#110473Installer e2e test cannot recover from asset load error
Description
Problem
The playwright installer specs fail sporadically on their very first
expectation, with the browser showing an empty, styled page:
bc. e2e-install/install-mysql.spec.ts:25 – install TYPO3 on MySQL (10.3s)
Error: expect(locator).toBeVisible() failed
Locator: getByText(‘Installing TYPO3’)
Timeout: 10000ms
Error: element(s) not found
Observed in nightly job 5788225 of pipeline 106643 on branch 14.3. The
trace attached to that job shows the cause. The document and the style
sheets are served with 200, about 45 ES modules load fine, and then five
module requests are aborted within a 13 millisecond window:
bc. CONSOLE error Failed to load resource: net::ERR_NETWORK_CHANGED (5 times)
11:20:05.999 storage/client.js
11:20:05.999 cache/dedupe-async-task.js
11:20:06.000 Contrib/lit-html/directives/class-map.js
11:20:06.001 Contrib/lit-html/directives/style-map.js
11:20:06.012 Contrib/@lit/reactive-element/decorators/base.js
Why the test cannot recover
The installer entry template renders an empty body and loads the first
step client side, so the visible headline the spec waits for only exists
after installer.js has run. Chromium does not retry a module request that
failed with ERR_NETWORK_CHANGED, so the module graph never resolves and
installer.js never executes. The body stays empty for the rest of the
test.
installer.js has no rejection handling at all on its request chain
(mainLayout, checkInstallerAvailable, checkEnvironmentAndFolders,
showEnvironmentAndFolders), so even a failure later in that chain leaves
a permanently blank page rather than an error message.
Note on the trigger
ERR_NETWORK_CHANGED is raised by the browser when the network
configuration changes while requests are in flight. Every container
attached to or detached from the container network reconfigures it, and
the CI runners execute several jobs with their own container networks at
the same time.
Forge issue 110472 reduces the container churn of the test setup by
probing all endpoints from a single container, but that does not make
the test itself resilient.
Possible directions
- Make the spec tolerate a failed initial asset load, for instance by
reloading once when the installer body did not render. - Add rejection handling to installer.js so a failed request renders an
error instead of an empty page. That is user facing behaviour, not only
a test concern.
Branches
main, 14.3.
Reviews 3
Related issues 1
History 25
-
Stefan Bürk changed Related to #110472 <span data-processor="relativetime" data-value="2026-08-13T11:34:24+00:00" > 08/13/26, 11:34 </span>
-
Gerrit Code Review pushed patch set 3 for main to change 95313 status
NewUnder Review <span data-processor="relativetime" data-value="2026-08-13T12:46:42+00:00" > 08/13/26, 12:46 </span> -
Gerrit Code Review pushed patch set 3 for 14.3 to change 95314 <span data-processor="relativetime" data-value="2026-08-13T12:46:50+00:00" > 08/13/26, 12:46 </span>
-
Gerrit Code Review pushed patch set 3 for 13.4 to change 95315 <span data-processor="relativetime" data-value="2026-08-13T12:46:58+00:00" > 08/13/26, 12:46 </span>
-
Stefan Bürk commented and changed status
Under ReviewResolved done0 %100 % <span data-processor="relativetime" data-value="2026-08-13T14:15:13+00:00" > 08/13/26, 14:15 </span>
Applied in changeset commit:8815be4fd013efcad46534942f8c9ec2ef4786fe.