Run builds and checks in a per-node build slice with CPU, IO and memory limits (issue 1798)

internal/builder/kill.go

@@ -42,6 +42,11 @@ func inItsOwnGroup(cmd *exec.Cmd) {
 	cmd.WaitDelay = KillWait
 }
 
+// CgroupParent is the slice every container a build or check starts joins (novox/hq issue 1798), so the node's own
+// weights and cap bind it: the build-agent module declares the slice with the machine's CPU weight, IO weight and
+// memory limits, and names it in MESH_BUILD_CGROUP_PARENT. Empty, a container joins the runtime's default.
+var CgroupParent string
+
 // Labelled is a Runner that marks every container a build starts with the build's id, so a kill
 // finds what outlived the docker client (novox/hq ADR 0219). Applied at the one place every command
 // passes rather than at each `docker run` the builder composes: a run added later is labelled too.
@@ -56,8 +61,11 @@ func LabelledArgs(name string, args []string, id string) []string {
 	if name != "docker" || len(args) == 0 || args[0] != "run" || id == "" {
 		return args
 	}
-	out := make([]string, 0, len(args)+2)
+	out := make([]string, 0, len(args)+4)
 	out = append(out, "run", "--label", BuildLabel+"="+id)
+	if CgroupParent != "" {
+		out = append(out, "--cgroup-parent", CgroupParent)
+	}
 	return append(out, args[1:]...)
 }
 

internal/builder/kill_test.go

@@ -143,3 +143,24 @@ func TestAVolumeAlreadyGoneIsNoFailure(t *testing.T) {
 		t.Fatalf("%d %v", n, err)
 	}
 }
+
+// A node's build slice limits everything a build or check starts (novox/hq issue 1798): with a parent
+// set, every `docker run` joins it, so the node's CPU and IO weights and memory cap bind the container;
+// with none set the arguments are as they always were.
+func TestEveryContainerJoinsTheNodesBuildSlice(t *testing.T) {
+	was := CgroupParent
+	defer func() { CgroupParent = was }()
+	CgroupParent = "mesh-build.slice"
+	got := LabelledArgs("docker", []string{"run", "--rm", "img", "sh"}, "build-1")
+	want := []string{"run", "--label", "mesh.build=build-1", "--cgroup-parent", "mesh-build.slice", "--rm", "img", "sh"}
+	if !reflect.DeepEqual(got, want) {
+		t.Errorf("1798/AC1: docker run became %v", got)
+	}
+	if got := LabelledArgs("docker", []string{"build", "."}, "build-1"); !reflect.DeepEqual(got, []string{"build", "."}) {
+		t.Errorf("a docker build became %v", got)
+	}
+	CgroupParent = ""
+	if got := LabelledArgs("docker", []string{"run", "img"}, "build-1"); !reflect.DeepEqual(got, []string{"run", "--label", "mesh.build=build-1", "img"}) {
+		t.Errorf("1798/AC1: with no slice, docker run became %v", got)
+	}
+}