We discovered that in a tab restore scenario we were recording view time
observations that were wrong - we'd record time deltas as-if user was
looking at the page while the browser wasn't running.
This happens because when we record a viewTime observation, we compare
current time with lastAccess time of the tab. In a restore scenario,
that lastAccess time happens to be from when the browser was last
running - which could be days ago.
The simplest solution was to not record a viewTime observation if the
url for a tab didn't change during a load event. To achieve this, we
needed to change which action we were using as a proxy for "navigation
events" - UpdateUrlAction contains the new url, allowing us to compare
against the current tab url.
Alternative solutions would be to keep using loading actions, but
dispatch a lastAccess event before performing a metadata update. This
would have worked, but would result in two lastAccess events being
dispatched for each navigation event instead of just one.
We only instrument onCreate because it's the only one with an
implementation.
While declaring this as a function that accepts a lambda is less
fragile, we've previously had issues with it such as suspected memory
leaks when used for telemetry. Therefore, we go with the simpler
approach.
This is a speculative fix for the intermittent issue. Typically, these
intermittents are caused by mocked lambdas but there is no mocked lambda
here. If this doesn't work, one `any()` argument fills in for a lambda:
it's possible that's causing the failure.
Unfortunately, I can't verify this fix easily because the "run test until
failure" option was removed from Android Studio.
See d396c9eb41298cc07fbf136f1de971010bc45d97 for a prior attempt to
address the intermittents in this class.
This is a bug we noticed after landing search term grouping.
An adapter can submit an empty list of items to the `ConcatAdapter`
early. This has the side-effect of triggering our `observeFirstInsert`
too soon and therefore updating the visibility to show the empty tray
placeholder and never switches back.
Our solution is to keep a constant observer on the adapter so we can
perform the visibility check on every insert/remove.
Co-authored-by: Roger Yang <royang@mozilla.com>
The test as it exists relies on the robolectric lifecycle, which is hard
to predict, so it doesn't seem worth fixing the test. Writing the test
any other way would require excessive mocking, which also seems
impractical.
We've noticed that geckoview_example is ~300ms faster than fenix in cold page
load tests on arewefastyet for the Pixel 2. We suspect the main
difference is because geckoview_example runs with conditioned profiles
and fenix does not.
This PR is foremost an experiment to see if that's true because, after bug
1587542, we cannot run get results for fenix perftest PRs (i.e. it needs to be
merged into main). If we find that the results are not noisy, however, we
could end up leaving this in the tree. We've previously seen excessive
noise with fenix start up tests with conditioned profiles which is why
conditioned profiles are not currently enabled.
