I submitted a pull request to pydata/sparse (I’m one of its maintainers) about always keeping a COO array in canonical form.

We (Matthew Rocklin, CJ Carey and I) were having a discussion there about whether it makes sense to allow sparse COO arrays to be in non-canonical order. In short, canonical form is:

Coordinates are sorted in lexographical order, i.e., the same order they would appear in in a C-contiguous array.
There are no duplicate coordinates allowed.
There are no stored coordinates with a zero value.

With this in mind, what are the SciPy developers’ opinions on this? Does allowing COO arrays to be in non-canonical form make sense? Feedback welcome on the PR or the mailing list.