Single-bar histograms don't respect manual bins specs when data lies in final bin

[andrewkfiedler]

Here are the relevant codepens to reproduce:
http://codepen.io/andrewkfiedler/pen/ENLgJg
http://codepen.io/andrewkfiedler/pen/zojKgK

Both show different behavior, but both show that manual binning gets ignored.

[etpinard]

What is the desired behavior here in your mind

[andrewkfiedler]

In my mind, and judging by the docs (see https://plot.ly/python/histograms/#colored-and-styled-histograms), if a user gives manual bins then data should be binned accordingly. In my example, the bins should be 0 to 20000, and 20000 to 40000 regardless of the data.

Without this behavior, it's hard to make useful overlaid histograms. I went ahead and made another codepen example showing so here: http://codepen.io/andrewkfiedler/pen/YpLBdR?editors=0010

Notice that in the second histogram, there doesn't appear to be an overlay when there should. The bin farthest to the right on the histogram should be twice as dark. However, the data is getting binned as seen in the first codepen I posted here: http://codepen.io/andrewkfiedler/pen/ENLgJg. Instead of getting binned into the 20000 to 40000 bin, it makes it's own bin from 29.9996k to 30.0004k.

My particular use case is similar to what you see here: https://plot.ly/r/shinyapp-linked-brush/

I have a dataset plotted on a histogram and from another view I'm allowing them to select specific results that make up the histogram. When they select those results, I'm trying to highlight the bins that those results belong to (and to what amount they make up those bins).

[andrewkfiedler]

Anything else I can provide to help out on this? I looked into plotly's code a bit but couldn't figure out where the bins are being overridden.

[etpinard]

Our apologies, we've been a little busy lately and haven't had the time to look at this report in details.

At first glance it seems like your example over-determines the bins specifications: how can a bin start at 0, end at 4000 and have a size of 2000?

To help us out a little more, would you mind looks at other cases e.g. cases with more than 1 bin?

[andrewkfiedler]

It's all good, I figured you guys are busy (especially with the holidays).

The bin specifications are fine (at least judging by docs: https://plot.ly/python/histograms/#colored-and-styled-histograms). In the example linked they use: start=-1.8, end=4.2, size=0.2
The start and end values correspond to the range of all bins, not a single bin. The size is what determines how to go from the start to the end.

I went ahead and made another codepen showing the behavior when more data is supplied:
http://codepen.io/andrewkfiedler/pen/WoYwaJ
The trace is [10000, 30000] with a manual binning specified as before start=0, end=40000, size=20000.
Yet the final histogram is binned from 2500 to 17500 and 22500 to 37500 (start=2500, end=37500, size=15000). So neither the start and end are respected, or the size.

[etpinard]

Ah I see. You had both bargap and bargroupgap set. Commenting those out as in http://codepen.io/etpinard/pen/woQWgE? gets you the desired result (I think).

[andrewkfiedler]

You're my hero, I was really hoping it was a configuration mistake!

I'll try it out in my app and make sure it works.

[andrewkfiedler]

Ah dang, I tried your pen with less data and it didn't work.
See http://codepen.io/andrewkfiedler/pen/MbzeOW.

I updated the trace to only have one point at 22000, but it still bins around 29.9996k to 30.0004k.

Interestingly, if I update the trace to only have one point at 5000 it works correctly, see http://codepen.io/andrewkfiedler/pen/JbeKMW.

[andrewkfiedler]

I think I've tracked down the issue. The manual binning appears to work correctly, so long as the data is not limited to only the final bin.

I've made some new codepens to demonstrate this. I've updated the manual binning to be start=0, end=80000, size=20000. So there are four possible bins. I've added a range on the xaxis from 0 to 80000 as well.
Data in Bin 1: http://codepen.io/andrewkfiedler/pen/ENOyLN?editors=0010
Data in Bin 2: http://codepen.io/andrewkfiedler/pen/ENOyRN?editors=0010
Data in Bin 3: http://codepen.io/andrewkfiedler/pen/VmVjdy?editors=0010
Data in Final Bin: http://codepen.io/andrewkfiedler/pen/BQGzVv?editors=0010

As you can see, it's only the final bin that appears to experience the issue.

Do note that I also added a manual range, as it appears the range gets determined from the data rather than the manual binning. Not sure if that's as big of an issue, since I can specify range to get around it.

[etpinard]

it appears the range gets determined from the data rather

That's correct. This is consistent with all our other trace types.

[etpinard]

Ok @andrewkfiedler looks like http://codepen.io/andrewkfiedler/pen/BQGzVv?editors=0010 does indeed demonstrate a bug. Big thanks for hunting that one down!

I'll change the issue title accordingly. Please note that single-bar histograms aren't very common, so this bug won't take up a high priority on our list.

[andrewkfiedler]

No problem, thanks for helping!

No big deal on the priority, a workaround is to specify the end of the bins one size more than you'll actually need (so nothing ends up in the "last" bin).

[etpinard]

a workaround is to specify the end of the bins one size more than you'll actually need (so nothing ends up in the "last" bin).

Great. Thanks again!

[alexcjohnson]

There have been a bunch of histogram-related fixes in the last couple of months (#2413, #2113, #2028, #1944...), it looks to me like one of them fixed this issue but between all the back-and-forth here I can't exactly tell. @andrewkfiedler or @etpinard can you confirm?

[etpinard]

http://codepen.io/andrewkfiedler/pen/BQGzVv?editors=0010 now looks ok to me:

after since 1.31.0, so I'll close this thing. @andrewkfiedler feel free to open another bug report if your use case isn't fully fixed. Thanks!

[datalifenyc]

Since the codepen examples are in JS, I thought a python note may be helpful for others who come across the same issue. xaxis_range can be used to specify the desired range. In the example below, I needed 8 bins (size = 1/8). Thanks to @andrewkfiedler and @etpinard for the clues.

fig = go.Figure()
fig.add_trace(go.Histogram(
    x = pred_prob, # numpy.ndarray
    xbins = dict(
        start = 0.0,
        end = 1.0,
        size = 1/8,
    )
))

fig.update_layout(
    title_text = 'Sampled Results',
    xaxis_title_text = 'Probability Range',
    yaxis_title_text = 'Count',
    xaxis_range = [0, 1]
)
fig.show()