This PR improves the integration tests in two ways:
1. Make the integration tests more reliable and thus less flakey
2. Make the integration tests faster (by introducing concurrency)
Tried a lot of different things to make sure that these tests are fast
and stable.
---
The biggest issue we noticed is that some tests are flakey, these are
tests with long running dev-mode processes where watchers are being used
and/or dev servers are created.
To solve this, all the tests that spawn a process look at stdout/stderr
and wait for a message from the process to know whether we can start
making changes.
For example, in case of an Astro project, you get a `watching for file
changes` message. In case of Nuxt project you can wait for an `server
warmed up in` and in case of Next.js there is a `Ready in` message.
These depend on the tools being used, so this is hardcoded per test
instead of a magically automatic solution.
These messages allow us to wait until all the initial necessary work,
internal watchers and/or dev servers are setup before we start making
changes to the files and/or request CSS stylesheets before the server(s)
are ready.
---
Another improvement is how we setup the dev servers. Before, we used to
try and get a free port on the system and use a `--port` flag or a
`PORT` environment variable. Instead of doing this (which is slow), we
rely on the process itself to show a URL with a port. Basically all
tools will try to find a free port if the default port is in use. We can
then use the stdout/stderr messages to get the URL and the port to use.
To reduce the amount of potential conflicts in ports, we used to run
every test and every file sequentially to basically guarantee that ports
are free. With this new approach where we rely on the process, I noticed
that we don't really run into this issue again (I reran the tests
multiple times and they were always stable)
<img width="316" alt="image"
src="https://github.com/user-attachments/assets/b75ddab4-f919-4995-85d0-f212b603e5c2"
/>
Note: these tests run Linux, Windows and macOS in this branch just for
testing purposes. Once this is done, we will only run Linux tests on PRs
and run all 3 of them on the `next` branch.
We do make the tests concurrent by default now, which in theory means
that there could be conflicts (which in practice means that the process
has to do a few more tries to find a free port). To reduce these
conflicts, we split up the integration tests such that Vite, PostCSS,
CLI, … tests all run in a separate job in the GitHub actions workflow.
<img width="312" alt="image"
src="https://github.com/user-attachments/assets/fe9a58a1-98eb-4d9b-8845-a7c8a7af5766"
/>
Comparing this branch against the `next` branch, this is what CI looks
like right now:
| `next` | `feat/improve-integration-tests` |
| --- | --- |
| <img width="594" alt="image"
src="https://github.com/user-attachments/assets/540d21eb-ab03-42e8-9f6f-b3a071fc7635"
/> | <img width="672" alt="image"
src="https://github.com/user-attachments/assets/8ef2e891-08a1-464b-9954-4153174ebce7"
/> |
There also was a point in time where I introduced sequential tests such
that all spawned processes still run after each other, but so far I
didn't run into issues if we keep them concurrent so I dropped that
code.
Some small changes I made to make things more reliable:
1. When relying on stdout/stderr messages, we split lines on `\n` and we
strip all the ANSI escapes which allows us to not worry about special
ANSI characters when finding the URL or a specific message to wait for.
2. Once a test is done, we `child.kill()` the spawned process. If that
doesn't work, for whatever reason, we run a `child.kill('SIGKILL')` to
force kill the process. This could technically lead to some memory or
files not being cleaned up properly, but once CI is done, everything is
thrown away anyway.
3. As you can see in the screenshots, I used some nicer names for the
workflows.
| `next` | `feat/improve-integration-tests` |
| --- | --- |
| <img width="276" alt="image"
src="https://github.com/user-attachments/assets/e574bb53-e21b-4619-9cdb-515431b255b9"
/> | <img width="179" alt="image"
src="https://github.com/user-attachments/assets/8bc75119-fb91-4500-a1d0-bd09f74c93ad"
/> |
They also look a bit nicer in the PR overview as well:
<img width="929" alt="image"
src="https://github.com/user-attachments/assets/04fc71fc-74b0-4e7c-9047-2aada664efef"
/>
The very last commit just filters out Windows and macOS tests again for
PRs (but they are executed on the `next` branch.
---
### Nest steps
I think for now we are in a pretty good state, but there are some things
we can do to further improve everything (mainly make things faster) but
aren't necessary. I also ran into issue while trying it so there is more
work to do.
1. More splits — instead of having a Vite folder and PostCSS folder, we
can go a step further and have folders for Next.js, Astro, Nuxt, Remix,
…
2. Caching — right now we have to run the build step for every OS on
every "job". We can re-use the work here by introducing a setup job that
the other jobs rely on. @thecrypticace and I tried it already, but were
running into some Bun specific Standalone CLI issues when doing that.
3. Remote caching — we could re-enable remote caching such that the
`build` step can be full turbo (e.g.: after a PR is merged in `next` and
we run everything again)
This ensures our glob hoisting mechanism (see #14896) works on Windows
when performing an upgrade.
---------
Co-authored-by: Jordan Pittman <jordan@cryptica.me>
Fixes: #14839Fixes: #14796
This PR fixes an issue in the Vite extension where we previously only
ran a small list of allow-listed plugins for the second stage transform
in the build step. This caused some CSS features to unexpectedly not
work in production builds (one such example is Vue's `:deep(...)`
selector).
To fix this, I changed the allow listed plugins that we do want to run
to a block list to filter out some plugins we know we don't want to run
(e.g. the Tailwind Vite plugin for example or some built-in Vite plugins
that are not necessary).
## Test plan
This PR adds a new integration test suite to test interop with a custom
Vite transformer that looks like this:
```js
{
name: 'recolor',
transform(code, id) {
if (id.includes('.css')) {
return code.replace(/red/g, 'blue')
}
},
}
```
I also validated that this does indeed fix the Vue `:deep(...)` selector
related issue that we were seeing by copying the repro of #14839 into
our playground:
You can see in the screenshot above that the `:deep()` selector
overwrites the scoped styles as expected in both the dev mode and the
prod build (screenshotted).
Furthermore I reproduced the issue reported in
https://github.com/tailwindlabs/tailwindcss/issues/14796 and was able to
confirm that in a production build, the styling works as expected:
<img width="517" alt="Screenshot 2024-11-06 at 14 26 50"
src="https://github.com/user-attachments/assets/ade6fe38-be0d-4bd0-9a9a-67b6fec05ae0">
Lastly, I created a repository out of the biggest known-to-me Vite
projects: [Astro, Nuxt, Remix, SolidStart, and
SvelteKit](https://github.com/philipp-spiess/tailwind-playgrounds) and
verified that both dev and prod builds show no issue and the candidate
list is properly appended in each case.
---------
Co-authored-by: Jordan Pittman <jordan@cryptica.me>
Co-authored-by: Adam Wathan <adam.wathan@gmail.com>
There are still instances in which CI is flaky after #14332. This PR
applies the same fix (that is, moving the file write into the retrying
block) to all `retryAssertion` callbacks.
We noticed that Nuxt projects were not working with the tailwindcss
project. The issue was traced down to the fact that Nuxt starts multiple
Vite dev servers and calling the experimental `waitForRequestsIdle()` on
one of the test servers would never resolve.
This was fixed upstream and is part of the latest Vite/Nuxt release:
https://github.com/vitejs/vite/issues/17980.
We still need to handle the fact that Vite can spawn multiple dev
servers. This is necessary because when we invalidate all roots, we need
to find that module inside all of the spawned servers. If we only look
at the _last server_ (what we have done before), we would not find the
module and thus could not invalidate it.
