Support building on target nodes

This partially addresses #33, and allows Colmena to be used more
easily on bandwidth-constrained hosts and macOS.

With `deployment.buildOnTarget = true;` deployment works fine from
macOS without designated builders, except when IFD is involved.

integration-tests/build-on-target/default.nix

@@ -0,0 +1,35 @@
+{ pkgs ? import ../nixpkgs.nix }:
+
+let
+  tools = pkgs.callPackage ../tools.nix {
+    deployers = [ "deployer" "alpha" "beta" ];
+    targets = [];
+  };
+in tools.makeTest {
+  name = "colmena-build-on-target";
+
+  bundle = ./.;
+
+  testScript = ''
+    # The actual build will be initiated on alpha
+    deployer.succeed("cd /tmp/bundle && ${tools.colmenaExec} apply --on alpha")
+
+    with subtest("Check that the new configurations are indeed applied"):
+        alpha.succeed("grep SUCCESS /etc/deployment")
+
+    alpha_profile = alpha.succeed("readlink /run/current-system")
+
+    with subtest("Check that the built profile is not on the deployer"):
+        deployer.fail(f"nix-store -qR {alpha_profile}")
+
+    with subtest("Check that we can override per-node settings and build locally"):
+        deployer.succeed("cd /tmp/bundle && ${tools.colmenaExec} build --on alpha --no-build-on-target")
+        deployer.succeed(f"nix-store -qR {alpha_profile}")
+
+    with subtest("Check that we can override per-node settings and build remotely"):
+        deployer.succeed("cd /tmp/bundle && ${tools.colmenaExec} apply --on beta --build-on-target")
+        beta.succeed("grep SUCCESS /etc/deployment")
+        profile = beta.succeed("readlink /run/current-system")
+        deployer.fail(f"nix-store -qR {profile}")
+  '';
+}

integration-tests/build-on-target/hive.nix

@@ -0,0 +1,33 @@
+let
+  tools = import ./tools.nix {
+    insideVm = true;
+    deployers = [ "deployer" "alpha" "beta" ];
+    targets = [];
+  };
+in {
+  meta = {
+    nixpkgs = tools.pkgs;
+  };
+
+  defaults = {
+    environment.etc."deployment".text = "SUCCESS";
+  };
+
+  deployer = tools.getStandaloneConfigFor "deployer";
+
+  alpha = {
+    imports = [
+      (tools.getStandaloneConfigFor "alpha")
+    ];
+
+    deployment.buildOnTarget = true;
+  };
+
+  beta = {
+    imports = [
+      (tools.getStandaloneConfigFor "beta")
+    ];
+
+    deployment.buildOnTarget = false;
+  };
+}

integration-tests/default.nix

@@ -1,6 +1,7 @@
 {
   apply = import ./apply {};
   apply-local = import ./apply-local {};
+  build-on-target = import ./build-on-target {};
   exec = import ./exec {};
   flakes = import ./flakes {};
   parallel = import ./parallel {};

integration-tests/tools.nix

@@ -55,6 +55,11 @@ let
         ];
       };
 
+      services.openssh.enable = true;
+      users.users.root.openssh.authorizedKeys.keys = [
+        sshKeys.snakeOilPublicKey
+      ];
+
       environment.systemPackages = with pkgs; [
         git # for git flake tests
 

manual/src/SUMMARY.md

@@ -10,6 +10,7 @@
     - [Secrets](./features/keys.md)
     - [Ad Hoc Evaluation](./features/eval.md)
     - [Parallelism](./features/parallelism.md)
+    - [Remote Builds](./features/remote-builds.md)
 - [Examples](./examples/index.md)
     - [Multi-Architecture Deployments](./examples/multi-arch.md)
 - [Reference](./reference/index.md)

manual/src/features/index.md

@@ -7,3 +7,4 @@ This section introduces the main features in Colmena:
 - **[Secrets](keys.md)** - Deploying sensitive files separate from the main configuration
 - **[Ad Hoc Evaluation](eval.md)** - Evaluating a Nix expression with access to your configuration
 - **[Parallelism](parallelism.md)** - Controlling how Colmena parallelizes the deployment process
+- **[Remote Builds](remote-builds.md)** - Building system profiles on remote machines

manual/src/features/remote-builds.md

@@ -0,0 +1,21 @@
+# Remote Builds
+
+If the host running Colmena is not powerful enough, consider offloading the actual builds to remote machines.
+Colmena supports two ways to achieve this:
+
+## Using Colmena's `deployment.buildOnTarget`
+
+If you set [`deployment.buildOnTarget = true;`](../reference/deployment.md#deploymentbuildontarget) for a node, then the actual build process will be initiated on the node itself.
+Colmena will evaluate the configuration locally before copying the derivations to the target node.
+You can temporarily enable this for all nodes by passing `--build-on-target` on the command line, or disable it with `--no-build-on-target`.
+
+This is most useful in scenarios where the machine running Colmena is bandwidth-constrained, or it's inconvenient to configure designated builders beforehand.
+With this method, the build results will _not_ be copied back to the local machine or otherwise shared across the target nodes.
+If you have custom packages used on multiple nodes, the work required to build those packages will be duplicated across the nodes.
+
+## Using the native distributed build feature in Nix
+
+When [distributed build](https://nixos.org/manual/nix/unstable/advanced-topics/distributed-builds.html) is enabled, Nix will transparently forward builds to the configured builders.
+After the builds are done, Nix will copy the results back to the local machine.
+
+Builders can either be configured globally or in your configuration with [`meta.machinesFile`](../reference/meta.md#machinesFile).

src/command/apply.rs

@@ -1,7 +1,7 @@
 use std::env;
 use std::path::PathBuf;
 
-use clap::{Arg, App, SubCommand, ArgMatches};
+use clap::{Arg, App, SubCommand, ArgMatches, ArgSettings};
 
 use crate::nix::deployment::{
     Deployment,
@@ -89,6 +89,20 @@ To upload keys without building or deploying the rest of the configuration, use
             .help("Do not use gzip")
             .long_help("Disables the use of gzip when copying closures to the remote host.")
             .takes_value(false))
+        .arg(Arg::with_name("build-on-target")
+            .long("build-on-target")
+            .help("Build the system profiles on the target nodes")
+            .long_help(r#"Build the system profiles on the target nodes themselves.
+
+If enabled, the system profiles will be built on the target nodes themselves, not on the host running Colmena itself.
+This overrides per-node perferences set in `deployment.buildOnTarget`.
+To temporarily disable remote build on all nodes, use `--no-build-on-target`.
+"#)
+            .takes_value(false))
+        .arg(Arg::with_name("no-build-on-target")
+            .long("no-build-on-target")
+            .set(ArgSettings::Hidden)
+            .takes_value(false))
         .arg(Arg::with_name("force-replace-unknown-profiles")
             .long("force-replace-unknown-profiles")
             .help("Ignore all targeted nodes deployment.replaceUnknownProfiles setting")
@@ -146,6 +160,12 @@ pub async fn run(_global_args: &ArgMatches<'_>, local_args: &ArgMatches<'_>) ->
             options.set_create_gc_roots(true);
         }
 
+        if local_args.is_present("no-build-on-target") {
+            options.set_force_build_on_target(false);
+        } else if local_args.is_present("build-on-target") {
+            options.set_force_build_on_target(true);
+        }
+
         options
     };
 

src/nix/deployment/mod.rs

@@ -31,6 +31,7 @@ use super::{
     Profile,
     ProfileDerivation,
     CopyDirection,
+    CopyOptions,
     key::{Key, UploadAt as UploadKeyAt},
 };
 use super::host;
@@ -229,8 +230,29 @@ impl Deployment {
         let mut futures = Vec::new();
 
         for (name, profile_drv) in profile_drvs.iter() {
-            let target = chunk.remove(name).unwrap();
-            futures.push(self.clone().deploy_node(parent.clone(), target, profile_drv.clone()));
+            let mut target = chunk.remove(name).unwrap();
+
+            if let Some(force_build_on_target) = self.options.force_build_on_target {
+                target.config.set_build_on_target(force_build_on_target);
+            }
+
+            let job_handle = parent.clone();
+            let arc_self = self.clone();
+            futures.push(async move {
+                let (target, profile) = {
+                    if target.config.build_on_target() {
+                        arc_self.clone().build_on_node(job_handle.clone(), target, profile_drv.clone()).await?
+                    } else {
+                        arc_self.clone().build_and_push_node(job_handle.clone(), target, profile_drv.clone()).await?
+                    }
+                };
+
+                if arc_self.goal.requires_activation() {
+                    arc_self.activate_node(job_handle, target, profile).await
+                } else {
+                    Ok(())
+                }
+            });
         }
 
         join_all(futures).await
@@ -273,14 +295,45 @@ impl Deployment {
         }).await
     }
 
-    /// Builds, pushes, and optionally activates a system profile on a node.
-    ///
-    /// This will also upload keys to the node.
-    async fn deploy_node(self: DeploymentHandle, parent: JobHandle, mut target: TargetNode, profile_drv: ProfileDerivation)
-        -> NixResult<()>
+    /// Builds a system profile directly on the node itself.
+    async fn build_on_node(self: DeploymentHandle, parent: JobHandle, mut target: TargetNode, profile_drv: ProfileDerivation)
+        -> NixResult<(TargetNode, Profile)>
+    {
+        let nodes = vec![target.name.clone()];
+
+        let permit = self.parallelism_limit.apply.acquire().await.unwrap();
+
+        let build_job = parent.create_job(JobType::Build, nodes.clone())?;
+        let (target, profile) = build_job.run(|job| async move {
+            if target.host.is_none() {
+                return Err(NixError::Unsupported);
+            }
+
+            let mut host = target.host.as_mut().unwrap();
+            host.set_job(Some(job.clone()));
+
+            host.copy_closure(
+                profile_drv.as_store_path(),
+                CopyDirection::ToRemote,
+                CopyOptions::default().include_outputs(true),
+                ).await?;
+
+            let profile = profile_drv.realize_remote(&mut host).await?;
+
+            job.success_with_message(format!("Built {:?} on target node", profile.as_path()))?;
+            Ok((target, profile))
+        }).await?;
+
+        drop(permit);
+
+        Ok((target, profile))
+    }
+
+    /// Builds and pushes a system profile on a node.
+    async fn build_and_push_node(self: DeploymentHandle, parent: JobHandle, mut target: TargetNode, profile_drv: ProfileDerivation)
+        -> NixResult<(TargetNode, Profile)>
     {
         let nodes = vec![target.name.clone()];
-        let target_name = target.name.clone();
 
         let permit = self.parallelism_limit.apply.acquire().await.unwrap();
 
@@ -292,21 +345,21 @@ impl Deployment {
             let mut builder = host::local(arc_self.nix_options.clone());
             builder.set_job(Some(job.clone()));
 
-            let profile = profile_drv.realize(&mut *builder).await?;
+            let profile = profile_drv.realize(&mut builder).await?;
 
             job.success_with_message(format!("Built {:?}", profile.as_path()))?;
             Ok(profile)
         }).await?;
 
         if self.goal == Goal::Build {
-            return Ok(());
+            return Ok((target, profile));
         }
 
         // Push closure to remote
         let push_job = parent.create_job(JobType::Push, nodes.clone())?;
         let push_profile = profile.clone();
         let arc_self = self.clone();
-        let mut target = push_job.run(|job| async move {
+        let target = push_job.run(|job| async move {
             if target.host.is_none() {
                 return Err(NixError::Unsupported);
             }
@@ -321,11 +374,22 @@ impl Deployment {
             Ok(target)
         }).await?;
 
-        if !self.goal.requires_activation() {
-            // We are done here :)
-            return Ok(());
+        drop(permit);
+
+        Ok((target, profile))
     }
 
+    /// Activates a system profile on a node.
+    ///
+    /// This will also upload keys to the node.
+    async fn activate_node(self: DeploymentHandle, parent: JobHandle, mut target: TargetNode, profile: Profile)
+        -> NixResult<()>
+    {
+        let nodes = vec![target.name.clone()];
+        let target_name = target.name.clone();
+
+        let permit = self.parallelism_limit.apply.acquire().await.unwrap();
+
         // Upload pre-activation keys
         let mut target = if self.options.upload_keys {
             let job = parent.create_job(JobType::UploadKeys, nodes.clone())?;
@@ -386,7 +450,7 @@ impl Deployment {
         }).await?;
 
         // Upload post-activation keys
-        if self.options.upload_keys {
+        let target = if self.options.upload_keys {
             let job = parent.create_job(JobType::UploadKeys, nodes.clone())?;
             job.run_waiting(|job| async move {
                 let keys = target.config.keys.iter()
@@ -396,7 +460,7 @@ impl Deployment {
 
                 if keys.is_empty() {
                     job.noop("No post-activation keys to upload".to_string())?;
-                    return Ok(());
+                    return Ok(target);
                 }
 
                 job.state(JobState::Running)?;
@@ -407,15 +471,21 @@ impl Deployment {
                 host.upload_keys(&keys, true).await?;
 
                 job.success_with_message("Uploaded keys (post-activation)".to_string())?;
-                Ok(())
-            }).await?;
-        }
+                Ok(target)
+            }).await?
+        } else {
+            target
+        };
 
         // Create GC root
         if self.options.create_gc_roots {
             let job = parent.create_job(JobType::CreateGcRoots, nodes.clone())?;
             let arc_self = self.clone();
             job.run_waiting(|job| async move {
+                if target.config.build_on_target() {
+                    job.noop("The system profile was built on target node itself".to_string())?;
+                }
+
                 if let Some(dir) = arc_self.hive.context_dir() {
                     job.state(JobState::Running)?;
                     let path = dir.join(".gcroots").join(format!("node-{}", &*target_name));

src/nix/deployment/options.rs

@@ -20,6 +20,9 @@ pub struct Options {
     /// directory if it exists.
     pub(super) create_gc_roots: bool,
 
+    /// Whether to override per-node setting to build on the nodes themselves.
+    pub(super) force_build_on_target: Option<bool>,
+
     /// Ignore the node-level `deployment.replaceUnknownProfiles` option.
     pub(super) force_replace_unknown_profiles: bool,
 }
@@ -41,6 +44,10 @@ impl Options {
         self.create_gc_roots = enable;
     }
 
+    pub fn set_force_build_on_target(&mut self, enable: bool) {
+        self.force_build_on_target = Some(enable);
+    }
+
     pub fn set_force_replace_unknown_profiles(&mut self, enable: bool) {
         self.force_replace_unknown_profiles = enable;
     }
@@ -61,6 +68,7 @@ impl Default for Options {
             gzip: true,
             upload_keys: true,
             create_gc_roots: false,
+            force_build_on_target: None,
             force_replace_unknown_profiles: false,
         }
     }

src/nix/hive/eval.nix

@@ -67,6 +67,11 @@ let
 
           See https://nixos.org/manual/nix/stable/#chap-distributed-builds
           for the machine specification format.
+
+          This option is ignored when builds are initiated on the remote nodes
+          themselves via `deployment.buildOnTarget` or `--build-on-target`. To
+          still use the Nix distributed build functionality, configure the
+          builders on the target nodes with `nix.buildMachines`.
         '';
         default = null;
         apply = value: if value == null then null else toString value;
@@ -138,6 +143,26 @@ let
           type = types.bool;
           default = false;
         };
+        buildOnTarget = lib.mkOption {
+          description = ''
+            Whether to build the system profiles on the target node itself.
+
+            When enabled, Colmena will copy the derivation to the target
+            node and initiate the build there. This avoids copying back the
+            build results involved with the native distributed build
+            feature. Furthermore, the `build` goal will be equivalent to
+            the `push` goal. Since builds happen on the target node, the
+            results are automatically "pushed" and won't exist in the local
+            Nix store.
+
+            You can temporarily override per-node settings by passing
+            `--build-on-target` (enable for all nodes) or
+            `--no-build-on-target` (disable for all nodes) on the command
+            line.
+          '';
+          type = types.bool;
+          default = false;
+        };
         tags = lib.mkOption {
           description = ''
             A list of tags for the node.

src/nix/host/ssh.rs

@@ -40,18 +40,17 @@ impl Host for Ssh {
         self.run_command(command).await
     }
     async fn realize_remote(&mut self, derivation: &StorePath) -> NixResult<Vec<StorePath>> {
-        // FIXME
-        let paths = self.ssh(&["nix-store", "--no-gc-warning", "--realise", derivation.as_path().to_str().unwrap()])
-            .capture_output()
-            .await;
+        let command = self.ssh(&["nix-store", "--no-gc-warning", "--realise", derivation.as_path().to_str().unwrap()]);
+
+        let mut execution = CommandExecution::new(command);
+        execution.set_job(self.job.clone());
+
+        let paths = execution
+            .capture_output()
+            .await?;
 
-        match paths {
-            Ok(paths) => {
         paths.lines().map(|p| p.to_string().try_into()).collect()
     }
-            Err(e) => Err(e),
-        }
-    }
     async fn upload_keys(&mut self, keys: &HashMap<String, Key>, require_ownership: bool) -> NixResult<()> {
         for (name, key) in keys {
             self.upload_key(name, key, require_ownership).await?;

src/nix/mod.rs

@@ -24,7 +24,7 @@ pub mod hive;
 pub use hive::{Hive, HivePath};
 
 pub mod store;
-pub use store::{StorePath, StoreDerivation};
+pub use store::{StorePath, StoreDerivation, BuildResult};
 
 pub mod key;
 pub use key::Key;
@@ -158,6 +158,10 @@ pub struct NodeConfig {
 
     #[serde(rename = "allowLocalDeployment")]
     allow_local_deployment: bool,
+
+    #[serde(rename = "buildOnTarget")]
+    build_on_target: bool,
+
     tags: Vec<String>,
 
     #[serde(rename = "replaceUnknownProfiles")]
@@ -223,6 +227,11 @@ impl NodeConfig {
     pub fn tags(&self) -> &[String] { &self.tags }
     pub fn allows_local_deployment(&self) -> bool { self.allow_local_deployment }
 
+    pub fn build_on_target(&self) -> bool { self.build_on_target }
+    pub fn set_build_on_target(&mut self, enable: bool) {
+        self.build_on_target = enable;
+    }
+
     pub fn to_ssh_host(&self) -> Option<Ssh> {
         self.target_host.as_ref().map(|target_host| {
             let username =

src/nix/profile.rs

@@ -10,6 +10,7 @@ use super::{
     NixError,
     StorePath,
     StoreDerivation,
+    BuildResult,
 };
 
 pub type ProfileDerivation = StoreDerivation<Profile>;
@@ -75,12 +76,18 @@ impl Profile {
 
         Ok(())
     }
+
+    fn from_store_path_unchecked(path: StorePath) -> Self {
+        Self(path)
+    }
 }
 
-impl TryFrom<Vec<StorePath>> for Profile {
+impl TryFrom<BuildResult<Profile>> for Profile {
     type Error = NixError;
 
-    fn try_from(paths: Vec<StorePath>) -> NixResult<Self> {
+    fn try_from(result: BuildResult<Self>) -> NixResult<Self> {
+        let paths = result.paths();
+
         if paths.is_empty() {
             return Err(NixError::BadOutput {
                 output: String::from("There is no store path"),
@@ -93,7 +100,9 @@ impl TryFrom<Vec<StorePath>> for Profile {
             });
         }
 
-        let path = paths.into_iter().next().unwrap();
-        Self::from_store_path(path)
+        let path = paths.into_iter().next()
+            .unwrap().to_owned();
+
+        Ok(Self::from_store_path_unchecked(path))
     }
 }

src/nix/store.rs

@@ -13,6 +13,19 @@ use super::{Host, NixCommand, NixResult, NixError};
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct StorePath(PathBuf);
 
+/// A store derivation (.drv) that will result in a T when built.
+#[derive(Debug)]
+pub struct StoreDerivation<T: TryFrom<BuildResult<T>>>{
+    path: StorePath,
+    _target: PhantomData<T>,
+}
+
+/// Results of a build/realization.
+pub struct BuildResult<T: TryFrom<BuildResult<T>>> {
+    results: Vec<StorePath>,
+    _derivation: PhantomData<T>,
+}
+
 impl StorePath {
     /// Returns the raw store path.
     pub fn as_path(&self) -> &Path {
@@ -41,7 +54,7 @@ impl StorePath {
     }
 
     /// Converts the store path into a store derivation.
-    pub fn into_derivation<T: TryFrom<Vec<StorePath>>>(self) -> NixResult<StoreDerivation<T>> {
+    pub fn into_derivation<T: TryFrom<BuildResult<T>>>(self) -> NixResult<StoreDerivation<T>> {
         if self.is_derivation() {
             Ok(StoreDerivation::<T>::from_store_path_unchecked(self))
         } else {
@@ -76,14 +89,21 @@ impl From<StorePath> for PathBuf {
     }
 }
 
-/// A store derivation (.drv) that will result in a T when built.
-#[derive(Debug, Clone)]
-pub struct StoreDerivation<T: TryFrom<Vec<StorePath>>>{
-    path: StorePath,
-    _target: PhantomData<T>,
+impl<T: TryFrom<BuildResult<T>>> Clone for StoreDerivation<T> {
+    fn clone(&self) -> Self {
+        Self {
+            path: self.path.clone(),
+            _target: PhantomData,
+        }
+    }
+}
+
+impl<T: TryFrom<BuildResult<T>>> StoreDerivation<T> {
+    /// Returns the store path.
+    pub fn as_store_path(&self) -> &StorePath {
+        &self.path
     }
 
-impl<T: TryFrom<Vec<StorePath>>> StoreDerivation<T> {
     fn from_store_path_unchecked(path: StorePath) -> Self {
         Self {
             path,
@@ -92,16 +112,38 @@ impl<T: TryFrom<Vec<StorePath>>> StoreDerivation<T> {
     }
 }
 
-impl<T: TryFrom<Vec<StorePath>, Error=NixError>> StoreDerivation<T> {
+impl<T: TryFrom<BuildResult<T>, Error=NixError>> StoreDerivation<T> {
     /// Builds the store derivation on a host, resulting in a T.
-    pub async fn realize(&self, host: &mut dyn Host) -> NixResult<T> {
+    pub async fn realize(&self, host: &mut Box<dyn Host>) -> NixResult<T> {
         let paths: Vec<StorePath> = host.realize(&self.path).await?;
-        paths.try_into()
+
+        let result = BuildResult {
+            results: paths,
+            _derivation: PhantomData,
+        };
+        result.try_into()
+    }
+
+    /// Builds the store derivation on a host without copying the results back.
+    pub async fn realize_remote(&self, host: &mut Box<dyn Host>) -> NixResult<T> {
+        let paths: Vec<StorePath> = host.realize_remote(&self.path).await?;
+
+        let result = BuildResult {
+            results: paths,
+            _derivation: PhantomData,
+        };
+        result.try_into()
     }
 }
 
-impl<T: TryFrom<Vec<StorePath>>> fmt::Display for StoreDerivation<T> {
+impl<T: TryFrom<BuildResult<T>>> fmt::Display for StoreDerivation<T> {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
         write!(f, "{:?}", self.path)
     }
 }
+
+impl<T: TryFrom<BuildResult<T>, Error=NixError>> BuildResult<T> {
+    pub fn paths(&self) -> &[StorePath] {
+        self.results.as_slice()
+    }
+}