Return "alleles" as a numpy ragged array

[hyanwong]

I thought I might have brought this up before, but can't find the issue. Is there a performance reason why we return the "alleles" for a variant as a tuple rather than a numpy array? It's handy to be able to get our the allelic states by doing variant.alleles[variant.genotypes] but at the moment this needs to be wrapped like this:

np.array(variant.alleles)[variant.genotypes]

which is a tad messy.

[benjeffery]

So it looks like we always have to copy the alleles, as although they are often one continuous block of ragged array memory (variant->user_alleles_mem) this is not always the case as when updating a site we point to memory in the derived state of the mutation table.
It would be possible to copy into a numpy array and not a tuple, but I don't expect it to increase performance and it may be a breaking change for some edge cases. In general in the API we return normal Pythonic objects except for arrays of numeric data.

[molpopgen]

The tuple is also nice because it signals that the data are immutable. (We can do that with arrays, but generally do not in tskit.)

[jeromekelleher]

Numpy strings are generally pretty icky as well, so the idea was to make the alleles "normal" python strings where possible. Unless there's a compelling use-case I can't see there being much motivation to change this as it would almost certainly break people's code.

[benjeffery]

Closing this, repopen if you feel strongly about it!

[hyanwong]

I think these are all good arguments, and it's easy to work around the issue as above, so I can see why we want to keep it as-is, so closing this seems right to me. I guess there's no way to return a tuple as we do at the moment but allow it to be indexed using a numpy array by default?

[benjeffery]

I think these are all good arguments, and it's easy to work around the issue as above, so I can see why we want to keep it as-is, so closing this seems right to me. I guess there's no way to return a tuple as we do at the moment but allow it to be indexed using a numpy array by default?

Not without doing something distasteful like subclassing tuple, sorry @hyanwong!