fix: bin width < half spread #3646

[Gornoka]

prevents edge-case when a bin only has elements with spread < 0.5*bin_width .

I would personally prefer this option, with ceil instead of min(1,..).
This would alter previous behavior for binning in general( creating an additional bin in ~50% of cases),I would be interested to know if that is acceptable here or if you would like to use the min approach instead.

[mwaskom]

What is the edge case? Is there an issue that you can link to? Is there a failing test that you could add?

[Gornoka]

What is the edge case? Is there an issue that you can link to? Is there a failing test that you could add?

this is related to #3646 , i can also add a test.

[mwaskom]

I would personally prefer this option, with ceil instead of min(1,..).

Do you mean max(1, ...)?

That sounds like a better approach to me — IIUC the linked issue, the edge case is entirely about cases where the default bin rule gives "less than 1 bin" and that doesn't make sense. Why would we want to change other plots?

[Gornoka]

I would personally prefer this option, with ceil instead of min(1,..).

Do you mean max(1, ...)?

That sounds like a better approach to me — IIUC the linked issue, the edge case is entirely about cases where the default bin rule gives "less than 1 bin" and that doesn't make sense. Why would we want to change other plots?

Mainly I would want to fix the issue with no bins, personally i would prefer the behaviour with ceil plots instead round plots. But I understand that some people may expect the old behavior, that is why I asked for opinions on that.

I will update this PR to use the max approach, the min was just a mixup, sorry for the confusion.

[mwaskom]

personally i would prefer the behaviour with ceil plots instead round plots

why though? do you have a reason?

[Gornoka]

personally i would prefer the behaviour with ceil plots instead round plots

why though? do you have a reason?

When I specify binwidth my assumption was that I would get a histogram with that bin width.
When I specify bins I would get a histogram with that number of bins.
When I specify the binrange I would get whatever is in that range depending on the other settings.

Example graphs of the current behavior below.

I was expecting bindwidth to create new bins until it fits my whole data range .
If I have data that does not fit perfectly in the bin width, I should get the last bin semi-empty,
resulting in a round up of n_bins. (like it does in the bottom row)

If the range is not specified explicitly seaborn will shrink or extend the bins to fit the bins to evenly over the data.
That is a good decision to make the histogram easier to read correctly, and when using round the shrunk/extended bin width is closer to what the user asked for. So probably my preference is wrong here for many applications and I will have to restort to setting both binrange and binwidth in the future.

[Gornoka]

are there any next steps for this pr, / any requested changes or do you want to kill it @mwaskom ?