A big thank you to Titus Brown, associate professor of population health and reproduction at UC Davis, for hosting a two-day Docker Workshop at BIDS. Below is a write-up of the event as well as Titus's takeaways from the workshop.
We just finished the second day of a workshop on Docker at the Berkeley Institute for Data Science. I was invited to organize the workshop after some Berkeley folk couldn't make our Davis workshop in November, and so I trundled on down for two days to give it a try. About 15 people showed up, from all walks of research and engineering. (The materials are freely available under a CC0 license.) I did the intro stuff, and Luiz Irber and Carl Boettiger both taught sections on their more advanced uses of Docker.
tl;dr? I think the workshop went well, in the sense that the attendees all got some hands-on experience with Docker and engaged its potential benefits (as well as some of its limitations).
A special thanks to BIDS for hosting me & putting me up overnight!
We delivered the workshop in a Software Carpentry style, with lots of command-line interaction and plenty of time for questions. People were welcome to follow along with what I was doing on the screen, or just watch; in either case, we tried to provide fairly thorough notes so that people could refer back to them in the future.
The workshop was attended by a variety of people, including software engineering folk, infrastructure folk, supercomputing folk, librarians, and grad students / postdocs. Some people had a fair bit of experience with Docker, while others had only vaguely heard of it; despite this spread in experience I am told everyone got something out of it. The experienced folk got to see how we used it, while the people new to Docker got a somewhat thorough overview of the technology and its potential applications.
The workshop focused on scientific applications of Docker, but we had to get through what Docker was first! Day onewas mostly about the basics - Docker containers and images, docker-machine, building docker images, data volumes, directory mirroring, and exposing ports. Day two went a bit further afield and talked about mybinder, docker-compose and docker-swarm, and some potential workflows for data-intensive applications.
A few random thoughts --
I still don't really know how to teach Docker, because it's a mix of super obvious (for the sysadmin and CS-types) and super detailed (to, you know, actually use it). I think the hands-on approach has to beat more conceptual or theoretical approaches here :). The concept is clear enough for CS people, but the details of the implementation (and how it all actually works) are really at the heart of putting Docker to use and figuring out where it fits.
This time, I started with local docker and then introduced docker-machine. I might introduce docker-machine first next time, to make it clear that local volume mapping is a super-special feature that you can't rely on, and to introduce the point that even on your Mac/Windows machine, the port mapping is going to be confusing!
Docker doesn't run on HPCs or other multi-tenant systems very well, which everyone recognizes as a problem. There was some discussion of Shifter, which is a Docker-like technology intended to address this deficit.
(I'm happy to sign on to support a grant proposal if someone can credibly claim to be able to deal with the problems of Docker running as root. I can also broker introductions to people who might pay for it :).
Docker still isn't great in terms of data management. Data volumes are poor vessels for communicating data. Carl Boettiger showed us an approach based on Flocker that "required only a little configuration." I should probably look into this...
We don't have a clear use case for docker-compose in science!
The ease with which you can get an RStudio Server or a Jupyter notebook running inside of Docker is just astonishing.
Everyone was pretty impressed with how easy and obvious mybinder.org is, and seemed to think that it would fit into an only slight evolved data science workflow. I really want to explore at the seams between long-running data-intensive tasks, and the all-data-fits-in-github short-run analysis supported by mybinder...
I put together two demos for mybinder, one of a Software Carpentry data set and one showing how to do some significant software installs.
It's a super cool service!
(There's two other similar services/packages I want to look at - JupyterHub and Everware. Pointers to intro tutorials welcome!)
I'm really interested in suggesting Docker as a way to deal with installation issues in workshops - if only as a backup. The two things standing in the way of that are (a) older laptops and (b) my inexperience with running Docker on Windows.
Some mixtures of jupyterhub/mybinder/everware might be an alternative solution. But then network becomes an issue. Heck, at a recent workshop on training, we actually brought up the idea of carting around a pile of Raspberry Pis to run local compute clusters for training purposes...
See orginal post here.