Implement DataFrame diff()

[skirui-source]

Fixes: #9604 and resolves #1271

This PR introduces diff() method for DataFrames of numeric types only. As a follow up, PR # #10212 will add support for non-numeric types (specifically timestamps/duration)
for both Series and DataFrame

[codecov]

Codecov Report

Merging #9817 (575c118) into branch-22.04 (a7d88cd) will increase coverage by 0.04%.
The diff coverage is 0.00%.

@@               Coverage Diff                @@
##           branch-22.04    #9817      +/-   ##
================================================
+ Coverage         10.42%   10.47%   +0.04%     
================================================
  Files               119      122       +3     
  Lines             20603    20496     -107     
================================================
- Hits               2148     2147       -1     
+ Misses            18455    18349     -106     

Impacted Files	Coverage Δ
python/cudf/cudf/core/_base_index.py	0.00% <ø> (ø)
python/cudf/cudf/core/column/categorical.py	0.00% <0.00%> (ø)
python/cudf/cudf/core/column/column.py	0.00% <0.00%> (ø)
python/cudf/cudf/core/column/datetime.py	0.00% <ø> (ø)
python/cudf/cudf/core/column/numerical.py	0.00% <0.00%> (ø)
python/cudf/cudf/core/column/string.py	0.00% <ø> (ø)
python/cudf/cudf/core/column/timedelta.py	0.00% <0.00%> (ø)
python/cudf/cudf/core/column_accessor.py	0.00% <0.00%> (ø)
python/cudf/cudf/core/dataframe.py	0.00% <0.00%> (ø)
python/cudf/cudf/core/frame.py	0.00% <0.00%> (ø)
... and 35 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update a25a2ec...575c118. Read the comment docs.

[vyasr]

Removing my review request since this PR looks to be pretty well-covered by other reviewers.

[karthikeyann]

rerun tests

[karthikeyann]

rerun tests

[skirui-source]

@bdice @isVoid in response to you review comment above, here's an example:

In [1]: import pandas as pd

In [2]: import cudf

In [3]: gdf = cudf.to_datetime(
   ...:     cudf.DataFrame({"year": [2015, 2016], "month": [2, 3], "day": [4, 5]})
   ...: )
   ...: gdf
Out[3]: 
0   2015-02-04
1   2016-03-05
dtype: datetime64[s]

In [4]: pdf = gdf.to_pandas()

In [5]: pdf.diff(periods=2)
Out[5]: 
0   NaT
1   NaT
dtype: timedelta64[ns]

In [6]: gdf.diff(periods=2)
---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
Input In [6], in <module>
----> 1 gdf.diff(periods=2)

File ~/cudf/python/cudf/cudf/core/series.py:3203, in Series.diff(self, periods)
   3198     raise AssertionError(
   3199         "Diff currently requires columns with no null values"
   3200     )
   3202 if not np.issubdtype(self.dtype, np.number):
-> 3203     raise NotImplementedError(
   3204         "Diff currently only supports numeric dtypes"
   3205     )
   3207 # TODO: move this libcudf
   3208 input_col = self._column

NotImplementedError: Diff currently only supports numeric dtypes

This PR doesn;t cover timestamps, so should I go ahead and get this merged first ( add a TO-DO/ file follow-up issue)?
I'm also unsure what the scope is, to add support for time data, any thoughts?

[bdice]

@skirui-source I see that this is already implemented for Series for only numeric types, so we can introduce this PR as-is without expanding the scope. However, we should consider a follow-up that relaxes that restriction (see code below) for both Series and DataFrame to support non-numeric types (specifically to support timestamps/durations, I'm not sure what else) in a follow-up PR. It shouldn't require anything fancy, since we already support binops on timestamp types and this just does a shift-and-subtract. A time-series diff is probably a very common user need, if you have timestamps and want the durations between rows.

These should be relaxed to support any type that can be subtracted:

cudf/python/cudf/cudf/core/series.py

Lines 3202 to 3205 in 82f941b

	if not np.issubdtype(self.dtype, np.number):
	raise NotImplementedError(
	"Diff currently only supports numeric dtypes"
	)

cudf/python/cudf/cudf/core/dataframe.py

Lines 2666 to 2669 in 82f941b

	if not all(is_numeric_dtype(i) for i in self.dtypes):
	raise NotImplementedError(
	"DataFrame.diff only supports numeric dtypes"
	)

[skirui-source]

@isVoid @bdice Cool, Thanks 👍🏾
in that case I think this PR is ready for another review/approval unless there's any more unresolved discussions

[bdice]

I opened #10212 about enabling broader type support for timestamps.

[bdice]

I have a few small requests to improve the tests. Otherwise LGTM.

[skirui-source]

@gpucibot merge