wiki:RebaselineServer

Version 3 (modified by mihaip@chromium.org, 13 years ago) (diff)

Rebaseline server: begin UI walkthrough

Rebaseline Server

The "Rebaseline Server" is a tool that makes it easy to do "mass" rebaselines (updating of pixel or text baselines for dozens or hundreds of layout tests). It makes it easy to review changes to baselines and update existing ones (and optionally move them, to handle OS transitions). It's known as a server because it runs a local HTTP server which makes it easy to have a cross-platform GUI for reviewing that is backed by Python code for doing the SCM (SVN/Git) operations on the filesystem.

UI Walkthrough

1. Test selectors

These menus let you choose which test(s) you want to rebaseline. The first menu breaks down tests by failure type. Within that failure type, the second menu filters by directory. The third menu lists individual tests. You can use the « and » buttons to move through the tests in the chosen directory (there are also keyboard shortcuts). The View test link shows the test's source on trac.webkit.org.

Keyboard shortcuts

The following keyboard shortcuts are supported:

  • left and right arrow: move through tests
  • q: add the current test to the rebaseline queue
  • x: remove the current test from the rebaseline queue
  • r: rebaseline all the tests in the queue

Sample Usages

Leopard to Snow Leopard Transition

Snow Leopard has a different text anti-aliasing algorithm (and other small rendering tweaks) which requires many pixel baselines to be updated. The flow for doing this would be.

  1. (On a Snow Leopard machine) run new-run-webkit-tests --tolerance 0 --pixel-tests [path/to/tests] (or omit the path to run all tests)
  2. Observe that many tests failed with imaged differences
  3. Run webkit-patch rebaseline-server WebKitBuild/{Debug|Release}/layout-test-results (or wherever your layout test results are localed)
  4. http://localhost:8127/ should be launched in your browser, navigate to it if not
  5. From the footer of the UI, set the baseline target to mac (since we want the default mac port baselines to reflect the newest shipping OS)
  6. Similarly, set the Move current baselines to: dropdown to mac-leopard, so that the existing baselines are still used for bots that are on Leopard
  7. Refer to the UI walkthrough for more details on how to use the GUI
  8. Commit the result patch with webkit-patch as usual

If working on the Chromium port, the above steps can be used there too, except the baseline target is chromium-mac and existing baselines should be moved to chromium-mac-leopard.

Skia changes

Skia is the graphics library used by the Chromium port on Windows and Linux; changes to it often require updates to pixel baselines.

  1. (On a Windows or Linux machine) run new-run-webkit-tests --tolerance 0 --pixel-tests [path/to/tests] (or omit the path to run all tests)
  2. Observe that many tests failed with imaged differences
  3. Run webkit-patch rebaseline-server WebKitBuild/{Debug|Release}/layout-test-results (or wherever your layout test results are localed)
  4. http://localhost:8127/ should be launched in your browser, navigate to it if not
  5. From the footer of the UI, set the baseline target to chromium-win or chromium-win as appropriate.
  6. Leave the Move current baselines to: dropdown set to Nowhere (replace)
  7. Refer to the UI walkthrough for more details on how to use the GUI
  8. Commit the result patch with webkit-patch as usual

Note that if you'd like to rebaseline both Windows and Linux in the same patch, you'll need to run the tests on both platforms by hand and copy the results file to the machine that's running the rebaseline server (see the caveats section).

Code location and design

The Python server lives at WebKitTools/Scripts/webkitpy/tool/commands/rebaselineserver.py, the GUI is at WebKitTools/Scripts/webkitpy/tool/commands/data/rebaselineserver/.

The Python server parses the unexpected_results.jsonfile generated by new-run-webkit-tests. It then starts an HTTP server. That server can respond with the list of failing tests, expected/actual output for a test (read from the test results directory), or to commands to rebaseline tests. The latter is done by using the scm.py wrapper that makes it agnostic to Git vs. SVN checkouts.

The UI populates various <select>s based on the test output, grouping results by failure type, directory and test. The "loupe" functionality is accomplished by rendering pixel tests into a <canvas> and then drawing the enlarged pixels into another <canvas> (see loupe.js). The queue is maintained as a <select> as well (see queue.js).

Caveats/Limitations

  • new-run-webkit-tests must be used (the tool currently relies on the JSON output that only NRWT produces)
  • The tool can currently only use local test results, it cannot pull from bots
  • The UI works best on 30 inch monitor

Attachments (4)

Download all attachments as: .zip