Automatic regression testing with CI

[munrocket]

In Three.js repo we have several PR in a day and right now is really hard to formally validate correctness of the new PRs for maintainers. This idea was suggested by mrdoob and discussed here #16941, #13017. Also it makes possible to create a gallery with all existed examples in one page #17957.

This POC heavily optimized and provides test coverage in 98%. Details you can find here: https://github.com/munrocket/puppeteer-three. Before merge we probably need to create new contribution guides and add handy command in package.json that generate one screenshot.

[munrocket]

2do list:

• fix code style
• migrate to serve from express and http-server
• add id=startButton in webgl_performance_nodes button
• rethink ci configs
• fix indent style
• add class .lbl in misc_animation_authoring
• print diffs in console
• update contribution guide
• add files with extensions
• color polyfill in travis console

[NicolasRannou]

Very cool, out of curiosity, is it the same approach (determinist timers and so on) injection other projects such as the model viewer? https://modelviewer.dev/

[munrocket]

@NicolasRannou probably not, but if they fix issue with offscreen-canvas or video tag we can take approach from there.

P.S. they doesn't have any randomness or timer hooks, I was inspired from gl-bench, this issue from Greggman and Miku-Miku from Takahirox.

[munrocket]

Example of error report in Travis terminal:

[Mugen87]

This PR has already 42 commits. Can you please force-push a single commit that contains all changes?

[munrocket]

Yep.

[munrocket]

I done different tests but after force push it gone:

• Stress tests with 10 attempts, 2 from 10 fails.
• Test to generate all screenshots from scratch.
• Test that should fail and print result in Travis console.

This PR ready for merge. Also this similar settings just works in Travis under OS Windows:

Note that I am not changed contribution guide in Wiki page.

[munrocket]

I fixed a problem with webgl_loader_svg witch is fails 2 times in previous stress test.
Updated results about robustness: 1/10 fail in Travis, 0/10 with CircleCI.

Under pressure Travis goes in infinite loop, because it can't resolve a promise.

[munrocket]

@mrdoob actually this PR ready for merge, but need a feedback. Especially I am confused about folder naming diff, and probably I need to exclude test-diff from test command in package.json because different GPUs give different results.

Code is complicated and exists small bug that appeared in 1/10 of tests in Travis under Linux but in most of the cases it's work, and this is better than nothing I guess.

[mrdoob]

migrate to serve from express and http-server

Oh, I missed this decision. What's the reasoning?

[munrocket]

Because concurrently with http-server become too noisy and I can't run it in silent mode https://github.com/kimmobrunfeldt/concurrently/issues/25 , so I need to use something like express from node script.

But having two similar package http-server and express in one project looks strange. serve is suitable for both cases: as a runner and from node.

[mrdoob]

Fine by me.

I personally do not use npm run dev anyway. I usually have a http-serve instance that I have installed globally.

I wonder if anyone uses npm run dev. As far as I remember it was added by @fernandojsg.

[mrdoob]

Merging this 🤞

[mrdoob]

Many thanks!

[munrocket]

@looeee do you used 8080 port somehow in your settings? serve by default with 5000 port.

[Mugen87]

Seems that I need to create a new PR, because this already merged?

Correct.

We've had some trouble with serve in the past and switched to http-server on purpose:, see #18424 (comment) and #14264. And yes, this problems occur now again. I suggest we move back to http-server and use serve only for the regression testing.

[Mugen87]

Besides, is it possible to just log a warning if the screenshot for an example is missing? Right now, it seems the complete build fails which seems not appropriated for this use case.

[munrocket]

@Mugen87 looks reasonable when you are prototyping, by the way we have exception list for Windows and Linux(CI) users.

Fixed http-server here #18711

[munrocket]

We can make a new fake stage in Travis with name coverall for example and allow fail to it. It will check how all examples are covered with screenshots.

[looeee]

@looeee do you used 8080 port somehow in your settings?

No, but I hadn't followed the recent switch to serve and port 5000. I'd prefer sticking with http-server unless people report issues with it.

I wonder if anyone uses npm run dev.

I do. Well, I use the alias npm run start.

By the way @mrdoob, it's not a big deal for me if this gets removed. However I would prefer if there's at least a watch script which runs rollup -c utils/build/rollup.config.js -w.

[takahirox]

Hi, after we switch to serve I face some troubles on Examples and Docs.

If I run npm run start and then access http://localhost:5000/examples or http://localhost:5000/examples/index.html, the examples don't appear because of ReferenceError: files is not defined. Somehow http://localhost:5000/examples/#webgl_animation_cloth works tho.

Similarly on http://localhost:5000/docs or http://localhost:5000/docs/index.html, the docs don't appear because of ReferenceError: list is not defined.

Can anyone here reproduce the same error? I use Windows 10 + Firefox/Chrome.

Update: Ah, sorry. I have missed the PR reverting to http-server. Never mind, thanks.

[munrocket]

@takahirox hi! We reverted http-server, but it’s take some time to merge.

[mrdoob]

Still one left:

ERROR! Diff wrong in 0.079 of pixels in file: css3d_youtube

https://travis-ci.org/mrdoob/three.js/builds/654693062

I guess we can ignore that one?

[munrocket]

Added css3d_youtube in ingnore list. But problem with non-determinism in video tag still exist #18760. In less than 1/25 of builds webgl_kinect fails.

[mrdoob]

Lets add webgl_kinect to the list then.

Also, I started to received emails every time the test pass. I used to only receive emails when test failed and when they get fixed. Any chance we can make it that way again?

[munrocket]

@mrdoob looks strange, I am not changed anything with emails, do you started receive this emails yesterday after #18745?

Here documentation from Travis

notifications:
  email:
    recipients:
      - one@example.com
      - other@example.com
    on_success: never # default: change
    on_failure: always # default: always

emails can be hided.

[mrdoob]

Yes, they started yesterday.

[munrocket]

Fixed #18776

[munrocket]

@Mugen87 because this stage exist in CI and usually we have ESLint in code editor. You have another cases?

[Mugen87]

I was just surprised that the familiar command was suddenly gone, that's all.

[munrocket]

Need help for good determinism in video/audio tag, other contributions also welcome.
In readme I am described current problems.

[mrdoob]

I think we can ignore these examples. These tests are to make sure we don't break the renderer, video/audio is handled by the browser.

[munrocket]

Sure that it can be solved with hooks, but #18782 for now.

[mrdoob]

I just ran npm run make-screenshot physics_cannon_instancing and I got this... 🤔

13a0fc2?short_path=45d17f6#diff-45d17f64c08d97c9b0ec35e778f822df

[munrocket]

Weird, did you ever try to run npm run test-e2e on your machine? I have a good screenshot.

[mrdoob]

Weird, did you ever try to run npm run test-e2e on your machine?

I didn't. I'll try later.

[mrdoob]

I think this should have been RAF( cb ).