API: ban mutation within groupby.apply

[jreback]

xref #8662
xref #12652

This makes pandas code jump thru hoops and is a complete anti-pattern for pandas. let's nuke it. users should never do this.

[jbrockmendel]

agreed this is a footgun. brainstorming how to implement:

• document that it isnt supported and be done with it
• maybe a contextmanager to make the relevant objects immutable within the relevant calls

[rhshadrach]

@jbrockmendel with the contextmanager, are you thinking doing a comparison after the apply call to see if the objects involved changed? Is there a way to avoid a copy here?

[jbrockmendel]

with the contextmanager, are you thinking doing a comparison after the apply call to see if the objects involved changed?

I am thinking of setting the array.flags.writeable to False for any ndarrays contained inside the DataFrame. We'd need to implement something analogous for ExtensionArrays.

[rhshadrach]

I think of this issue applying to all UDF functions (apply, aggregate, transform, filter) on all pandas objects. @jbrockmendel - would you support just adding to the docs without the context manager? This would perhaps raise some awareness and I think doesn't require a deprecation, and gives a place we can point to when issues are reported.

[jbrockmendel]

would you support just adding to the docs without the context manager

I think thats probably our best bet. Something to point out when people inevitably open issues about this.

[rhshadrach]

I'm +1 on closing this, I don't believe the context manager is necessary.

[MarcoGorelli]

Looks like this has been clearly documented in https://pandas.pydata.org/pandas-docs/stable/user_guide/gotchas.html#gotchas-udf-mutation by #39762 - closing then, but please let me know if I've misunderstood, or if you believe that just documenting this isn't enough of a solution