We can accurately downsample so I don't think we should error here. An aliased downsampled image is still zoomable, at which point the aliasing will go away. If the image is above 2**23 we should just subsample by a factor of 2 and warn.

👍 - for subsampling, would you recommend just doing arr[::2]?

I think we want arr[::int(np.ceil(Ny/2**23)), :][:,::int(np.ceil(Nx/2**23))]?

Yep, my comment was modulo finding the right stride, but just wanted to double check that striding was the correct approach.

The right approach is open to debate, in that you could anti-alias the data before hitting it with a stride. But for such a massive subsample, I think a simple stride is better than an unspecified overflow/wrap!

Why do we bother to downsample at all? We cannot display more than N pixels and should refuse to accept more (In the face of ambiguity refuse the temptation to guess). If a user has more data, it's their responsibility and competence to reduce them.

I'm happy to go either way - perhaps worth discussing on a dev call and coming back with a decision?

We almost always resample so I don't see the problem with reducing the data before trying the final resample step. This will be aliased, but a max-sized image will also be aliased by the final downsample, so I doubt it would be detectable.

I am 👍🏻 in favor of this.

As discussed on the call, we should go ahead and subsample and warn, which I think is what this PR does.

👍 - I still need to add a test for resampling the columns that's similar to the existing one for rows

This should be good for review now. Tests and an API note added.

Needs rebase for docs to work.

power-cycled to pick up the CI config from main.