From #9536.

The current Table behavior was based on how numpy behaves both with respect to sorting a single array in place and sorting a structured array (with tbl.sort(order='sort_col')). In the latter case in particular a ref to a column is not lost during the sort operation.

My guess was that pandas would do the same, but I just tried df.sort_values('sort_col', inplace=True) and found that it does indeed make new column references.

So there is not an unambiguous precedent, though numpy and astropy Table have historically kept references. BTW, for Table it used to be the case that replacing a column in the table was slow, but no longer.

One thing I that I just appreciated is that take is actually make a deep copy of all the info attributes (e.g. meta). From that perspective we are already paying for a full independent copy of the Column or mixin, so maybe just using the new column always is better?

This would be a real API change.