BUG: Make DTI/TDI/PI argsort match their underlying arrays #37965
Conversation
pandas/core/groupby/grouper.py
Outdated
| # use stable sort to support first, last, nth | ||
| indexer = self.indexer = ax.argsort(kind="mergesort") | ||
| # TODO: why does putting na_position="first" fix datetimelike cases? | ||
| indexer = self.indexer = ax.argsort(kind="mergesort", na_position="first") |
There was a problem hiding this comment.
def argsort(
self,
ascending: bool = True,
kind: str = "quicksort",
na_position: str = "last",
in arrays/base.py so this is strange
There was a problem hiding this comment.
i think it must involve the fact that on older numpys, dt64.argsort used to put NaT at the beginning instead of the end
jreback
added
API - Consistency
|
@mroeschke will you have time to take a look at this in the not-too-distant future? if not ill mothball, since its stuck in the mud. |
| if (self.sort or sort) and not ax.is_monotonic: | ||
| # use stable sort to support first, last, nth | ||
| indexer = self.indexer = ax.argsort(kind="mergesort") | ||
| # TODO: why does putting na_position="first" fix datetimelike cases? |
There was a problem hiding this comment.
I am not really familiar with the history of NaTs with respect to resample. I would have assumed NaT would/should have been dropped by this point since we just exclude them from the result #13164.
I'm okay with this solution as a stop gap solution.
|
lgtm. any user facing change? |
Changes DTi/TDI/PI.argsort, that should be it. |
3 participants
black pandasgit diff upstream/master -u -- "*.py" | flake8 --diffThe fix to keep resample working is a kludge I'd like to make unnecessary before merging; cc @mroeschke