Scripting with Nix

NPM was originally a rudimentary package manager for Javascript and node.js, but these days it seems to be getting used as a general way of fetching dependencies for things like shell scripts.

I’ve not played with node.js since the very early days, and never really used NPM. However, I do see the need for modular, composable shell scripts.

Nix for Scripting

Personally, I’ve been using Nix in a similar way, since it also has nice features like caching, laziness, splicing into indented strings, dependency management, etc. For example, if you have a Nix expression stored in my-script.nix you can use the following (e.g. in my-script.sh) to invoke it:

nix-instantiate --read-write-mode --show-trace --eval -E 'import ./my-script.nix'

The --eval tells Nix to evaluate an expression, rather than build a package. -E is the expression to evaluate (in this case, importing our “script” file). --read-write-mode allows the script to add things to the Nix store (which is normally read-only). --show-trace will help us to debug, by showing a backtrace if a runtime error occurs.

Note that Nix makes a distinction between “evaluation time” and “build time” (similar to the compile/run time distinction in many other languages). The above command will evaluate the contents of my-script.nix, which should be a Nix expression, but most of the time that’s not our actual script (after all, Nix is a pure language, and hence unsuitable for many scripting applications).

Instead, we usually use Nix expressions to compose a “derivation” (basically a package), into which we can put all sorts of scripts.

If my-script.nix evaluates to a derivation, we can use this command instead of the above, to build it:

nix-build --show-trace -E 'import ./my-script.nix'

This creates a symlink called result, so if our derivation is a script we can then do ./result to run it.

A cleaner way would be to use the --no-out-link option to prevent cluttering up the working directory with symlinks, capture the stdout using e.g. SCRIPT=$(...), and running the resulting string using "$SCRIPT" (since it will contain the Nix store path to the result).

By the way, if you want to evaluate Nix expressions interactively, the nix-repl command is much easier to use than nix-instantiate (although it doesn’t seem to offer a --show-trace option).

Also note that we can use nix-shell shebangs to fetch dependencies of a script at run time; although I find that separate Nix files are usually needed to work around the limitations of writing expressions in a shebang line, and it can often be quite slow to have nix-shell invoked on each run, when we could instead build a script with its dependencies once, and invoke them as many times as we like with little overhead.

Strings

The reason I like Nix’s treatment of strings is that we can write “indented strings”, which means that long strings (such as scripts) can be embedded inside other expressions quite naturally, without getting filled with whitespace. For example, here’s a Nix expression:

runCommand "foo"
           { buildInputs = [ python imagemagick ]; }
           ''
             I am an indented string
             I will be executed as a bash script, with the following
             dependencies available:
              - python
              - imagemagick

             Since these lines, and those sentences above beginning with "I",
             have the least indentation, they will appear flush to the left in
             the resulting file. The "list" above will hence be indented by 1
             space, rather than the 14 spaces which appear here.
           ''

Secondly, splicing allows Nix expressions to be embedded inside strings. A splice begins with ${ and ends with }. The expression should either evaluate to a string, which is inserted as-is, or a “derivation” (e.g. a package), which gets “instantiated” (i.e. installed) and its installation directory is inserted into the resulting string. Splices can be nested too.

For example, instead of giving python as a dependency in the buildInputs, we could splice the full path into a string, e.g.

''
  "${python}/bin/python" my_script.py
''

Although this is probably a bad idea, since there may be transitive dependencies, etc. missing when the script gets executed.

Splices aren’t just a simple hack for resolving variable names; they drop us into a complete Nix expression context, where we can write arbitrarily complicated expressions, including other strings, containing other splices, etc.

Decomposing

If we want to build up a result incrementally, with each step getting cached, we can use runCommand, and write the results to “$out”. For example:

with import <nixpkgs> {};
with builtins;
with rec {

  # Takes a script and runs it with jq available (Nix functions are curried)
  runJq = runCommand "jq-cmd" { buildInputs = [ jq ]; };

  step1 = runJq ''
    echo "I am step 1" 1>&2
    echo '[{"name": "foo"}, {"name": "bar"}]' | jq 'map(.name)' > "$out"
  '';
  step2 = runJq ''
    echo "I am step 2" 1>&2
    I won't be executed, because Nix is lazy and nothing calls me
  '';
  step3 = runJq ''
    echo "I am step 3" 1>&2
    jq 'length' < "${step1}" > "$out"
  '';
};

import step3

When evaluated, this gives the following:

$ ./go.sh
building path(s) ‘/nix/store/5ks08zbvmgzbhg9kr0k4g75nf2ymsqsr-jq-cmd’
I am step 1
building path(s) ‘/nix/store/v1svcqq6cmi4xc9650qz9w2x177w4pfr-jq-cmd’
I am step 3
2
$ ./go.sh
2

The results are cached, and will be re-used as long as the commands aren’t edited, and their dependencies don’t change (e.g. if we update nixpkgs and a newer version of jq is available, they’ll be re-run with that version).

In this case, each “step” represents the data, which is common in lazy languages. Alternatively, we can use writeScript to write more ‘traditional’ process-oriented scripts:

with import <nixpkgs> {};
with builtins;
with rec {

  # Takes a script and runs it with jq available (Nix functions are curried)
  runJq = runCommand "jq-cmd" { buildInputs = [ jq ]; };

  step1 = writeScript "step-1" ''
    echo "I am step 1" 1>&2
    echo '[{"name": "foo"}, {"name": "bar"}]' | jq 'map(.name)'
  '';

  step2 = writeScript "step-2" ''
    echo "I am step 2" 1>&2
    I won't be executed, because Nix is lazy and nothing calls me
  '';

  step3 = writeScript "step-3" ''
    echo "I am step 3" 1>&2
    "${step1}" | jq 'length'
  '';
};

import (runJq ''"${step3}" > "$out"'')

Of course, we need something to invoke these scripts, which is why I used runJq in the final expression. When run, we get:

$ ./go.sh
building path(s) ‘/nix/store/fnw68cmkib5fkmhls4fkdhx0vb2cyka8-step-1’
building path(s) ‘/nix/store/1kiwa6m11d0apxfjbwpqq3vl6jbv3sdx-step-3’
building path(s) ‘/nix/store/9hv1jcrglyx8x6xa64pnds6vzcp35zl5-jq-cmd’
I am step 3
I am step 1
2
$ ./go.sh
2

This time the scripts are cached, but we execute them both together in a normal pipe. The overall result of the “runJq” call is still cached though. This is how you’d run non-bash scripts too: by using writeFile to save your code to disk, and runCommand to invoke it with a bash one-liner. For example, if we want step4 to use Haskell we might do the following:

  runJq = runCommand "jq-cmd" { buildInputs = [ jq ghc ]; };

  ...

  hsScript = script: writeScript "hs-cmd" ''
    runhaskell "${writeScript "hs-script" script}"
  '';

  ...

  step4 = hsScript ''
    doTimes :: (Show a) => a -> String -> String
    doTimes str n = show (replicate (read n) str)

    hello = "hello world"

    main = interact (doTimes hello)
  '';
};

import (runJq ''"${step3}" | "${step4}" > "$out"'')

This reads the length given by jq, and writes out a list of that many “hello world”s:

$ ./go.sh
building path(s) ‘/nix/store/2d7wrd78dk1ilj84adnyq8ddgzy6m2rr-hs-cmd’
building path(s) ‘/nix/store/haqcwssfbzbj5s4ampv322qbpll1gw1h-jq-cmd’
I am step 3
I am step 1
"[\"hello world\",\"hello world\"]"

Unfortunately, this can end up separating the code from its dependencies, i.e. we needed to give ghc as a dependency to whichever script invokes step4 (via runJq), rather than being able to add it in hsScript. If we used the original data-oriented approach, this wouldn’t be an issue.

To mitigate this, my Nix config provides a wrap function (documented here) which lets us write scripts which “bake in” their dependencies:

myScript = wrap {
  name   = "my-script";
  paths  = [ bash ghc jq python ];
  vars   = { someEnvVar = "Arbitrary content"; };
  script = ''
    #!/usr/bin/env bash
    ...do something here with "$someEnvVar" set, and a PATH containing bash,
    ghc, jq and python...
  '';
};

Since I often use wrap to make scripts which I then expose via bin/, I’ve also made a function mkBin which takes the same arguments as wrap, but instead of just building a script, it will put that script into a bin/ subdirectory of its output, with the filename taken from name (e.g. bin/my-script in this case).

Note that we can also write our scripts in standalone files, and reference them from Nix by giving the path to the file, e.g. ./myPythonScript.py. Since these are just regular files, we can use syntax highlighting, etc. when editing them. We also avoid having to care about escaping, but note that we can’t use string interpolation this way.

We also end up exposing the “raw” script (i.e. without its dependencies). If you want to target non-Nix users, that can be useful; otherwise, you might want to make it clear that it’s an internal detail which shouldn’t be run as-is (e.g. you could put it in a subdirectory called internal or src or something).

I/O

It’s also pretty easy to transfer data between the Nix language and the processes we’re invoking, using readFile, readDir, fromJSON and toJSON inside a splice; although Nix doesn’t support floats, so you might need to turn them into strings first. This is useful for doing tricky transformations on small amounts of data, which may be error-prone in bash, but where invoking a full-blown language like Haskell or Python would be overkill. It can also be useful for things like assertions.

Note that my preferred way of accessing the result of a derivation from within Nix is to use import rather than readFile. One advantage is that we’re guaranteed that Nix will build the result as needed, rather than saying ‘no such file’ (e.g. if we’ren ot in read/write mode).

Another is that Nix will not propagate dependencies of that derivation into the value that it reads; for example if we have a derivation foo which depends on Python and results in a file containing the string "hello", then import foo will give the string hello without tracking the Python dependency. This can be important if we have a fast-moving dependency which might cause lots of rebuilding, but doesn’t actually affect the contents of its output: we can generate some intermediate value using this dependency, which will get rebuilt a lot, but by importing that value for use in the subsequent steps, those steps won’t need to be rebuilt as long as that intermediate value doesn’t change.

Of course, the tricky part is writing data into Nix files such that they can be imported in the first place. Nix’s syntax is quite simple, but it can sometimes help to use fromJSON, etc. to make life easier.