dimension scale creation method is in wrong (confusing) location

[equaeghe]

Currently, dimension scales are created as dataset.dims.create_scale(scale_dataset, 'scale name'), i.e., using the create_scale method of the dims attribute of a Dataset. However, as far as I can tell, dimension scale creation only acts on scale_dataset and has nothing to do with dataset, so it would be more logical to have a create_scale method (which can take a scale name as an argument) on Dataset or perhaps a static method that takes a scale dataset (and a scale name) as an argument.

The current practice is confusing, as it may lead people to believe that the scale is created specifically for dataset, which is not the case, I think. (At least I was confused.)

The change I propose does not need to lead to an API incompatibility: the new method could be added and the old one kept (perhaps in some way deprecated or advised against).

[tacaswell]

This looks like it could be converted into a static method on the dims proxy, however it is using an instance method to encode the name which may pull information from the underlying object (but it looks like this code path does not).

Having the create_scale method on the dims proxy makes a fair amount of sense to me, it provides a name space so we don't have too many loose functions floating around.

At a minimum it should be better documented.

[takluyver]

I agree with @equaeghe that it's somewhat confusing to have this method associated with an uninvolved dataset.

Instead of this:

f['data'].dims.create_scale(f['x2'], 'x2 name')

It would conceptually align with HDF5 better to do something like this:

f['x2'].make_scale('x2 name')

Or it could be an option when you create a dataset - this makes the intent clearer to my mind, though it's somewhat less flexible.

f.create_dataset('x2', data=np.arange(10), as_scale='x2 name')

[rmvanhees]

I looked at it in a different way: I use dimension scales as follows: row = np.arange(0, 256, dtype='u2') # define data for my dimension ds_row = fid.create_dataset('row', data=row) # create a dataset ds_row.dims.create_scale(fid['row'], 'some description') # turn this dataset into a dimension dset = fid.create_dataset('frame', data=frame) # create a real dataset dset.dims[0].attach_scale(ds_row) # attach one of its dimensions I think the problem is in the third line, where I have to use fid['row'] again. Reading the previous emails I realize that I create a dimension-scale by saying that it is its own dimension scale, which I can then use it for other datasets. This approach works well and is also accepted also by netCDF4 S/W. Thus I would also prefer the construction as suggested by Thomas Kluyver: `ds_row.make_scale('some description')`. Then it is clear that I can use ds_row as dimension scale and attach it to other datasets in the file. Greetings, Richard

…

On Wed, 24 Oct 2018 at 13:46, Thomas Kluyver ***@***.***> wrote: I agree with @equaeghe <https://github.com/equaeghe> that it's somewhat confusing to have this method associated with an uninvolved dataset. Instead of this: f['data'].dims.create_scale(f['x2'], 'x2 name') It would conceptually align with HDF5 better to do something like this: f['x2'].make_scale('x2 name') Or it could be an option when you create a dataset - this makes the intent clearer to my mind, though it's somewhat less flexible. f.create_dataset('x2', data=np.arange(10), as_scale='x2 name') — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#830 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADi7qt2fM0lZPGGMRkkCENo9TOAFVKhxks5uoFMugaJpZM4LpX6P> .

[takluyver]

I've opened PR #1212 to address this.