If you haven't seen mybinder.org, you should go check it out. It's a site that runs IPython/Jupyter Notebooks from GitHub for free, and I think it's a solution to publishing reproducible computational work.
For a really basic example, take a look at my demo Software Carpentry lesson. Clicking on that link brings up a Python notebook that will load and plot some data; it's interactive and you can go ahead and execute the cells.
There are a couple of really appealing things about this.
First, it's simple -- the source repository is just a directory that contains a Jupyter notebook and some data. It's pretty much what you'd have if you were actually, you know, doing data analysis with data in a repo.
Second, it's built on top of solid technology. mybinder's foundation is Docker and Jupyter Notebook, both of which are robust and have large, supportive communities. Sure, behind the scenes, mybinder is using some bleeding edge tech (Kubernetes running on Google Compute Engine) to provision and load balance the compute. But I bet it would take me no more than 30 minutes to cobble together something that would let me run any mybinder-compatible repository on my laptop, or on any compute server that can run Docker.
Third, it's super configurable -- because it's based on Docker, you can install pretty much anything you want. I tested this out myself by installing R and an R kernel for the Jupyter notebook -- see this mybinder link and repo. The Dockerfile that installs things is pretty straightforward (given that installing all that stuff isn't straightforward in the first place ;).
Fourth, it's open source. Go ahead, do what you want with it!
I believe mybinder is the hitherto missing link for publishing reproducible computation. Before this, we could give people our data and our data analysis scripts, but not our compute environment - or at least not in a way that was easy to run. We can now provide readers with a link that they can use to instantly go from my paper to a running instantiation of my paper's data analysis.
What's missing and easy?
Authentication and Cost: right now, mybinder is running on someone else's dime -- specifically, the Freeman Lab's dime. This can't continue forever, especially since it's totes open for abuse. I expect that sooner or later someone will gin up a way to run this on many cloud compute services; once that works, it should be pretty easy to build a site that redirects you to your favorite compute service and lets you enter your own credentials, then boom there you go. Insta-compute.
(I think current mybinder limitations are in the several GB of RAM range, with a few dozen MB of disk space.)
Executing anything other than GitHub public repos: mybinder only runs off of public GitHub repositories right now, but that's not going to be too tricky to fix, either. BitBucket and GitLab and private repos can be served almost as easily.
RStudio and RMarkdown: currently, it's all Project Jupyter, but it should be pretty easy to get RStudio Server or some sort of RMarkdown application (Shiny?) working on it.
Pre-built Docker images: you need to derive your Dockerfile from the andrewosh/binder-base image, which means you can't use pre-built images from the Docker Hub (even if they were built from binder-base). I predict that will go away soon enough, since it's not a fundamental requirement of the way they're doing things.
What's missing and not easy?
Big files and big/long compute: Coordinating the delivery of big files, big compute, and big memory (potentially for a long period of time) is something that's out of easy reach, I think. Docker isn't ready for really data-intensive stuff, from what I can tell. More, I'm not sure when (if ever) people will completely automate the heavyweight data analysis - surely it's further away than most people using Jupyter or RStudio for data analysis.
The split that my lab has made here is to use a workflow engine (e.g. make, pydoit, or snakemake) for the compute & data intensive stuff, and then feed those intermediate results (assembly and mapping stats, quantification, etc.) into analysis notebooks. For mybinder purposes, there should be no problem saving those intermediate results into a github repo for us and everyone else to analyze and reanalyze.
What are the next steps?
I bet this will evolve pretty fast, because it's an easy way for scientists to deliver compute to friends, colleagues, and students.
In terms of learning how to make use of mybinder, Software Carpentry already teaches scientists how to do much of the necessary stuff -- scripting, git, and Jupyter Notebooks are all part of the standard curriculum. It's not a big leap to show students some simple Docker files. There's virtually no unnecessary configuration required at the moment, which is a big win; we can talk about why everything is actually important for the end result, which is <ahem> not always the case in computing.
Beyond that, I'm not sure where the gaps are - as usual we're missing incentives for adopting this, so I'd be interested in thinking about how to structure things there. But I look forward to seeing how this evolves!
--titus
p.s. Everware does something similar to mybinder. I'm aware of it, I just haven't had the chance to try it out - I'd love to hear from people who have used it!
p.p.s. For a real, and super cool, example of mybinder, see: Min Ragan-Kelly's LIGO/gravitational waves tutorial.
Comments !