Flammie

Flammie A Pirinen on github pages

Project maintained by flammie Hosted on GitHub Pages — Theme by mattgraham

Flammie on API design (for scientific software)

Firstly a brief note about what API is though, in 2020’s young programmers have mostly learnt the word in reference to web APIs (REST, json, GraphQL), but the word really just means “application programming interface”, or the way you call your code, traditionally from a programming language or a script instead of like designated set of HTTP addresses and response types, of course the idea is generalisable over both but what I talk about here about is I have an application programmer who is using programming language and a libraries in mind.

I mainly work as an academic who knows software engineering and linguistics. There is a long history of scientific software being all pre-alpha quality scripts that were just workable enough to run an experiment to write a paper. Very few academic software are successful enough that many other people will use them, but as the field (here: NLP) gets more popular and profitable, more requests will crop up. And I think funders also want to see products as result of funding. So a few research software like this have ended up in use by many or even most of the researchers in the field, perhaps some commercial groups and end-users too. And the API’s always suck, I have nearly never seen anything that would be nice to use. Always:

too many steps, classes and thingamajigs to load up
confusing naming
optional params and stuff that aren’t optional and obligatory that shouldn’t
too many ways to do the same thing
the most obvious way is missing (you can use CachingPoolWorkerAsync or AbstractFactoryLoader over UnbufferedQuantumStreamFifo but no doIt on Strings?)
you have to keep dragging weird long list of arguments through everywhere

It’s understandable, these were built organically during experimentation, however, the rest of the world do not want to do the experiments first, they want to use the SOTA model you showed at the EMNLP and get the same results first. (And I can guarantee you there is only 90 % of required parameters in that paper, with very different names as in the API).