Undesired auto sort when adding two multiindexes

[Aurthes]

idx0 =

timestamp sequence attribute
2014-11-12 17:02:23.635000-06:00 0 price
count
quantity

idx1 =

timestamp sequence attribute
2014-11-12 17:02:24.060000-06:00 1 price
count
quantity

idx0 + idx1 =

timestamp sequence attribute
2014-11-12 23:02:23.635000 0 count
price
quantity
2014-11-12 23:02:24.060000 1 count
price
quantity

Any way to solve this?

[TomAugspurger]

Can you clean up your example to make it copy-pasteable into a REPL so that it's clearer what the problem is?

[Aurthes]

Hi,

Here's an example code:

idx1 = pd.MultiIndex.from_product([[1],[3],['c','b','a']])
idx2 = pd.MultiIndex.from_product([[2],[4],['f','e','d']])
print idx1
print idx2
idx = idx1 + idx2
print idx

You'll see how the 3rd level got sorted automatically. Is there a way I can prevent this?

[jreback]

@Aurthes you almost ALWAYS want the multi-index to sort to do any type of indexing. What is your usecase? (you can always do .sortlevel() to specifcally sort on a certain level(s))

[shoyer]

@Aurthes In the current version of pandas, the + operator calculates the set union of two indexes (note that the alias + for union is currently deprecated).

If you simply want to concatenate two multi-indexes, you can do this with numpy:

idx3 = pd.MultiIndex.from_tuples(np.concatenate([idx1.values, idx2.values]))

To get the unique (non-sorted) elements (which I think is what you want here), you can then use .unique(), e.g.,

idx4 = pd.MultiIndex.from_tuples(idx3.unique())

This is slightly awkward, but pandas provides all the tools to do it efficiently, and I do believe this sort of use is pretty niche -- we have lots of shortcuts that make this easier when working with dataframes or series.

[Aurthes]

Problem solved, very efficient. Thanks! Pandas is great :)

[Aurthes]

Thx all !