I suppose could just disable this. numpy doesn't undertsand object array searchsorted generally

In [4]: pandas.MultiIndex([[0],["a"]], [[0],[0]]).values.searchsorted((1,"b"))
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-4-b9352c7b8bea> in <module>()
----> 1 pandas.MultiIndex([[0],["a"]], [[0],[0]]).values.searchsorted((1,"b"))

TypeError: unorderable types: tuple() > str()

maybe just raise a NotImplementedError. This is pretty much a useless operation anyhow, you always to search by levels via indexing

Is this issue still available to fix ?

yes,

note that we should simply define .searchsorted in pandas/indexes/multi.py and use the direct indexers, .get_indexer, which is way more efficient (as its hashtable based).

@jreback I am new to pandas, could you throw some light on how i can use .get_indexer to implement .searchsorted ?

.get_indexer returns the indexer , IOW the location of the point. -1 marks not found items. This works on any multi-index. Note these don't even have to be sorted (but is more efficient if they are).

In [2]: i = pd.MultiIndex.from_tuples([(0, 'a'), (0, 'b'), (1, 'a')])

In [3]: i
Out[3]: 
MultiIndex(levels=[[0, 1], ['a', 'b']],
           labels=[[0, 0, 1], [0, 1, 0]])

In [4]: i.values
Out[4]: array([(0, 'a'), (0, 'b'), (1, 'a')], dtype=object)

In [5]: i.get_indexer([(0,'b'), (1, 'a'), (2, 'c')])
Out[5]: array([ 1,  2, -1])

Here's what searchsorted does; I am using an interger array because numpy doesn't play nice with tuples. It returns the indexer of the match (IOW where it is in the array). Note if something is not found it returns the last index before that (which is really unintuitve!)

In [6]: np.array([1, 2, 3]).searchsorted([2, 3])
Out[6]: array([1, 2])

In [7]: np.array([1, 2, 3]).searchsorted([2, 3, 5])
Out[7]: array([1, 2, 3])

@jreback this issue still open ?

yes

@TomAugspurger @jreback i am a little bit confused to what should i do exactly, i mean should i simply raise a NotImplementedError as said by @jreback ? or should i replace the searchsorted by get_indexer and somehow get a value for even tuples as said is issue example

If supporting searchsorted is an option that makes sense. #14833 (comment) indicates that that can be done using get_indexer.

Since there was no activity since March, I would be interested in working on this issue. It might be nice to implement searchsorted as suggested by @jreback, but the issue I see is that numpy's searchsorted can give the location of an element in a sorted array that would keep the sort order even if the element does not exist. Here is an example:

>>> np.searchsorted([1, 3, 5], [2])
array([1])

From what I understand .get_indexer will simply return -1 (element not found). Naive implementation of numpy's behaviour might use bisect module from Python's standard library for the not-found elements, but it would be rather inefficient.

Interestingly, numpy's searchsorted can also work with tuples if we define an appropriate dtype:

>>> dtype = [("int", 'i8'), ("str", "U1" )]
>>> arr = np.array([(0, 'a'), (0, 'b'), (1, 'c')], dtype=dtype)
>>> arr.searchsorted(np.array([(1, 'a')], dtype=dtype))
array([2])

A possible implementation of searchsorted would then coerce the multiindex to an ndarray with adapted dtype, and use the numpy's builtin searchsorted. What do you think?

so this might be easier now as the implementation of MI was recently refactored to directly keep the underlying codes in the cython table

take

take

Hi, is anyone still working on this, or may I take it up?
if the answer is NO, i.e. no one is working on this, then i have a couple of question:

btel suggested in #14833 (comment)

This can be one way to handle it. but it assumes the input array to be of 2-dimensional.
Should the input array be restricted to 2 dimensional ?
May i go ahead with this implementation, or should i just simple raise NotImplementedError ?

take