Storage optimization: Difference between revisions

Revision as of 01:03, 22 May 2021

A recurring problem with NixOS is lack of space on /. Even if you only ocasionally use Nix, it is easy for /nix/store to grow beyond 50GiB. Here are generic notes on how to not run out of space too often.

Optimizing the store

The option and command below save space by hardlinking store files:

Automatically

nix.autoOptimiseStore = true;

But this option only applies to new files: you'll still need to manually optimise your store once, after you enable this option.

Manually

Run nix-store --optimise. This will take some time to complete.

Garbage collection

The Nix store sometimes contains entries which are no longer useful.^{[cf. 1]} They can be deleted with nix-collect-garbage -d ^{[cf. 2]} or nix-store --gc.^{[cf. 3]}

Note that if a result file still exists in the file system, and your Nix configuration has both keep-outputs = true and keep-derivations = true, all the dependencies used to build it will be kept. To see which result files prevent garbage collection, run:

$ nix-store --gc --print-roots
/home/danbst/dev/test-shell/.shell.drv -> /nix/store/4diqwczyjipdqyi7aj34wfagblbhfjr9-nixops-1.4
/home/danbst/dev/test-shell/.shell.drv-2 -> /nix/store/62h3c4d6rdnlxichixqg8h9jxi8nhxk0-stdenv
/home/danbst/dev/test-shell/.shell.drv-2-doc -> /nix/store/14gnv1q1w0n9qwa3q23idsqvn51354y8-bash-4.3-p42-doc
/home/danbst/stack/new/website/server/result -> /nix/store/1jhmp6vl364p32r8bjigk65qh1xa562f-server-0.1.0.0
/home/danbst/testing/.nix-gc-roots/shell.drv -> /nix/store/v3vqf48awjjzjivrx15kfqdh1d7cg4mq-sshpass-1.05
...
/home/danbst/testing/.nix-gc-roots/shell.drv-12 -> /nix/store/a2li4sl9pxh9aflqia2gp7w88ayvjwci-bash-4.3-p42
/home/danbst/testing/.nix-gc-roots/shell.drv-12-doc -> /nix/store/kcswyb1d8zimkym0pjfi2fj1dly1w34w-bash-4.3-p42-doc
/home/danbst/testing/.nix-gc-roots/shell.drv-12-info -> /nix/store/njb817fwiafswzwvj9skw7w7k6b3fnbi-bash-4.3-p42-info
/home/ec2-user/result -> /nix/store/q35aq2sh5dbyka6g6f6qb7b8msxwds5m-nixos-system-iron-16.03.1299.a8e0739
/nix/var/nix/profiles/per-container/analyt/system-3-link -> /nix/store/snrj72189wh9va23fawl3v80v92xnxlm-nixos-system-iron-16.03.1291.efe2d64
/nix/var/nix/profiles/per-container/d-live/system-6-link -> /nix/store/cp2c58hnczsjk5h69ksajq5xfhsyhl6v-nixos-system-iron-16.03.1299.a8e0739
/nix/var/nix/profiles/per-container/d-test/system-4-link -> /nix/store/n1w7ywjg65x8iimchznxcyygbgmyfh55-nixos-system-iron-16.03.1287.6ac7ffd
/nix/var/nix/profiles/per-container/dashboard/system-41-link -> /nix/store/7qk19pkwgq0h3a1q9dcql3nks40rr75s-nixos-system-iron-16.03.1340.5a090dd
...
/nix/var/nix/profiles/per-container/ttt/system-1-link -> /nix/store/1kj9qs5gl3421jlkl3jfc2kqdsl8akwr-nixos-system-ttt-16.03.977.1da05df
/nix/var/nix/profiles/per-user/danbst/channels-1-link -> /nix/store/s0qay9qyqrn92zayldbvvj3zrfcl7a72-user-environment
/nix/var/nix/profiles/per-user/danbst/profile-28-link -> /nix/store/69ds606146dqml04sm0fbpqwnv2w8i3q-user-environment
/nix/var/nix/profiles/per-user/ec2-user/profile-7-link -> /nix/store/y2hc7zsnkzys9ba6xaijvjhff03rcgpy-user-environment
/nix/var/nix/profiles/per-user/root/channels-4-link -> /nix/store/254b6pkhhnjywvj5c0lp2vdai8nz4p0g-user-environment
/nix/var/nix/profiles/system-398-link -> /nix/store/wmndyzzrbc9fyjw844jmvzwgwgcinq7s-nixos-system-iron-16.0916.09pre.custom
/root/forkstat/result -> /nix/store/i5glmg3wk2a48x52rhd92zip1cmc0kq9-forkstat-git
/run/booted-system -> /nix/store/8jkrl9jyq7hqxb6xpwcaghpdm26gq98j-nixos-system-iron-16.0916.09pre.custom
/run/current-system -> /nix/store/wmndyzzrbc9fyjw844jmvzwgwgcinq7s-nixos-system-iron-16.0916.09pre.custom

GC roots can be found in /nix/var/nix/gcroots. The following script demonstrates how this directory can be used to (for example) query the state of manually made result symlinks:

find -H /nix/var/nix/gcroots/auto -type l | xargs -I {} sh -c 'readlink {}; realpath {}; echo'

This acts a simpler (but faster) version of --print-roots and could be implemented as a bash alias for convenience.

Run as root

nix-collect-garbage -d operates only for the current user. To clear system profiles, run it with root privileges.

Look for `result` symlinks

If you use nix-build, but not --no-build-output, your file system will be filled with result symlinks to various derivations. In the example above, note the following symlinks:

/home/danbst/stack/new/website/server/result -> /nix/store/1jhmp6vl364p32r8bjigk65qh1xa562f-server-0.1.0.0
/home/ec2-user/result -> /nix/store/q35aq2sh5dbyka6g6f6qb7b8msxwds5m-nixos-system-iron-16.03.1299.a8e0739
/root/forkstat/result -> /nix/store/i5glmg3wk2a48x52rhd92zip1cmc0kq9-forkstat-git

How much space do these (apparently) abandoned derivations use?

$ du -sch $(nix-store -qR /root/forkstat/result /home/ec2-user/result /home/danbst/stack/new/website/server/result)
...
3.4G    total

Not all of the derivations are garbage in this case, but quite a few are:

# rm /root/forkstat/result /home/ec2-user/result /home/danbst/stack/new/website/server/result
# nix-collect-garbage -d
...
690 store paths deleted, 1817.99 MiB freed

Look for system derivations in particular. Those are created on many occasions, for example when running nixos-rebuild build-vm

Reboot

As you see, the reference in /run/booted-system is a GC root, so it won't be cleared until reboot. If you don't want to reboot, just rm /run/booted-system that link and rerun sudo nix-collect-garbage.

Pinning

Running the following command:

$ nix-instantiate shell.nix --indirect --add-root ./.nix-gc-roots/shell.drv ...

Will create a persistent snapshot of your shell.nix dependencies, which then won't be garbage collected, as long as you have configured keep-outputs = true (and haven't changed the default of keep-derivations = true). This is useful if your project has a dependency with no substitutes available, or you don't want to spend time waiting to re-download your dependencies every time you enter the shell.

You need to re-run that nix-instantiate command any time your shell.nix changes.

And there is a subtle gotcha if your shell.nix happens to evaluate to more than one derivation: nix-instantiate will number each derivation sequentially, so if you change your shell.nix to contain fewer derivations, such that (for example) the name of the last GC root starts with shell.drv-7, then shell.drv-{8,9,10,11,12...} will be dangling and unused.

The easiest way to get around this is to delete the ./.nix-gc-roots directory periodically (i.e., any time you re-run the nix-instantiate command).

Don't forget to periodically check your GC roots, and remove any that you no longer need.

Automation

It is possible to enable periodic automatic GC,^{[cf. 4]} for example like this:

nix.gc = {
  automatic = true;
  dates = "weekly";
  options = "--delete-older-than 30d";
};

This can result in redownloads (tarballs fetched with import (builtins.fetchTarball ...) for example are not referenced anywhere and removed on GC), but it frees you from runnning GC manually.

It is also possible to automatically run garbage collection whenever there is not enough space left.^{[cf. 5]} For example, to free up to 1GiB whenever there is less than 100MiB left:

nix.extraOptions = ''
  min-free = ${toString (100 * 1024 * 1024)}
  max-free = ${toString (1024 * 1024 * 1024)}
'';

This is particularly useful when the store is on its own partition, see below.

Moving the store

/nix can reside on another device. This is useful if your root device is very small, and you have another, larger drive available.

If the second mountpoint is on the same device, some benefit can still be gained by formatting the partition it points to with a different file system. For example: on a Raspberry Pi, f2fs could possibly be used for a gain in I/O throughput.

Regardless of /nix's filesystem, it can also be mounted with noatime (as seen in the example below). This will reduce metadata writes, improving I/O and the device's lifespan.

This is easiest to set up while installing NixOS, but /nix can be moved on a live system:

Create a new partition and mount it over /mnt
rsync -aAxv everything from /nix to /mnt
Bind /mnt to /nix (e.g. using mount).
Restart nix daemon with systemctl restart nix-daemon.service systemctl restart nix-daemon.socket .

Rerun nixos-rebuild switch with something like

fileSystems."/nix" = { device = "/dev/disk/by-label/nix";  neededForBoot = true; options = [ "noatime" ]; };

Reboot to be sure /nix/store is properly mounted.
Once you are sure everything works, you can delete the old store doing a bind mount / to /old_root, and remove /old_root/nix.

Keep in mind that all commands like mount and bash point to some executable in /nix/store, so never mount an empty disk over /nix or /nix/store, otherwise you will be locked out until reboot!

@@ Line 94: / Line 94: @@
 === Pinning ===
-When you invoke <code>nix-shell</code> with:
+Running the following command:
 <syntaxhighlight lang="console">
-$ nix-instantiate shell.nix --indirect --add-root $DIR/.nix-gc-roots/shell.drv ...
+$ nix-instantiate shell.nix --indirect --add-root ./.nix-gc-roots/shell.drv ...
 </syntaxhighlight>
-Then you'll have a persistent environment which won't be garbage collected as long as you have configured <code>keep-outputs = true</code>. This is useful when you don't want to spend time waiting for redownloads every time you enter the shell.
+Will create a persistent snapshot of your <code>shell.nix</code> dependencies, which then won't be garbage collected, as long as you have configured <code>keep-outputs = true</code> (and haven't changed the default of <code>keep-derivations = true</code>). This is useful if your project has a dependency with no substitutes available, or you don't want to spend time waiting to re-download your dependencies every time you enter the shell.
-A little problem exists though. GC roots are numbered sequentially, so if you change <code>shell.nix</code> to contain '''fewer''' derivations, such that the name of the last GC root starts with <code>shell.drv-7</code>, then <code>shell.drv-{8,9,10,11,12}*</code> will be dangling and unused. To overcome this problem you should remove GC roots dir periodically (or just before running <code>nix-shell</code>).
+You need to re-run that <code>nix-instantiate</code> command any time your <code>shell.nix</code> changes.
-Obviously, you should remove the GC roots directory for projects you don't plan to work on.
+And there is a subtle gotcha if your <code>shell.nix</code> happens to evaluate to more than one derivation: <code>nix-instantiate</code> will number each derivation sequentially, so if you change your <code>shell.nix</code> to contain ''fewer'' derivations, such that (for example) the name of the last GC root starts with <code>shell.drv-7</code>, then <code>shell.drv-{8,9,10,11,12...}</code> will be dangling and unused.
+The easiest way to get around this is to delete the <code>./.nix-gc-roots</code> directory periodically (i.e., any time you re-run the <code>nix-instantiate</code> command).
+Don't forget to periodically check your GC roots, and remove any that you no longer need.
 === Automation ===