Install R packages using docker file

Viewed 32729

I have installed R using below line in my docker file. Please suggest how do I specify now packages to be installed in my docker file.

RUN yum -y install R-core R-devel

I'm doing something like this:

RUN R -e "install.packages('methods',dependencies=TRUE, repos='http://cran.rstudio.com/')"\
    && R -e "install.packages('jsonlite',dependencies=TRUE, repos='http://cran.rstudio.com/')" \
    && R -e "install.packages('tseries',dependencies=TRUE, repos='http://cran.rstudio.com/')" 

Is this the right way to do?

8 Answers

As suggested by @Cameron Kerr's comment, Rscript does not give you a build failure. As of now, the recommended way is to do as the question suggests.

RUN R -e "install.packages('methods',dependencies=TRUE, repos='http://cran.rstudio.com/')"
RUN R -e "install.packages('jsonlite',dependencies=TRUE, repos='http://cran.rstudio.com/')"
RUN R -e "install.packages('tseries',dependencies=TRUE, repos='http://cran.rstudio.com/')" 

If you're fairly certain of no package failures then use this one-liner -

RUN R -e "install.packages(c('methods', 'jsonlite', 'tseries'),
                           dependencies=TRUE, 
                           repos='http://cran.rstudio.com/')"

EDIT: If you're don't use the Base-R image, you can use rocker-org's r-ver or r-studio or tidyverse images. Here's the repo. Here's an example Dockerfile -

FROM rocker/tidyverse:latest

# Install R packages
RUN install2.r --error \
    methods \
    jsonlite \
    tseries

The --error flag is optional, it makes install.packages() throw an error if the package installation fails (which will cause the docker build command to fail). By default, install.packages() only throws a warning, which means that a Dockerfile can build successfully even if it has failed to install the package.

All rocker-org's basically installs the littler package for the install2.R functionality

This is ugly but it works - see below for real world example of why it's worth doing.

# install packages and check installation success, install.packages itself does not report fails
RUN R -e "install.packages('RMySQL');     if (!library(RMySQL, logical.return=T)) quit(status=10)" \
 && R -e "install.packages('devtools');   if (!library(devtools, logical.return=T)) quit(status=10)" \
 && R -e "install.packages('data.table'); if (!library(data.table, logical.return=T)) quit(status=10)" \
 && R -e "install.packages('purrr');      if (!library(purrr, logical.return=T)) quit(status=10)" \
 && R -e "install.packages('tidyr');      if (!library(tidyr, logical.return=T)) quit(status=10)"

Real world example: devtools install starts failing because it suddenly needs libgit2-dev. install.packages() prints informative info. about the failure, but without a non-zero exit code, that just scrolls away as docker build continues.

The R -e "install.packages..." approach does not always produce an error when package installation fails.

I wrote a script based on Cameron Kerr's answer here, which produces an error if the package cannot be loaded, and interrupts the Docker build process. It installs packages from either an R package repo, from GitHub, or from source given a full URL. It also prints the time taken to install, to help plan which packages to group together in one command.

Example usage in Dockerfile:

# Install from CRAN repo:
RUN Rscript install_packages_or_die.R https://cran.rstudio.com/ Cairo
RUN Rscript install_packages_or_die.R Cairo # Uses default CRAN repo
RUN Rscript install_packages_or_die.R jpeg png tiff # Multiple packages

# Install from GitHub:
RUN Rscript install_packages_or_die.R github ramnathv/htmlwidgets
RUN Rscript install_packages_or_die.R github timelyportfolio/htmlwidgets_spin spin

# Install from source given full URL of package:
RUN Rscript install_packages_or_die.R https://cran.r-project.org/src/contrib/Archive/curl/curl_4.0.tar.gz curl

Here's the script:

#!/usr/bin/env Rscript

# Install R packages or fail with error.
#
# Arguments:
#   - First argument (optional) can be one of:
#       1. repo URL
#       2. "github" if installing from GitHub repo (requires that package 'devtools' is
#          already installed)
#       3. full URL of package from which to install from source; if used, provide package
#          name in second argument (e.g. 'curl')
#     If this argument is omitted, the default repo https://cran.rstudio.com/ is used.
#   - Remaining arguments are either:
#       1. one or more R package names, or
#       2. if installing from GitHub, the path containing username and repo name, e.g.
#          'timelyportfolio/htmlwidgets_spin', optionally followed by the package name (if
#          it differs from the GitHub repo name, e.g. 'spin').

arg_list = commandArgs(trailingOnly=TRUE)

if (length(arg_list) < 1) {
  print("ERROR: Too few arguments.")
  quit(status=1, save='no')
}

if (arg_list[1] == 'github' || grepl("^https?://", arg_list[1], perl=TRUE)) {
  if (length(arg_list) == 1) {
    print("ERROR: No package name provided.")
    quit(status=1, save='no')
  }
  repo = arg_list[1]
  packages = arg_list[-1]
} else {
  repo = 'https://cran.rstudio.com/'
  packages = arg_list
}

for(i in seq_along(packages)){
    p = packages[i]

    start_time <- Sys.time()
    if (grepl("^https?://[A-Za-z0-9.-]+/.+\\.tar\\.gz$", repo, perl=TRUE)) {
      # If 'repo' is URL with path after domain name, treat it as full path to a package
      # to be installed from source.
      install.packages(repo, repo=NULL, type="source");
    } else if (repo == "github") {
      # Install from GitHub.
      github_path = p
      elems = strsplit(github_path, '/')
      if (lengths(elems) != 2) {
        print("ERROR: Invalid GitHub path.")
        quit(status=1, save='no')
      }
      username = elems[[1]][1]
      github_repo_name = elems[[1]][2]
      if (!is.na(packages[i+1])) {
        # Optional additional argument was given specifying the R package name.
        p = packages[i+1]
      } else {
        # Assume R package name is the same as GitHub repo name.
        p = github_repo_name
      }

      library(devtools)
      install_github(github_path)
    } else {
      # Install from R package repository.
      install.packages(p, dependencies=TRUE, repos=repo);
    }
    end_time <- Sys.time()

    if ( ! library(p, character.only=TRUE, logical.return=TRUE) ) {
      quit(status=1, save='no')
    } else {
      cat(paste0("Time to install ", p, ":\n"))
      print(end_time - start_time)
    }

    if (repo == "github") {
      break
    }
}

The best solution I found is with install2.r from the littler package.

  • First install littler
RUN R -e "install.packages('littler', dependencies=TRUE)"
  • Then you can use it from bash in your Dockerfile
RUN install2.r --error --deps TRUE methods
RUN install2.r --error --deps TRUE jsonlite
RUN install2.r --error --deps TRUE tseries

The --error flag makes the build quit if the package has not been installed correctly. The --deps TRUE flag is for automatically installing the dependencies for the package

# corrected used of package name ('litter' to 'littler')

Are these repositories a solution to this problem?

My solution in this repository is to create two Docker images: The "install image": The first image consists only of the prerequisites for the projects. When running a container from this image it can install R packages in the format it needs inside the container and save them to {renv}'s cache on the host through a mount. The "final image": The second image copies the project along with dependencies from the host into the image.

I would like to recommend the rocker/tidyverse image, on top of which you can install other packages like this:

RUN R -e "install.packages('bigrquery',dependencies=TRUE, repos='http://cran.rstudio.com/')"

The same installation from r-base was followed by an issue with Rserve, which, probably, was preinstalled in r-base image. I found nothing about this on the page about r-base, so I do not recommend r-base as an easy solution.

Installation of R packages could be also done with apt-get install r-cran-*, but maintainers of rocker/tidyverse do not recommend it for this particular image because this will lead to the installation of another R version. However, you may check it out and find out it is fine for your task.

Related