REGR: FutureWarning issued and empty DataFrame returned where no numeric types to aggregate

[simonjayhawkins]

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the master branch of pandas.

Reproducible Example

import numpy as np
import pandas as pd
frame = pd.DataFrame({"a": np.random.randint(0, 5, 50), "b": ["foo", "bar"] * 25})
frame[["b"]].groupby(frame["a"]).mean()

Issue Description

the code sample was raising DataError: No numeric types to aggregate in 1.2.5 but in 1.3.x this now issues a warning that it will raise in the future. (and also appears that the stacklevel is incorrect)

on master

>>> frame[["b"]].groupby(frame["a"]).mean()
sys:1: FutureWarning: Dropping invalid columns in DataFrameGroupBy.mean is deprecated. In a future version, a TypeError will be raised. Before calling .mean, select only columns which should be valid for the function.
Empty DataFrame
Columns: []
Index: [0, 1, 2, 3, 4]

The change to a empty dataframe was in #41706 with the warning being added in #43154

cc @jbrockmendel @Dr-Irv

Expected Behavior

same as 1.2.5

>>> frame[["b"]].groupby(frame["a"]).mean()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/simon/miniconda3/envs/pandas-1.2.5/lib/python3.9/site-packages/pandas/core/groupby/groupby.py", line 1496, in mean
    return self._cython_agg_general(
  File "/home/simon/miniconda3/envs/pandas-1.2.5/lib/python3.9/site-packages/pandas/core/groupby/generic.py", line 1015, in _cython_agg_general
    agg_mgr = self._cython_agg_blocks(
  File "/home/simon/miniconda3/envs/pandas-1.2.5/lib/python3.9/site-packages/pandas/core/groupby/generic.py", line 1121, in _cython_agg_blocks
    raise DataError("No numeric types to aggregate")
pandas.core.base.DataError: No numeric types to aggregate

Installed Versions

pd.show_versions()

[Dr-Irv]

@simonjayhawkins It's not clear to me how you're suggesting this should be changed. Are you saying that the FutureWarning should be removed and the old behavior (raising a DataError) restored for 1.3.x? Are you also saying that we need to decide what will happen in 1.4 and beyond?

[simonjayhawkins]

If it was raising DataError in 1.2.5 and we want to raise TypeError in the future, we could probably make that change without a warning (and a change in behavior to return an empty DataFrame in the interim)

I think that changing the exception type may not be an api change. changing the behavior is.

so for 1.3.x we could either restore the DataError (or maybe change to TypeError)

[jbrockmendel]

and also appears that the stacklevel is incorrect

there's now a find_stack_level that can avoid these headaches

IIRC this was the best option available for ensuring internal consistency until the deprecation is enforced. e.g. we want frame[["b"]].groupby(frame["a"]).mean() to match frame[["b"]].groupby(frame["a"]).agg("mean") and frame[["b"]].groupby(frame["a"]).apply(np.mean)

[Dr-Irv]

Here's a table contrasting the behaviors using the examples @jbrockmendel listed between 1.2.5 and 1.3.2, and master:

Code	1.2.5 Result	1.3.2 Result	1.4.0.dev0+647.g0072fa8cb5 Result
frame[["b"]].groupby(frame["a"]).mean()	DataError exception	Empty DataFrame	FutureWarning and Empty DataFrame
frame[["b"]].groupby(frame["a"]).agg("mean")	DataError exception	Empty DataFrame	FutureWarning and Empty DataFrame
frame[["b"]].groupby(frame["a"]).apply(np.mean)	Empty DataFrame	FutureWarning and Empty DataFrame	FutureWarning and Empty DataFrame
frame[["b"]].groupby(frame["a"]).sum()	DataFrame with values	Empty DataFrame	DataFrame with values
frame[["b"]].groupby(frame["a"]).agg("sum")	DataFrame with values	Empty DataFrame	DataFrame with values
frame[["b"]].groupby(frame["a"]).apply(np.sum)	DataFrame with values	DataFrame with values	DataFrame with values

With the fix in #43154 , we now have restored the 1.2.x behavior in the case of sum . I think the DataError was correct in 1.2.x for mean, but then we have to figure out what to do in the case of apply(np.mean), because returning an empty DF in that case could be considered a bug (i.e., we should have raised). I'm not sure how easy that would be to fix. OTOH, putting in FutureWarning for apply(np.mean) makes sense, but maybe for the .mean() and .agg("mean") cases, we should restore the DataError behavior and not put in a warning.

Results above produced with following program:

long python code. click to expand

import pandas as pd
import numpy as np

frame = pd.DataFrame({"a": np.random.randint(0, 5, 50), "b": ["foo", "bar"] * 25})

print("Using mean")


print("calling mean directly")
try:
    print(frame[["b"]].groupby(frame["a"]).mean())
except Exception as e:
    print("exception raised ", str(e))

print("calling agg('mean')")
try:
    print(frame[["b"]].groupby(frame["a"]).agg("mean"))
except Exception as e:
    print("exception raised ", str(e))

print("calling apply(np.mean)")
try:
    print(frame[["b"]].groupby(frame["a"]).apply(np.mean))
except Exception as e:
    print("exception raised ", str(e))

print("calling sum directly")
try:
    print(frame[["b"]].groupby(frame["a"]).sum())
except Exception as e:
    print("exception raised ", str(e))

print("calling agg('sum')")
try:
    print(frame[["b"]].groupby(frame["a"]).agg("sum"))
except Exception as e:
    print("exception raised ", str(e))

print("calling apply(np.sum)")
try:
    print(frame[["b"]].groupby(frame["a"]).apply(np.sum))
except Exception as e:
    print("exception raised ", str(e))

[jbrockmendel]

I think the DataError was correct in 1.2.x for mean, but then we have to figure out what to do in the case of apply(np.mean), because returning an empty DF in that case could be considered a bug (i.e., we should have raised)

The apply-that-is-an-agg part of this seems like something to ask @rhshadrach about.

[rhshadrach]

The relevant parts of the algorithm works in two steps:

• Subset the columns to numeric only
• Operate on the resulting DataFrame

To separate these, consider changing the example in the question to:

frame = pd.DataFrame({"a": np.random.randint(0, 5, 50), "b": ["foo", "bar"] * 25})
res = frame[[]].groupby(frame["a"]).mean()
print(res)

This raises on 1.2.x, but not on master. I believe raising in this case is a bug - it is inconsistent with other reductions (e.g. sum, min, max, prod). Making it so this op doesn't raise from 1.2 -> 1.3 seems appropriate to me.

Now consider the example:

frame = pd.DataFrame({"a": np.random.randint(0, 5, 50), "b": ["foo", "bar"] * 25})
res = frame[["b"]].groupby(frame["a"]).mean(numeric_only=True)
print(res)

If we agree that not raising on an empty frame is correct, then it seems to me to be reasonable to have the same behavior when numeric_only subsets the frame to become empty. In other words, I think this shouldn't raise as well.

Finally, back to the original example which is the same as above but numeric_only is left unspecified, it did raise in 1.2.x, it no longer raises but will raise a TypeError in the future. This is one case where I could see reverting to raising instead, but also think it would be okay to leave this behavior as is.

I think the DataError was correct in 1.2.x for mean, but then we have to figure out what to do in the case of apply(np.mean), because returning an empty DF in that case could be considered a bug (i.e., we should have raised)

I think we currently treat .apply(np.mean) et al the same as any other UDF - i.e. as a black box (correct me if this is wrong!). With this, I don't think any special consideration should be given that we don't give to other UDFs. In other words, we should not go out of our way to ensure apply(np.mean) is consistent with .mean() or .agg('mean'). With this last statement - I most certainly think there is room for disagreement. That is, I think there is an argument to make that numpy functions are special and should be given special consideration. My opinion is that, at least right now, the extra logic and maintenance burden is not worth it.