Skip to main content

Testing

Testing in Hyperstack is core to the framework. Imagine that first the testing experience was sketched out, and then Hyperstack was built around that. This is way booting a Hyperstack app and tearing it down is very easy, and it has different booting modes for every use case.

For example, to run a full app, Hyperstack boots with all pieces: controllers, models, workers, and more. But, for example, to run a workers test, Hyperstack boots without controllers.

General concept and idea

Core app testing pillars:

  • Unit - discrete pieces of logic
  • Models - most logic here
  • Requests - integration and interaction

And then supporting pillars:

  • Workers
  • Mailers
  • Tasks

Each of these can be tested separately and has its own test framework helper so you can write as little set up / teardown code as possible.

Models

A few notes here:

  • Located in tests/models/{model-name}.spec.ts
  • Use snapshots liberally, and assist with redaction util
import { PopBand } from '../../app/models/pop-band'
import { test } from 'hyperstack'
import { root } from '../../config/settings'

const { models } = test(root)
describe('models', () => {
describe('pop-band', () => {
models('should find', async (_app) => {
await PopBand.create({
frontman: 'bill steer',
email: 'bill@example.com',
})
const c = await PopBand.count()
expect(c).toEqual(1)
})
//:
//:
})
})

And the supporting configuration, annotated:

  database: {
uri: 'postgres://localhost:5432/tie_test',
ssl: false,
native: true,

synchronize: true, // sync the schema to latest
truncate: true, // clean up and reset DB pk counting

dropSchema: false, // no need to drop unless you want to test that
migrate: false, // no need for migration unless you want to test that
logging: console.log, // set to 'false' if you don't want to see what the DB is doing during tests.
},

Taken care for you:

  • Booting the app
  • Setting up the DB and connections for testing
  • Truncate, sync for DB schema
  • Teardown: closing connections and shutting down the app cleanly for next tests.

Fixtures

Fixtures are plain typescript files, imported normally into test code. It is recommended to write those in __fixtures__/index.ts.

If you want to split fixtures to separate files, do that, but have the root index.ts re-export as needed so you can just say from './__fixtures__' everywhere.

Because of this best practice, the actual fixture can be data loaded from JSON, hardcoded entities, a central database, faker.js based fixtures, or hand-rolled logic.

Use the following convention:

async <create><Model><For><UseCase>

E.g.

import { createUsersForRegistration } from './__fixtures__'

Also, have each fixture creator be async. A fixture is to be used inside a test kit, and so it can assume that all entities and app are initialized.

models('some test', async () => {
await createUsersForRegistration() // uses Users.create safely
// :
// :
// test is tearing down all data here automatically.
})

Requests

A few pointers here:

  • Located in tests/requests/{controller-name}/{scenario}.spec.ts
  • These are integration tests, playing out various scenarios, and it's OK to stage data in models, and other side effects; or to verify those if needed whilst testing.
  • Use snapshots liberally, specifically requestWithSnapshot which redacts some ever-changing fields like dates, etags and such, creating a stable snapshot between tests.
  • You have a convenience helper request which binds supertest to the current app for you. You can perform multiple requests during a given scenario: each request() will bind an app again.
  • When you need more power, app is a central access point to the entire app objects which includes context, express app, and logger, and server is the currently configured Express app.

An example test:

import { test } from 'hyperstack'
import { PopBand } from '../../../app/models/pop-band'
import { root } from '../../../config/settings'

const {
requests,
matchers: { requestWithSnapshot },
} = test(root)

describe('requests', () => {
describe('/auth', () => {
requests('should send reset with email', async (_app, request, _server) => {
await PopBand.create({
email: 'bill@example.com',
frontman: 'bill steer',
})
await requestWithSnapshot(
200,
request().post('/auth/reset?email=bill@example.com')
)
})
})
})

Often endpoints perform several other tasks other than saving data, such as sending email. Registration is often done with sending a welcome email for verification. This is still steamlined for your testing experience. See below:

requests(
'should register and send welcome email',
async (request) => {
const { User } = appContext.models()
await matchRequestWithSnapshot(
200,
request().post('/auth/register').send({
username: 'evh@example.com',
password: 'mypass-should-login',
name: 'Eddie Van Halen',
}),
serialize
)
const user = await User.findOne({
where: { username: 'evh@example.com' },
})

expect(serializeUser(user)).toMatchSnapshot()

matchDeliveriesWithSnapshot(1)
}
)

Highlights:

  • Emails are captured inmemory for validation
  • Before each test, inbox is emptied
  • request is a fully rigged app request. You don't need to know or understand which app, which host or port this calls
  • matchDeliveriesWithSnapshot covers all your bases when it comes to email testing: email count and shape for this test, look at your snapshots for the outgoing emails and approve those.

Testing workers

Conceptually there are 3 types of workers:

  1. Just an async caller of a method on a model. This worker contains no logic at all, just hustles data from the queue, and calls User.updateStuff(data). Because of that, there's nothing really to test here, which is a Good Thing. You can test this worker as a pure unit.
  2. A workflow orchestrator of models and data: for example, generate daily reports and insert into a reports table for a dashboard controller.
  3. An "elbow grease worker": grabbing models, getting data, uploading it to S3, downloading some material, parsing it, cleaning up disk, performing Vaccum / maintenance on data store and more. This requires a full on integration test, bringing up containers, staging some things on disk and more.

Notes:

  • Located at tests/workers/{worker-name}/{scenario}.spec.ts - use {scenario} to separate between the various types, where for example type (3) requires some heavy machinery, bringing up containers and such.
  • These are integration tests, playing out various scenarios, and it's OK to stage data in models, and other side effects; or to verify those if needed whilst testing.
  • Use snapshots liberally

Testing workers in-process

This is fastest testing strategy for workers, because workers perform the job in process. In addition, workers can be viewed as just units if you test with performNow instead of performLater, and this simplifies or removes completely the need for full integration tests.

Here's an example:

import { test } from 'hyperstack'
const {
workers,
} = test(root)
// ...
describe('calculator', () => {
workers('should calculate, inprocess', async (_app) => {
const res = await Calculator.performNow({ number: 30 })
expect(res).toMatchSnapshot()

// this is not different, because the 'later' part is immediate
const res2 = await Calculator.performLater({ number: 30 })
expect(res2).toMatchSnapshot()
})
// ...

Testing workers with Redis

The best way to test with Redis is that you'd have a live instance of it locally. Most CI systems support that these days. In this mode, you can activate the truncate flag which will automatically truncate the current Redis database before each test (as well as truncating the Postgres database).

Testing workers using 'testcontainers'

This technique inclues all moving parts as close to real life as possible, and brings up a full blown Redis containers for each test, and is slower. Use this when you really need to bring up some heavy duty infrastructure around the tests.

Here's an example:

import { test } from 'hyperstack'
const {
workers,
containers: { withRedis },
} = test(root)

// compose on top of the workers helper
const workersWithRedis = withRedis(workers, {})

// ...
describe('downloader', () => {
workersWithRedis('should download', async (_app) => {
const magicNumber = Math.random().toString()
Downloader.magicNumber = magicNumber
Downloader.performLater('foobar')
await sleep(1000)
expect(await Downloader.backend.connection.get('downloaded')).toEqual(
magicNumber
)
})
// ...

Since testing at scale can challenge Docker, it's really better to just have a live Redis instance and not use testcontainers. But, if you have some sort of fancy configuration, testcontainers might be the only choice for you.

Docs

Get StartedBuilding APIsModeling DataBackground JobsTestingDeploymentFAQ

More