Enhancement: Add dropna to Float64Index

[michaelaye]

AttributeError: 'Float64Index' object has no attribute 'dropna'

Any reason why it could not have a dropna()?

[jreback]

it could

keep in mind that the indexing code for Nan's is really hairy - so maybe a good idea!

[michaelaye]

yeah, this was the result of prematurely making something the index while I should have dropped the NaNs before assigning it as an index. So maybe it's better to teach the user to prevent NaNs in the index? Possibly would avoid a lot of other issues.

[jreback]

the problem with nan in an index is that selection basically goes out the window and indexes are almost useless because since nan != nan and so ordering is shot

can you post a usecase?

what we maybe could do is have an index that has a fill value so that ordering / selection is maintained, but treat it as missing - just a thought

[michaelaye]

i'm not sure i CAN defend a use case, now that I remember how bad NaNs are in the index?
As I said I had the following happening: dataframe with some columns having NaNs at the tail end. Creating a new dataframe with 2 columns of this dataframe without dropping NaNs. Assigning one of the new columns as an index to the new dataframe, ending up with NaNs in the index.
Maybe the real 'issue' here could be that I didn't notice? Maybe a warning would be appropriate if an array containing NaNs is made an Index?

[jreback]

no warning as setting an index possibly containing nan's is allowed. Its really the indexing of it that is the problem, which does raise an exception in some cases.

Their needs to be some more work in this area in any event. If you do find a case where nans are useful (or a nuiscance) in a Float64Index, pls post.