BUG: df.boxplot fails to use existing axis/subplot (#3578)

[onesandzeroes]

Alright, I think I have a fix for issue #3578. In plotting._grouped_plot_by_column, there was no check being done for whether ax was None, which was the main source of the issue. When the boxplot is of multiple columns, I don't think there's anything sensible that can be done with the ax argument, so that
now raises a ValueError.

I've written some tests, but I'm not sure if they really get to the heart of the problem, so any insight on how to improve them would be appreciated.

Testing code adapted from the original bug report to demonstrate the new behaviour:

import matplotlib.pyplot as plt
import pandas
from pandas import DataFrame, Series

data = {'day': Series([1, 1, 1, 2, 2, 2, 3, 3, 3]),
        'group2': Series([2, 2, 2, 2, 2, 1, 1, 1, 1]),
        'val': Series([3, 4, 5, 6, 7, 8, 9, 10, 11]),
        'val2': Series([8, 9, 10, 11, 12, 13, 14, 15])}
df = pandas.DataFrame(data)

# Single-column using existing axis: should create a single plot
plt.figure()
plt.subplot(2, 2, 1)
plt.plot([1, 2, 3])
plt.subplot(2, 2, 4)
ax = plt.gca()
fig = ax.get_figure()
axes = df.boxplot('val', 'day', ax=ax)
print("Testing identity of returned axes: (should be True)")
print(id(axes) == id(ax.get_axes()))
plt.show()

# Multiple column, not using existing axis: should create two plots
plt.figure()
plt.subplot(2, 2, 1)
plt.plot([1, 2, 3])
plt.subplot(2, 2, 4)
ax = plt.gca()
axes = df.boxplot(['val', 'val2'], 'day')
print("Testing identity of returned axes: (should be False)")
print(id(axes) == id(ax.get_axes()))
plt.show()


# Multiple column using existing axis: should raise an exception,
# since it's hard to know what to do: we need an axis for each
# column and ax is just a single axis object
plt.figure()
plt.subplot(2, 2, 1)
plt.plot([1, 2, 3])
plt.subplot(2, 2, 4)
ax = plt.gca()
try:
    df.boxplot(['val', 'val2'], 'day', ax=ax)
except ValueError:
    print("Raising exception as expected")
plt.show()

[jreback]

@TomAugspurger @cpcloud

[cpcloud]

@onesandzeroes can you rehash and force push to run travis ci again? thx

here's a line to do just that

git commit --amend -C HEAD && git push --force

[onesandzeroes]

Alright, I've just pushed it up now. I have a feeling that this might overlap with the new #7035 pull request, which is fixing a pretty closely related issue, so if that gets merged first I might have to go back over and check this one.

[onesandzeroes]

Sorry, was I supposed to rebase as well? The build still seems to be failing for unrelated reasons (reading finance data from Yahoo), is this likely to be fixed if I rebase on master?

[onesandzeroes]

I've rebased on master and the tests seem to be passing now.

[jreback]

@TomAugspurger @cpcloud ?

[jreback]

@onesandzeroes if you can squash would be great, see https://github.com/pydata/pandas/wiki/Using-Git

[TomAugspurger]

Could you add a test for when by is None? In this case we return a dict (which may be changing in the nearish future, but it would still be good to test it for now). Something like

import itertools

fig, ax = plt.subplots()
d = df.boxplot(ax=ax)
lines = list(chain.from_iterable(d.values()))
self.assertEqual(len(ax.get_lines()), len(lines))

[TomAugspurger]

We could handle this case, assuming that ax is an array with size equal to ngroups. Want to open a separate issue for that?

[TomAugspurger]

This is out of scope for this issue, but do we have a specified way for when a plot call with None for figure / ax creates a new fig / ax vs. updates an existing one? I'll pay more attention, but I seem to remember it being inconsistent.

---

Just that one additional test for when by is None and then I think this is ready.

[onesandzeroes]

Added the new test, rebased and squashed it down to a single commit, so it should be good to go.

I've also created a new issue for allowing multiple axes in the ax argument as @TomAugspurger mentioned.

[TomAugspurger]

Merged via f851b57

Thanks!