cube.transpose is an in-place operation

[pp-mo]

I think it's pretty nasty and unexpected to have cube.transpose() modify in-place
That is, especially, as it seems to mimic numpy "transpose", which is not in-place.
Plus, we just don't have many in-place operations, largely by design (I think).

Thus in #2604 ...

• I had a test cube and a sample result array which should match a derived coordinate.
• I applied transpose operations to both, but the cube behaves differently to the result, when it should have been the same.

I think we should review this as it's peculiar and unexpected, and breaks the general simlilarity with numpy.
Arguably the numpy function could more clearly be named "transposed" (a la Python "sorted").
Obviously, that ship has sailed, but we could use that naming for a cube method.

Propose:

• provide copying version called "cube.transposed()" instead
• deprecating the existing "cube.transpose" -- or simply remove it from Iris 2.0 ??

[cpelley]

+1 for cube.transpose returning a cube, however it is important to us that it does not consume any more memory for the data or coordinates i.e. the cube returned should be a view of the original cube. This would also be in keeping with ndarray.transpose behaviour. Not sure where transposed comes into this as numpy uses transpose.

[pelson]

Taking this out of the v2.0.0 milestone - we aren't going to hit this in time, but can very readily add this as a future flag during the v2.x series. Of course, we will need to figure out data sharing before we can do this though (again, not going to happen for v2.0.0).

[github-actions]

In order to maintain a backlog of relevant issues, we automatically label them as stale after 500 days of inactivity.

If this issue is still important to you, then please comment on this issue and the stale label will be removed.

Otherwise this issue will be automatically closed in 28 days time.

[github-actions]

This stale issue has been automatically closed due to a lack of community activity.

If you still care about this issue, then please either:

• Re-open this issue, if you have sufficient permissions, or
• Add a comment pinging @SciTools/iris-devs who will re-open on your behalf.

[monishanatchiar]

I agree that in-place operations need to be reduced, or we should at least have an option like in_place=False. Currently, I have to create a copy of the iris.cube.Cube to not transpose it in-place.

[cpelley]

I appreciate numpy is not consistent itself here but 'typically' method calls are in-place and calling an operation via function call would return a copy (numpy.ndarray.transpose method call actually returns a view... rather than internally transposing a view on the object itself i.e. ~in-place).

i.e. anyway, so for iris what about:

return copy

iris.transpose(cube, ...)

in-place transpose of cube

cube.transpose(...)

...we should at least have an option like in_place=False...

Numpy makes use of the optional out=None parameter for this purpose and is more flexible:
e.g., see cumsum

I think a contract is required on the UI. Ideally consistency, predictability AND flexibility.

Cheers

[trexfeathers]

Xref #5007