Web frontend black-box Puppeteer tests

While our node test suite is the preferred way to test most frontend code because they are easy to write and maintain, some code is best tested in a real browser, either because of navigation (E.g. login) or because we want to verify the interaction between Zulip logic and browser behavior (E.g. copy/paste, keyboard shortcuts, etc.).

Running tests

You can run this test suite as follows:

    tools/test-js-with-puppeteer

See tools/test-js-with-puppeteer --help for useful options, especially running specific subsets of the tests to save time when debugging.

The test files live in frontend_tests/puppeteer_tests and make use of various useful helper functions defined in frontend_tests/puppeteer_lib/common.js.

How puppeteer tests work

The Puppeteer tests use a real Chromium browser (powered by puppeteer), connected to a real Zulip development server. These are black-box tests: Steps in a Puppeteer test are largely things one might do as a user of the Zulip webapp, like “Type this key”, “Wait until this HTML element appears/disappears”, or “Click on this HTML element”.

For example, this function might test the x keyboard shortcut to open the compose box for a new private message:

async function test_private_message_compose_shortcut(page) {
    await page.keyboard.press("KeyX");
    await page.waitForSelector("#private_message_recipient", {visible: true});
    await common.pm_recipient.expect(page, "");
    await close_compose_box(page);
}

The test function presses the x key, waits for the #private_message_recipient input element to appear, verifies its content is empty, and then closes the compose box. The waitForSelector step here (and in most tests) is critical; tests that don’t wait properly often fail nonderministically, because the test will work or not depending on whether the browser updates the UI before or after executing the next step in the test.

Black-box tests are fantastic for ensuring the overall health of the project, but are also slow, costly to maintain, and require care to avoid nondeterministic failures, so we usually prefer to write a Node test instead when both are options.

They also can be a bit tricky to understand for contributors not familiar with async/await.

Debugging Puppeteer tests

The following questions are useful when debugging Puppeteer test failures you might see in continuous integration:

  • Does the flow being tested work properly in the Zulip browser UI? Test failures can reflect real bugs, and often it’s easier and more interactive to debug an issue in the normal Zulip development environment than in the Puppeteer test suite.

  • Does the change being tested adjust the HTML structure in a way that affects any of the selectors used in the tests? If so, the test may just need to be updated for your changes.

  • Does the test fail deterministically when you run it locally using E.g. ./tools/test-js-with-puppeteer 03? If so, you can iteratively debug to see the failure.

  • Does the test fail nondeterministically? If so, the problem is likely that a waitForSelector statement is either missing or not waiting for the right thing. Tests fail nondeterministically much more often on very slow systems like those used for Continuous Integration (CI) services because small races are amplified in those environments; this often explains failures in CI that cannot be easily reproduced locally.

These tools/features are often useful when debugging:

  • You can use console.log statements both in Puppeteer tests and the code being tested to print-debug.

  • Zulip’s Puppeteer tests are configured to generate screenshots of the state of the test browser when an assert statement fails; these are stored under var/puppeteer/*.png and are extremely helpful for debugging test failures.

  • TODO: Mention how to access Puppeteer screenshots in CI.

  • TODO: Add an option for using the headless: false debugging mode of puppeteer so you can watch what’s happening, and document how to make that work with Vagrant.

  • TODO: Document --interactive.

  • TODO: Document how to run 100x in CI to check for nondeterminstic failures.

  • TODO: Document any other techniques/ideas that were helpful when porting the Casper suite.

  • The Zulip server powering these tests is just run-dev.py with some extra Django settings from zproject/test_extra_settings.py to configure an isolated database so that the tests will not interfere/interact with a normal development environment. The console output while running the tests includes the console output for the server; any Python exceptions are likely actual bugs in the changes being tested.

See also puppeteer upstream’s debugging tips; some tips may require temporary patches to functions like run_test or ensure_browser in frontend_tests/puppeteer_lib/common.js.

Writing Puppeteer tests

Probably the easiest way to learn how to write Puppeteer tests is to study some of the existing test files. There are a few tips that can be useful for writing Puppeteer tests in addition to the debugging notes above:

  • Run just the file containing your new tests as described above to have a fast debugging cycle.

  • When you’re done writing a test, run it 100 times in a loop to verify it does not fail nondeterminstically (see above for notes on how to get CI to do it for you); this is important to avoid introducing extremely annoying nondeterministic failures into master.

  • With black-box browser tests like these, it’s very important to write your code to wait for browser’s UI to update before taking any action that assumes the last step was processed by the browser (E.g. after you click on a user’s avatar, you need an explicit wait for the profile popover to appear before you can try to click on a menu item in that popover). This means that before essentially every action in your Puppeteer tests, you’ll want to use waitForSelector or similar wait function to make sure the page or element is ready before you interact with it. The puppeteer docs site is a useful reference for the available wait functions.

  • When using waitForSelector, you always want to use the {visible: true} option; otherwise the test will stop waiting as soon as the target selector is present in the DOM even if it’s hidden. For the common UI pattern of having an element always be present in the DOM whose presence is managed via show/hide rather than adding/removing it from the DOM, waitForSelector without visible: true won’t wait at all.

  • The test suite uses a smaller set of default user accounts and other data initialized in the database than the normal development environment; specifically, it uses the same setup as the backend tests. To see what differs from the development environment, check out the conditions on options["test_suite"] in zilencer/management/commands/populate_db.py.