Constants for dimensions and variables

[eric-czech]

Can we live without constants for dimension names in datasets (cf. https://github.com/pystatgen/sgkit/pull/10)?

Specifically I mean these in api.py:

DIM_VARIANT = "variants"
DIM_SAMPLE = "samples"
DIM_PLOIDY = "ploidy"
DIM_ALLELE = "alleles"
DIM_GENOTYPE = "genotypes"

Using these to build/append a dataset involves something like this:

data_vars = { "variant/contig": ([DIM_VARIANT], variant_contig)}

What I'm wondering is if we're going to use constants, why stop at the dimension names? It could look like this:

data_vars = { f"{VAR_GROUP_VARIANT}/{VAR_NAME_CONTIG}": ([DIM_VARIANT], variant_contig)}

but I doubt any of us would prefer it. I think the two biggest advantages of the constants are:

1. Preventing typos
2. Allowing us to change the names

Of these, 2 seems unlikely to be important (and we'll probably not use the constants in examples/documentation anyhow) and 1 might eventually be solved with things like python/typing#28 (comment) and pydata/xarray#3967. I'm not going to hold my breath for that, but I do think it's worth asking whether or not we would all prefer this instead:

data_vars = { "variant/contig": (["variants"], variant_contig)}

As a user, I think I would be happy to make my own constants or use mypy Literal types if I wanted some static safety and to just live with the risk of typos otherwise. As a contributor, I'm not so sure but I'm leaning towards preferring the latter. Any thoughts?

[jeromekelleher]

I don't have a strong opinion here.

[alimanfoo]

Similar, no strong opinion. If just using strings, any typos should get caught in unit tests.

[eric-czech]

Thanks guys. I've stopped using them in the functions I'm working on and will stick with that from now on.