This patch adds a general-purpose grid framework to parallelize
Playwright across multiple agents.
This patch adds two CLI commands to manage grid:
- `npx playwright experimental-grid-server` - to launch grid
- `npx playwrigth experimental-grid-agent` - to launch agent in a host
environment.
Grid server accepts an `--agent-factory` argument. A simple
`factory.js` might look like this:
```js
const child_process = require('child_process');
module.exports = {
name: 'My Simple Factory',
capacity: Infinity, // How many workers launch per agent
timeout: 10_000, // 10 seconds timeout to create agent
launch: ({agentId, gridURL, playwrightVersion}) => child_process.spawn(`npx`, [
'playwright'
'experimental-grid-agent',
'--grid-url', gridURL,
'--agent-id', agentId,
], {
cwd: __dirname,
shell: true,
stdio: 'inherit',
}),
};
```
With this `factory.js`, grid server could be launched like this:
```bash
npx playwright experimental-grid-server --factory=./factory.js
```
Once launched, it could be used with Playwright Test using env variable:
```bash
PW_GRID=http://localhost:3000 npx playwright test
```
This is an attempt to improve video performance when encoding
does not keep up with frames. This situation can be reproduced
by running multiple encoders at the same time.
Added `utils/video_stress.js` to manually reproduce this issue.
Observing ffmpeg logs, it does not do any encoding initially and
instead does "input analysis / probing" that detects fps and other
parameters. By the time it starts encoding (launches vpx and creates
the video file), we already have many frames in the buffer.
Reducing probing helps:
`-avioflags direct -fpsprobesize 0 -probesize 32 -analyzeduration 0`
Another issue observed is questionable default `-threads` value.
We compile without threads support, so logs say "using emulated threads".
For some reason, setting explicit `-threads 1` (or any other value)
makes it better when cpu is loaded.
- Each overload, e.g. for `page.evaluate`, shows a nice autocomplete doc,
not only the first one.
- We can have multiple overloads directly on the docs page, e.g.
`test.skip(title, fn)` and `test.skip(condition, description)`.
These overloads are internally named `Test.skip#1` and all aliased
to `test.skip`.
- Uses some auto fixtures to set default options and instrumentation on BrowserType.
- Moves screenshot, trace and video to worker-scoped fixtures.
- Throws in page/context when used from beforeAll/afterAll.
- Plumbs around BrowserType to be accessible from Browser and BrowserContext.
Each hook gets its own test scope. This is not too useful for
object fixtures like `page` (although one can use a page in
`beforeAll` to save storage state), but much more useful for option
fixtures like `viewport`.
