In the following scenario, we were adding SIGINT handler twice, but
removing it just once:
- Task runner starts testing, creates SigIntWatcher, installs SIGINT
handler.
- Press Ctrl+C, task runner interrupts, disarms SigIntWatcher, SIGINT
handler is not removed due to 1000ms cooldown.
- Task runner starts cleanup, creates SigIntWatcher, installs another
SIGINT handler.
- Cleanup finishes, SigIntWatcher disarms, could remove or not remove
SIGINT handler based on timing (same 1000ms cooldown). In any case, we
have one or two SIGINT handlers still on.
- HTML reporter hangs in onExit, while we still have SIGINT handler up,
so Ctrl+C does not exit.
Regressed in #24265.
A separate process is `spawnSync`'ed on process exit to cleanup
temporary directories, introduced in #13769 that followed up after
#13343.
A separate process might stall for various fs-related issues, which
prevents the original process from exiting.
With the recent changes, we always gracefully close and cleanup after
all launched executables before calling `process.exit()`, and so it
should only be possible to leave temp directories when using Playwright
and calling `process.exit()` programmatically without closing browsers.
We can now drop the extra process and rely on `rimraf.sync` for
last-resort cleanup in these rare circumstances.
This patch brings in antialiasing tests from `looks-same` project for
our experimental `ssim-cie94` comparator.
One of the new tests found a bug in our implementation.
References https://github.com/microsoft/playwright/issues/24312
Since
99ff08340f
search event is disabled in cocoa.
---------
Signed-off-by: Max Schmitt <max@schmitt.mx>
Co-authored-by: github-actions <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Max Schmitt <max@schmitt.mx>
Migrate to version 4 which returns a promise rather than leverages a
callback. -> https://www.npmjs.com/package/rimraf?activeTab=readme
- contains its own types, eliminate "@types/rimraf"
- Parameter `maxBusyTries` changed to `maxRetries`
We store shard number in the report metadata event and then sort shard
files by shard number. This guarantees that within each project sharded
events will always go in stable order.
Usually, we can just chain two locators with `>>` to implement
`Locator.locator(locator)`. However, this does not play nicely with more
advanced inner locators like `or` and `and`:
```ts
const child = page.locator('input').or(page.locator('button'));
page.locator('parent').locator(child);
```
One would expect the above to locate "input or button" inside a
"parent". However, currently it locates "input inside a parent" or
"button", because it's translated to `parent >> input >>
internal:or="button"`.
To fix this, we have to wrap inner locator into `internal:chain` and
query it separately from the parent.
Fixes#23724.
Partial fix for https://github.com/microsoft/playwright/issues/6319
After this fix, the following scenario won't leak and the context state
(cookies, storage, etc) can be reused by the new page sessions:
```js
for (let i = 0; i < 1000; ++i) {
const page = await context.newPage();
await page.goto('...');
await page.close('...');
}
```
Fixes#24144.
Previously, we only matched by url, which confuses GET and HEAD requests
where the latter is usually zero-sized.
Also make sure that resources are sorted by their monotonicTime, since
that's not always the case in the trace file, where they are sorted by
the "response body retrieved" time.
So, the following will work:
```
page.frameLocator('iframe').locator('span').or(page.frameLoactor('iframe').locator('div'))
```
The following will not work, because frame locators are not exactly the
same:
```
page.frameLocator('#iframe1').locator('span').or(page.frameLoactor('#iframe2').locator('div'))
```
Also improve the error message to be more readable and include the
locator.
Fixes#23697.
When rendering snapshot, disregard earlier resources with the same url,
because it's most likely that the latest one was used for rendering.
An example would be reloading the page before the stylesheet has
finished loading. In this case, the stylesheet will be requested twice,
and the second copy that was not aborted should be used for the
snapshot.
Fixes#23709.
For linux tests without tracing blob-report-1.zip takes 19M, while
unpacked size is 228 MB. That size is counted for GitHub artifact
billing:
<img width="434" alt="image"
src="https://github.com/microsoft/playwright/assets/9798949/5bc32511-6686-4581-a348-acb6a54cd99b">
We zip individual .jsonl reports so that they still have unique names
and can be easily uploaded into the same artifacts directory without
name collisions.
- properly annotate continued requests
- nest `attach` steps inside the related `expect` step
- fix primary-id-to-non-primary-id mapping
- make sure images in trace are not draggable
Fixes#23693
---------
Signed-off-by: Andrey Lushnikov <aslushnikov@gmail.com>
Co-authored-by: Max Schmitt <max@schmitt.mx>
Chained selectors where the second part starts with a scope did not work
before:
```ts
page.locator('div').locator(':scope + span')
page.locator('div >> +span')
```
Previously, such requests were skipped because they never receive
`Fetch.requestPaused` as there was no real network.
Also cleanup some redundant tests and move them from chromium-only file.
Fixes#23424.
This fixes following error in the merged reports:
```
(index):9825 TypeError: Cannot read properties of null (reading 'toFixed')
at msToString ((index):18917:15)
at TestFilesView ((index):19373:9)
at Uh ((index):9372:7)
at kj ((index):10444:7)
at Uk ((index):12508:86)
at Tk ((index):12134:11)
```
