The problem was happening in component integration tests on the rendering stage, sometimes the rendering would never finish.
Using time moments in the future when faking time solves the problem. Unfortunately, I don't know why exactly it helps. It was just a lucky guess after some hours I spent trying to figure out what's going on. But I've done a lot of testings, so looks like it really works. I'll be monitoring builds for some time after merging this anyway.
Unit tests seem to work alright with moments in the past. And we don't fake time in acceptance tests at the moment but I guess they would very likely be flaky with time moments from the past since they also do rendering.
I'm actually thinking of moving all fake time moments to the future (including moments in unit tests) to decrease the chances of flakiness. But I don't want to do everything in one PR, because I can accidentally introduce new flakiness.
A pretty easy way of picking time moments in the future for tests is to use the 2100 year. It has the same calendar as 2021. If a day is Monday in 2021 it's Monday in 2100 too.
Using arrow functions changes `this` context, which is undesired in tests, e.g. it makes it impossible to setup things like pretender (`this.server`) in `beforeEach` hooks.
Ember guides always use classic functions in examples (e.g. https://guides.emberjs.com/release/testing/test-types/), and that's what it uses in its own test suite, as do various addons and ember apps.
It was also already used in Discourse where `this` was required. Moving forward, it will be needed in more places as we migrate toward ember-cli.
(I might later add a custom rule to eslint-discourse-ember to enforce this)
We used many global functions to handle tests when they should be
imported like other libraries in our application. This also gets us
closer to the way Ember CLI prefers our tests to be laid out.
This is where they should be as far as ember is concerned. Note this is
a huge commit and we should be really careful everything continues to
work properly.