensure that these functions fulfill their promise of a cohesive unit built of a number of subcomponents: the filesystem, a database, an API, or the screen, along with data transformations.

Two types of functions, two types of test approaches.

Unit testing for data transformations. Integration testing for workflow.


In my “A modest async proposal” post, I described a different way to structure asynchronous workflow code in Node.js. But it can be refactored further.

How might you split this getUser function into one workflow function and one data transformation function?

NewReferencesProcess.prototype.getUser = function getUser() {
  var _this = this;

  return this.db.getUser(this._userId)
    .then(function(user) {
      if (!user) {
        return Bluebird.reject(new Error('Could not find user! ' + _this._userId));

      _this._user = user;

      return user;

You’ll note that you can’t be pure with this function, since it’s changing a value on this. We made that sacrifice to simplify data sharing across the asynchronous workflow. It’s the exception, rather than the rule, in my code. I’ve mostly discarded Object-Oriented design principles in favor of functional design:

  • Plain data instead of classes. Classes would combine logic and data in one package.
  • Pure functions which operate on that data. No side effects, always returning the same result for the same inputs.
  • Application state stored in very few places, instead of every instance of a class.

Functional design, in my experience, makes everything easier to reason about. It’s easier to determine how a given change happened, even if only part of an application is using these principles.

A clearly separated future?

I know I’ve found some good benefits with this preference for functional design, and this workflow vs. data transformation distinction. How might you separate the two types of functions in your code? Would it make things easier? Harder? Are there tricky parts which aren’t easily split up?

I’d love to hear about your experiences with these techniques!

Bringing functional design to Javascript:

I won't share your email with anyone. See previous emails.


From tech person to people person 2016 May 17

I’ve been on both sides: made fun of as a geeky computer enthusiast, and derided as useless project management overhead. Both sides of the technical divide are frustrated and unhappy. How to bridge... Read more »


The state of thehelp 2016 May 03

The collection of node modules and client libraries I released under the thehelp family name has been available now for about a year and a half. It’s been a good run, but now it’s time to take... Read more »

It's me!
Hi, I'm Scott. I've written both server and client code in many languages for many employers and clients. I've also got a bit of an unusual perspective, since I've spent time in roles outside the pure 'software developer.' You can find me on Mastodon.