Distributed build

From NixOS Wiki
Jump to: navigation, search

Sometimes you want to use a faster machine for building a nix derivation you want to use on a slower one. If you have ssh access to a machine where Nix (not necessarily NixOS) is installed, you can offload building to this machine.

There is a dedicated chapter in the Nix Manual.

This is a step-by-step guide to setting up distributed builds, that includes some SSH tips that are out of scope for the Nix Manual chapter.

Prerequisites

SSH access to the remote builder must be set up to enable distribution of builds to it. The table below lists SSH access requirements for most scenarios.

Local Machine Remote Builder Requirements
NixOS / System-wide Nix installation NixOS / System-wide Nix installation Local machine's root user needs SSH access to a user on the remote machine.
Single-user Nix installation NixOS / System-wide Nix installation Your user on your Local machine needs SSH access to a user on the remote machine.
NixOS / System-wide Nix installation Single-user Nix installation Local machine's root user needs SSH access to the user on the remote machine, with Nix installed with their UID (see Nix manual page).
Single-user Nix installation NixOS / System-wide Nix installation Your user on your Local machine needs SSH access to the user on the remote machine, with Nix installed with their UID (see Nix manual page).

In any case, the definitive test for SSH access for Nix is:

nix store ping --store ssh://<REMOTE-BUILDER>

Where <REMOTE-BUILDER> is the remote builder's IP address, host address or whatever name you configure in ~/.ssh/config or /root/.ssh/config, including the user@ prefix.

An alternative check is:

ssh <REMOTE-BUILDER> 'type nix-store'

The following sections guide you how to setup such authentication, with security in mind, and maximal comfort, assuming basic knowledge about SSH authentication keys.

General best practices

It is recommended to not allow root access to the remote machine, even if only via an SSH public/private key pair. Especially because it's not required in any of the 4 scenarios described in the table above.

In all of the cases above it is recommended to create an SSH public / private key pair without a passphrase, so that you won't have to run ssh-add along with ssh-agent prior to using the remote builder. When the local machine has NixOS / System-wide installation of Nix you'd probably need to spawn eval $(ssh-agent) while you are logged in as root. Not using a passphrase for the SSH key allows other users to enjoy the remote builder.

Since the access to the remote machine doesn't have to be privileged, you can choose to login to a weakly privileged, and password-locked user on the remote machine, which may help you feel comfortable with the fact root can access it without a passphrase.

When the remote machine doesn't have NixOS / System-wide Nix installation, the only option is to allow access without passphrase and with an SSH key to the user with Nix installed for them.

Recommended System-wide Nix –> System-wide Nix setup

For the common case where your local Nix is installed system-wide, create a user on the remote machine that will have an unwriteable home directory, with a ~/.ssh/authorized_keys in it, that will allow SSH access to that user without a passphrase. The steps are:

- ssh to the remote builder. - Run (requires privileges) useradd -m -L nixremote where -L locks the user such that nobody will be able to su to it, and -m makes sure a home directory is created for the nixremote user. - Run (requires privileges) mkdir ~nixremote/.ssh.

If your remote builder has Nix installed system-wide, but without NixOS, you may need to add something like the following to your /etc/ssh/sshd_config:

Breeze-text-x-plain.png
/etc/ssh/sshd_config
SetEnv PATH=/nix/var/nix/profiles/default/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin


So that straight away when logging in, your $PATH will include nix' executables' directory.

Then, on your local machine, create the private / public key pair without a passphrase, as root:

# ssh-keygen -f /root/.ssh/nixremote

Copy the contents of /root/.ssh/nixremote.pub from your local machine to the remote builder ~nixremote/.ssh/authorized_keys.

Then to further harden the setup, remove write permissions from everyone on the remote host's nixremote home directory:

# chmod -R a-w ~nixremote

Now you want to make it easy for root on your local machine to connect to nixremote@builder. You can do that by creating the following /root/.ssh/config on the local machine:

Breeze-text-x-plain.png
/root/.ssh/config
Host builder # Replace by IP address, or add a ProxyCommand, see `man ssh_config` for full docs.
        # Prevent using ssh-agent or another keyfile, useful for testing
        IdentitiesOnly yes
        IdentityFile /root/.ssh/nixremote
        # The weakly privileged user on the remote builder – if not set, 'root' is used – which will hopefully fail
        User nixremote


You may also want to make nix on the remote machine trust that new user by adding it to nix.settings.trusted-users if it's using NixOS, or by manually adding trusted-users = nixremote to /etc/nix/nix.conf.

You may also want to tell Nix on your local machine that the remote builder is available, and tell it what are its supported features (see missing nix documentation issue). If your local machine uses NixOS, you can use the documented nix.buildMachines NixOS option.

Here's an example NixOS configuration that may be a useful inspiration:

Breeze-text-x-plain.png
/etc/nixos/configuration.nix
{ config, pkgs, ... }:

{
	nix.buildMachines = [ {
	 hostName = "builder";
	 system = "x86_64-linux";
                   protocol = "ssh-ng";
	 # if the builder supports building for multiple architectures, 
	 # replace the previous line by, e.g.
	 # systems = ["x86_64-linux" "aarch64-linux"];
	 maxJobs = 1;
	 speedFactor = 2;
	 supportedFeatures = [ "nixos-test" "benchmark" "big-parallel" "kvm" ];
	 mandatoryFeatures = [ ];
	}] ;
	nix.distributedBuilds = true;
	# optional, useful when the builder has a faster internet connection than yours
	nix.extraOptions = ''
	  builders-use-substitutes = true
	'';
}

Remote builders' features

Each builder is declared with a set of supportedFeatures. When a builder lacks one of the requiredSystemFeatures of a derivation, it will be ignored. Here are some features used in nixpkgs:

Feature Derivations requiring it
kvm Everything which builds inside a vm, like NixOS tests
nixos-test Machine can run NixOS tests
big-parallel kernel config, libreoffice, evolution, llvm and chromium
benchmark Machine can generate metrics (means the builds usually takes the same amount of time)

Non-standard Nix installations

If you are not root on the remote builder and have used nix-user-chroot or PRoot to install nix there (see Nix Installation Guide) then nix is not available on the PATH of the remote builder. We describe a solution for nix-user-chroot which is easily adapted to PRoot.

  • Create a script ~/bin/nix_wrapper.sh as follows:
#!/bin/sh
exec ~/bin/nix-user-chroot ~/.nix bash -c '
. ~/.nix-profile/etc/profile.d/nix.sh
exec $SSH_ORIGINAL_COMMAND
'

Of course, adapt this script to the location of the store and nix-user-chroot. Make the script executable.

  • In ~/.ssh/authorized_keys, locate the line corresponding to ~/.ssh/nixremote.pub and prepend this: command="/home/something/bin/nix_wrapper.sh".

Now ssh will transparently run nix-user-chroot when you connect to the remote builder with the specified ssh key.

Using remote builders

Local builder

Your local machine is still a builder, so when connecting to remote builders fails, nix will fall back to building locally. To never use the local machine, set the max-jobs nix option to 0:

$ nix-build -j0 blah

Using remote builders as substituters

If you have two remote builders A and B (where A has higher speed than B), if a derivation foo.drv is already built on B, and your local machine needs to build foo.drv, then it will:

  • build (possibly remotely) all the build dependencies of foo.drv
  • build foo.drv on A

Even if foo.drv is also on A, you will still have to build the build dependencies of foo.drv before sending the build to A, which will build it instantly since it is in cache.

To solve this problem, you can set up your remote builders as substituters. Every time (the local machine's) nix considers building a derivation, it will connect to the remote builders to check whether it is already available there. Here is how to set this up via ssh. See also Binary Cache for an alternative using http and nix-serve.

1. On the remote builder, create a binary cache key:

$ nix-store --generate-binary-cache-key builder-name cache-priv-key.pem cache-pub-key.pem

The private key must be readable only by the user running the build: ??? on multi-user installs, and the owner of /nix on single-user installs. builder-name is only here for your convenience to distinguish several public keys, it has no functional meaning.

2. On the remote builder, set up nix to sign all store paths it builds: in the nix configuration (/etc/nix/nix.conf on multi-user installs and ~/.config/nix/nix.conf on single user installs), add the following line:

secret-key-files = /path/to/cache-priv-key.pem

If necessary, restart the nix daemon.

3. The previous point does not retroactively sign existing paths in the store of the builder. To do so, run

$ nix sign-paths --all -k /path/to/cache-priv-key.pem

4. In the nix configuration of the local machine, append the content of cache-pub-key.pem to the option trusted-public-keys. Also append ssh-ng://builder to the option substituters. If you only want to use the remote builder occasionally as a substituter, use trusted-substituters instead of substituters. Then, when you want to use the builder, pass --option extra-substituters ssh-ng://builder to the nix command you run.

Troubleshooting

  • How do I know if I'm distributing my build at all?
    • Run nix build with --max-jobs 0.
  • How do I know why my builds aren't being distributed?
    • Run nix build -vvvvvvvvv 2>&1 | less and search for decline.
  • I can nix store ping but the build doesn't distribute.
    • If on NixOS, Check that nix store ping command works when run as root.
    • If you configured builders on the command line (with --builders), make sure your account is in nix.trustedUsers in /etc/nixos/configuration.nix. Only /etc/nix/nix.conf is taken into account otherwise.
  • I can ping the store as root, but I'm getting "broken pipe" errors when trying to distribute.
    • You may have hit bug Issue.png#46038. Add nix.distributedBuilds = true; to configuration.nix and nixos-rebuild switch.

See also