API Consistency: validations on no-ops

[jbrockmendel]

We are not consistent about whether we do validation checks on no-ops

Examples of no-ops:

• datetimelike_array[empty_indexer] = empty_values_wrong_dtype # <-- we do not raise
• series_dt64[empty_indexer] = empty_object_dtype # <-- we cast to object, meaning we did raise internally
• arr_with_no_nas.fillna(invalid_fill_value) # <-- we mostly dont raise, see below
• interval_index.take(indexer_with_no_negative_ones, fill_value=something_invalid) # <-- we do not raise

pd.array([1, 4, 6]).fillna("foo")
pd.array([1.0, 4.9, 6.0]).fillna("foo")

pd.IntervalIndex.from_breaks(range(3)).fillna("foo")

dti = pd.date_range("2016-01-01", periods=3)
dti.fillna("foo")
(dti - dti).fillna("foo")
dti.to_period("D").fillna("foo")

pd.Categorical(dti).fillna("foo")   <-- raises

I think we should (with deprecation cycles where necessary) aim to be consistently strict in this genre of check.

Update: listing related issues as I find them

• BUG: drop(level=.) doesn't raise if label is absent #18561 BUG: drop(level=.) doesn't raise if label is absent

[jreback]

+1 we should raise ultimately if we have an invalid value

so deprecation fine for now

[jorisvandenbossche]

It would indeed be nice to make this more consistent. Moving towards always raising on a invalid fill value sounds good.

What would you do for the second example? (series_dt64[empty_indexer] = empty_object_dtype # <-- we cast to object, meaning we did raise internally) Keep that behaviour?