Choose configuration library

[hammer]

@eric-czech likes how Dask does configuration, and Matt Rocklin pointed us to https://github.com/pytroll/donfig, which is Dask's configuration system factored out into a separate repo.

There's been some recent conversation about bringing this project under the dask organization: dask/dask#3853. We should keep an eye on that and potentially voice our support.

[jeromekelleher]

Just out of curiosity, what do we need a configuration library for?

[hammer]

I defer to @eric-czech for a more detailed explanation, but you can imagine that we'll need some way for users to specify which backends are available and what options should be passed to those backends.

[eric-czech]

I wouldn't say there is an urgent need for any of these now, but here are some of the things I had in mind when I was trying to think about how our GWAS pipelines will evolve:

• Switch array backends, useful when e.g.:
  • Benchmarking a whole pipeline on CPUs vs GPUS
  • Forcing a pipeline that defaults to dask implementations to use numpy implementations instead
• Set thresholds for a dispatcher that chooses between array backends intelligently
• Configure behavior related to expensive assertions
• Set default variable names for arrays created internally
• Display options for visualizations

I liked the dask config model for these things since any of the above could be scoped to one function, a block of code in a context manager, or globally (and optionally read from a file).