BUG: fixes pd.Grouper for non-datetimelike groupings #8866 #8964
Conversation
returns basegrouper class for non-datetime groupings while still returning a datetime index for datetime grouping
|
this needs a test. This should be a much simpler change. |
|
I thought it was going to be simpler, maybe I've missed something but I couldn't get around that I needed to pass a basegrouper to aggregate and a datetime index to resample. |
dbdc3f3 to
2162586
Compare
|
no, this just needs a small change in when |
|
Ah I see, I presumed TimeGrouper should only be used for datetimes, (was that originally the case?). So anything that gets passed into groupby as a grouper will be handled by the |
2162586 to
00eb085
Compare
|
I've had a look again and it's easy enough to return a |
00eb085 to
139342e
Compare
139342e to
1aebb81
Compare
|
no this should be virtually identical to setting the index then resampling (in fact that can be w test) |
|
how can we resample on a non-datetime index? |
|
I mean groupby. |
|
turns out this was a trivial issue, see #9008 thanks for the effort |
closes #8866
The bug was caused by a couple of things, firstly when specifying only a level, the axis default was None. This meant it failed to create the internal grouper breaking on line 254 of groupby.py:
Secondly, the aggregate function is set up to take a basegrouper class whereas for datetime resampling we need to pass a datetime index. I split the
_set_groupermethod to return a datetime index when a resampling frequency is specified and aBaseGrouperclass otherwise. I opted to split that out into two functions to reduce complexity with much of the_set_basegrouperfunction being borrowed from the_get_grouperfunction. Let me know if there's a better way to fix this.