z - Advanced topic: Running R or Shell Code in Nix from R

Testing code in evolving software dependency environments with confidence

Adhering to sound versioning practices is crucial for ensuring the reproducibility of software. Despite the expertise in software engineering, the ever-growing complexity and continuous development of new, potentially disruptive features present significant challenges in maintaining code functionality over time. This pertains not only to backward compatibility but also to future-proofing. When code handles critical production loads and relies on numerous external software libraries, it’s likely that these dependencies will evolve. Infrastructure-as-code and other DevOps principles shine in addressing these challenges. However, they may appear less approachable and more labor-intensive to set up for the average R developer.

Are you ready to test your custom R functions and system commands in a a different environment with isolated software builds that are both pure at build and at runtime, without leaving the R console?

Let’s introduce with_nix(). with_nix() will evaluate custom R code or shell commands with command line interfaces provided by Nixpkgs in a Nix environment, and thereby bring the read-eval-print-loop feeling. Not only can you evaluate custom R functions or shell commands in Nix environments, but you can also bring the results back to your current R session as R objects.

Two operational modes of computations in environments: ‘System-to-Nix’ and ‘Nix-to-Nix’

We aim to accommodate various use cases, considering a gradient of declarativity in individual or sets of software environments based on personal preferences. There are two main modes for defining and comparing code running through R and system commands (command line interfaces; CLIs)

  1. ‘System-to-Nix’ environments: We assume that you launch an R session with an R version defined on your host operating system, either from the terminal or an integrated development environment like RStudio. You need to make sure that you actively control and know where you installed R and R packages from, and at what versions. You may have interactively tested that your custom function pipeline worked for the current setup. Most importantly, you want to check whether you get your computations running and achieve identical results when going back to a Nix revision that represent either newer or also older versions of R and package sources.
  2. ‘Nix-to-Nix’ environments: Your goals of testing code are the same as in 1., but you want more fine-grained control in the source environment where you launch with_nix() from, too. You are probably on the way of getting a passionate Nix user.

Case study 1: Evolution of base R

Carefully curated software improves over time, so does R. We pick an example from the R changelog, the following literal entry in R 4.2.0:

The goal is to illustrate this change in behavior from R versions 4.1.3 and before to R versions 4.2.0 and later.

Setting up the (R) software environment with Nix

First, we write a default.nix file containing Nix expressions that pin R version 4.1.3 from Nixpkgs.

library("rix")

path_env_1 <- file.path(".", "_env_1_R-4-1-3")

rix(
  r_ver = "4.1.3",
  overwrite = TRUE,
  project_path = path_env_1
)
#> 
#> ### Bootstrapping isolated, project-specific, and runtime-pure R setup via Nix ###
#> 
#> ==> Existing isolated nix-R project folder:
#>  /tmp/RtmpBQTLhD/Rbuild191dc694ac4ec/rix/vignettes/_env_1_R-4-1-3 
#> 
#> * current R session running inside Nix environment and not from RStudio
#> 
#> ==> Added `.Rprofile` file and code lines for new R sessions launched from:
#> /tmp/RtmpBQTLhD/Rbuild191dc694ac4ec/rix/vignettes/_env_1_R-4-1-3
#> 
#> * Added the location of the Nix store to `PATH` environmental variable for new R sessions on host/docker RStudio:
#> /nix/var/nix/profiles/default/bin
#> 
#> ==> Also adjusting `PATH` via `Sys.setenv()`, so that system commands can invoke key Nix commands like `nix-build` in this RStudio session outside Nix
#> 
#> 
#> ### Successfully generated `default.nix` and `.Rprofile` ###

The following expression is written to default.nix in the subfolder ./_env_1_R-4-1-3/.

#> # This file was generated by the {rix} R package v0.12.4 on 2024-10-02
#> # with following call:
#> # >rix(r_ver = "6e3a86f2f73a466656a401302d3ece26fba401d9",
#> #  > project_path = path_env_1,
#> #  > overwrite = TRUE)
#> # It uses nixpkgs' revision 6e3a86f2f73a466656a401302d3ece26fba401d9 for reproducibility purposes
#> # which will install R version 4.1.3.
#> # Report any issues to https://github.com/ropensci/rix
#> let
#>  pkgs = import (fetchTarball "https://github.com/NixOS/nixpkgs/archive/6e3a86f2f73a466656a401302d3ece26fba401d9.tar.gz") {};
#>      
#>   system_packages = builtins.attrValues {
#>     inherit (pkgs) 
#>       glibcLocales
#>       nix
#>       R;
#>   };
#>   
#> in
#> 
#> pkgs.mkShell {
#>   LOCALE_ARCHIVE = if pkgs.system == "x86_64-linux" then "${pkgs.glibcLocales}/lib/locale/locale-archive" else "";
#>   LANG = "en_US.UTF-8";
#>    LC_ALL = "en_US.UTF-8";
#>    LC_TIME = "en_US.UTF-8";
#>    LC_MONETARY = "en_US.UTF-8";
#>    LC_PAPER = "en_US.UTF-8";
#>    LC_MEASUREMENT = "en_US.UTF-8";
#> 
#>   buildInputs = [    system_packages   ];
#>   
#> }

This also includes a custom .Rprofile file that ensure that this subshell will not load any packages installed to the user’s library of packages.

Defining and interactively testing custom R code with function(s)

We now have set up the configuration for R 4.1.3 set up in a default.nix file in the folder ./_env_1_R-4-1-3. Since you are sure you are using an R version higher 4.2.0 available on your system, you can check what that as.vector.data.frame() S3 method returns a list.

df <- data.frame(a = 1:3, b = 4:6)
as.vector(x = df, mode = "list")
#> $a
#> [1] 1 2 3
#> 
#> $b
#> [1] 4 5 6

This is different for R versions 4.1.3 and below, where you should get an identical data frame back.

Run functioned up code and investigate results produced in pure Nix Rsoftware environments

To formally validate in a ‘System-to-Nix’ approach that the object returned from as.vector.data.frame() is before R < 4.2.0, we define a function that runs the computation above.

df_as_vector <- function(x) {
  out <- as.vector(x = x, mode = "list")
  return(out)
}
(out_system_1 <- df_as_vector(x = df))
#> $a
#> [1] 1 2 3
#> 
#> $b
#> [1] 4 5 6

Then, we will evaluate this test code through a nix-shell R session. This adds both build-time and run-time purity with the declarative Nix software configuration we have made earlier. with_nix() leverages the following principles under the hood:

  1. Computing on the Language: Manipulating language objects using code.

  2. Static Code Analysis: Detecting global objects and package environments in the function call stack of ‘expr’. This involves utilizing essential functionality from the ‘codetools’ package, which is recursively iterated.

  3. Serialization of Dependent R objects: Saving them to disk and deserializing them back into the R session’s RAM via a temporary folder. This process establishes isolation between two distinct computational environments, accommodating both ‘System-to-Nix’ and ‘Nix-to-Nix’ computational modes. Simultaneously, it facilitates the transfer of input arguments, dependencies across the call stack, and outputs of expr between the Nix-R and the system’s R sessions.

This approach guarantees reproducible side effects and effectively streams messages and errors into the R session. Thereby, the {sys} package facilitates capturing standard outputs and errors as text output messages. Please be aware that with_nix() will invoke nix-shell, which will itself run nix-build in case the Nix derivation (package) for R version 4.1.3 is not yet in your Nix store. This will take a bit of time to get the cache. You will see in your current R console the specific Nix paths that will be downloaded and copied into your Nix store automatically.

# now run it in `nix-shell`; `with_nix()` takes care
# of exporting global objects of `df_as_vector` recursively
out_nix_1 <- with_nix(
  expr = function() df_as_vector(x = df), # wrap to avoid evaluation
  program = "R",
  project_path = path_env_1,
  message_type = "simple" # you can do `"verbose"`, too
)

# compare results of custom codebase with indentical
# inputs and different software environments
identical(out_system_1, out_nix_1)

# should return `FALSE` if your system's R versions in
# current interactive R session is R >= 4.2.0

Syntax option for specifying function in expr argument of with_nix()

In the previous code snippet we wrapped the top-level expr function with function() or function(){}. As an alternative, you can also provide default arguments when assigning the function used as expr input like this:

df_as_vector <- function(x = df) {
  out <- as.vector(x = x, mode = "list")
  return(out)
}

Then, you just supply the name of the function to evaluate with default arguments.

out_nix_1_b <- with_nix(
  expr = df_as_vector, # provide name of function
  program = "R",
  project_path = path_env_1,
  message_type = "simple" # you can do `"verbose"`, too
)

It yields the same results.

Reduce(f = identical, list(out_nix_1, out_nix_1_b))

Comparing as.vector.data.frame() for both R versions 4.1.3 and 4.2.0 from Nixpkgs

Here follows an example a Nix-to-Nix solution, with two subshells to track the evolution of base R in this specific case. We can verify the breaking changes in case study 1 in more declarative manner when we use both R 4.1.3 and R 4.2.0 from Nixpkgs. Since we already have defined R 4.1.3 in the env_1_R-4-1-3 subshell, we can use it as a source environment where with_nix() is launched from. Accordingly, we define the R 4.2.0 environment in a env_1_2_R-4-2-0using Nix via rix::rix(). The latter environment will be the target environment where df_as_vector() will be evaluated in.

library("rix")
path_env_1_2 <- file.path(".", "_env_1_2_R-4-2-0")

rix(
  r_ver = "4.2.0",
  overwrite = TRUE,
  project_path = path_env_1_2,
  shell_hook = "R"
)

list.files(path_env_1_2)
#> [1] "default.nix"

Now, initiate a new R session as development environment using nix-shell. Open a new terminal at the current working directory of your R session. The provided expression default.nix. defines R 4.1.3 in a “subfolder per subshell” approach. nix-shell will use the expression by default.nix and prefer it over any other .nix files, except when you put a shell.nix file in that folder, which takes precedence.

nix-shell --pure ./_env_1_R-4-1-3

After some time downloading caches and doing builds, you will enter an R console session with R 4.1.3. You did not need to type in R first, because we set up a R shell hook via rix::rix(). Next, we define again the target function to test in R 4.2.0, too.

# current Nix-R session with R 4.1.3
df_as_vector <- function(x) {
  out <- as.vector(x = x, mode = "list")
  return(out)
}
(out_nix_1 <- df_as_vector(x = df))
out_nix_1_2 <- with_nix(
  expr = function() df_as_vector(x = df),
  program = "R",
  project_path = path_env_1_2,
  message_type = "simple" # you can do `"verbose"`, too
)

You can now formally compare the outputs of the computation of the same code in R 4.1.3 vs. R 4.2.0 environments controlled by Nix.

identical(out_nix_1, out_nix_1_2)
# yields FALSE

Case study 2: Breaking changes in {stringr} 1.5.0

We add one more layer to the reproducibility of the R ecosystem. User libraries from CRAN or GitHub, one thing that makes R shine is the huge collection of software packages available from the community.

There was a change introduce in {stringr} 1.5.0; in earlier versions, this line of code:

stringr::str_subset(c("", "a"), "")

would return the character "a". However, this behaviour is unexpected: it really should return an error. This was addressed in versions after 1.5.0:

out_system_stringr <- tryCatch(
  expr = {
    stringr::str_subset(c("", "a"), "")
  },
  error = function(e) NULL
)

Since the code returns an error, we wrap it inside tryCatch() and return NULL instead of an error (if we wouldn’t do that, this vignette could not compile!).

Let’s build a subshell with the latest version of R, but an older version of {stringr}:

library("rix")

path_env_stringr <- file.path(".", "_env_stringr_1.4.1")

rix(
  r_ver = "4.3.1",
  r_pkgs = "stringr@1.4.1",
  overwrite = TRUE,
  project_path = path_env_stringr
)

list.files(path = path_env_stringr, all.files = TRUE)
#> [1] "."           ".."          ".Rprofile"   "default.nix"

We can now run the code in the subshell

out_nix_stringr <- with_nix(
  expr = function() stringr::str_subset(c("", "a"), ""),
  program = "R",
  project_path = path_env_stringr,
  message_type = "simple"
)

Here are the last few lines printed on screen:

==> `expr` succeeded!

### Finished code evaluation in `nix-shell` ###

* Evaluating `expr` in `nix-shell` returns:
[1] "a"

Not only do we see the result of evaluating the code in the subshell, we also have access to it: out_nix_stringr holds this result.

We can now compare the two: the result of the code running in our main session with the latest version of {stringr} and the result of the code running in the subshell with the old version of {stringr}:

identical(out_system_stringr, out_nix_stringr)

As expected, the result is FALSE.

Case study 3: Using a subshell to get hard to install dependencies

Nix subshells are quite useful in cases where you need to use a package that might be difficult to install, such as {arrow}, or other packages that must be compiled. Depending on your operating system you need to compile {arrow} from source, which can be a frustrating experience, especially if you only need it to load data and bring it down to a manageable size (using select() and filter() for instance). This use cases illustrates how to achieve this.

Let’s start by building a subshell that is based on a distinct revision of nixpkgs, for which we know that arrow compiles on both linux and macOS (darwin).

library("rix")

path_env_arrow <- file.path("env_arrow")

rix(
  r_ver = "4.1.1",
  r_pkgs = c("dplyr", "arrow"),
  overwrite = TRUE,
  project_path = path_env_arrow
)

This specific revision of R contains {arrow} 13. Let’s now suppose that you already have a script with some code to load and transform some data using {arrow}. It may look something like this:

library(arrow)
library(dplyr)

arrow_cars <- arrow_table(cars)

arrow_cars %>%
  filter(speed > 10) %>%
  as.data.frame()

To run this code in a subshell, we recommend wrapping it inside a function:

arrow_script <- function() {
  library(arrow)
  library(dplyr)

  arrow_cars <- arrow_table(cars)

  arrow_cars %>%
    filter(speed > 10) %>%
    as.data.frame()
}

Which we can then run in the subshell:

out_nix_arrow <- with_nix(
  expr = function() arrow_script(),
  program = "R",
  project_path = path_env_arrow,
  message_type = "simple"
)

This will run the function in the subshell, and its output will be saved in the out_nix_arrow variable, for further manipulation in your main shell/session.