Test failure with matplotlib 1.5.2rc2 on Debian

[olebole]

When running the test suite after the package was built, we get the following failure:

ERROR: test_plot (pandas.tests.test_graphics.TestDataFramePlots)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/$BUILD/debian/tmp/usr/lib/python2.7/dist-packages/pandas/tests/test_graphics.py", line 1322, in test_plot
    df.plot.line(blarg=True)
  File "/$BUILD/debian/tmp/usr/lib/python2.7/dist-packages/pandas/tools/plotting.py", line 3758, in line
    return self(kind='line', x=x, y=y, **kwds)
[...]
  File "/$BUILD/debian/tmp/usr/lib/python2.7/dist-packages/pandas/tools/plotting.py", line 1340, in _plot
    return ax.plot(*args, **kwds)
  File "/usr/lib/python2.7/dist-packages/matplotlib/__init__.py", line 1821, in inner
    return func(ax, *args, **kwargs)
  File "/usr/lib/python2.7/dist-packages/matplotlib/axes/_axes.py", line 1432, in plot
    for line in self._get_lines(*args, **kwargs):
  File "/usr/lib/python2.7/dist-packages/matplotlib/axes/_base.py", line 381, in _grab_next_args
    for seg in self._plot_args(remaining, kwargs):
  File "/usr/lib/python2.7/dist-packages/matplotlib/axes/_base.py", line 369, in _plot_args
    seg = func(x[:, j % ncx], y[:, j % ncy], kw, kwargs)
  File "/usr/lib/python2.7/dist-packages/matplotlib/axes/_base.py", line 276, in _makeline
    seg = mlines.Line2D(x, y, **kw)
  File "/usr/lib/python2.7/dist-packages/matplotlib/lines.py", line 380, in __init__
    self.update(kwargs)
  File "/usr/lib/python2.7/dist-packages/matplotlib/artist.py", line 859, in update
    raise AttributeError('Unknown property %s' % k)
AttributeError: Unknown property blarg

Full log here (from the Debian bug report).

pandas.show_versions() gives:

INSTALLED VERSIONS
------------------
commit: None
python: 2.7.12.final.0
python-bits: 64
OS: Linux
OS-release: 4.6.0-1-amd64
machine: x86_64
processor: 
byteorder: little
LC_ALL: C
LANG: C

pandas: 0.18.0+git114-g6c692ae
nose: 1.3.7
pip: None
setuptools: 20.10.1
Cython: None
numpy: 1.11.1rc1
scipy: 0.17.1
statsmodels: 0.6.1
xarray: None
IPython: None
sphinx: 1.4.4
patsy: 0.4.1
dateutil: 2.4.2
pytz: 2015.7
blosc: None
bottleneck: None
tables: 3.2.2
numexpr: 2.6.0
matplotlib: 1.5.2rc2
openpyxl: 2.3.0
xlrd: 1.0.0
xlwt: 0.7.5
xlsxwriter: None
lxml: 3.6.0
bs4: 4.4.1
html5lib: 0.999
httplib2: None
apiclient: None
sqlalchemy: None
pymysql: None
psycopg2: None
jinja2: 2.8
boto: None
pandas_datareader: None

[sinhrks]

Thanks for the report. Looks to be caused by mpl which raises AttributeError rather than TypeError. Should change the test on pandas side.

• https://github.com/pydata/pandas/blob/master/pandas/tests/test_graphics.py#L1317

Appreciated to submit a PR.

[olebole]

line 1333

Shall this be compatible with matplotlib < 1.5.2rc? Then this is not that simple, since tm.assertRaises() only accepts a singe exception, and one then would need to create another mpl_ge_1_5_2 comparison specifically for this test.
Is this really worth it? What is the rationale of the test?

[jorisvandenbossche]

cc @tacaswell Just to be sure, the change from TypeError to AttributeError, was this on purpose? (it's raised in set_lineprops)

[tacaswell]

I do not think that was intentional and I am not sure off the top of my head why this changed. The Artist.update method has raised AttributeError since 2004.

[tacaswell]

So, I broke this: matplotlib/matplotlib#6175

It looks like we used to have Artist.set which raised TypeError and Artist.update which raised AttributeError (because history). These two code-paths got merged in matplotlib/matplotlib#5599 (also my fault) and we missed the API change in the the exceptions.

Unfortunately I just tagged 1.5.2 last weekend, but have not posted it to pypi or publicized it yet 😞. I am inclined to just document this as an API change, but if you want to lobby for a 1.5.3 fixing this, this is the time to do it!

[jorisvandenbossche]

I don't think it is that an important change, so leaving (and documenting) it is fine for me (for pandas it is only a test that is broken, and I think you will have a better idea of how big this change is for matploltib users). In any case, I think the AttributeError is more logical.

[olebole]

@sinhrks @jorisvandenbossche For me the question here is still, what that test is for?

[sinhrks]

can u add _mpl_ge_1_5_2() function like below, then:

if _mpl_ge_1_5_2():
    with tm.assertRaises(AttributeError):
        ...
else:
    with ...

• https://github.com/pydata/pandas/blob/master/pandas/tools/plotting.py#L129

[jorisvandenbossche]

For me the question here is still, what that test is for?

Well, it tests that if you pass a wrong keyword, that this raises an error, and is not swallowed somewhere in the implementation. So it's not a huge important test, but still useful (we actually have to many functions in pandas that silently swallow invalid arguments)