Please keep within 80 characters line width as per PEP8.

Shouldn't this be log_base ** -1? (i.e. 0.1 only makes sense when the log base is 10).

Yes, but how do I access the base of log scale axis?

Ah, I see. I missed the fact that hist doesn't take a base as argument. hist is a bit of an anomaly -- it takes a 'log' argument, whereas most other things (except bar*) work by requiring a subsequent call to set_xscale or set_yscale, which of course take an optional "base". Making this also work that way would be a little bit of work, because this minimum adjustment stuff you've added here would have to be moved to the draw phase. Ideally, that's where I'd like to see this go, but I don't know if that's worth abandoning this fix, and we've have to maintain backward compatiblity anyway -- this "log" kwarg goes back at least 5 years, as far as I can tell.

We could either add an additional keyword "log_base", or extend "log" so it can be "True" (meaning base 10 as it does now), or a number, meaning a base. I would also recommend updating bar* to be in line with that.

However, I'm aware that none of this is premised on any real world use case -- just my own desire for consistency across uses of log scale, so I could be persuaded to leave this as-is by further discussion.

This might be kind of hard to set the minimum correctly if log base can be changed between to calls. Consider this

ax.hist( x1, log=10)
ax.hist(x2, log=2)

The first hist call would have no idea that eventually the base of log scale will be changed to 2.

But, I do not think it would be hard for use to do another call to set_ylim though. The histogram will be painted all the way down (see line 7892) anyway.

Yes -- that, in a nutshell, is why I'm not a fan of "log" being a kwarg to this method. You may want to hold of on implementing any sort of fix for this issue until we discuss it -- it may be best to just leave as-is.

@efiring, @WeatherGod, @pelson: any thoughts?

w.r.t. not being able to access the axes's log base, we are assuming base 10 throughout this part of the code, and we have always done so before. We could always look to future PRs to address the handling of arbitrary log bases. I see this more as a good bug fix to an outstanding issue.

I guess it would be smart, though, for future developers, to somehow comment the spots that are assuming base 10 so that they know to change it if/when we extend the interface in the future to accept arbitary log bases.

I tend to agree with @WeatherGod here: let's leave this assuming base 10 for now.

Wrap these comment lines to less than 75 characters

75? I thought we did 80?

Actually, I was wrong, in PEP8, docstrings are recommended to be at most 72 characters.

My comments above wrt multiple calls to hist() goes here.

This is kind of the part I don't fully understand. What function does 0.9 serve and where should I get the old bound. What's the difference between self.datalim.bound and self.datalim.intervalx/y?

I don't have time for a full explanation as I am about to head off on
vacation, but perhaps I can provide some clues for you. I think this code
part has been neglected. Other functions like plot() and scatter() utilize
pads for the data limits so that the axes wouldn't be extremely tight
around the data being plotted. This value used to be a hard-coded 0.05 on
each endside (for a total of 0.1). In plot() and scatter() (and possibly
bar()) we now handle this a lot more intelligently, but I think hist() was
left behind in this regard.

Note that data limits don't necessarially mean the same as view limits. I
am not sure what the distinction is between data limits and data interval
is, though.

I hope this helps!
Ben Root