Runnable's approach to writing code and tests.
We use standard
in most projects. If it passes standard
, anything else is up to the developer. No questions asked.
Let's discuss various tests, what they should do and cover, and how best to organize them.
There are three types of tests we should be concerned about.
- Unit
- Integration
- Functional
Unit tests are pretty obvious (or should be). They test exactly one unit of functionality. Most likely this should follow the boundaries of functions or at largest, modules. If we're finding we're writing a lot unit tests for a specific function, it may need to be broken up.
An important note: unit tests should never touch an external service (e.g. database).
Integration tests make sure that one or more components function correctly together. This can be multiple functions within a module, multiple modules interacting with each other, or a module interacting with a database or external service.
Functional tests should align with product ideas. If a user goes through a given flow, they should finish with some outcome. We've found these tests to be pretty difficult to work with as of late, and are trying to get rid of some of these.
There will always be discussion on where the lines between (for example) unit and integration tests. For a unit test in a mongoose environment, mocking out the layer right at mongoose and simply returning sample objects is sufficient. The developer should have faith (and understanding) that mongoose will correctly return objects. If there's more paculiar logic that you may not be sure of or is more complex (like a map-reduce or aggrigation in mongo, or making sure populate is doing the correct thing), you would move to an integration test and write it there.
These definitely aren't hard-and-fast lines in the sand. A good goal would be that no external services are required to get unit tests off the ground and running, but integration tests may need more setup (like a database). When we find various tests that don't fit into the classification they're housed in, we can always fix the dependency or move the test when appropriate.
In most cases, we should attempt to organize our file structure as follows:
repository/
lib/ (or something like it)
tests/
unit/
integration/
functional/
Additionally, the package.json
should support the following commands:
npm test
: runs a code linter and all three sets of testsnpm unit
: runs the unit testsnpm integration
: runs the integration testsnpm functional
: runs the functional tests
One of the most powerful tools we have (and are still learning to use well) is sinon
. sinon
is a stubbing and spying library that helps us assert that functions are called with what we expect.
sinon
's docs are quite good, but we will talk about some guidelines to follow when using it:
- In all tests where we are going to stub or spy on something, do so in a
before*
- In all tests where we do stub or spy on something,
- if it is on an object that is regenerated every test, no action is necessary
- if it is on a module or something persistent, always restore the function in an
after*
- Stub modules and functions in the correct scopes (meaning, as close to the test that uses it as possible, but hierarchical stubbing is fine)
- When stubbing functions, we may stub the function to a 'truthy' behavior; doing so reduces code duplication, though the developer will need to adjust the stub(s) to test error cases
- Collecting stubs into logical blocks and using the language to define what is being set up is encouraged (i.e. use
describe
blocks to collect and describe stubs,it
blocks to say what happens)
Now that we have tests all squared away and written, let's talk about our assertions.
First, a few options we have been using are code
and chai.assert
. code
is pretty straight forward, using the expect
form of assertions. chai
's assertion library has worked well, especially with promises, and we have been using chai's assert
which helps keep straight which library with which we are working.
We also have started using sinon
's assert features as well (e.g. sinon.assert.calledWith
) which provides much nicer error messages than alternatives.
A few notes to help guide assertions:
- All tests must assert something
- When testing with a stub in the same logical group, we must assert something about the stub in at least one test. Should include
- if it was called, or not
- what it was called with, if applicable
- Stubs should (most likely) include checks from this list:
sinon.assert.called
sinon.assert.notCalled
sinon.assert.calledOnce
,sinon.assert.calledTwice
, etc.sinon.assert.callOrder
sinon.assert.calledWith
sinon.assert.calledWithExactly
(preferred overcalledWith
)
- When testing with promises (and
chai-as-promised
), always start with something such asreturn assert.isFulfilled(myPromise())
, and always remember to return your promise.
An interesting read on how 'value' could be perceived in testing is one of the posts on Google's Testing Blog. It's not super long, so I encourage anyone to read. They very much encourage unit and integration tests over end-to-end (I read it as functional) tests as they help the developers much more and have the exact same value to the customer as any end-to-end test: a bug fix. If you can write a unit test for your bug rather than an end-to-end test, wouldn't you rather do that?
Repositories that need different configurations should have configuration files in the config
directory. Files should be prefixed with .env
and postfixed with the environment. Example .env.production-gamma
. The .env
file contains defaults shared across all environments.
Here is the order of precedence with number 1 being the highest priority.
- Environment variables
.env.[environment]
.env
Configuration files should include everything needed to configures the application per environment with the exception of secret keys / passwords and connections. Only .env.development
and .env.test
can have these in them.