Reproducibility affected by package versions

nanxstats commented 9 years ago

Reproducibility design problem raised by @tengfei .

To increase reproducibility, instead of just using the latest versions of (system and R) packages (since the packages are constantly being updated), we should make sure that users use certain versions of packages. But how? See below discussion.

Discussion on Best Practice of Full Reproducibility on Jul 20, 2015 in the balcony with lake view:

For full reproducibility, users need to make and push their own containers.
How to select base image. rocker/r-base only has LATEST images, which means this is not fully reproducible when users pull the r-base image later. Instead, BioC has daily snapshots. -- (Is it ok for users to select their own base image (like some images with the LATEST tag), because the snapshot we store in Docker hub is a somehow 'full' image with the LATEST image at that time point? If so, we can use rocker/r-base:LATEST) THIS NEEDS CONFIRMATION
The big process. liftr will pack up local packages and make image, push container to Docker hub, the pushed container ID (tag) will be returned by a liftr function. (We need to make functions for pushing containers to Docker hub or other registries. Probably need a Docker Hub API package). Then we write out a new Rmd, only containing the pushed from: imagename:version. This need a new design of the workflow of liftr (more functions).
The use of Docker VOLUME will affect reproducibility (obviously not able to build once, run anywhere). We need to change this, probably copying all the files in the liftr working directory to the container.
~~(extract the versions (or let the user specify) and write them to the Dockerfile?)~~

nanxstats commented 9 years ago

CRAN package version persistence supported in c43c744 . (v0.3 submitted to CRAN)

nanxstats commented 7 years ago

Remote package support added: https://github.com/road2stat/liftr/commit/51cdd169b2d988480a29e948d1ad962d9f2d1078

Currently investigating the possibility of replacing all low-level Docker calls using docker-r, which will eventually solve the docker push problem.

Closing the issue for now.

nanxstats / liftr

Reproducibility affected by package versions #4