declaring the origin of an array

[d-v-b]

There are many applications where it's useful to declare the origin of an array relative to the indexing domain defined by the chunks. Origin metadata allows arrays to "underfill" chunks at the beginning of the chunk grid. Currently arrays can only underfill at the end of the chunk grid.

I think the chunk_grid field is a good place to declare the origin. An additional origin field, with a type array[int], in the configuration of the regular chunk grid would be sufficient to express this. I'm not sure if we need a completely new chunk grid or a new key added to the regular chunk grid configuration. The latter is viable only if naive zarr implementations error when they encounter the new key. I would also be interested in having an origin field in the rectilinear chunk grid configuration.

But maybe there is a better place to express the origin? Curious to hear thoughts from others.

cc @alexamici

[jbms]

This is a good idea --- there are actually two different origins we can consider.

The chunk grid origin is the position in the array index space that corresponds to the start of chunk 0 in each dimension (normally 0).

Then there is the lower bound of the array indexing space (normally 0).

Changing the grid origin of an existing array would shift the position of existing data relative to the array grid.

It would also be useful to allow negative indices both in the array index space and in the chunk coordinates in order to allow expansion in both directions.

[LDeakin]

Then there is the lower bound of the array indexing space (normally 0).

Do you have any ideas on how to express this in metadata? I think it is quite natural to write negative blocks with some chunk grids, but the standard array_shape is not sufficient for expressing their existence.

Perhaps a Zarr 3.2 expansion of the field

"array_shape": [[-10, 100], ...],

[d-v-b]

explicitly declaring the extent of the array as a bounding box is an interesting idea. The shape is then the length of each edge of the bounding box. But the bounding box representation allows an array's axes to have orientation, e.g. [[-1, 1]] vs [[1, -1]], which allows expressing reversal of an array along an axis. However I'm not aware of a use case for this.

[jbms]

Handling reversing of dimensions might better be supported via top-level (before chunking) array to array codecs (not yet supported but would make sense).