Posted by Osvaldo
July 24, 2016
Although the terms are often used interchangeably, there is an important difference between reproducibility and replicability. The former is the ability to repeat some analysis thanks to the fact that sufficient detail has been shared by those who performed it first. Replicability is the probability that an independent experiment will reach the same results and conclusions that were first reported.
A data scientist is typically responsible for the analysis of data that have been produced in the lab (or in the field) by someone else, so he/she is primarily responsible to guarantee its reproducibility. This shouldn’t be too hard, since all you have to do is 1) saying which data were analyzed and 2) how exactly. You share both and you are done; certainly easier than replicate a complex and usually expensive experiment. Rmarkdown documents (by the way, this blog is written in Rmarkdown) and Python notebook have become popular in recent years also because they address this need.
For more complex projects, where multiple tools are needed, it can be more complicated than that, and it is recognized that much of the scientific literature is hard to reproduce. Reasons for this are largely to be found in a lack of right incentives, but there are several remarkable technical challenges posed by the large number of factors that influence the results. Even if we only consider the analysis part, the huge variety of available systems makes reproducibility challenging. You performed the analysis on a Linux cluster, will it run the same on a Mac OS X? If you ran Python 3.4, will it still work and give the same results with Python 3.5.1? Do I have to update that library, or is the one I have already installed recent enough? And I’m still missing that dependency…
Virtual machines (VM), which I understand were created for entirely different reasons, offer a possible way to make replication easier. The researcher can setup an environment to run the analysis, then make an image of this and share it. Another user who wants to reproduce the analysis can launch this image with a VM and it will be like sitting at the same computer: everything exactly the same.
This solution presents a few disadvantages. Images are pretty big objects, often in the order of several gigabytes. Moving them around is fairly inconvenient, to begin with. Further, a VM takes in the order of minutes to launch. Not too much for a one time analysis, still slightly annoying. In general, VMs tend to be quite heavy on the hardware.
Containers offer a light-weight option to VMs. They are usually smaller and, above all, they are launched in a second or less.
I’ve been recently developing a pipeline for the analysis of DNA sequences that relies on many external tools. As it is customary in academia, the plan is to “advertise” it with a scientific paper and hope that many other users will find it useful, use it and cite it. But installing ten other tools is certainly a disincentive.
So, I recently looked into Docker, which is now probably the most famous containerisation software.
A developer usually starts from one of the images found in Docker Hub and adds the necessary configuration by writing a
Dockerfile. Each instruction in this file is a layer, in docker terms, and the engine builds the image by stacking these layers in an optimized way. The resulting image can be made available to others via Docker Hub. In this way one can make available to the community an environment in which the application is certainly running as expected. The advantages of Docker over VMs are not limited to being light and fast. The presence of a central directory and the possibility to develop images easily and in an open format both make Docker interesting. It must be said that at the same time this poses a small risk: making dozens of images for every possible little project that one makes and pushing it onto the hub.