Skip to content

Using renv with Docker

Source: vignettes/docker.Rmd
docker.Rmd

While renv can help capture the state of your R library at some point in time, there are still other aspects of the system that can influence the runtime behavior of your R application. In particular, the same R code can produce different results depending on:

  • The operating system in use,
  • The compiler flags used when R and packages are built,
  • The LAPACK / BLAS system(s) in use,
  • The versions of system libraries installed and in use,

And so on. Docker is a tool that can help solve this problem through the use of containers. Very roughly speaking, one can think of a container as a small, self-contained system within which different applications can be run. Using Docker, one can declaratively state how a container should be built (what operating system it should use, and what system software should be installed within), and use that system to run applications. (For more details, please see https://environments.rstudio.com/docker.)

Using Docker and renv together, one can then ensure that both the underlying system, alongside the required R packages, are fixed and constant for a particular application.

The main challenges in using Docker with renv are:

  • Ensuring that the renv cache is visible to Docker containers, and

  • Ensuring that required R package dependencies are available at runtime.

This vignette will assume you are already familiar with Docker; if you are not yet familiar with Docker, the Docker Documentation provides a thorough introduction. To learn more about using Docker to manage R environments, visit environments.rstudio.com.

We’ll discuss two strategies for using renv with Docker:

  1. Using renv to install packages when the Docker image is generated;
  2. Using renv to install packages when Docker containers are run.

We’ll also explore the pros and cons of each strategy.

Creating Docker images with renv

With Docker, Dockerfiles are used to define new images. Dockerfiles can be used to declaratively specify how a Docker image should be created. A Docker image captures the state of a machine at some point in time – e.g., a Linux operating system after downloading and installing R 4.5. Docker containers can be created using that image as a base, allowing different independent applications to run using the same pre-defined machine state.

First, you’ll need to get renv installed on your Docker image. For example, you could install the latest release of renv from CRAN:

RUN R -e "install.packages('renv', repos = c(CRAN = 'https://cloud.r-project.org'))"

Alternatively, if you need to use the development version of renv, you could use:

RUN R -e "install.packages('renv', repos = 'https://rstudio.r-universe.dev')"

Next, we’ll copy renv.lock into the container:

WORKDIR /project
COPY renv.lock renv.lock

Now, we renv::restore() to install those packages. At this stage, you’ll need to decide which of R’s library paths you’d like to use for pacakge installation. (See ?.libPaths for more information.) There are a couple of options available:

Use the default library paths

This method is appropriate if you’d like these packages to be visible to all R processes launched using this image, and can be done via:

RUN R -e "renv::restore()"

Note that this method may fail if R’s default library paths are not on a writable volume in the Docker image. If this is the case, consider one of the alternatives below.

Use the default project library path

If you want to use renv’s default project-local library path, you’ll need to initialize the project within the Docker container as an renv project. This can be done with:

RUN R -s -e "renv::init(bare = TRUE)"
RUN R -s -e "renv::restore()"

Or, alternatively, if you already have a project autoloader + settings available – e.g. because you’re creating a Docker image from an existing renv project – you could use:

RUN mkdir -p renv
COPY .Rprofile .Rprofile
COPY renv/activate.R renv/activate.R
COPY renv/settings.json renv/settings.json
RUN R -s -e "renv::restore()"

Note that in this mode, the installed packages would only be visible to R sessions launched using /project as the working directory. This will be the default behavior as long as WORKDIR is not changed, but it’s important to keep this in mind.

Use a custom library path

If you’d like to fully customize the library path used, the simplest approach is likely to use the RENV_PATHS_LIBRARY environment variable. This mimics the above approach, but customizes the library paths used by renv. For example:

ENV RENV_PATHS_LIBRARY=renv/library
RUN R -s -e "renv::init(bare = TRUE)"
RUN R -s -e "renv::restore()"

Alternatively, you could manage the library paths yourself via .libPaths() – see ?.libPaths in R for more inforamtion.

Speeding up package installations

The previously-described approaches are useful if you have multiple applications with identical package requirements. In this case, a single image containing this identical package library could serve as the parent image for several containerized applications.

However, renv::restore() is slow – it needs to download and install packages, which can take some time. Thus, some care is required to efficiently make use of the renv cache for projects that require:

  1. Building an image multiple times (e.g., to debug the production application as source code is updated), or

  2. Calling renv::restore() each time the container is run.

The former process can be sped up using multi-stage builds, the latter by dynamically provisioning R Libraries, as described below.

Multi-stage builds

For projects that require repeatedly building an image, multi-stage builds can be used to speed up the build process. With multi-stage builds, multiple FROM statements are used in the Dockerfile and files can be copied across build stages.

This approach can be leveraged to generate more efficient builds by dedicating a first stage build to package synchronization and a second stage build to copying files and executing code that may need to be updated often across builds (e.g., code that needs to be debugged in the container).

To implement a two stage build, the following code could be used as part of a Dockerfile.

FROM <parent-image> AS base

# intialize the project; assuming renv infrastructure available
WORKDIR /project
RUN mkdir -p renv
COPY renv.lock renv.lock
COPY .Rprofile .Rprofile
COPY renv/activate.R renv/activate.R
COPY renv/settings.dcf renv/settings.dcf

# change default location of cache to project folder
RUN mkdir renv/.cache
ENV RENV_PATHS_CACHE=renv/.cache

# restore 
RUN R -s -e "renv::restore()"

The above code uses FROM <parent-image> AS <name> to name the first stage of the build base. Here, <parent-image> should be replaced with an appropriate image name.

Subsequently, the code uses approach 2 (described above) to copy the auto-loader to the project directory in the image. It additionally creates the renv/.cache directory that is to be used as the renv cache.

The second stage of the build is defined by adding the following code to the same Dockerfile, below the previous code chunk.

FROM <parent-image>

WORKDIR /project
COPY --from=base /project .

# add commands that need to be debugged below

Here, <parent-image> could be the same as the parent image of base, but does not have to be (see documentation for more details).

The key line is the COPY command, which specifies that the contents of /project directory from the base image are copied into the /project directory of this image.

Any commands that will change frequently across builds could be included below the COPY command. If only this code associated with the second stage build is updated then renv::restore() will not be called again at build time. Instead, the layers associated with the base image will be loaded from Docker’s cache, thereby saving significant time in build process.

In fact, renv::restore() will only be called when the base image needs to be rebuilt (e.g., when changes are made to renv.lock). Docker’s cache system is generally good at understanding the dependencies of images. However, if you find that the base image is not updating as expected, it is possible to manually enforce a clean build by including the --no-cache option in the call to docker build.

Dynamically Provisioning R Libraries with renv

However, on occasion, one will have multiple applications built from a single base image, but each application will have its own independent R package requirements. In this case, rather than including the package dependencies in the image itself, it would be preferable for each container to provision its own library at runtime, based on that application’s renv.lock lockfile.

In effect, this is as simple as ensuring that renv::restore() happens at container runtime, rather than image build time. However, on its own, renv::restore() is slow – it needs to download and install packages, which could take prohibitively long if an application needs to be run repeatedly.

The renv package cache can be used to help ameliorate this issue. When the cache is enabled, whenever renv attempts to install or restore an R package, it first checks to see whether that package is already available within the renv cache. If it is, that instance of the package is linked into the project library. Otherwise, the package is first installed into the renv cache, and then that newly-installed copy is linked for use in the project.

In effect, if the renv cache is available, you should only need to pay the cost of package installation once – after that, the newly-installed package will be available for re-use across different projects. At the same time, each project’s library will remain independent and isolated from one another, so installing a package within one container won’t affect another container.

However, by default, each Docker container will have its own independent filesystem. Ideally, we’d like for all containers launched from a particular image to have access to the same renv cache. To accomplish this, we’ll have to tell each container to use an renv cache located on a shared mount.

In sum, if we’d like to allow for runtime provisioning of R package dependencies, we will need to ensure the renv cache is located on a shared volume, which is visible to any containers launched. We will accomplish this by:

  1. Setting the RENV_PATHS_CACHE environment variable, to tell the instance of renv running in each container where the global cache lives;

  2. Telling Docker to mount some filesystem location from the host filesystem, at some location (RENV_PATHS_CACHE_HOST), to a container-specific location (RENV_PATHS_CACHE_CONTAINER).

For example, if you had a container running a Shiny application:

# the location of the renv cache on the host machine
RENV_PATHS_CACHE_HOST=/opt/local/renv/cache

# where the cache should be mounted in the container
RENV_PATHS_CACHE_CONTAINER=/renv/cache

# run the container with the host cache mounted in the container
docker run --rm \
    -e "RENV_PATHS_CACHE=${RENV_PATHS_CACHE_CONTAINER}" \
    -v "${RENV_PATHS_CACHE_HOST}:${RENV_PATHS_CACHE_CONTAINER}" \
    -p 14618:14618 \
    R -s -e 'renv::restore(); shiny::runApp(host = "0.0.0.0", port = 14618)'

Note that the invocation above assumes that the project has already been initialized either via calling renv::init() or by copying the requisite renv project infrastructure. With this, any calls to renv APIs within the created docker container will have access to the mounted cache. The first time you run a container, renv will likely need to populate the cache, and so some time will be spent downloading and installing the required packages. Subsequent runs will be much faster, as renv will be able to reuse the global package cache.

The primary downside with this approach compared to the image-based approach is that it requires you to modify how containers are created, and requires a bit of extra orchestration in how containers are launched. However, once the renv cache is active, newly-created containers will launch very quickly, and a single image can then be used as a base for a myriad of different containers and applications, each with their own independent package dependencies.

Handling the renv autoloader

When R is launched within a project folder, the renv auto-loader (if present) will attempt to download and install renv into the project library if it’s not available. Depending on how your Docker container is configured, this could fail. For example:

Error installing renv:
======================
ERROR: unable to create '/usr/local/pipe/renv/library/master/R-4.0/x86_64-pc-linux-gnu/renv'
Warning messages:
1: In system2(r, args, stdout = TRUE, stderr = TRUE) :
  running command ''/usr/lib/R/bin/R' --vanilla CMD INSTALL -l 'renv/library/master/R-4.0/x86_64-pc-linux-gnu' '/tmp/RtmpwM7ooh/renv_0.12.2.tar.gz' 2>&1' had status 1
2: Failed to find an renv installation: the project will not be loaded.
Use `renv::activate()` to re-initialize the project.

Bootstrapping renv into the project library might be unnecessary for you. If that is the case, then you can avoid this behavior by launching R with the --vanilla flag set; for example:

R --vanilla -s -e 'renv::restore()'