I was able to track the issue down to the changes made in this patch:
https://gerrit.wikimedia.org/r/331563
This patch is mostly a simplification of ce8a2d4
(I9a73b3086fc8fb0edd897a347b5497d5362e20ef):
- Don't make wait#wait() abortable. This adds complexity and isn't
needed since the token is rechecked in the .then() of
actions#linkDwell(). The request is permitted to continue and fetch if
that token still matches and is never issued otherwise.
Once a request has been issued, that request is still abortable.
However, note that calling XHR.abort() is just a request to abort and
may not be granted. Whether or not XHR.catch() is invoked is what
dispatches the FETCH_ABORTED action (or doesn't if the request to
abort was denied).
- Remove the abort tests for wait#wait() since the functionality is no
longer provided.
- Pass the preview token in the FETCH_ABORTED action and reduce
FETCH_ABORTED the same way as ABANDON_COMPLETE in reducers/preview.
The follow-up patch, I3597b8025f7a12db0cf5d83cce5a77abace9bae3, adds
integration tests for the specific bug fix. Note that these Selenium
tests are incompatible with the content proxy, so it is probably best to
simply unset $wgPopupsRestGatewayEndpoint and $wgPopupsGateway and allow
the defaults to be used. Also note that the tests are incompatible with
recent versions of Node.js (so use NVM) and emit many deprecation
warnings (so set deprecationWarnings to false in
tests/selenium/wdio.conf.js) An example run of the tests looks something
like:
chromedriver --url-base=wd/hub --port=4444 &
# If any changes are made locally, also run `npm -s start &`.
MW_SERVER=http://localhost:8181 \
MEDIAWIKI_USER=foo \
MEDIAWIKI_PASSWORD=bar \
DISPLAY= \
npm -s run selenium-test
Live testing may be performed as well. Remember that RESTBase requests
are incompatible with MobileFrontend's content proxy hack so ensure to
comment it out if $wgPopupsGateway is configured for RESTBase (see
T218159).
1. Open the DevTools network tab.
2. Disable the browser cache. Chromium, at least, won't abort requests
coming form the cache.
3. Hover back and forth quickly over a preview. In Chromium, canceled
requests are labeled and appear red. This is a good scenario to test.
With the patch, a preview should always be shown when ultimately
resting on a link. Without the patch, it is possible to rest upon the
link with no preview showing. This may require several attempts.
Bug: T219434
Change-Id: I9da84b0296dd14e9ce69cb35f1ca475272fb249a
Whenever an HTTP request sequence is started, i.e. wait for the fetch
start time, issue a network request, and return the result, abort the
process if the results are known to no longer be needed. This occurs
when a user has dwelt upon one link and then abandoned it either during
the fetch start wait time or during the fetch network request itself.
This change is accomplished by preserving the pending promises in two
actions, LINK_DWELL and FETCH_START, and whenever the ABANDON_START
action is issued, it now aborts any previously pending XHR-like promise,
called a "AbortPromise" which is just a thenable with an abort() method.
There is a similar concept in Core:
ecc812f06e/resources/src/mediawiki.api/index.js.
Aborting pending requests has big implications for client and server
logging as requests are quickly canceled, especially on slower
connections. These differences can be observed on the network tab of
DevTools and the log in Redux DevTools.
Consider, for instance, the scenario of dwelling upon and quickly
abandoning a single link prior to this patch:
BOOT EVENT_LOGGED LINK_DWELL FETCH_START ABANDON_START FETCH_END STATSV_LOGGED ABANDON_END EVENT_LOGGED FETCH_COMPLETE
And after this patch when the fetch timer is canceled (prior to an
actual network request):
BOOT EVENT_LOGGED LINK_DWELL ABANDON_START ABANDON_END EVENT_LOGGED
In the above sequence, FETCH_* and STATSV_LOGGED actions never occur.
And after this patch when the network request itself is canceled:
BOOT EVENT_LOGGED LINK_DWELL FETCH_START ABANDON_START FETCH_FAILED STATSV_LOGGED FETCH_COMPLETE ABANDON_END EVENT_LOGGED
FETCH_FAILED occurs intentionally, STATSV_LOGGED and FETCH_COMPLETE
still happen even though the fetch didn't complete successfully, and
FETCH_END doesn't.
Additionally, since less data is transmitted, it's possible that the
timing and success rate of logging will improve on low bandwidth
connections.
Also, this patch tries to revise the JSDocs where possible to support
type checking and fix a call to the missing assert.fail() function in
changeListener.test.js.
Bug: T197700
Change-Id: I9a73b3086fc8fb0edd897a347b5497d5362e20ef
Remove usages of deprecated methods like .done which make jquery
promises fall back to non-standard behavior
Additional changes:
* Rename var promise to a more descriptive name in tests
Bug: T173819
Change-Id: I7b041d0a7a8c42a8eac947295d265e898085c60a
Simplify all our tests to return to promises
Use catch rather than fail when testing error cases.
Bug: T170812
Change-Id: I37c4e3f86343052c946d8586f8ff840a81f631f8
Why: Because they are the approved standard by TC39 and Ecma for
JavaScript modules.
Changes:
* Wrap mw-node-qunit in run.js to register babel to transpile modules
for node v6
* Change all sources in src/ to use ES modules
* Change constants.js to be able to run without
jQuery.bracketedDevicePixelRatio given ES modules are hoisted to
the top by spec so we can't patch globals before importing it
* Change all tests in tests/node-qunit/ to use ES modules
* Drop usage of mock-require given ES modules are easy to stub with
sinon
Additional changes:
* Rename tests/node-qunit/renderer.js to renderer.test.js to follow
the convention of all the other files
* Make npm run test:node run only .test.js test files so that it
doesn't run the stubs.js or run.js file.
Bug: T171951
Change-Id: I17a0b76041d5e2fd18e2d54950d9d7c0db99a941
