Improve use of install.packages()

[coatless]

• From 5 install commands down to 1 by listing packages in pkgs variable.
• Use ncores to parallelize compiles as each package is installed from source.

[choldgraf]

Thanks for the improvements! One note (maybe @minrk can confirm) is that I think increasing the number of cores won't make a difference on mybinder.org since users only get 50% of a single CPU. If we confirm this, could you write in a comment to this effect just so users aren't confused if it still takes a while to build.

On that note, if you have any idea for how we can avoid building from source on linux, I'd love to hear it :-)

[coatless]

@choldgraf as the image is using rocker/tidyverse:3.5.1, which in turn is built off of debian:stretch. You could use apt-get r-cran-<pkgname>.

For example, ggplot2 could be installed with:

apt-get update && apt-get install -y r-cran-ggplot2

c.f. https://cran.r-project.org/bin/linux/debian/#administration-and-maintenance
http://dirk.eddelbuettel.com/blog/2017/12/13/#013_faster_package_installation_two

[cboettig]

@coatless Actually it's probably not a good idea to install r-cran-* packages on the rocker/tidyverse image. Those debian binaries will depend on r-base debian binary, and so install another, older version of R on your image. See https://github.com/rocker-org/rocker-versioned/#notes. Definitely preferable to use install.packages() in this context, though I realize it does increase the compile time.

If you're using rocker/tidyverse, ggplot2 & rest of tidyverse, httr & rmarkdown will already be installed; though perhaps you're referring rather to the generic question of installing packages. I'm not sure what the best solution is there; but it may help to use a FROM image in the Dockerfile that simply has more common dependencies pre-installed -- that's why rocker/binder builds on rocker/geospatial, for instance.

[choldgraf]

@cboettig thanks for the thoughts! We've also been discussing this topic over in jupyterhub/repo2docker#412 ... we hadn't realized that were no binaries built for the linux world. It seems that some people are actually just using the anaconda distribution w/ the conda-forge repository to install R packages, which seems like a hacky-but-workable solution for linux. Seems like there are lots of sub-optimal solutions out there but nothing that's obviously the way to go :-/

[eddelbuettel]

we hadn't realized that were no binaries built for the linux world

Incorrect as stated. Ie the Rutter PPAs mentioned on the CRAN README have ~ 3000 of the 13000 CRAN packages as binaries. Of course, those are current as opposed to snapshotted a la MRAN, but saying "oh, linux has no binaries, let's just switch to anaconda" is not entirely correct in its reasoning.

[coatless]

@coatless Actually it's probably not a good idea to install r-cran-* packages on the rocker/tidyverse image. Those debian binaries will depend on r-base debian binary, and so install another, older version of R on your image. See https://github.com/rocker-org/rocker-versioned/#notes. Definitely preferable to use install.packages() in this context, though I realize it does increase the compile time.

Ewk! I missed that portion. Thanks for pointing that out @cboettig.