Multiple regressions in display of charts in scalars tab.

[iansf]

Environment information (required)

Please run diagnose_tensorboard.py (link below) in the same
environment from which you normally run TensorFlow/TensorBoard, and
paste the output here:

Run on Google's internal infrastructure using latest prebuilt tensorboard binaries.

For browser-related issues, please additionally specify:

• Browser type and version: Chrome 76.0.3809.100 (Official Build) (64-bit)
• Screenshot, if it’s a visual issue:

Issue description

About a week ago, multiple regressions appeared in the scalars tab. One such regression is shown in the attached screenshot. This occurs intermittently when refreshing with the tensorboard refresh button.

Another regression is that, when refreshing, charts often fail to automatically expand their x dimension to accommodate new data.

Another regression is that closing and reopening a particular panel of charts used to refresh the charts. Now the charts remain cached across panel closes. This removed the easiest way to fix situations where multiple charts had been zoomed or panned, or were not updating correctly. Instead, the only way I have found to reset the charts (besides a full page reload) is by entering a random value into the Runs regex field such that no runs are selected, and then clearing out that value. This is much less convenient than just closing and reopening the problematic panel.

[wchargin]

screenshot [of scalar charts scrunched up to a too-small size]

See #190 and #204. This used to be much more common than it is now; my
understanding was that #204 had just about fixed it (for scalars), and I
can’t reproduce this via the usual process of switching dashboards while
the data is in the middle of reloading. I can reproduce this by
expanding a collapsed category that has never been expanded, then
quickly collapsing it again, then waiting for the data to load fully
before re-expanding it (guesswork); or by reloading data while a pane
that has previously been expanded is collapsed—but in either of these
cases triggering an in-page data reload via the button in the top right
fixes the problem.

Can you be more specific about what you’re doing when this occurs?

Tested on both latest public and latest Google-internal TensorBoard.

Another regression is that, when refreshing, charts often fail to
automatically expand their x dimension to accommodate new data.

See #1946. This is not a recent regression, but I do think that it
used to work better (in 2017) than it does now.

closing and reopening a particular panel of charts used to refresh the
charts. Now the charts remain cached across panel closes

True: collapsing and re-expanding currently displays the chart if no
reload has taken place; prior to #2392 (merged about 5 weeks ago), there
would be a delay after re-expanding, but the data would not be
re-fetched over the network. But this by itself is not a bug.

Instead, the only way I have found to reset the charts (besides a full
page reload) is by entering a random value into the Runs regex field
such that no runs are selected, and then clearing out that value

As an easier workaround, changing the horizontal axis from “step” to
“relative” and back (or any other pair) should fix this for all charts
at once.