Could you test this on #51411 as well? This should address the problem for CoW

Could you test this on #51411 as well? This should address the problem for CoW

I tested locally on top of your branch in #51411 and it seems to work. What "problem" are you referring to?

Performance, we are avoiding the copy on the block level there

@phofl - hmmm... clearly this doesn't play well with #51411

Not sure what I did yesterday to convince myself it did... sorry

actually, what is the expected behavior for COW wrt pyarrow? Since arrow arrays are immutable, ArrowExtensionArray.copy doesn't actually copy the underlying pyarrow array. I think tests like this (which are failing) are expected to fail:

import pandas as pd
import numpy as np

arr = pd.array([1, 2, 3], dtype="int64[pyarrow]")
arr2 = arr.copy()

assert not np.shares_memory(arr, arr2)  # -> raises

CoW handles the copies globally on the block level, hence we can remove the check above.

You can adjust the tests, eas made to many defensive copies without CoW, you got rid of one of them, which is a good thing

Ok will do. Is the comment you added still relevant? (# TODO(CoW): Not necessary anymore when CoW is the default)

I would think this fastpath can still apply after CoW is the default (e.g. for whatever reason CoW is not being used)

We are not planning on having a non CoW mode when CoW is the default

Merge when ready @phofl

Could you rebase @lukemanley

Thanks @lukemanley