REGR: fix inplace arithmetic operation on Series no longer updating parent DataFrame

[jorisvandenbossche]

Closes #36373

See discussion at #36498 (comment) for reverting the original fix (#30501 for #30484)

[jreback]

lgtm merge on green.

[jbrockmendel]

@jorisvandenbossche i have no objection if you're just looking for the quickest available fix, but another option is to update NDFrame._inplace_method:

def _inplace_method(self, other, op):
    result = op(self, other)

+    if self.ndim == 1 and result._indexed_same(self) and result.dtype == self.dtype:
+        self._values[:] = result._values
+        return self

    self._reset_cacher()
    [...]

i havent run the full test suite on this yet, but it passes the new test this PR adds and one of the two that this PR xfails.

[jorisvandenbossche]

i have no objection if you're just looking for the quickest available fix

That's indeed the case, as I want to have this fixed for 1.1.4, and I am almost off to bed ;-)
But certainly feel free to push that change here (or in another PR), if it passes the test we can merge that instead.

[jreback]

yeah let's just revert this as is to avoid blocking. @jbrockmendel feel free to push your change in a new PR

[simonjayhawkins]

@jbrockmendel feel free to push your change in a new PR

or could add the change to _maybe_cache_changed from #36498. The blockers on that PR weren't related to that change.

[jbrockmendel]

im inclined to think this test is wrong.

orig = DataFrame(zip(range(3), range(3)), columns=["a", "b"])
df = orig.copy()

ser = df.loc[:, "a"]
ser += 2

tm.assert_frame_equal(df, orig)

df.loc[:, "a"] is a view on df, and += is specifically an inplace op (as long as it doesnt result in a dtype change). I think df should be altered.

[jorisvandenbossche]

im inclined to think this test is wrong.

Yep, I had the same feeling: #36498 (comment)

[jorisvandenbossche]

Closing in favor of #37508, which fixes it without reverting the fix in drop (thanks @jbrockmendel !)