Download as png a small size chart does not show all data

[fparga]

When downloading as png a chart (in that case an histogram), the downloaded image doesn't display all the data visible on the webpage chart.

Chart on the webpage:

Downloaded png chart:

Jsbin: https://jsbin.com/gomina/1/edit?js,output

[etpinard]

Bug confirmed. Thanks very much for providing a jsbin example.

After some investigation, the bug seems to be only present in 1-data pt histogram traces.

Possible workaround: modify all the 1-data pt histogram traces to bars with y: [1].

[alexcjohnson]

The browser vs download difference here is really one of rounding and a peculiarity of how we have to draw bar charts. With bargap=0 (as is the default for histograms) or when you stack bars (also the case here) we need to disable antialiasing (css shape-rendering: crispEdges;) otherwise you get horrible artifacts where bars meet. But then when you draw with low resolution, some items simply disappear.

It's kind of an interesting question actually, what we should do with bar widths in this plot. We don't currently have any way to explicitly set the width of bars in a bar chart / histogram, they just get set as wide as they can with the constraints that 1) they're all the same within one trace, 2) they don't overlap each other within one trace, and 3) if barmode is 'stack' or 'group', all the traces get the same bar width.

In this plot you've explicitly set bin widths, which SHOULD be the same as the bar widths if we want to interpret histograms in the normal way. BUT because barmode='stack' all the widths have to be the same, so we have conflicting demands. What to do?

[alexcjohnson]

@etpinard you still want to call this a bug? I suppose could imagine a rule like "if we're going to rasterize, AND a bar (stack) has width < 1 px AND there's no bar to draw on at least one side of it, expand it to at least one px width. But that's super complicated and I still think it sounds sketchy - what if for example the plot is a comb, consisting of 0.1 px bars separated by 0.1 px? Then you have to figure out that there's a bar near (but not quite touching) and not expand so much that you overlap it... and even then you give a misleading picture, because it would look like the space is filled when really it's only half full.

The real problem here seems to me the conflicting demands on bar widths. How do we handle grouped or stacked histograms with explicitly different bin widths? I'd almost like to make that an error, but at the very least we should let the user know somehow, and harmonize them before binning, not during plotting, so we can ensure that bin width and bar width are always the same.

[fparga]

The thing is, those bins are the ones use by plotly by default. I didn't specified them I juste took the data directly from plotly in my example. Which is related to #50
See: https://jsbin.com/gugerelemu/1/edit?js,console,output

That's why the image output is exactly the same even when no bin size is specified.

[fparga]

My question might be naive, but does the downloaded image necessary needs to be the same size as the displayed plot?

[alexcjohnson]

yup, #50 is the crux - what we should really be doing here is calculating the histogram autobins taking all histogram traces into account at the same time. Then you'd probably get a bin size of 1 or 2 and this would all be moot.

From the plot.ly workspace you can download your image at arbitrary (higher) resolution, but the version on the plot itself is not so full featured.

[fparga]

Any chance this might change in the future? That's quite a big limitation..

[etpinard]

@alexcjohnson @Ni-Hil

you still want to call this a bug?

No. This issues that @Ni-Hil pointed out are now split into #50 and #83 .

yup, #50 is the crux - what we should really be doing here is calculating the histogram autobins taking all histogram traces into account at the same time.

We would have to be careful here. Calculating histogram autobins from all traces might break some plots that maneuver around that problem.

[alexcjohnson]

Migrated part of this discussion to #50 and @etpinard opened #83 with another part - is there anything left here or should we close it?

[fparga]

I guess there is nothing more. Fixing #50 should fix this, so I'm closing.

[amirmog]

Hello folks,
I'm still encountering this issue.
I'm drawing bar charts and some of the bars are missing.

Any ideas how to fix it?

Thanks,