Hashable DataFrames

[hayd]

See this SO answer, they want to use memoisation.

OP points out this gets different results from (presumably it does it off id)

hash(pd.DataFrame([1,2,3])) 

Should they be hashable or should hash raise? (does it defeat the point of hashing if hashing is expensive?) cc @cpcloud

[cpcloud]

i've been thinking about this off and on. a somewhat related issue is that of the empty frame, i.e., DataFrame(). i think the PandasObject NDFrame should raise in all cases since that's what numpy does (overriding where it makes sense and is useful). i guess u could have the empty DataFrame be hashable but that seems like it's not worth the effort it would take to do, who needs to hash empty DataFrames?

[jreback]

series raises on __hash__ as should all NDFrame, because they are mutable hashing is meaningless. OTOH, index are hashable, as they are immutable

[cpcloud]

Indexes are currently not hashable, since they try to hash the underlying ndarray.

[jreback]

yes..you are right...oh well my argument is bad then!

[hayd]

Ah, you're right, I didn't even check series, it's just DataFrame which should raise.

Easy fix (raise __hash__ for generics) pr on the way.

[cpcloud]

could implement this for indices...thoughts?

[cpcloud]

in that case u should probably hash the name, number of levels, class, and dtype

[jreback]

still have the mutability issue

though I suppose if the user accepts this it would be nice to deal with it

I would table to 0.12 for now

[hayd]

So, at the moment I've put this in NDFrame.

Maybe it should go in PandasObject, and then have objects which should hash override it (like if we can get indices to hash using that clever method). Are there any besides Index/MultiIndex?

[cpcloud]

i vote for default to not hashable. better to alert the user to non-hashability rather than possibly giving misleading ideas about the hashability of things

[jreback]

agree....not hashability is/should be default until we change API

[hayd]

ok I've moved it to PandasObject, removes repeated code too. :)