Free a build node's disk after every build and refuse work under a floor of free space (issue 1834)

A full / made every check on shanks end 'store never became ready' with the reason erased by --rm.
Builds now remove their local image tag, narrowed source and context clone; a build or check
refuses under 20 GiB free on / and the workspace, saying so; the throwaway store is no longer
started with --rm and a store that does not come up reports its exit code and last 20 log lines.

internal/builder/build_source_test.go

@@ -86,11 +86,14 @@ func TestAnImageCompilingGoIsHandedItsBuildSourceAndSaysIt(t *testing.T) {
 	}
 	for file, want := range map[string]bool{"proxy/main.go": true, "internal/shared/s.go": true, "go.mod": true,
 		"cmd/ctl/main.go": false, "internal/only/o.go": false, "README.md": false} {
-		_, err := os.Stat(filepath.Join(workspace, "narrow-server", filepath.FromSlash(file)))
+		held := slices.Contains(r.builtFrom, file)
-		if (err == nil) != want {
+		if held != want {
-			t.Errorf("%s handed to the recipe: %v, wanted %v", file, err == nil, want)
+			t.Errorf("%s handed to the recipe: %v, wanted %v", file, held, want)
 		}
 	}
+	if _, err := os.Stat(filepath.Join(workspace, "narrow-server")); err == nil {
+		t.Error("the narrowed build source was left in the workspace")
+	}
 
 	// Said per repository: its own, the manifest and the recipe; the context's, the closure.
 	if len(got.Sources) != 2 {

internal/builder/builder.go

@@ -219,6 +219,11 @@ func build(ctx context.Context, run Runner, publish Publisher,
 	if err := os.MkdirAll(workspace, 0o755); err != nil {
 		return Result{}, err
 	}
+	// **A node with no room takes no build** (novox/hq issue 1834).
+	if err := RoomToWork("/", workspace); err != nil {
+		say("clone", "refused: %v", err)
+		return Result{}, err
+	}
 	// The credential is a file git reads, never an argument: a URL carrying a password in argv
 	// would be readable by anything that can list processes for as long as a clone runs.
 	credentials, forget, err := storeCredential(workspace, forge)
@@ -738,11 +743,34 @@ func logging(log Log) func(step, format string, args ...any) {
 	}
 }
 
+// keepsImageLocally says the publisher leaves the image in the machine's runtime as the artifact itself.
+func keepsImageLocally(publish Publisher) bool {
+	switch publish.(type) {
+	case Local, *Local:
+		return true
+	}
+	return false
+}
+
+// releaseImage removes the local image tag a build made, with a context of its own: the build's may be cancelled. A
+// tag that was never made (the build failed before it) is no fault; one that cannot be removed is said.
+func releaseImage(ctx context.Context, run Runner, tag string, say func(step, format string, args ...any)) {
+	removing, done := context.WithTimeout(context.WithoutCancel(ctx), time.Minute)
+	defer done()
+	if _, err := run(removing, "", "docker", "rmi", tag); err != nil {
+		if !strings.Contains(err.Error(), "No such image") {
+			say("image", "the local image %s could not be removed, and goes with the week's sweep: %v", tag, err)
+		}
+		return
+	}
+	say("image", "the local image %s is removed: the registry holds it", tag)
+}
+
 // contextFrom clones an image artifact's own build context, when it names one apart from this
 // module's own repository — a fresh tree, the same way the module's own is, keyed by artifact
 // name so two artifacts of one module naming different contexts do not collide.
 func contextFrom(ctx context.Context, run Runner, workspace, artifact, credentials string,
-	from catalogue.ArtifactContext, seats map[string]string, say func(step, format string, args ...any)) (string, error) {
+	from catalogue.ArtifactContext, seats map[string]string, say func(step, format string, args ...any)) (_ string, err error) {
 	url, err := contextURL(from, seats)
 	if err != nil {
 		return "", err
@@ -752,6 +780,12 @@ func contextFrom(ctx context.Context, run Runner, workspace, artifact, credentia
 	if err := os.RemoveAll(dir); err != nil {
 		return "", err
 	}
+	// A clone that failed half-way is no context to leave behind (novox/hq issue 1834).
+	defer func() {
+		if err != nil {
+			os.RemoveAll(dir)
+		}
+	}()
 	if _, err := run(ctx, workspace, "git", cloneWith(credentials, "clone", "--quiet", url, dir)...); err != nil {
 		return "", fmt.Errorf("cannot clone %s: %w", url, err)
 	}
@@ -1143,6 +1177,8 @@ func one(ctx context.Context, run Runner, publish Publisher,
 				src.contexts[a.Name] = t
 			}
 			buildDir = cloned
+			// Let go of when the image is built, whatever became of it (novox/hq issue 1834).
+			defer os.RemoveAll(cloned)
 			if t, _ := trunkOf(ctx, run, cloned, "HEAD"); t == "" || !atTrunkHead(ctx, run, cloned, "HEAD", t) {
 				src.contextOffHead()
 			}
@@ -1180,6 +1216,7 @@ func one(ctx context.Context, run Runner, publish Publisher,
 				src.ownHas(paths...)
 			}
 			buildDir = narrowed
+			defer os.RemoveAll(narrowed)
 		} else if a.Context != nil {
 			src.readWhole(*a.Context)
 		}
@@ -1205,6 +1242,12 @@ func one(ctx context.Context, run Runner, publish Publisher,
 		}
 		invocation = append(invocation, ".")
 		say("image", "docker build -f %s", recipePath)
+		// **The local tag goes with the build** (novox/hq issue 1834): an image published to a registry is held there, and
+		// a tag per module and commit left on the node was the 96 GB of a full disk. Only a node publishing nowhere (a
+		// mesh being raised) keeps it, since the local image is then the artifact.
+		if !keepsImageLocally(publish) {
+			defer releaseImage(ctx, run, local, say)
+		}
 		if _, err := run(ctx, buildDir, "docker", invocation...); err != nil {
 			return catalogue.Built{}, fmt.Errorf("%s: building %s failed: %w", module, a.Name, err)
 		}

internal/builder/builder_test.go

@@ -38,6 +38,9 @@ type recorded struct {
 	stamped time.Time
 	// tree is what git names as the tree of the module's directory; empty answers as for a commit.
 	tree string
+	// builtFrom is the files (slash-separated, relative) of the folder the last `docker build` ran in, read as it ran:
+	// the build removes that folder when it ends (novox/hq issue 1834).
+	builtFrom []string
 }
 
 func (r *recorded) run(_ context.Context, dir, name string, args ...string) (string, error) {
@@ -53,6 +56,16 @@ func (r *recorded) run(_ context.Context, dir, name string, args ...string) (str
 			break
 		}
 	}
+	if name == "docker" && len(args) > 0 && args[0] == "build" {
+		r.builtFrom = nil
+		_ = filepath.WalkDir(dir, func(p string, d os.DirEntry, err error) error {
+			if err == nil && !d.IsDir() {
+				rel, _ := filepath.Rel(dir, p)
+				r.builtFrom = append(r.builtFrom, filepath.ToSlash(rel))
+			}
+			return nil
+		})
+	}
 	switch {
 	case name == "git" && isClone:
 		repository := args[len(args)-2]

internal/builder/check.go

@@ -392,6 +392,12 @@ func Check(ctx context.Context, run Runner, spec CheckSpec, workspace, registry
 	ctx, stop := context.WithTimeout(ctx, CheckTimeout)
 	defer stop()
 
+	// **A node with no room takes no check** (novox/hq issue 1834): said as the check's verdict, never a throwaway store
+	// that dies at once on a full disk.
+	if err := RoomToWork("/", workspace); err != nil {
+		say("check", "refused: %v", err)
+		return CheckVerdict{}, err
+	}
 	root := filepath.Join(workspace, "check")
 	if err := os.RemoveAll(root); err != nil {
 		return CheckVerdict{}, err
@@ -1631,7 +1637,9 @@ func BusImage(version string) string {
 
 // throwaway starts a container publishing one port on loopback, and answers where it is reached.
 func throwaway(ctx context.Context, run, plain Runner, name, image string, port int, opts, args []string) (string, error) {
-	invocation := []string{"run", "-d", "--rm", "--name", name, "-p", fmt.Sprintf("127.0.0.1::%d", port)}
+	// Never with --rm (novox/hq issue 1834): a store that dies at once took its log with it. The check removes what it
+	// raised, with its volumes, after it has read why one stopped (RemoveContainersOf).
+	invocation := []string{"run", "-d", "--name", name, "-p", fmt.Sprintf("127.0.0.1::%d", port)}
 	invocation = append(invocation, opts...)
 	invocation = append(invocation, image)
 	invocation = append(invocation, args...)
@@ -1652,17 +1660,45 @@ func throwaway(ctx context.Context, run, plain Runner, name, image string, port
 
 // waitFor runs a readiness command in a container until it answers, for a minute.
 func waitFor(ctx context.Context, run Runner, name string, ready []string) error {
-	for i := 0; i < 60; i++ {
+	return waitReady(ctx, run, name, ready, 60)
+}
+
+// waitReady is waitFor for a number of tries, a second apart. A container that never answers is not left a mystery
+// (novox/hq issue 1834): its state and exit code and the last 20 lines of its log are in the error, read before the
+// check removes it.
+func waitReady(ctx context.Context, run Runner, name string, ready []string, attempts int) error {
+	for i := 0; i < attempts; i++ {
 		if _, err := run(ctx, "", "docker", append([]string{"exec", name}, ready...)...); err == nil {
 			return nil
 		}
+		if i == attempts-1 {
+			break
+		}
 		select {
 		case <-ctx.Done():
 			return ctx.Err()
 		case <-time.After(time.Second):
 		}
 	}
-	return fmt.Errorf("%s never became ready", name)
+	return fmt.Errorf("%s never became ready%s", name, whyItStopped(ctx, run, name))
+}
+
+// whyItStopped is ": <state>, exit code N; its last log lines: …" for a container, or "" when docker says nothing.
+func whyItStopped(ctx context.Context, run Runner, name string) string {
+	reading, done := context.WithTimeout(context.WithoutCancel(ctx), 30*time.Second)
+	defer done()
+	var said []string
+	if out, err := run(reading, "", "docker", "inspect", "--format",
+		"{{.State.Status}}, exit code {{.State.ExitCode}}", name); err == nil && strings.TrimSpace(out) != "" {
+		said = append(said, strings.TrimSpace(out))
+	}
+	if out, err := run(reading, "", "docker", "logs", "--tail", "20", name); err == nil && strings.TrimSpace(out) != "" {
+		said = append(said, "its last log lines: "+strings.TrimSpace(out))
+	}
+	if len(said) == 0 {
+		return ""
+	}
+	return ": " + strings.Join(said, "; ")
 }
 
 // dialable waits for an address to take a connection, for a minute.

internal/builder/disk_1834_test.go

@@ -0,0 +1,174 @@
+package builder
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// Issue 1834: a build node's root disk filled and every check there ended "the store never became ready". A node
+// frees what each build and check leaves, refuses to take work under a floor of free space, and a store that does
+// not come up says why.
+
+func init() {
+	// The disk the tests run on is none of their business; the tests of the floor say their own.
+	FreeBytes = func(string) (uint64, error) { return 500 << 30, nil }
+}
+
+func withFree(t *testing.T, bytes uint64) {
+	t.Helper()
+	was := FreeBytes
+	FreeBytes = func(string) (uint64, error) { return bytes, nil }
+	t.Cleanup(func() { FreeBytes = was })
+}
+
+// 1834/AC2: given 3 GiB free, when a check or a build is asked, then it refuses at once, saying so, running nothing.
+func TestACheckUnderTheFloorOfFreeSpaceRefusesSayingSo(t *testing.T) {
+	withFree(t, 3<<30)
+	var ran []string
+	run := func(_ context.Context, _ string, name string, args ...string) (string, error) {
+		ran = append(ran, name+" "+strings.Join(args, " "))
+		return "", nil
+	}
+	_, err := Check(context.Background(), run, CheckSpec{ID: "build-1"}, t.TempDir(), "registry.invalid", GitCredential{}, nil)
+	if err == nil || !strings.Contains(err.Error(), "the build node has 3 GiB free on /, under the floor of 20 GiB") {
+		t.Fatalf("a check on a full disk said %v", err)
+	}
+	if len(ran) != 0 {
+		t.Errorf("a refused check ran %v", ran)
+	}
+}
+
+func TestABuildUnderTheFloorOfFreeSpaceRefusesSayingSo(t *testing.T) {
+	withFree(t, 3<<30)
+	r := &recorded{}
+	_, err := Build(context.Background(), r.run, r, "https://forge.invalid/x.git", "", "", t.TempDir(), nil, Npmrc{}, GitCredential{}, nil)
+	if err == nil || !strings.Contains(err.Error(), "under the floor of 20 GiB") {
+		t.Fatalf("a build on a full disk said %v", err)
+	}
+	if len(r.ran) != 0 {
+		t.Errorf("a refused build ran %v", r.ran)
+	}
+}
+
+// The workspace's filesystem is read too, and a workspace not made yet is read at its nearest folder.
+func TestTheWorkspacesFilesystemIsReadToo(t *testing.T) {
+	was := FreeBytes
+	t.Cleanup(func() { FreeBytes = was })
+	missing := filepath.Join(t.TempDir(), "not", "yet")
+	FreeBytes = func(p string) (uint64, error) {
+		if p == "/" {
+			return 100 << 30, nil
+		}
+		if _, err := os.Stat(p); err != nil {
+			t.Errorf("read the free space of %s, which is not there", p)
+		}
+		return 1 << 30, nil
+	}
+	err := RoomToWork("/", missing)
+	if err == nil || !strings.Contains(err.Error(), "1 GiB free on "+missing) {
+		t.Errorf("the workspace's disk, under the floor, said %v", err)
+	}
+	FreeBytes = func(string) (uint64, error) { return 0, errors.New("no") }
+	if RoomToWork("/") == nil {
+		t.Error("a disk that cannot be read was taken as having room")
+	}
+}
+
+// 1834/AC3: given a store that stops at once, when the check waits for it, then its exit code and last log lines are
+// in the error and the container is removed afterwards, not by --rm before anyone could read it.
+func TestAStoreThatDiesSaysWhyItDied(t *testing.T) {
+	var ran []string
+	run := func(_ context.Context, _ string, name string, args ...string) (string, error) {
+		line := name + " " + strings.Join(args, " ")
+		ran = append(ran, line)
+		switch {
+		case args[0] == "exec":
+			return "", errors.New("Error response from daemon: container is not running")
+		case args[0] == "inspect":
+			return "exited exit code 1\n", nil
+		case args[0] == "logs":
+			return "initdb: error: could not create directory: No space left on device\n", nil
+		}
+		return "", nil
+	}
+	err := waitReady(context.Background(), run, "build-1-store", []string{"pg_isready"}, 1)
+	for _, want := range []string{"build-1-store never became ready", "exit code 1", "No space left on device"} {
+		if err == nil || !strings.Contains(err.Error(), want) {
+			t.Errorf("the verdict %q lacks %q", err, want)
+		}
+	}
+	logs := ""
+	for _, l := range ran {
+		if strings.HasPrefix(l, "docker logs") {
+			logs = l
+		}
+	}
+	if logs != "docker logs --tail 20 build-1-store" {
+		t.Errorf("the log was read as %q", logs)
+	}
+}
+
+func TestTheThrowawayStoreIsNotStartedWithRm(t *testing.T) {
+	var started []string
+	run := func(_ context.Context, _ string, name string, args ...string) (string, error) {
+		if args[0] == "run" {
+			started = args
+		}
+		return "127.0.0.1:40000\n", nil
+	}
+	if _, err := throwaway(context.Background(), run, run, "build-1-store", "postgres:17-alpine", 5432, nil, nil); err != nil {
+		t.Fatal(err)
+	}
+	for _, a := range started {
+		if a == "--rm" {
+			t.Errorf("started with --rm, which erases why it died: %v", started)
+		}
+	}
+}
+
+// 1834/AC1: given an image built for a check or a walk and published to the registry, when the build ends, then the
+// local image tag, the narrowed source and the context clone it left are gone; an image published nowhere (a mesh
+// being raised) is kept, since it is the artifact.
+func TestABuildRemovesWhatItLeftOnTheNode(t *testing.T) {
+	r := &recorded{
+		contents:  map[string]string{"modules/m/" + ManifestName: anImageWithAContext, "modules/m/Dockerfile": "FROM scratch\n"},
+		secondary: map[string]map[string]string{"https://forge.invalid/ctl.git": {"go.mod": "module x\n"}},
+	}
+	workspace := t.TempDir()
+	if _, err := Build(context.Background(), r.run, r, "https://forge.invalid/c.git", "modules/m", "", workspace, nil, Npmrc{},
+		GitCredential{}, nil); err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(strings.Join(r.ran, "\n"), "docker rmi m-server:c0ffeec0") {
+		t.Errorf("the image built for it was not removed: %v", r.ran)
+	}
+	left, _ := filepath.Glob(filepath.Join(workspace, "*"))
+	if len(left) != 0 {
+		t.Errorf("the build left %v in its workspace", left)
+	}
+}
+
+const anImageWithAContext = `{"module":"m","version":"1",
+  "build":{"artifacts":[{"name":"server","kind":"image","from":"Dockerfile",
+    "context":{"repository":"https://forge.invalid/ctl.git","ref":"main"}}]}}`
+
+func TestAnImageKeptLocallyIsNotRemoved(t *testing.T) {
+	r := &recorded{contents: map[string]string{"modules/m/" + ManifestName: anImageWithAContext, "modules/m/Dockerfile": "FROM scratch\n"}}
+	local := Local{Run: func(ctx context.Context, dir, name string, args ...string) (string, error) {
+		if args[0] == "image" {
+			return "sha256:" + strings.Repeat("a", 64) + "\n", nil
+		}
+		return r.run(ctx, dir, name, args...)
+	}}
+	if _, err := Build(context.Background(), local.Run, local, "https://forge.invalid/c.git", "modules/m", "", t.TempDir(), nil, Npmrc{},
+		GitCredential{}, nil); err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(strings.Join(r.ran, "\n"), "docker rmi") {
+		t.Errorf("the artifact of a mesh being raised was removed: %v", r.ran)
+	}
+}

internal/builder/floor.go

@@ -0,0 +1,46 @@
+package builder
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"syscall"
+)
+
+// FreeSpaceFloor is the free space a build node must have, on / and on its workspace's filesystem, before it takes
+// a build or a check (novox/hq issue 1834): under it, a throwaway store cannot initialise its data directory and
+// dies at once, and the check ended "never became ready" instead of saying the disk was full.
+const FreeSpaceFloor = 20 << 30
+
+// FreeBytes reads the bytes an unprivileged writer can still use on the filesystem holding path. A variable so a
+// test answers for a disk it does not have.
+var FreeBytes = func(path string) (uint64, error) {
+	var s syscall.Statfs_t
+	if err := syscall.Statfs(path, &s); err != nil {
+		return 0, err
+	}
+	return s.Bavail * uint64(s.Bsize), nil
+}
+
+// RoomToWork refuses, saying how much is free, when any of the paths' filesystems is under FreeSpaceFloor. A path
+// that does not exist yet (a workspace not made) is read at its nearest folder that does.
+func RoomToWork(paths ...string) error {
+	for _, p := range paths {
+		at := p
+		for {
+			if _, err := os.Stat(at); err == nil || filepath.Dir(at) == at {
+				break
+			}
+			at = filepath.Dir(at)
+		}
+		free, err := FreeBytes(at)
+		if err != nil {
+			return fmt.Errorf("the free space on %s cannot be read: %w", p, err)
+		}
+		if free < FreeSpaceFloor {
+			return fmt.Errorf("the build node has %d GiB free on %s, under the floor of %d GiB",
+				free>>30, p, FreeSpaceFloor>>30)
+		}
+	}
+	return nil
+}