CLN/BUG: Better validation of levels/labels/names

[jtratner]

Improved validation of levels, labels and names (plus fleshed out test
cases). Fixes #4794.

Changed an assert_almost_equal to assert_series_equal to better reflect
actual return value.

Also makes set_levels, set_labels and set_names match behavior of rename to not return self if inplace.

[jreback]

look ok to me

[jtratner]

took a bit of wrangling to test correctly given the slightly weird internal representation of levels and labels. Do you think this needs to test that levels are all integers? I feel like most of the time people should not be calling MI directly.

[jtratner]

note that 0.12 and earlier wasn't doing that either.

[jreback]

u could prob validate that it's int64 dtype
I agree users shouldn't really touch mi internals

[jtratner]

I wonder if I can just call Int64Index and that'll be good enough? I'll try

[jreback]

I guess there are internal pandas methods that may create labels/levels (that re not in index) - eg reshape/ concat

[jtratner]

huh? but I think the internal representation depends on levels being integers no? It's really simple actually, just change _ensure_index to _ensure_int64index. Going to see what happens.

[jtratner]

of course it's labels that's actually supposed to be integer :P.

[jtratner]

I'm going to leave it - there's a lot of complexity with checking that it's not a float and such, given that it's internal-ish (and we don't really have a good fastpath yet), probably better to just not figure out something complicated to check. Unless you feel strongly, I'll leave it for 0.14 and the index refactor (where it should be easier to reason about and check). Internal calls all use _ensure_int64 anyways.

[jtratner]

for posterity - I think this was because sometimes assert_almost_equal was being passed a Series. but no part of the library deals with that anymore anyways.

[jtratner]

@jreback okay?

[jreback]

yep