BUG: Series.get with Float64Index and numpy nan returns empty Series

[willpan]

I tried to use series.get(x) (where x is np.nan in some instance) and the result was an empty series. It seems like it should just return None if it can't be found.

import pandas as pd
import numpy as np

series = pd.Series(range(0, 10))
series.get(np.nan) 
# returns None

series = series.reindex(pd.Float64Index(range(0, 10)))

series.get(10) 
# This correctly returns None

series.get(np.nan) 
# returns empty series with empty Float64Index

series.get(np.nan, np.nan) 
# same thing

[jreback]

indexing with np.nan on any index doesn't make much sense. what are you trying to do?

maybe s[s.isnull()]?

[willpan]

Well, it's being called like this:

some_series = some_series.map(lambda x: some_other_series.get(x))

[jreback]

and why would you do that? again simply use s.isnull() that's what it is for.

[willpan]

The code (I ran across the difference while regression testing, I'm not the original author) is basically trying to merge a Series into a DataFrame so the actual code was something like:

df[new_col] = df[lookup_col].map(lambda x: some_other_series.get(x))

I just noticed the behavior was inconsistent between versions (I'm going from 0.9 to 14.1).

[jreback]

well, that is quite inefficient, and not what you want to do anyhow, is their a reason you are not using
a) merge, or b) df[lookup_col].map(some_other_series) ?

[willpan]

I'm going to end up replacing it with the second option. I work with a lot of people who write code as a part of their job, but who aren't programmers, and they don't always make the best choices.

Anyways, I thought that the result of series.get(np.nan) was unusual, but feel free to close if this isn't worth it. Thanks.

[jreback]

I do appreciate the report (esp since it's an apparent change from a long ago version)

I may link this to another issue
technically a test like thks is ok

let me see