Support pad/backfill/nearest reindexing even for unsorted indexes by storing a sorted index?

[shoyer]

Recently, I've been working on adding a 'nearest' method to reindexing: #9258

It occurs to me that we could easily extend reindexing/get_indexer methods to work with unordered indexes if we were willing to do a sort operation on the index if necessary. This would probably entail saving the sorted result on the parent index, similarly to how get_indexer is currently supported on MultiIndex by creating a tuple index internally.

I think this would be a nice usability gain over the current implementation, and not be too surprising. Sorting indexes (once) is pretty fast, for anything up to millions of rows.

Thoughts?

[jreback]

xref is #3539

you could do this in an index, as we can easily store meta data, but not currently for a column (that would require a meta dict in the BlockManager).

Note have to be sure to invalidate these on any setting operations.

Note that IIRC the index cython object DOES cache these things, e.g. is_monotonic, so not sure how much this would actually change it. (I am talking about if the index IS sorted, then you don't have to recompute it again). This issue is different though.

[mroeschke]

Looks like this never took off so closing. Can reopen if there's interest