ENH: Pairwise dropna in PairGrid. Closes #407.

[jseabold]

PR for #407.

[mwaskom]

It would be ideal to wrap up any of these changes with those discussed in #361.

[jseabold]

Bump minimum pandas version on travis or fix to support 0.12.0 (released 7/13)?

[jseabold]

#361 is a bit of an orthogonal issue AFAICT, but I can take a look at that too. This preserves the current behavior, but now handles NA (to the extent the test suite covers the existing sorts in axisgrid.py).

[mwaskom]

Bump minimum pandas version on travis or fix to support 0.12.0 (released 7/13)?

The aim is to support Debian stable, though I always have a hard time figuring out what that is and usually wait for @yarikoptic to yell at me :)

[jseabold]

Fair enough, I'll fix it up.

[mwaskom]

Well based on this I think "debian stable backports" is running 0.13 (now?). Would inplace=False work there?

[jseabold]

No, I don't think this was settled by the 0.13.x series either. I've worked around this before IIRC. Will fix up later (got to run now).

[mwaskom]

So from what I can tell for the narrow version of this PR that preserves current behavior, wouldn't something like

hue_names = np.sort(data[hue].dropna().unique())

work?

[jseabold]

I think so, yeah. The only benefit of this version is that it also supports the new categorical dtype in an expected way, if a desired change. The only break would be if I changed np.unique to pd.unique or vice versa without a sort, which I don't think I did.

[mwaskom]

I'd say just do the simpler thing for now and not worry about categoricals, which I'll handle with a broader change to how factor orders are determined before 0.6 goes out.