I basically agree about bringing modern software development into science, and I teach grad students and postdocs all over the country as well. But I'm looking for a better way.
A major problem is constant churn. We used to teach Travis CI. That seems to have fallen out of favor, so now we teach GHA. But what about all those projects still using Travis? What happens if/when Github or GHA fall out of favor?
Same with python packaging. setup.py vs. setup.cfg, pypi vs. conda vs. poetry vs. whatever else. Various ways of specifying requirements. Plus docker on top for some people.
How to build/host docs is yet another thing to figure out. GH Pages, netlify, or RTD? Depends on the language? Now lots of projects are both C++ and Python, and cross language documentation is hacky at best.
And what about git? Lots of non-science software developers have issues understanding that, what chance do informally-trained scientists have?
And now we are starting to see people wanting to use Julia and Rust. Or to program web apis and web pages (html/css/js).
Each of these issues may not be too bad on its own, but they all add up. When you already know about all these things, you drastically underestimate how difficult they can be to understand for someone new.
Tech companies and startups can keep up with this stuff. I'm not sure scientists' time should be spent doing this.
[PS: Not sure why you think my name is Jonathan, unless I said that somewhere. If I did, I was lying :) ]
To git specifically, git is hard for everyone, anyone that says otherwise is either lying or doesn't know what they don't know. I said as much at the software carpentry classes I helped with. The lead scientist trainer didn't really like me saying that, but it is the truth.
Humans have a fixed decision budget, we can only handle so much complexity before our brain turns off. Do we want to spend that solving domain problems or on huge software stacks?
I guess this is where the folks like us could do a better job to make simple, obvious ways to do things. There is a lot of value in standardization and a reduced amount of choice. Hobgoblins and all.
I knew there was a lot of complexity in our tech stacks, trying to bootstrap smart people from basically nothing made it even more apparent. But I don't think the answer is to get scientists out of writing code. Code is the mirror to science as science is the mirror to our understanding. It rigorously and formally, when used correctly, closes the loop.
For all the problems that Python notebooks have, I think they make an excellent way to convey ideas, but they need history and a way to make them reproducible.
Apologies for the mistaken identity, I thought you were a different seagull.
Bingo.
Sounds like we are fairly in agreement. Scientists sometimes have to write code, there's now way around it. But I wish there was more help for all the ancillary stuff.
I want scientists working on their simulation, not debugging why their docs build pipeline fails once in a while, or have to investigate some warnings about needing to upgrade their chosen distro in the CI pipeline.
> There is a lot of value in standardization and a reduced amount of choice.
The only issue here is today's standard is tomorrow's "old-fashioned way" - and I'm only half-serious about "tomorrow". So even if they get everything set up the standard way, they have to know enough at some point to migrate it. I still don't have a good answer for that.
Everything is containerized. All the batteries are included. Each commit also records all the tests to assists in any upgrades that might occur over the life of the project.
The stacks would be versioned, but also include a mechanism to update individual components w/o breaking the whole system. The root-node containers would have source, and one should be able to "rebuild world", so we aren't relying on cached packages from a distro that is long dead. I am trying to revive some software that was last installable on Debian 4.
The whole workflow's ultimate product is a paper and an executable notebook in a way that is easily reproducible by anyone. The full software stack is rebuildable from scratch using only source (like a FreeBSD build world). One should be able to take a paper from an archive 50 years from now and recreate a bit for bit copy using one command.
Pretty sure that is a reference to the book by Richard Bach.
Jonathan, you can use a google doc, or a gist, or bug in a github project to get your message out. I recommend you should.
I am a failed scientist that programmer in tech. I have helped grad students train up on unit testing, version control, api design, etc. I think modern software development needs to be brought into science so that science is reproducible. Without reproduction, we have nuthing. So I disagree, but I would love to hear what you have to say. Probably that our definitions of modern don't match and that you are drowing in sprint planning and story points, 1 beaker to 3 beakers.