This post is part of the R x Nix Nexus: Crafting Cohesive Environments series.

This post describes how to set up a transparent automated setup for reproducible R workflows using nixpkgs, niv, and lorri. The explanatory example used throughout the post is one of setting up the rethinking package and running some examples from the excellent second edition of “Statistical Rethinking” by Richard McElreath.

Background

As detailed in an earlier post1, I had set up Nix to work with non-CRAN packages. If the rest of this section is unclear, please refer back to the earlier post.

Setup

For the remainder of the post, we will set up a basic project structure:

1mkdir tryRnix/

Now we will create a shell.nix as2:

 1# shell.nix
 2{ pkgs ? import <nixpkgs> { } }:
 3with pkgs;
 4let
 5  my-r-pkgs = rWrapper.override {
 6    packages = with rPackages; [
 7      ggplot2
 8      tidyverse
 9      tidybayes
10      tidybayes.rethinking
11      (buildRPackage {
12        name = "rethinking";
13        src = fetchFromGitHub {
14          owner = "rmcelreath";
15          repo = "rethinking";
16          rev = "d0978c7f8b6329b94efa2014658d750ae12b1fa2";
17          sha256 = "1qip6x3f6j9lmcmck6sjrj50a5azqfl6rfhp4fdj7ddabpb8n0z0";
18        };
19        propagatedBuildInputs = [ coda MASS mvtnorm loo shape rstan dagitty ];
20      })
21    ];
22  };
23in mkShell {
24  buildInputs = with pkgs; [ git glibcLocales openssl which openssh curl wget ];
25  inputsFrom = [ my-r-pkgs ];
26  shellHook = ''
27    mkdir -p "$(pwd)/_libs"
28    export R_LIBS_USER="$(pwd)/_libs"
29  '';
30  GIT_SSL_CAINFO = "${cacert}/etc/ssl/certs/ca-bundle.crt";
31  LOCALE_ARCHIVE = stdenv.lib.optionalString stdenv.isLinux
32    "${glibcLocales}/lib/locale/locale-archive";
33}

So we have:

1tree tryRnix
tryRnix
└──shell.nix
0directories,1file

Introspection

At this point:

  • I was able to install packages (system and R) arbitrarily
  • I was able to use project specific folders
  • Unlike npm, pipenv, poetry, conda and friends, my system was not bloated by downloading and setting up the same packages every-time I used them in different projects

However, though this is a major step up from being chained to RStudio and my system package manager, it is still perhaps not immediately obvious how this workflow is reproducible. Admittedly, I have defined my packages in a nice functional manner; but someone else might have a different upstream channel they are tracking, and thus will have different packages. Indeed the only packages which I could be sure of were the R packages I built from Github, since those were tied to a hash. Finally, the setup described for each project is pretty onerous, and it is not immediately clear how to leverage fantastic tools like direnv for working through this.

Towards Reproducible Environments

The astute reader will have noticed that I mentioned that the R packages were reproducible since they were tied to a hash, and might reasonable argue that the entire Nix ecosystem is about hashing in the first place. Once we realize that, the rest is relatively simple3.

Niv and Pinning

Niv essentially keeps track of the channel from which all the packages are installed. Setup is pretty minimal.

1cd tryRnix/
2nix-env -i niv
3niv init

At this point, we have:

1tree tryRnix
tryRnix
├──nix
│  ├──sources.json
│  └──sources.nix
└──shell.nix
1directory,3files

We will have to update our shell.nix to use the new sources.

 1let
 2  sources = import ./nix/sources.nix;
 3  pkgs = import sources.nixpkgs { };
 4  stdenv = pkgs.stdenv;
 5  my-r-pkgs = pkgs.rWrapper.override {
 6    packages = with pkgs.rPackages; [
 7      ggplot2
 8      tidyverse
 9      tidybayes
10    ];
11  };
12in pkgs.mkShell {
13  buildInputs = with pkgs;[ git glibcLocales openssl which openssh curl wget my-r-pkgs ];
14  shellHook = ''
15    mkdir -p "$(pwd)/_libs"
16    export R_LIBS_USER="$(pwd)/_libs"
17  '';
18  GIT_SSL_CAINFO = "${pkgs.cacert}/etc/ssl/certs/ca-bundle.crt";
19  LOCALE_ARCHIVE = stdenv.lib.optionalString stdenv.isLinux
20    "${pkgs.glibcLocales}/lib/locale/locale-archive";
21}

We could inspect and edit these sources by hand, but it is much more convenient to simply use niv again when we need to update these.

1cd tryRnix/
2niv update nixpkgs -b nixpkgs-unstable

At this stage we have a reproducible set of packages ready to use. However it is still pretty annoying to have to go through the trouble of writing nix-shell and also waiting while it rebuilds when we change things.

Lorri and Direnv

In the past, I have made my admiration for direnv very clear (especially for python-poetry). However, though direnv does allow us to include arbitrary bash logic into our projects, it would be nice to have something which has some defaults for nix. Thankfully, the folks at TweagIO developed lorri to scratch that itch.

The basic setup is simple:

1nix-env -i lorri
2cd tryRnix/
3lorri init

1tree -a tryRnix/
tryRnix/
├──.envrc
├──nix
│  ├──sources.json
│  └──sources.nix
└──shell.nix
1directory,4files

We can and should inspect the environment lorri wants us to load with direnv file:

1cat tryRnix/.envrc

1$(lorri direnv)

In and of itself that is not too descriptive, so we should run that on our own first.

  1EVALUATION_ROOT="$HOME/.cache/lorri/gc_roots/407bd4df60fbda6e3a656c39f81c03c2/gc_root/shell_gc_root"
  2
  3watch_file "/run/user/1000/lorri/daemon.socket"
  4watch_file "$EVALUATION_ROOT"
  5
  6#!/usr/bin/env bash
  7# ^ shebang is unused as this file is sourced, but present for editor
  8# integration. Note: Direnv guarantees it *will* be parsed using bash.
  9
 10function punt () {
 11    :
 12}
 13
 14# move "origPreHook" "preHook" "$@";;
 15move() {
 16    srcvarname=$1 # example: varname might contain the string "origPATH"
 17    # drop off the source variable name
 18    shift
 19
 20    destvarname=$1 # example: destvarname might contain the string "PATH"
 21    # drop off the destination variable name
 22    shift
 23
 24    # like: export origPATH="...some-value..."
 25    export "${@?}";
 26
 27    # set $original to the contents of the variable $srcvarname
 28    # refers to
 29    eval "$destvarname=\"${!srcvarname}\""
 30
 31    # mark the destvarname as exported so direnv picks it up
 32    # (shellcheck: we do want to export the content of destvarname!)
 33    # shellcheck disable=SC2163
 34    export "$destvarname"
 35
 36    # remove the export from above, ie: export origPATH...
 37    unset "$srcvarname"
 38}
 39
 40function prepend() {
 41    varname=$1 # example: varname might contain the string "PATH"
 42
 43    # drop off the varname
 44    shift
 45
 46    separator=$1 # example: separator would usually be the string ":"
 47
 48    # drop off the separator argument, so the remaining arguments
 49    # are the arguments to export
 50    shift
 51
 52    # set $original to the contents of the the variable $varname
 53    # refers to
 54    original="${!varname}"
 55
 56    # effectfully accept the new variable's contents
 57    export "${@?}";
 58
 59    # re-set $varname's variable to the contents of varname's
 60    # reference, plus the current (updated on the export) contents.
 61    # however, exclude the ${separator} unless ${original} starts
 62    # with a value
 63    eval "$varname=${!varname}${original:+${separator}${original}}"
 64}
 65
 66function append() {
 67    varname=$1 # example: varname might contain the string "PATH"
 68
 69    # drop off the varname
 70    shift
 71
 72    separator=$1 # example: separator would usually be the string ":"
 73    # drop off the separator argument, so the remaining arguments
 74    # are the arguments to export
 75    shift
 76
 77
 78    # set $original to the contents of the the variable $varname
 79    # refers to
 80    original="${!varname:-}"
 81
 82    # effectfully accept the new variable's contents
 83    export "${@?}";
 84
 85    # re-set $varname's variable to the contents of varname's
 86    # reference, plus the current (updated on the export) contents.
 87    # however, exclude the ${separator} unless ${original} starts
 88    # with a value
 89    eval "$varname=${original:+${original}${separator}}${!varname}"
 90}
 91
 92varmap() {
 93    if [ -f "$EVALUATION_ROOT/varmap-v1" ]; then
 94        # Capture the name of the variable being set
 95        IFS="=" read -r -a cur_varname <<< "$1"
 96
 97        # With IFS='' and the `read` delimiter being '', we achieve
 98        # splitting on \0 bytes while also preserving leading
 99        # whitespace:
100        #
101        #    bash-3.2$ printf ' <- leading space\0bar\0baz\0' \
102        #                  | (while IFS='' read -d $'\0' -r x; do echo ">$x<"; done)
103        #    > <- leading space<
104        #    >bar<
105        #    >baz<```
106        while IFS='' read -r -d '' map_instruction \
107           && IFS='' read -r -d '' map_variable \
108           && IFS='' read -r -d '' map_separator; do
109            unset IFS
110
111            if [ "$map_variable" == "${cur_varname[0]}" ]; then
112                if [ "$map_instruction" == "append" ]; then
113                    append "$map_variable" "$map_separator" "$@"
114                    return
115                fi
116            fi
117        done < "$EVALUATION_ROOT/varmap-v1"
118    fi
119
120
121    export "${@?}"
122}
123
124function declare() {
125    if [ "$1" == "-x" ]; then shift; fi
126
127    # Some variables require special handling.
128    #
129    # - punt:    don't set the variable at all
130    # - prepend: take the new value, and put it before the current value.
131    case "$1" in
132        # vars from: https://github.com/NixOS/nix/blob/92d08c02c84be34ec0df56ed718526c382845d1a/src/nix-build/nix-build.cc#L100
133        "HOME="*) punt;;
134        "USER="*) punt;;
135        "LOGNAME="*) punt;;
136        "DISPLAY="*) punt;;
137        "PATH="*) prepend "PATH" ":" "$@";;
138        "TERM="*) punt;;
139        "IN_NIX_SHELL="*) punt;;
140        "TZ="*) punt;;
141        "PAGER="*) punt;;
142        "NIX_BUILD_SHELL="*) punt;;
143        "SHLVL="*) punt;;
144
145        # vars from: https://github.com/NixOS/nix/blob/92d08c02c84be34ec0df56ed718526c382845d1a/src/nix-build/nix-build.cc#L385
146        "TEMPDIR="*) punt;;
147        "TMPDIR="*) punt;;
148        "TEMP="*) punt;;
149        "TMP="*) punt;;
150
151        # vars from: https://github.com/NixOS/nix/blob/92d08c02c84be34ec0df56ed718526c382845d1a/src/nix-build/nix-build.cc#L421
152        "NIX_ENFORCE_PURITY="*) punt;;
153
154        # vars from: https://www.gnu.org/software/bash/manual/html_node/Bash-Variables.html (last checked: 2019-09-26)
155        # reported in https://github.com/target/lorri/issues/153
156        "OLDPWD="*) punt;;
157        "PWD="*) punt;;
158        "SHELL="*) punt;;
159
160        # https://github.com/target/lorri/issues/97
161        "preHook="*) punt;;
162        "origPreHook="*) move "origPreHook" "preHook" "$@";;
163
164        *) varmap "$@" ;;
165    esac
166}
167
168export IN_NIX_SHELL=impure
169
170if [ -f "$EVALUATION_ROOT/bash-export" ]; then
171    # shellcheck disable=SC1090
172    . "$EVALUATION_ROOT/bash-export"
173elif [ -f "$EVALUATION_ROOT" ]; then
174    # shellcheck disable=SC1090
175    . "$EVALUATION_ROOT"
176fi
177
178unset declare
179
180Jun 06 19:02:32.368 INFO lorri has not completed an evaluation for this project yet, expr: $HOME/Git/Github/WebDev/Mine/haozeke.github.io/content-org/tryRnix/shell.nix
181Jun 06 19:02:32.368 WARN `lorri direnv` should be executed by direnv from within an `.envrc` file, expr: $HOME/Git/Github/WebDev/Mine/haozeke.github.io/content-org/tryRnix/shell.nix

Upon inspection, that seems to check out. So now we can enable this.

1direnv allow

Additionally, we will need to stick to using a pure environment as much as possible to prevent unexpected situations. So we set:

1# .envrc
2eval "$(lorri direnv)"
3nix-shell --run bash --pure

There’s still a catch though. We need to have lorri daemon running to make sure the packages are built automatically without us having to exit the shell and re-run things. We can turn to the documentation for this. Essentially, we need to have a user-level systemd socket file and service for lorri.

 1# ~/.config/systemd/user/lorri.socket
 2[Unit]
 3Description=Socket for Lorri Daemon
 4
 5[Socket]
 6ListenStream=%t/lorri/daemon.socket
 7RuntimeDirectory=lorri
 8
 9[Install]
10WantedBy=sockets.target

 1# ~/.config/systemd/user/lorri.service
 2[Unit]
 3Description=Lorri Daemon
 4Requires=lorri.socket
 5After=lorri.socket
 6
 7[Service]
 8ExecStart=%h/.nix-profile/bin/lorri daemon
 9PrivateTmp=true
10ProtectSystem=strict
11ProtectHome=read-only
12Restart=on-failure

With that we are finally ready to start working with our auto-managed, reproducible environments.

1systemctl --user daemon-reload && \
2systemctl --user enable --now lorri.socket

Rethinking

As promised, we will first test the setup to see that everything is working. Now is also a good time to try the tidybayes.rethinking package. In order to use it, we will need to define the rethinking package in a way so we can pass it to the buildInputs for tidybayes.rethinking. We will modify new shell.nix as follows:

 1# shell.nix
 2let
 3  sources = import ./nix/sources.nix;
 4  pkgs = import sources.nixpkgs { };
 5  stdenv = pkgs.stdenv;
 6  rethinking = with pkgs.rPackages;
 7    buildRPackage {
 8      name = "rethinking";
 9      src = pkgs.fetchFromGitHub {
10        owner = "rmcelreath";
11        repo = "rethinking";
12        rev = "d0978c7f8b6329b94efa2014658d750ae12b1fa2";
13        sha256 = "1qip6x3f6j9lmcmck6sjrj50a5azqfl6rfhp4fdj7ddabpb8n0z0";
14      };
15      propagatedBuildInputs = [ coda MASS mvtnorm loo shape rstan dagitty ];
16    };
17  tidybayes_rethinking = with pkgs.rPackages;
18    buildRPackage {
19      name = "tidybayes.rethinking";
20      src = pkgs.fetchFromGitHub {
21        owner = "mjskay";
22        repo = "tidybayes.rethinking";
23        rev = "df903c88f4f4320795a47c616eef24a690b433a4";
24        sha256 = "1jl3189zdddmwm07z1mk58hcahirqrwx211ms0i1rzbx5y4zak0c";
25      };
26      propagatedBuildInputs =
27        [ dplyr tibble rlang MASS tidybayes rethinking rstan ];
28    };
29  rEnv = pkgs.rWrapper.override {
30    packages = with pkgs.rPackages; [
31      ggplot2
32      tidyverse
33      tidybayes
34      devtools
35      modelr
36      cowplot
37      ggrepel
38      RColorBrewer
39      purrr
40      forcats
41      rstan
42      rethinking
43      tidybayes_rethinking
44    ];
45  };
46in pkgs.mkShell {
47  buildInputs = with pkgs; [ git glibcLocales which ];
48  inputsFrom = [ rEnv ];
49  shellHook = ''
50    mkdir -p "$(pwd)/_libs"
51    export R_LIBS_USER="$(pwd)/_libs"
52  '';
53  GIT_SSL_CAINFO = "${pkgs.cacert}/etc/ssl/certs/ca-bundle.crt";
54  LOCALE_ARCHIVE = stdenv.lib.optionalString stdenv.isLinux
55    "${pkgs.glibcLocales}/lib/locale/locale-archive";
56}

The main thing to note here is that we need the output of the derivation we create here, i.e. we need to use inputsFrom and NOT buildInputs for rEnv.

Let us try to get a nice graphic for the conclusion.

 1library(magrittr)
 2library(dplyr)
 3library(purrr)
 4library(forcats)
 5library(tidyr)
 6library(modelr)
 7library(tidybayes)
 8library(tidybayes.rethinking)
 9library(ggplot2)
10library(cowplot)
11library(rstan)
12library(rethinking)
13library(ggrepel)
14library(RColorBrewer)
15
16theme_set(theme_tidybayes())
17rstan_options(auto_write = TRUE)
18options(mc.cores = parallel::detectCores())
19
20
21set.seed(5)
22n = 10
23n_condition = 5
24ABC =
25  tibble(
26    condition = factor(rep(c("A","B","C","D","E"), n)),
27    response = rnorm(n * 5, c(0,1,2,1,-1), 0.5)
28  )
29
30mtcars_clean = mtcars %>%
31  mutate(cyl = factor(cyl))
32
33m_cyl = ulam(alist(
34    cyl ~ dordlogit(phi, cutpoint),
35    phi <- b_mpg*mpg,
36    b_mpg ~ student_t(3, 0, 10),
37    cutpoint ~ student_t(3, 0, 10)
38  ),
39  data = mtcars_clean,
40  chains = 4,
41  cores = parallel::detectCores(),
42  iter = 2000
43)
44
45cutpoints = m_cyl %>%
46  recover_types(mtcars_clean) %>%
47  spread_draws(cutpoint[cyl])
48
49# define the last cutpoint
50last_cutpoint = tibble(
51  .draw = 1:max(cutpoints$.draw),
52  cyl = "8",
53  cutpoint = Inf
54)
55
56cutpoints = bind_rows(cutpoints, last_cutpoint) %>%
57  # define the previous cutpoint (cutpoint_{j-1})
58  group_by(.draw) %>%
59  arrange(cyl) %>%
60  mutate(prev_cutpoint = lag(cutpoint, default = -Inf))
61
62fitted_cyl_probs = mtcars_clean %>%
63  data_grid(mpg = seq_range(mpg, n = 101)) %>%
64  add_fitted_draws(m_cyl) %>%
65  inner_join(cutpoints, by = ".draw") %>%
66  mutate(`P(cyl | mpg)` =
67    # this part is logit^-1(cutpoint_j - beta*x) - logit^-1(cutpoint_{j-1} - beta*x)
68    plogis(cutpoint - .value) - plogis(prev_cutpoint - .value)
69  )
70
71
72data_plot = mtcars_clean %>%
73  ggplot(aes(x = mpg, y = cyl, color = cyl)) +
74  geom_point() +
75  scale_color_brewer(palette = "Dark2", name = "cyl")
76
77fit_plot = fitted_cyl_probs %>%
78  ggplot(aes(x = mpg, y = `P(cyl | mpg)`, color = cyl)) +
79  stat_lineribbon(aes(fill = cyl), alpha = 1/5) +
80  scale_color_brewer(palette = "Dark2") +
81  scale_fill_brewer(palette = "Dark2")
82
83png(filename="../images/rethinking.png")
84plot_grid(ncol = 1, align = "v",
85  data_plot,
86  fit_plot
87)
88dev.off

Finally we will run this in our environment.

1Rscript tesPlot.R

Conclusions

This post was really more of an exploratory follow up to the previous post, and does not really work in isolation. Then again, at this point everything seems to have worked out well. R with Nix has finally become a truly viable combination for any and every analysis under the sun. Some parts of the workflow are still a bit janky, but will probably resolve themselves over time.

Update: There is a final part detailing automated ways of reloading the system configuration

  1. My motivations were laid out in the aforementioned post, and will not be repeated ↩︎

  2. For why these are the way they are see the this is written, see the aforementioned post ↩︎

  3. Christine Dodrill has a great write up on using these tools as well ↩︎

Series info

R x Nix Nexus: Crafting Cohesive Environments series

  1. Nix with R and devtools
  2. Statistical Rethinking and Nix <-- You are here!
  3. Emacs for Nix-R