MultiIndex promotes float64 dtype to object

[dalejung]

import pandas as pd
import numpy as np
print pd.__version__
#0.11.0.dev-184f2db

lev1 = np.arange(10)
lev2 = np.linspace(0, 9, 10)
mi = pd.MultiIndex.from_arrays([lev1, lev2])

# ints work fine
assert lev1.dtype == mi.levels[0].dtype

# failure
assert lev2.dtype == mi.levels[1].dtype
print lev2.dtype, mi.levels[1].dtype
# float64 object

Ran into this when using np.isclose on the level values. Numpy fails due to casting rules and object dtype.

[dalejung]

Ah, just noticed that this is more of an pd.Index thing.

#236

btw, how do you properly reference another issue?

[jreback]

this is correct behavior. have floats in an index is funny in any event w.r.t. indexing. do you have a particular situation that this is showing a problem in?

np.isclose requires a numeric dtype which is why that would fail (if comparing to a float dtype), which you know

[dalejung]

I ran into the issue when running tests on parameter distributions created by something like np.linspace. The resulting DataFrame had the parameters in the columns and I was grabbing a subset to plot.

In either case, I adjusted for this behavior on my end. This can be closed if it's invalid.

[jreback]

thanks