Compare commits

..
Author SHA1 Message Date
jochen 635363adbd Run the controller as a Go bundle the host starts as a process (hq issue 213)
The controller is a Go program and was the one piece of the mesh's own Go
code still shipped and run as an image (novox/hq issue 213; ADR 0188 §1:
a module's own code is bundles; §3: a service bundle is a process).

The manifest now builds one Go bundle, `controller`, and runs it as the
process `mesh-controller` (`./mesh-controller serve`) under an account
the module declares. What the container gave it, replaced:

- host network: a process is on the host's network; nothing it reads
  names a container network
- user 65534: the account `mesh-controller`, which owns its secrets and
  its state directory
- the eight mounts: the env names the host paths the mesh already places
  (the store, broker and bus files under the state directory, the
  broker's certificate under /var/lib/mesh-broker-tls); the `broker`
  mount was read by nothing and is gone with the others
- `container-runtime` is no longer required on its machine

Its preparation is the same binary with `prepare`, as a run-once process,
and the process `replaces` the container `server`: the host keeps the
container answering until the process is running (mesh-host). Needs the
previous commit live in the running controller, and the host's
`replaces` on the controller's machine, before it is registered.

No image is built by the mesh any more. The Dockerfile stays for genesis
and the lab (`make image`, its Go base now pinned in the Makefile).
2026-10-04 01:11:31 +02:00
21 changed files with 54 additions and 1939 deletions
+1 -6
View File
@@ -1,11 +1,6 @@
# The Go it builds with, pinned here because genesis builds this file with no arguments (novox/hq
# issue 223) — the Makefile passes the same digest. A tag older than go.mod asks for is how
# `make image` broke once before (issue 146).
ARG GO_BASE=golang@sha256:8ac98ca534ac3f51e1f420a1dd2c15e74c75cfa0f23f3ad27eb5d7236c349a0c
ARG GO_BASE=golang:1.25-alpine
# The control plane's image — for genesis and the lab only. The mesh runs the controller as a Go
# bundle the host starts as a process (module.json; novox/hq issue 213), and builds no image of it.
# Genesis builds this file and raises it as the container the process replaces on the first push
# (mesh-host internal/bootstrap, novox/hq issue 223).
#
# novox/hq ADR 0006: this image is pinned by digest in the bundle the host carries, fetched on a
# machine where no mesh exists yet, and run before there is anything to check it against. So it
-4
View File
@@ -558,10 +558,6 @@ func takeIn(ctx context.Context, inv *inventory.Inventory, result link.BuildResu
}
return manifest, kept, err
}
// The keep set just moved, and new bytes just landed (novox/hq ADR 0189). Asked here rather
// than on a timer of its own: this is the only moment either is true. Never fatal — the build
// worked and the module is registered.
collect(ctx, inv)
return manifest, kept, nil
}
-108
View File
@@ -1,108 +0,0 @@
package main
import (
"context"
"errors"
"fmt"
"os"
"time"
"github.com/novox/mesh-controller/internal/artifacts"
"github.com/novox/mesh-controller/internal/inventory"
)
// Letting the artifact store go of what the mesh no longer keeps (novox/hq ADR 0189, issue 108).
//
// **Run where the records change.** A build is the moment new bytes landed in the store and the
// moment the keep set moved, so it is the moment to say what may go — and it needs no timer of
// its own. Reclaiming the bytes is the store's own nightly step; this only decides.
//
// Never fatal to a build. The build succeeded, the module is registered, and a store that could
// not be reached is a thing to say rather than a reason to undo any of that. The next build asks
// again, and the references it could not collect are still uncollected, so nothing is lost by
// having failed.
// collect asks the store to let go of everything the mesh made and no longer keeps, and records
// what it let go of. Says what it did and what it could not; returns nothing, because nothing
// upstream should branch on it.
func collect(ctx context.Context, inv *inventory.Inventory) {
references, err := inv.ToCollect(ctx)
if err != nil {
fmt.Fprintf(os.Stderr, "could not work out what the artifact store may let go of: %v\n", err)
return
}
if len(references) == 0 {
return
}
shelf, err := inv.Catalogue(ctx)
if err != nil {
fmt.Fprintf(os.Stderr, "could not read the catalogue to find the artifact store: %v\n", err)
return
}
// As the mesh reaches it from the network. Empty means the store is not on the network — on a
// mesh being raised it is not yet, and there the store holds one build of anything and has
// nothing to collect.
address, err := artifactStoreAddress(ctx, inv, shelf, "")
if err != nil || address == "" {
if err != nil {
fmt.Fprintf(os.Stderr, "could not find the artifact store to collect from: %v\n", err)
}
return
}
// **Bounded, because this runs inside somebody's build.** The first sweep of a mesh that has
// never collected has the whole history to get through, and a person waiting on `build` should
// not pay for it. Two bounds, and what is left over is simply offered again next time —
// builds are frequent, and the point is that the store stops growing, not that it empties
// tonight.
within, stop := context.WithTimeout(ctx, sweepBudget)
defer stop()
store := artifacts.Store{Address: address}
var done []string
var left int
for i, reference := range references {
if i >= mostPerSweep || within.Err() != nil {
left = len(references) - i
break
}
err := store.LetGo(within, reference)
if err == nil || errors.Is(err, artifacts.Gone) {
// Gone is the outcome wanted, already true. Recorded so the next sweep does not ask
// again for ever.
done = append(done, reference)
continue
}
// **Stopped at the first refusal, not pushed through.** A store that refuses one refuses
// all of them — deletion disabled, the store down, the network gone — so going on would
// be a hundred identical failures and a hundred identical log lines in front of whoever
// was building something.
fmt.Fprintf(os.Stderr, "the artifact store kept %s, so nothing more was asked of it: %v\n",
reference, err)
left = len(references) - i
break
}
if len(done) > 0 {
// Recorded outside `within`: the deletions happened, and losing the record of them because
// the sweep ran out of budget would mean asking about them again for ever.
if err := inv.MarkCollected(ctx, done); err != nil {
fmt.Fprintf(os.Stderr, "the store let go of %d artifact(s) and the record of it did not keep: %v\n",
len(done), err)
return
}
fmt.Fprintf(os.Stderr, "the artifact store let go of %d artifact(s) the mesh no longer keeps\n",
len(done))
}
if left > 0 {
fmt.Fprintf(os.Stderr, "%d more to collect; the next build asks again\n", left)
}
}
// mostPerSweep is how many artifacts one sweep will ask about. Enough that a mesh building
// several times a day converges within days of this landing; small enough that no single build
// waits on the whole backlog.
const mostPerSweep = 200
// sweepBudget is the longest a sweep will keep a build waiting.
const sweepBudget = 60 * time.Second
-99
View File
@@ -1,99 +0,0 @@
// Package artifacts speaks to the mesh's artifact store over its own door.
//
// Only what the mesh needs that nothing else does: letting go of something it put there
// (novox/hq ADR 0189, issue 108). Pushing is the builder's, through the container runtime; reading
// is every machine's, through its runtime. This is the one operation that belongs to the thing
// holding the records, because it is the only one that is a decision rather than a transfer.
package artifacts
import (
"context"
"fmt"
"net/http"
"strings"
"time"
"github.com/novox/mesh-controller/internal/catalogue"
)
// Store is the artifact store at an address, as this machine reaches it.
type Store struct {
// Address is `host:port` — the store as the caller reaches it now, composed and never
// recorded (novox/hq 04-ISSUES/102).
Address string
// HTTP is the client used; nil is a client with a modest timeout.
HTTP *http.Client
}
// Gone is the answer when the store does not hold it: the outcome wanted, already true.
var Gone = fmt.Errorf("the store does not hold it")
// LetGo asks the store to drop one artifact the mesh recorded making.
//
// Takes a reference as the mesh records it — `artifact-store://<module>/<artifact>@sha256:…` for
// an image, `…/blobs/sha256:…` for an archive — because that is the identity every record uses,
// and composes the address here at the moment of use.
//
// Returns Gone when the store answers that it does not have it. That is not a failure: the sweep
// wants the artifact absent, and it is. It is distinguished from success only so a caller can say
// which of the two happened.
func (s Store) LetGo(ctx context.Context, reference string) error {
path, kept := catalogue.InArtifactStore(reference)
if !kept {
// Nothing the mesh put in its own store. Refused rather than attempted: composing a
// delete for a reference of unknown shape is how a sweep reaches something that is not
// the mesh's.
return fmt.Errorf("%s is not a reference into the mesh's artifact store", reference)
}
if s.Address == "" {
return fmt.Errorf("this mesh has no artifact store on its network to ask about %s", reference)
}
repository, kind, digest, err := split(path)
if err != nil {
return err
}
url := "http://" + s.Address + "/v2/" + repository + "/" + kind + "/" + digest
request, err := http.NewRequestWithContext(ctx, http.MethodDelete, url, nil)
if err != nil {
return err
}
client := s.HTTP
if client == nil {
client = &http.Client{Timeout: 30 * time.Second}
}
response, err := client.Do(request)
if err != nil {
return err
}
defer response.Body.Close()
switch response.StatusCode {
case http.StatusAccepted, http.StatusOK, http.StatusNoContent:
return nil
case http.StatusNotFound:
return Gone
case http.StatusMethodNotAllowed:
// The registry was started without deletion enabled. Said plainly, because the remedy is
// a setting on the store's module and not anything about this artifact.
return fmt.Errorf(
"the artifact store refuses deletion: its server was started without it enabled "+
"(REGISTRY_STORAGE_DELETE_ENABLED), so nothing can be collected until the store "+
"module is applied again (novox/hq ADR 0189). Asking about %s", reference)
default:
return fmt.Errorf("the artifact store answered %s for %s", response.Status, reference)
}
}
// split reads a recorded path into the repository, which endpoint names the thing, and the digest.
//
// Two shapes, which are the two the mesh records: `<repository>@sha256:<hex>` is a manifest, and
// `<repository>/blobs/sha256:<hex>` is a blob.
func split(path string) (repository, kind, digest string, err error) {
if before, after, ok := strings.Cut(path, "@sha256:"); ok {
return before, "manifests", "sha256:" + after, nil
}
if before, after, ok := strings.Cut(path, "/blobs/sha256:"); ok {
return before, "blobs", "sha256:" + after, nil
}
return "", "", "", fmt.Errorf("%q names nothing the store holds by digest", path)
}
-94
View File
@@ -1,94 +0,0 @@
package artifacts
import (
"context"
"errors"
"net/http"
"net/http/httptest"
"strings"
"testing"
"github.com/novox/mesh-controller/internal/catalogue"
)
// Asking the store to let go of what the mesh no longer keeps (novox/hq ADR 0189, issue 108).
//
// A fake store records what it was asked to delete, so what is asserted is the mesh's decision
// and the shape of the request — not the registry's behaviour, which is the registry's to test.
func fakeStore(t *testing.T, answer int) (Store, *[]string) {
t.Helper()
var asked []string
server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
if r.Method != http.MethodDelete {
t.Errorf("the store was asked %s %s; collecting is a delete", r.Method, r.URL.Path)
}
asked = append(asked, r.URL.Path)
w.WriteHeader(answer)
}))
t.Cleanup(server.Close)
return Store{Address: strings.TrimPrefix(server.URL, "http://")}, &asked
}
func TestAnImageAndAnArchiveAreAskedForAtTheirOwnEndpoints(t *testing.T) {
// The two shapes the mesh records: a manifest by digest, and a blob by digest. They are
// different endpoints, and asking at the wrong one answers 404 — which this would then
// record as collected, leaving the bytes on disk for ever while the record says otherwise.
store, asked := fakeStore(t, http.StatusAccepted)
ctx := context.Background()
image := catalogue.ArtifactStoreScheme + "web/app@sha256:abc123"
archive := catalogue.ArtifactStoreScheme + "web/config/blobs/sha256:def456"
if err := store.LetGo(ctx, image); err != nil {
t.Fatal(err)
}
if err := store.LetGo(ctx, archive); err != nil {
t.Fatal(err)
}
want := []string{"/v2/web/app/manifests/sha256:abc123", "/v2/web/config/blobs/sha256:def456"}
if len(*asked) != 2 || (*asked)[0] != want[0] || (*asked)[1] != want[1] {
t.Fatalf("the store was asked %v; want %v", *asked, want)
}
}
func TestAStoreThatDoesNotHaveItAnswersGone(t *testing.T) {
// The outcome wanted, already true. Told apart from success only so the sweep can say which
// happened; both are recorded, because retrying for ever is the thing to avoid.
store, _ := fakeStore(t, http.StatusNotFound)
err := store.LetGo(context.Background(), catalogue.ArtifactStoreScheme+"web/app@sha256:abc123")
if !errors.Is(err, Gone) {
t.Fatalf("a store that does not hold it answered %v, want Gone", err)
}
}
func TestAStoreWithDeletionOffSaysSoAndNamesTheRemedy(t *testing.T) {
// The registry answers 405 when it was started without deletion enabled. The remedy is a
// setting on the store's module, and saying "405" would send somebody to the wrong place.
store, _ := fakeStore(t, http.StatusMethodNotAllowed)
err := store.LetGo(context.Background(), catalogue.ArtifactStoreScheme+"web/app@sha256:abc123")
if err == nil {
t.Fatal("a store that refuses deletion was read as success")
}
if !strings.Contains(err.Error(), "REGISTRY_STORAGE_DELETE_ENABLED") {
t.Fatalf("the refusal does not name the remedy: %v", err)
}
}
func TestAReferenceThatIsNotTheMeshsOwnIsNeverAsked(t *testing.T) {
// The whole safety of the sweep is that it names only what the mesh recorded putting there.
// A reference of another shape — a vendor's image, a package version — is refused rather
// than composed into a delete somewhere that is not the mesh's store.
store, asked := fakeStore(t, http.StatusAccepted)
for _, reference := range []string{
"docker.io/library/registry@sha256:abc123",
"registry@sha256:abc123",
"1.4.2",
} {
if err := store.LetGo(context.Background(), reference); err == nil {
t.Errorf("%s was asked about; it is not a reference into the mesh's store", reference)
}
}
if len(*asked) != 0 {
t.Fatalf("the store was asked about %v", *asked)
}
}
+2 -7
View File
@@ -215,7 +215,7 @@ func Build(ctx context.Context, run Runner, publish Publisher,
sort.Slice(artifacts, func(i, j int) bool { return artifacts[i].Name < artifacts[j].Name })
for _, a := range artifacts {
say("artifact", "%s (%s%s) — starting", a.Name, a.Kind, langSuffix(a))
made, err := one(ctx, run, publish, manifest.Module, within, workspace, commit, credentials, a, args, held, npmrcPath, npmrc, seatBases, say)
made, err := one(ctx, run, publish, manifest.Module, within, workspace, commit, credentials, a, args, held, npmrcPath, seatBases, say)
if err != nil {
say("artifact", "%s FAILED: %v", a.Name, err)
return Result{}, err
@@ -443,7 +443,7 @@ func wantsPackages(manifest catalogue.Manifest, within string) bool {
func one(ctx context.Context, run Runner, publish Publisher,
module, tree, workspace, commit, credentials string, a catalogue.Artifact, args []string,
held map[string]string, npmrc string, registry Npmrc, seats map[string]string,
held map[string]string, npmrc string, seats map[string]string,
say func(step, format string, args ...any)) (catalogue.Built, error) {
switch a.Kind {
@@ -582,11 +582,6 @@ func one(ctx context.Context, run Runner, publish Publisher,
"holds no copy of it. Build %s first",
module, a.Name, chain.Language, chain.Base, chain.Artifact, chain.Base)
}
// The module's own packages first, where the compiler and the bundler resolve them from
// (dependencies.go); nothing at all for a module whose package.json names only the SDK.
if err := installOwn(ctx, run, tree, chain, base, registry, say); err != nil {
return catalogue.Built{}, fmt.Errorf("%s: %s: %w", module, a.Name, err)
}
say("bundle", "compiling %s in %s's toolchain", a.Language, chain.Base)
compiled, err := compile(ctx, run, tree, chain, base, a)
if err != nil {
-147
View File
@@ -1,147 +0,0 @@
package builder
import (
"context"
"encoding/json"
"errors"
"fmt"
"os"
"path/filepath"
"sort"
"strings"
)
// A module's own packages, installed before its bundle is compiled, so the bundler inlines them.
//
// **A bundle could only import what the toolchain happened to carry.** The compiler and the bundler
// resolve an import by walking up from the module's source: the module's own directory first, then
// the toolchain image's node_modules. Nothing ever put anything in the first, so a module needing a
// database driver (`pg`, `mongodb`, `mssql`) could not be a bundle at all, and kept a container whose
// recipe installed it by hand (novox/hq ADR 0198 §4: "the backend's own driver inside the bundle").
// Now the module's `package.json` says what it depends on, as any Node package does, and the build
// installs exactly that into the module's own directory before compiling.
//
// **The SDK the toolchain carries is the one a bundle is built with, whatever the module says**
// (novox/hq issue 212: the toolchain is rebuilt on every SDK release and every bundle after it). A
// module's `package.json` names `@novox/mesh-sdk` with a range — it has to, to type-check on a
// workstation — and installing that range would shadow the toolchain's copy for this module alone:
// one module compiled against an older SDK than its neighbours, chosen by a caret nobody re-reads.
// So the SDK is taken out of what is installed (and never fetched), and any copy something else
// pulls in is removed afterwards; every import of it resolves past the module's node_modules to the
// toolchain's. A module therefore cannot pin a different SDK, by design: the toolchain is the pin.
//
// **Correctness before speed.** Every build installs afresh into a fresh clone, from the lockfile
// when the module has one (`npm ci`, exact) and from its ranges otherwise; nothing installed is kept
// between builds. What is shared is npm's own download cache, a named volume, which is
// content-addressed and verified by integrity on every read — it saves the network, never the
// install. Install scripts do not run: the build node runs nobody's postinstall, and what a script
// would build natively could not be inlined into one file anyway.
// sdkPackage is the package a TypeScript bundle's launcher serves through, and the one package a
// module's own dependencies never supply (above).
const sdkPackage = "@novox/mesh-sdk"
// npmCache is the named volume npm's download cache lives in across builds on one build node.
const npmCache = "mesh-builder-npm-cache"
// ownDependencies is what a module's package.json depends on beyond the SDK, sorted; nothing when
// the module has no package.json or depends on nothing else — which builds exactly as before.
func ownDependencies(tree string) ([]string, error) {
raw, err := os.ReadFile(filepath.Join(tree, "package.json"))
if errors.Is(err, os.ErrNotExist) {
return nil, nil
}
if err != nil {
return nil, err
}
var p struct {
Dependencies map[string]string `json:"dependencies"`
}
if err := json.Unmarshal(raw, &p); err != nil {
return nil, fmt.Errorf("the module's package.json is not JSON: %w", err)
}
var names []string
for name := range p.Dependencies {
if name != sdkPackage {
names = append(names, name)
}
}
sort.Strings(names)
return names, nil
}
// installSteps is the script run inside the toolchain image, from the module's own directory ($0).
// It works in a scratch copy so the module's package.json and lockfile are never rewritten, takes
// the SDK out of what is installed, installs production dependencies only, removes any copy of the
// SDK something pulled in, and puts the result at the module's node_modules.
const installSteps = `set -e
work="$(mktemp -d)"
cp "$0/package.json" "$work/"
if [ -f "$0/package-lock.json" ]; then cp "$0/package-lock.json" "$work/"; fi
cd "$work"
node -e '
const fs = require("fs"), sdk = process.argv[1];
const p = JSON.parse(fs.readFileSync("package.json", "utf8"));
for (const k of ["dependencies", "peerDependencies", "optionalDependencies"]) if (p[k]) delete p[k][sdk];
delete p.devDependencies; delete p.scripts;
fs.writeFileSync("package.json", JSON.stringify(p));
' "$1"
shift
if [ -f package-lock.json ]; then
npm ci --omit=dev --omit=peer --ignore-scripts --no-audit --no-fund "$@"
else
npm install --omit=dev --omit=peer --ignore-scripts --no-audit --no-fund --no-package-lock "$@"
fi
find node_modules -depth -type d -path "*/node_modules/@novox/mesh-sdk" -exec rm -rf {} +
rm -rf "$0/node_modules"
cp -a node_modules "$0/node_modules"
`
// installOwn installs a TypeScript module's own production dependencies into its directory, in the
// toolchain image, before the compile — or does nothing at all for a module that has none.
func installOwn(ctx context.Context, run Runner, tree string, chain Toolchain, base string,
registry Npmrc, say func(step, format string, args ...any)) error {
if chain.Language != "typescript" {
return nil
}
deps, err := ownDependencies(tree)
if err != nil || len(deps) == 0 {
return err
}
scoped := strings.TrimSpace(registry.Scope)
if !registry.Enabled() {
// **No registry, no scoped package.** Without the mesh's registry a scoped name resolves on
// the public one, where anybody may have published it: a dependency that installs is not
// the dependency the module meant.
for _, d := range deps {
if strings.HasPrefix(d, "@novox/") {
return fmt.Errorf("the module depends on %s, and this build knows no package registry "+
"for its scope; it would resolve from the public registry, which is not where the "+
"mesh publishes it", d)
}
}
}
const within = "/app/modules/module"
invocation := []string{"run", "--rm",
"--volume", tree + ":" + within,
"--volume", npmCache + ":/root/.npm",
"--workdir", within}
var flags []string
if registry.Enabled() {
// The registry is reached where the binding says it is, which may be this machine's own
// loopback — the reason an image build that resolves packages runs on the host network too.
invocation = append(invocation, "--network", "host")
reg := strings.TrimSpace(registry.Registry)
if !strings.HasSuffix(reg, "/") {
reg += "/"
}
flags = append(flags, "--"+scoped+":registry="+reg)
}
invocation = append(invocation, base, "sh", "-c", installSteps, within, sdkPackage)
invocation = append(invocation, flags...)
say("bundle", "installing the module's own packages: %s", strings.Join(deps, ", "))
if _, err := run(ctx, tree, "docker", invocation...); err != nil {
return fmt.Errorf("installing the module's own packages (%s): %w", strings.Join(deps, ", "), err)
}
return nil
}
-131
View File
@@ -1,131 +0,0 @@
package builder
import (
"context"
"strings"
"testing"
)
// A module's own packages (dependencies.go): installed into its own directory, in the toolchain,
// before the compile, so the bundler inlines them — the SDK always the toolchain's.
func buildWithPackageJSON(t *testing.T, pkg string, extra map[string]string, registry Npmrc) (*recorded, error) {
t.Helper()
files := map[string]string{"index.ts": "console.log(1)"}
if pkg != "" {
files["package.json"] = pkg
}
for k, v := range extra {
files[k] = v
}
r, workspace := aRepository(t, aBundle, files)
held := map[string]string{"mesh-tools/build": "registry.invalid/mesh-tools/build@sha256:" + strings.Repeat("b", 64)}
_, err := Build(context.Background(), compiling{r}.run, r,
"https://forge.invalid/greeter.git", "", "", workspace, held, registry, GitCredential{}, nil)
return r, err
}
func installs(r *recorded) []string {
var out []string
for _, line := range r.ran {
if strings.HasPrefix(line, "docker run") && strings.Contains(line, "npm ci") {
out = append(out, line)
}
}
return out
}
func compileIndex(r *recorded) int {
for i, line := range r.ran {
if strings.Contains(line, "--outDir") {
return i
}
}
return -1
}
func TestAModulesOwnPackagesAreInstalledInTheToolchainBeforeTheCompile(t *testing.T) {
r, err := buildWithPackageJSON(t, `{"type":"module","dependencies":{"@novox/mesh-sdk":"^0.1.0","pg":"^8"},"devDependencies":{"typescript":"^5"}}`,
nil, Npmrc{Scope: "@novox", Registry: "https://forge.invalid/api/packages/novox/npm"})
if err != nil {
t.Fatal(err)
}
got := installs(r)
if len(got) != 1 {
t.Fatalf("want one install of the module's own packages:\n%s", strings.Join(r.ran, "\n"))
}
line := got[0]
for _, want := range []string{
"mesh-tools/build@sha256:", // in the toolchain image
":/app/modules/module", // into the module's own directory
"--workdir /app/modules/module", //
npmCache + ":/root/.npm", // npm's verified download cache, and only that
"--omit=dev", "--ignore-scripts", // production packages, no build-node scripts
"npm ci", "npm install", "--no-package-lock", // the lockfile when there is one, else the ranges
"--@novox:registry=https://forge.invalid/api/packages/novox/npm/", // the scope from the mesh's registry
"--network host",
"@novox/mesh-sdk", // named, to be taken out of what is installed
} {
if !strings.Contains(line, want) {
t.Errorf("the install lacks %q:\n%s", want, line)
}
}
// The SDK is the toolchain's: never installed from the module's range, and any copy removed.
if !strings.Contains(line, `delete p[k][sdk]`) || !strings.Contains(line, `-path "*/node_modules/@novox/mesh-sdk" -exec rm -rf`) {
t.Errorf("the module's own SDK range could shadow the toolchain's SDK:\n%s", line)
}
if i, c := strings.Index(strings.Join(r.ran, "\n"), "npm ci"), compileIndex(r); c < 0 ||
i > strings.Index(strings.Join(r.ran, "\n"), "--outDir") {
t.Fatalf("the install did not run before the compile:\n%s", strings.Join(r.ran, "\n"))
}
}
// **A module with nothing beyond the SDK builds exactly as before**: the same commands, no install.
func TestAModuleDependingOnlyOnTheSDKBuildsExactlyAsBefore(t *testing.T) {
without, err := buildWithPackageJSON(t, "", nil, Npmrc{})
if err != nil {
t.Fatal(err)
}
for _, pkg := range []string{
`{"type":"module","dependencies":{"@novox/mesh-sdk":"^0.1.0"},"devDependencies":{"typescript":"^5"}}`,
`{"type":"module"}`,
} {
with, err := buildWithPackageJSON(t, pkg, map[string]string{"package-lock.json": "{}"}, Npmrc{Scope: "@novox", Registry: "https://forge.invalid/npm/"})
if err != nil {
t.Fatal(err)
}
if strings.Contains(strings.Join(with.ran, "\n"), "npm ") {
t.Fatalf("a module depending on nothing but the SDK ran npm:\n%s", strings.Join(with.ran, "\n"))
}
if len(with.ran) != len(without.ran) {
t.Fatalf("a module depending only on the SDK built differently from one with no package.json:\n%s\n---\n%s",
strings.Join(with.ran, "\n"), strings.Join(without.ran, "\n"))
}
}
}
// Without the mesh's registry a scoped package would resolve on the public one: refused by name.
func TestAScopedPackageWithNoRegistryIsRefused(t *testing.T) {
r, err := buildWithPackageJSON(t, `{"dependencies":{"@novox/mesh-sdk":"^0.1.0","@novox/other":"^1"}}`, nil, Npmrc{})
if err == nil || !strings.Contains(err.Error(), "@novox/other") {
t.Fatalf("a scoped package was installed with no registry for its scope: %v", err)
}
if strings.Contains(strings.Join(r.ran, "\n"), "--outDir") {
t.Fatal("the compile ran after the refusal")
}
// A public package installs without one, from the public registry and nothing else.
r, err = buildWithPackageJSON(t, `{"dependencies":{"mssql":"^11"}}`, nil, Npmrc{})
if err != nil {
t.Fatal(err)
}
if got := installs(r); len(got) != 1 || strings.Contains(got[0], ":registry=") || strings.Contains(got[0], "--network host") {
t.Fatalf("a public package's install: %v", got)
}
}
func TestAnUnreadablePackageJSONIsRefusedByName(t *testing.T) {
_, err := buildWithPackageJSON(t, `{"dependencies":`, nil, Npmrc{})
if err == nil || !strings.Contains(err.Error(), "package.json") {
t.Fatalf("a broken package.json was not refused by name: %v", err)
}
}
+2 -3
View File
@@ -67,9 +67,8 @@ type Toolchain struct {
// `package.json` saying `"type": "module"` — Node reads a bare `.js` as CommonJS otherwise, so a
// bundle with its dependencies and without that line still fails to start — and the pruned,
// production-only node_modules the runtime itself ships with: the SDK's and the runtime's
// dependencies, and nothing module-specific (a module's own npm dependencies are installed into
// its own directory before the compile and inlined by the bundler: dependencies.go). Empty for a
// language whose bundle carries its own —
// dependencies, and nothing module-specific yet (novox/hq ADR 0188 §5: a skeleton; a module's
// own npm dependencies are a later step). Empty for a language whose bundle carries its own —
// a Go binary is static, a Python bundle is installed with its dependencies.
//
// A toolchain image without the directory fails the build by name rather than packing a bundle
+4 -12
View File
@@ -46,7 +46,7 @@ func boundUsed(content string) [][2]string {
// Three facts the mesh states about any provision, plus whatever the provider said it serves. A
// module may not reach a binding it does not have — the same boundary as a secret, for the same
// reason.
func knownFor(m Manifest, needs []Needed, node string) (map[string]map[string]string, error) {
func knownFor(m Manifest, needs []Needed, node string) map[string]map[string]string {
out := map[string]map[string]string{}
for _, want := range m.Wants() {
for i := range needs {
@@ -54,20 +54,12 @@ func knownFor(m Manifest, needs []Needed, node string) (map[string]map[string]st
if n.Name != want || n.For != m.Module {
continue
}
as := ConsumerIdentity(node, IdentitySource(m.Slug, m.Module))
values := map[string]string{
"at": n.At,
"from": n.From,
"as": as,
"as": ConsumerIdentity(node, IdentitySource(m.Slug, m.Module)),
}
// What the provider derives for this consumer rather than for all of them
// (novox/hq ADR 0201). Filled here, the one place a provision and the module
// requiring it are both in hand.
served, err := ServedTo(n.Serves, as)
if err != nil {
return nil, fmt.Errorf("%s requires %s: %w", m.Module, want, err)
}
for key, value := range served {
for key, value := range n.Serves {
// The provider's own vocabulary. Rendered plainly: a port is 5432, not 5432.000000,
// which is what a float would write and what a connection string would refuse.
values[key] = plainly(value)
@@ -75,7 +67,7 @@ func knownFor(m Manifest, needs []Needed, node string) (map[string]map[string]st
out[want] = values
}
}
return out, nil
return out
}
// withOwnNames adds a module's own composed names to what it may name from one binding:
-311
View File
@@ -1,311 +0,0 @@
package catalogue
import (
"fmt"
"regexp"
"sort"
"strings"
)
// What a provider derives for one consumer, said once in the provider's definition and delivered
// to both ends (novox/hq ADR 0201, issue 124).
//
// A `serves` block is otherwise literal: the same values for every consumer. Where the provider
// *names the resource* — a bucket, a database, a vhost — the name is derived from who is asking,
// and before this the mesh had no channel for it. The provider recomputed it in its own code and
// every consumer transcribed it into its own definition by hand, which is a copy of somebody
// else's rule kept in agreement by nobody. One of three transcriptions was wrong for months.
//
// **The mesh learns no protocol here; it spells its own name in an alphabet it already knows.**
// The only fact a served value may name is the identity the mesh itself minted for the consumer,
// in one of two alphabets: as it was minted, and as a DNS label. Everything a provider wants
// around it — a prefix, a suffix, a separator — it writes around the placeholder, because a
// served value is a string.
// consumerFact is `${consumer:<fact>}` or `${consumer:<fact>:<alphabet>}`.
var consumerFact = regexp.MustCompile(`\$\{consumer:([a-z][a-z0-9-]*)(?::([a-z][a-z0-9-]*))?\}`)
// consumerFacts are what a served value may name about the consumer it is being derived for.
// One entry, deliberately: the identity is the one thing about a consumer the mesh itself chose,
// so it is the one thing the mesh can hand to a provider without either end guessing.
var consumerFacts = []string{"as"}
// consumerAlphabets are the ways the mesh will write that identity. `dns` is the mesh's own
// identifier with its separator written `-` instead of `_` — the whole of the difference between
// the alphabet the mesh mints in and the one buckets, vhosts and hostnames accept.
var consumerAlphabets = []string{"dns"}
// ServedTo fills a provider's served values for one consumer.
//
// `as` is the identity the mesh minted for that consumer — the same string it is told to present
// as a login. Values with no placeholder are returned exactly as they were, and a block with no
// placeholder at all is returned unchanged, so this costs nothing for the providers that derive
// nothing.
//
// Only strings carry placeholders. A number, a boolean or a nested object is a value the provider
// stated outright, and is left alone.
func ServedTo(serves map[string]any, as string) (map[string]any, error) {
if len(serves) == 0 {
return serves, nil
}
var out map[string]any
for _, key := range sortedAnyKeys(serves) {
text, ok := serves[key].(string)
if !ok || !strings.Contains(text, "${consumer:") {
continue
}
filled, err := consumerInto(text, as)
if err != nil {
return nil, fmt.Errorf("the value served as %q: %w", key, err)
}
if out == nil {
// Copied only once something actually changes: the caller's map is the manifest's,
// and a provider that derives nothing must not have it rewritten underneath it.
out = make(map[string]any, len(serves))
for k, v := range serves {
out[k] = v
}
}
out[key] = filled
}
if out == nil {
return serves, nil
}
return out, nil
}
// consumerInto replaces every `${consumer:…}` in one value.
//
// **A fact or an alphabet the mesh does not have is refused, not left standing.** Written through,
// the literal `${consumer:as}` would reach a configuration file and be read as a bucket name,
// failing somewhere that names neither the module nor the mesh — the same reasoning `${bound:…}`
// is refused by (boundInto).
func consumerInto(value, as string) (string, error) {
var failed error
out := consumerFact.ReplaceAllStringFunc(value, func(match string) string {
parts := consumerFact.FindStringSubmatch(match)
fact, alphabet := parts[1], parts[2]
if fact != "as" {
if failed == nil {
failed = fmt.Errorf(
"says %s, and the mesh states %s about a consumer", match, orNothing(consumerFacts))
}
return match
}
switch alphabet {
case "":
return as
case "dns":
return asDNSLabel(as)
default:
if failed == nil {
failed = fmt.Errorf(
"says %s, and the mesh writes an identity as %s", match, orNothing(consumerAlphabets))
}
return match
}
})
if failed != nil {
return "", failed
}
return out, nil
}
// asDNSLabel writes a minted identity as a DNS label.
//
// The mesh's identities are already lower-case letters, digits and `_` (ConsumerIdentity), and
// already short enough for the tightest backend they reach (CheckIdentity, twenty characters). So
// this is the separator and nothing else — no lower-casing of what is already lower case, no
// truncation to a limit the identity is already inside, no padding of a name that is already long
// enough. Each of those would be the mesh guessing at a rule it has not been given.
func asDNSLabel(as string) string {
return strings.ReplaceAll(as, "_", "-")
}
// CheckServes refuses a `serves` block that names a consumer fact or an alphabet the mesh does not
// have, when the definition is parsed rather than when a consumer is resolved.
//
// A provision nobody consumes yet still has its rule read: a definition that would be refused the
// first time somebody required it is a definition that is wrong now.
func CheckServes(m Manifest) []string {
var problems []string
for _, provision := range sortedServes(m.Serves) {
for _, key := range sortedAnyKeys(m.Serves[provision]) {
text, ok := m.Serves[provision][key].(string)
if !ok {
continue
}
// A probe identity, because what is checked is the shape of the statement and not
// what any consumer is called.
if _, err := consumerInto(text, "mesh_node_module"); err != nil {
problems = append(problems, fmt.Sprintf(
"%s serves %s, and the value it serves as %q %s", m.Module, provision, key, err))
}
}
}
return problems
}
func sortedServes(serves map[string]map[string]any) []string {
out := make([]string, 0, len(serves))
for k := range serves {
out = append(out, k)
}
sort.Strings(out)
return out
}
func sortedAnyKeys(values map[string]any) []string {
out := make([]string, 0, len(values))
for k := range values {
out = append(out, k)
}
sort.Strings(out)
return out
}
// derivedFor is what the provider on this machine derives for one consumer of one provision
// (novox/hq ADR 0201).
//
// Settled first, then derived: an operator may set a prefix on what the provider serves and the
// mesh still fills the consumer's half of it ([ADR 0174]). Only the keys that actually name the
// consumer are returned — the rest of a `serves` block is the same for every consumer and is
// already in the provider's own definition, so repeating it here would be a second copy to go
// stale.
//
// The first module in the resolved order that says it serves the provision answers, which is the
// choice servedOnThisMachine makes for the consumer's half. Nothing serving it on this machine is
// not an error: a contribution can reach a machine whose provider is a record or an adapter, and
// then there is nothing derived to tell.
func (r Resolution) derivedFor(provision, as, consumer, local string, settings SettingsBy) (map[string]any, error) {
for _, m := range r.Modules {
serves, said := m.Serves[provision]
if !said {
continue
}
var names map[string]any
for key, value := range serves {
if text, ok := value.(string); ok && strings.Contains(text, "${consumer:") {
if names == nil {
names = map[string]any{}
}
names[key] = value
}
}
if names == nil {
return nil, nil
}
// **A consumer that keeps several holders of this provision is refused** — this is issue
// 124's own failure one case to the side, and it would be just as quiet.
//
// Each holder gets its own login, `…_<local>` (ADR 0094), and a provider derives from the
// login, so it would make one resource per holder. The consumer's side has no such
// dimension: one binding file per provision, one `${bound:<provision>:<key>}`, both
// derived from the un-suffixed identity. So the provider would create the holder's
// resource and the consumer would be configured against a name nothing made — it would
// authenticate successfully and be refused on every object, which reads like a credential
// fault and is not one.
//
// Lifting this means giving the consumer's side a local dimension. That is a decision,
// not an omission, and until it is taken the mesh says so rather than guessing.
if local != "" {
return nil, fmt.Errorf(
"%s keeps several holders of %s (this one is %q), and %s derives %s for each "+
"consumer from the login the mesh minted. Each holder has its own login, and a "+
"consumer is told one value per requirement — so the two ends would name "+
"different things and nothing would compare them (novox/hq ADR 0201)",
consumer, local, provision, m.Module, orNothing(sortedAnyKeys(names)))
}
settled, err := Settle(names, settings[m.Module])
if err != nil {
return nil, fmt.Errorf("%s serving %s: %w", m.Module, provision, err)
}
derived, err := ServedTo(settled, as)
if err != nil {
return nil, fmt.Errorf("%s serving %s to %s: %w", m.Module, provision, as, err)
}
return derived, nil
}
return nil, nil
}
// notTranscribed refuses a consumer's file that writes out the value its provider derives for it,
// instead of asking for it (novox/hq ADR 0201, issue 124).
//
// **What would have caught the one wrong instance.** The object store's three consumers each wrote
// their bucket into their own configuration by hand. One of them named a predecessor's bucket, and
// nothing compared it to what the provider would actually create: the module would have
// authenticated successfully and been refused on every object, which reads like a credential fault
// and is not one. It looked authoritative for months.
//
// The test is exact and costs one string search: a definition whose file already contains the
// value the mesh is about to derive for it has written down somebody else's rule. It cannot be a
// coincidence — a derived value carries the identity the mesh minted for this very consumer on
// this very machine, which nothing else would spell out — and it cannot be checked afterwards,
// because after substitution every consumer's file contains it legitimately.
//
// Only values that actually name the consumer are judged. A provider that serves a constant under
// the same key serves the same constant to everyone, and a consumer repeating it is redundant
// rather than wrong.
func notTranscribed(resource map[string]any, known map[string]map[string]string, module string) error {
if fmt.Sprint(resource["type"]) != "file" {
return nil
}
content, ok := resource["content"].(string)
if !ok || content == "" {
return nil
}
for _, provision := range sortedKnown(known) {
values := known[provision]
identity := values["as"]
if identity == "" {
continue
}
for _, key := range sortedStringKeys(values) {
if key == "as" {
// The login is not derived from itself, and a consumer that must present it in a
// connection string legitimately has it from `${bound:…}` — which is what it will
// be after substitution, so this would judge the substitution, not the module.
continue
}
value := values[key]
if value == "" || !namesTheConsumer(value, identity) {
continue
}
if !strings.Contains(content, value) {
continue
}
return fmt.Errorf(
"%s writes %q into %v, and that is exactly what %s derives for it — a definition "+
"keeping its own copy of somebody else's naming rule is one that can disagree "+
"with it, silently. Say ${bound:%s:%s} and be told",
module, value, resource["id"], provision, provision, key)
}
}
return nil
}
// namesTheConsumer is whether a derived value was built from this consumer's identity — in the
// alphabet it was minted in, or as a DNS label. A value that does not contain it was not derived
// from it, whatever else it may be.
func namesTheConsumer(value, identity string) bool {
return strings.Contains(value, identity) || strings.Contains(value, asDNSLabel(identity))
}
func sortedKnown(known map[string]map[string]string) []string {
out := make([]string, 0, len(known))
for k := range known {
out = append(out, k)
}
sort.Strings(out)
return out
}
func sortedStringKeys(values map[string]string) []string {
out := make([]string, 0, len(values))
for k := range values {
out = append(out, k)
}
sort.Strings(out)
return out
}
+5 -50
View File
@@ -645,16 +645,7 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
if err != nil {
return nil, err
}
as := ConsumerIdentity(r.Node, IdentitySource(m.Slug, m.Module))
// What the provider derives for THIS consumer, filled here where the consumer is
// known (novox/hq ADR 0201). The same fill knownFor does below, so the binding file
// and the module's `${bound:…}` substitutions cannot say different things.
told := *found
told.Serves, err = ServedTo(told.Serves, as)
if err != nil {
return nil, fmt.Errorf("%s is told about %s: %w", m.Module, to, err)
}
file, err := boundFile(told, m.Binds[to], as, own)
file, err := boundFile(*found, m.Binds[to], ConsumerIdentity(r.Node, IdentitySource(m.Slug, m.Module)), own)
if err != nil {
return nil, err
}
@@ -728,10 +719,7 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
return nil, err
}
// And what its bindings say, for the half of a connection that is not secret.
known, err := knownFor(m, r.Needs, r.Node)
if err != nil {
return nil, err
}
known := knownFor(m, r.Needs, r.Node)
// A requirement answered on this same machine is not in r.Needs — its binding file is
// written from `here` (above) — and so `${bound:…}` could not name it, though the file
// beside it said the same facts. Filled from the same answer, so the two cannot disagree.
@@ -748,11 +736,7 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
}
local := *answered
local.For = m.Module
here, err := knownFor(m, []Needed{local}, r.Node)
if err != nil {
return nil, err
}
for provision, values := range here {
for provision, values := range knownFor(m, []Needed{local}, r.Node) {
known[provision] = values
}
}
@@ -777,17 +761,6 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
// And the machine underneath, which no binding of its own can tell it.
thisMachine := machineFacts(r, with.Names, with.MeshRange)
// **A definition that already holds the answer transcribed it** (novox/hq ADR 0201).
// Judged over what the module itself declares, and before anything is substituted: the
// mesh's own generated files — the binding, the contributions — legitimately carry the
// derived value, and after substitution so does every consumer's file, so this is the one
// moment the two can be told apart.
for _, own := range m.Resources {
if err := notTranscribed(own, known, m.Module); err != nil {
return nil, err
}
}
// Which of this module's files carry a secret, for the rule that a container may not read
// one of them as its environment without saying so (ADR 0086, issue 041).
secretFiles := secretFilesOf(resources)
@@ -1202,19 +1175,6 @@ type Contribution struct {
// requirement's name — everything providing `reverse-proxy` understands the same shape, which
// is what makes swapping one for another cost nothing.
Values map[string]any `json:"values"`
// Derived is what this provider's own definition said it derives for this consumer, already
// derived (novox/hq ADR 0201).
//
// **The provider is told, rather than recomputing it.** A served value may name the consumer's
// identity — a bucket named for who is asking, a database prefixed with it — and before this
// the rule lived twice: once in the provisioner's code, once transcribed into every consumer's
// definition. The mesh fills the provider's own statement here and delivers the same filled
// value to the consumer, so the two cannot disagree: there is no second computation to
// disagree with.
//
// Only the keys that are per-consumer. The rest of what the provider serves is the same for
// everyone and is in its own definition, where it already is.
Derived map[string]any `json:"derived,omitempty"`
}
// grantPath is where one consumer's sealed credential lands on the providing machine.
@@ -1306,17 +1266,12 @@ func (r Resolution) contributions(settings SettingsBy, grants []Grant,
// told about it and withdraws the login on its next pass.
continue
}
as := holderAs(ConsumerIdentity(g.Consumer, IdentitySource(g.Slug, g.From)), g.Local)
derived, err := r.derivedFor(g.Provision, as, g.From, g.Local, settings)
if err != nil {
return nil, err
}
out[g.Provision] = append(out[g.Provision], Contribution{
From: g.From, Node: g.Consumer, At: g.At, Values: g.Values, Derived: derived,
From: g.From, Node: g.Consumer, At: g.At, Values: g.Values,
// One holder per local name: the identity the consumer is known by, and the local name
// after it where the module keeps several (ADR 0094). Not a login any backend checks —
// a secret is not a login — so the identity limit does not apply to the suffix.
As: as,
As: holderAs(ConsumerIdentity(g.Consumer, IdentitySource(g.Slug, g.From)), g.Local),
Secret: grantPath(directories[g.Provision], g.Consumer, holderAs(g.From, g.Local)),
})
if granted[g.Provision] == nil {
@@ -1,343 +0,0 @@
package catalogue
import (
"encoding/json"
"strings"
"testing"
)
// What a provider derives for each consumer, said once and delivered to both ends
// (novox/hq ADR 0201, issue 124).
//
// The failure these are written against: the object store's provisioner derived each consumer's
// bucket from the login the mesh minted, in its own code, and the mesh had no channel to tell the
// consumer which bucket that was — so all three consumers wrote the answer into their own
// definitions by hand. Two were right. One named a predecessor's bucket and would have
// authenticated successfully and been refused on every object. Each of them also named the
// machine the module happens to run on, which a definition may not do.
// store is an object store in the shape minio has: it serves a region and a port to everyone, and
// a bucket named for whoever is asking.
func store() Manifest {
return Manifest{
Module: "store", Version: "1",
Provides: FromAnywhere("s3-bucket"),
Listens: []Listening{{Port: 9000, Protocol: "tcp", From: FromMesh}},
Serves: map[string]map[string]any{"s3-bucket": {
"region": "eu-west",
"bucket": "${consumer:as:dns}",
}},
Receives: map[string]string{"s3-bucket": "/var/lib/store/grants/mesh.json"},
Grants: map[string]string{"s3-bucket": "/var/lib/store/grants"},
Resources: []map[string]any{{
"id": "server", "type": "container", "name": "store", "ports": []any{"9000"},
}},
}
}
// files is a consumer that writes the bucket into its own configuration — which is the thing it
// could not do before, and had to transcribe.
func files() Manifest {
return Manifest{
Module: "files", Version: "1", Slug: "files",
Requires: []string{"s3-bucket"},
Binds: map[string]string{"s3-bucket": "/var/lib/files/store.json"},
Secrets: map[string]string{"s3-bucket": "/var/lib/files/store.secret"},
Resources: []map[string]any{{
"id": "env", "type": "file", "path": "/var/lib/files/env", "mode": "0600",
"content": "BUCKET=${bound:s3-bucket:bucket}\nREGION=${bound:s3-bucket:region}\n",
}},
}
}
// pics is a second consumer of the same provider on the same machine: two derivations, neither
// the other's.
func pics() Manifest {
return Manifest{
Module: "pics", Version: "1", Slug: "pics",
Requires: []string{"s3-bucket"},
Binds: map[string]string{"s3-bucket": "/var/lib/pics/store.json"},
Secrets: map[string]string{"s3-bucket": "/var/lib/pics/store.secret"},
Resources: []map[string]any{{
"id": "env", "type": "file", "path": "/var/lib/pics/env", "mode": "0600",
"content": "BUCKET=${bound:s3-bucket:bucket}\n",
}},
}
}
// The three places the derived value lands must agree, because agreeing is the whole point: the
// consumer's own file, the binding it reads as JSON, and the provider's contributions entry.
func TestADerivedValueReachesBothEndsAndAgrees(t *testing.T) {
r, err := Resolve(shelf(store(), files()), []string{"store", "files"}, reachable(), World{})
if err != nil {
t.Fatal(err)
}
out, err := r.Declaration(Rendering{Grants: []Grant{{
Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
Values: map[string]any{}, Sealed: "c2VhbGVk",
}}})
if err != nil {
t.Fatal(err)
}
// The mesh minted this identity for the consumer; the bucket is that identity as a DNS label.
// Derived here with the mesh's own function, so the test cannot agree with a wrong rule.
as := ConsumerIdentity("workstation", IdentitySource("files", "files"))
want := strings.ReplaceAll(as, "_", "-")
if want == as || !strings.Contains(as, "_") {
t.Fatalf("the mesh's identity %q has no separator to rewrite; this test proves nothing", as)
}
env := fileNamed(out, "files.env")
if env == nil {
t.Fatalf("the consumer was given no file: %v", out)
}
if got := env["content"].(string); !strings.Contains(got, "BUCKET="+want+"\n") {
t.Errorf("the consumer's own file was not told the bucket:\n%s\nwant BUCKET=%s", got, want)
}
binding := fileNamed(out, "files.bound-s3-bucket")
if binding == nil {
t.Fatalf("the consumer was given no binding: %v", out)
}
var said struct {
Serves map[string]any `json:"serves"`
}
if err := json.Unmarshal([]byte(binding["content"].(string)), &said); err != nil {
t.Fatal(err)
}
if said.Serves["bucket"] != want {
t.Errorf("the binding says the bucket is %q, want %q", said.Serves["bucket"], want)
}
// And what is the same for everybody is still the same for everybody.
if said.Serves["region"] != "eu-west" {
t.Errorf("the binding lost what the provider serves to all: %v", said.Serves)
}
given := storeGrants(t, out)
if len(given) != 1 {
t.Fatalf("the provider was told about %d consumer(s): %v", len(given), given)
}
if given[0].Derived["bucket"] != want {
t.Errorf("the provider was told the bucket is %v, and the consumer was told %q — "+
"the two ends disagree, which is the whole failure", given[0].Derived["bucket"], want)
}
// Only the per-consumer half. The region is the same for everyone and is already in the
// provider's own definition; repeating it here would be a copy to go stale.
if _, carried := given[0].Derived["region"]; carried {
t.Errorf("the provider was handed back what it already says for everyone: %v", given[0].Derived)
}
}
// Two consumers of one provider on one machine get two buckets, and neither gets the other's.
func TestTwoConsumersOfOneProviderGetTheirOwnDerivation(t *testing.T) {
r, err := Resolve(shelf(store(), files(), pics()),
[]string{"store", "files", "pics"}, reachable(), World{})
if err != nil {
t.Fatal(err)
}
out, err := r.Declaration(Rendering{Grants: []Grant{
{Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
Values: map[string]any{}, Sealed: "c2VhbGVk"},
{Provision: "s3-bucket", Consumer: "workstation", From: "pics", Slug: "pics",
Values: map[string]any{}, Sealed: "c2VhbGVk"},
}})
if err != nil {
t.Fatal(err)
}
forFiles := strings.ReplaceAll(ConsumerIdentity("workstation", IdentitySource("files", "files")), "_", "-")
forPics := strings.ReplaceAll(ConsumerIdentity("workstation", IdentitySource("pics", "pics")), "_", "-")
if forFiles == forPics {
t.Fatal("the two consumers were given the same identity; this test proves nothing")
}
if got := fileNamed(out, "files.env")["content"].(string); !strings.Contains(got, "BUCKET="+forFiles+"\n") {
t.Errorf("files was not given its own bucket:\n%s", got)
}
if got := fileNamed(out, "pics.env")["content"].(string); !strings.Contains(got, "BUCKET="+forPics+"\n") {
t.Errorf("pics was not given its own bucket:\n%s", got)
}
var buckets []any
for _, g := range storeGrants(t, out) {
buckets = append(buckets, g.Derived["bucket"])
}
if len(buckets) != 2 || buckets[0] == buckets[1] {
t.Errorf("the provider was told %v; it must be told one bucket per consumer", buckets)
}
}
// An operator may still set what the provider serves, and the mesh still derives the rest: the
// setting is laid on first, then the consumer's half is filled.
func TestASettingComposesWithADerivedValue(t *testing.T) {
r, err := Resolve(shelf(store(), files()), []string{"store", "files"}, reachable(), World{})
if err != nil {
t.Fatal(err)
}
out, err := r.Declaration(Rendering{
Settings: SettingsBy{"store": {{From: "the operator",
Values: map[string]any{"bucket": "team-${consumer:as:dns}"}}}},
Grants: []Grant{{Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
Values: map[string]any{}, Sealed: "c2VhbGVk"}},
})
if err != nil {
t.Fatal(err)
}
want := "team-" + strings.ReplaceAll(ConsumerIdentity("workstation", IdentitySource("files", "files")), "_", "-")
if got := fileNamed(out, "files.env")["content"].(string); !strings.Contains(got, "BUCKET="+want+"\n") {
t.Errorf("the operator's prefix did not survive the derivation:\n%s\nwant BUCKET=%s", got, want)
}
if given := storeGrants(t, out); given[0].Derived["bucket"] != want {
t.Errorf("the provider was told %v, the consumer %q", given[0].Derived["bucket"], want)
}
}
// A fact or an alphabet the mesh does not have is refused where the definition is, not where a
// consumer happens to be resolved — and the refusal says what may be said instead.
func TestAServedValueNamingSomethingTheMeshDoesNotHaveIsRefused(t *testing.T) {
for _, c := range []struct{ value, says string }{
{"${consumer:node}", "as"},
{"${consumer:as:punycode}", "dns"},
} {
m := store()
m.Serves["s3-bucket"]["bucket"] = c.value
raw, err := json.Marshal(m)
if err != nil {
t.Fatal(err)
}
_, err = ParseManifest(raw)
if err == nil {
t.Fatalf("%s was accepted", c.value)
}
if !strings.Contains(err.Error(), c.value) {
t.Errorf("the refusal of %s does not quote it: %v", c.value, err)
}
if !strings.Contains(err.Error(), c.says) {
t.Errorf("the refusal of %s does not say what may be said (%q): %v", c.value, c.says, err)
}
}
}
// `dns` is checked against an identity the mesh actually mints, not an invented string.
func TestTheDNSAlphabetIsTheMintedIdentityWithItsSeparatorRewritten(t *testing.T) {
as := ConsumerIdentity("anchor", IdentitySource("ncloud", "nextcloud"))
if err := CheckIdentity("anchor", IdentitySource("ncloud", "nextcloud")); err != nil {
t.Fatalf("the mesh would not mint this identity at all: %v", err)
}
label := asDNSLabel(as)
if strings.Contains(label, "_") {
t.Errorf("%q is not a DNS label", label)
}
if strings.ReplaceAll(label, "-", "_") != as {
t.Errorf("%q is not %q with its separator rewritten", label, as)
}
}
// The check that would have caught the one wrong instance: a consumer that writes the derived
// value into its own definition instead of asking for it is refused, whether it transcribed the
// right answer or a predecessor's.
func TestAConsumerThatTranscribesWhatItsProviderDerivesIsRefused(t *testing.T) {
as := ConsumerIdentity("workstation", IdentitySource("files", "files"))
transcribed := strings.ReplaceAll(as, "_", "-")
m := files()
m.Resources = []map[string]any{{
"id": "env", "type": "file", "path": "/var/lib/files/env", "mode": "0600",
// Exactly what the provider will create — correct today, and a copy of a rule that is
// not this module's.
"content": "BUCKET=" + transcribed + "\n",
}}
r, err := Resolve(shelf(store(), m), []string{"store", "files"}, reachable(), World{})
if err != nil {
t.Fatal(err)
}
_, err = r.Declaration(Rendering{Grants: []Grant{{
Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
Values: map[string]any{}, Sealed: "c2VhbGVk",
}}})
if err == nil {
t.Fatal("a definition holding its own copy of the provider's naming rule was accepted")
}
if !strings.Contains(err.Error(), "${bound:s3-bucket:bucket}") {
t.Errorf("the refusal does not say what to write instead: %v", err)
}
// And a constant the provider serves to everyone is not a transcription: repeating it is
// redundant, not wrong, and refusing it would be the mesh policing style.
m.Resources = []map[string]any{{
"id": "env", "type": "file", "path": "/var/lib/files/env", "mode": "0600",
"content": "REGION=eu-west\n",
}}
r, err = Resolve(shelf(store(), m), []string{"store", "files"}, reachable(), World{})
if err != nil {
t.Fatal(err)
}
if _, err := r.Declaration(Rendering{Grants: []Grant{{
Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
Values: map[string]any{}, Sealed: "c2VhbGVk",
}}}); err != nil {
t.Errorf("a value the provider serves to everyone was judged a transcription: %v", err)
}
}
func storeGrants(t *testing.T, out []map[string]any) []Contribution {
t.Helper()
for _, r := range out {
if r["path"] != "/var/lib/store/grants/mesh.json" {
continue
}
var parsed struct {
Given []Contribution `json:"given"`
}
if err := json.Unmarshal([]byte(r["content"].(string)), &parsed); err != nil {
t.Fatal(err)
}
return parsed.Given
}
t.Fatalf("the provider was given no contributions file: %v", out)
return nil
}
// A consumer that keeps SEVERAL holders of one provision is refused, rather than told one thing
// while its provider is told another.
//
// **This is issue 124's own failure, one case to the side.** The mesh gives each holder its own
// login — `mesh_node_mod_<local>` (ADR 0094) — and the provider derives from the login, so it
// would make one resource per holder. The consumer's side has no such dimension: there is one
// binding file per provision and one `${bound:<provision>:<key>}`, both derived from the
// un-suffixed identity. So the provider would create `…-mod-cold` and the consumer would be
// configured against `…-mod`: it would authenticate successfully and be refused on every object,
// which is exactly the fault this whole record exists to end.
//
// Refused, loudly, at the one place that can see both halves. Lifting it means giving the
// consumer's side a local dimension, which is a decision and not an omission.
func TestAConsumerWithSeveralHoldersOfADerivingProviderIsRefused(t *testing.T) {
m := files()
// Two holders of the one provision, the shape ADR 0094 gives a module that keeps several.
m.Secrets = nil
m.SecretsMany = map[string]map[string]string{"s3-bucket": {
"hot": "/var/lib/files/hot.secret",
"cold": "/var/lib/files/cold.secret",
}}
m.Resources = []map[string]any{{
"id": "env", "type": "file", "path": "/var/lib/files/env", "mode": "0600",
"content": "BUCKET=${bound:s3-bucket:bucket}\n",
}}
r, err := Resolve(shelf(store(), m), []string{"store", "files"}, reachable(), World{})
if err != nil {
t.Fatal(err)
}
_, err = r.Declaration(Rendering{Grants: []Grant{
{Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
Local: "hot", Values: map[string]any{}, Sealed: "c2VhbGVk"},
{Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
Local: "cold", Values: map[string]any{}, Sealed: "c2VhbGVk"},
}})
if err == nil {
t.Fatal("a consumer with several holders of a deriving provider was accepted; " +
"its two ends would have disagreed in silence")
}
for _, want := range []string{"files", "s3-bucket", "bucket"} {
if !strings.Contains(err.Error(), want) {
t.Errorf("the refusal does not name %q: %v", want, err)
}
}
}
-29
View File
@@ -1,29 +0,0 @@
package catalogue
import (
"os"
"regexp"
"testing"
)
// novox/hq issue 223: genesis raises the controller as a container built from this repository's own
// Dockerfile, with no build arguments — the manifest no longer builds an image, so nothing passes a
// base in. The Dockerfile's own default must therefore be a Go that builds this module, pinned by
// digest, and the replacement the manifest's process names must be the container genesis raises.
func TestGenesisCanBuildTheControllersImageAsItStands(t *testing.T) {
raw, err := os.ReadFile("../../Dockerfile")
if err != nil {
t.Fatal(err)
}
if !regexp.MustCompile(`(?m)^ARG GO_BASE=golang@sha256:[0-9a-f]{64}$`).Match(raw) {
t.Fatal("the Dockerfile's default Go base is not pinned by digest; genesis builds it with no arguments")
}
makefile, err := os.ReadFile("../../Makefile")
if err != nil {
t.Fatal(err)
}
pin := regexp.MustCompile(`golang@sha256:[0-9a-f]{64}`)
if string(pin.Find(raw)) != string(pin.Find(makefile)) {
t.Errorf("the Dockerfile and the Makefile build on different Go: %s, %s", pin.Find(raw), pin.Find(makefile))
}
}
-79
View File
@@ -1428,10 +1428,6 @@ func ParseManifest(raw []byte) (Manifest, error) {
"%s serves %q to whoever requires it, and does not provide it", m.Module, to))
}
}
// A served value may be derived for the consumer it is served to (novox/hq ADR 0201). Read
// here, where the definition is, rather than when somebody first requires it: a rule that
// would be refused at the first consumer is wrong from the moment it is written.
problems = append(problems, CheckServes(m)...)
for to, where := range m.Binds {
if !placedOrAbsolute(where) {
problems = append(problems, fmt.Sprintf(
@@ -1626,13 +1622,6 @@ func ParseManifest(raw []byte) (Manifest, error) {
}
}
}
// **A scheduled step may hold this module's own containers still while it runs**
// (novox/hq ADR 0189). What the host judges is the declaration it receives — whether each
// id is a container placed on that machine; what belongs here is what only the definition
// shows: that the ids are this module's, that they are containers, and that the step is
// scheduled. A module naming a neighbour's container would be a module that can stop the
// mesh, and the manifest is where that is visible.
problems = append(problems, whileStoppedProblems(m, r, hasSchedule(r))...)
}
for name, own := range m.OwnSecrets {
if !placedOrAbsolute(own.Path) {
@@ -2164,71 +2153,3 @@ func (o OwnSecrets) Paths() map[string]string {
// InstancesInterchangeable is the one value of a definition's `instances`: the module is the same
// on every machine, so any instance may answer for the module.
const InstancesInterchangeable = "interchangeable"
// WhileStopped is the resource key naming the containers a scheduled step holds still while it
// runs (novox/hq ADR 0189). Carried to the host unchanged, like `schedule`.
const WhileStopped = "while-stopped"
// hasSchedule is whether a resource declares a cadence, as a string.
func hasSchedule(r map[string]any) bool {
s, _ := r["schedule"].(string)
return s != ""
}
// whileStoppedProblems judges one container's maintenance window against its own definition
// (novox/hq ADR 0189).
//
// Three things the manifest is the only place to see: that the step is scheduled (a one-time
// offline job says *before* rather than *instead of* — at apply the host already has a window,
// because the declaration is applied in order and a run-once step gates what follows); that every
// id it names is **this module's own** container; and that it does not name itself.
//
// The host checks the fourth — that the container is actually placed on that machine — because
// that is a fact about the declaration and not about the definition.
func whileStoppedProblems(m Manifest, r map[string]any, scheduled bool) []string {
raw, present := r[WhileStopped]
if !present {
return nil
}
ids, ok := raw.([]any)
if !ok {
return []string{fmt.Sprintf(
"%s declares %s on %v as a %T; it is a list of this module's container ids",
m.Module, WhileStopped, r["id"], raw)}
}
var problems []string
if len(ids) > 0 && !scheduled {
problems = append(problems, fmt.Sprintf(
"%s declares %s on %v, which has no schedule. A maintenance window is for a recurring "+
"step: at apply the mesh already has one, because a run-once step gates what is "+
"declared after it (novox/hq ADR 0189)", m.Module, WhileStopped, r["id"]))
}
containers := map[string]bool{}
for _, own := range m.Resources {
if fmt.Sprint(own["type"]) == "container" {
containers[fmt.Sprint(own["id"])] = true
}
}
for _, each := range ids {
id, ok := each.(string)
if !ok {
problems = append(problems, fmt.Sprintf(
"%s declares %s on %v naming a %T; each entry is a container's id",
m.Module, WhileStopped, r["id"], each))
continue
}
if id == fmt.Sprint(r["id"]) {
problems = append(problems, fmt.Sprintf(
"%s declares %s on %v naming itself", m.Module, WhileStopped, r["id"]))
continue
}
if !containers[id] {
problems = append(problems, fmt.Sprintf(
"%s declares %s on %v naming %q, which is not a container this module declares. "+
"A step may hold still its own module's containers and nobody else's — one "+
"that could quiesce a neighbour could stop the mesh",
m.Module, WhileStopped, r["id"], id))
}
}
return problems
}
-89
View File
@@ -1,89 +0,0 @@
package catalogue
import (
"encoding/json"
"strings"
"testing"
)
// A scheduled step may hold its module's own containers still while it runs (novox/hq ADR 0189).
//
// The host judges what it receives — whether each id is a container on that machine. What the
// definition is the only place to see is judged here, near whoever wrote it.
func aStoreManifest(step map[string]any) []byte {
m := map[string]any{
"module": "distribution", "version": "1",
"resources": []any{
map[string]any{"id": "store", "type": "container", "name": "mesh-registry",
"image": "registry@sha256:" + strings.Repeat("a", 64)},
step,
},
}
raw, _ := json.Marshal(m)
return raw
}
func TestAMaintenanceWindowOnItsOwnModulesContainerIsAccepted(t *testing.T) {
raw := aStoreManifest(map[string]any{
"id": "collect", "type": "container", "name": "mesh-registry-collect",
"image": "registry@sha256:" + strings.Repeat("a", 64),
"schedule": "30 3 * * *", "while-stopped": []any{"store"},
})
if _, err := ParseManifest(raw); err != nil {
t.Fatalf("a step holding its own module's container still was refused: %v", err)
}
}
func TestAMaintenanceWindowIsRefusedWhereTheDefinitionShowsItCannotMean(t *testing.T) {
for _, c := range []struct {
name string
step map[string]any
says string
}{
{
"on a step with no schedule",
map[string]any{"id": "collect", "type": "container", "name": "c",
"image": "registry@sha256:" + strings.Repeat("a", 64),
"while-stopped": []any{"store"}},
"gates what is declared after it",
},
{
"on a run-once step, which already has order",
map[string]any{"id": "collect", "type": "container", "name": "c",
"image": "registry@sha256:" + strings.Repeat("a", 64),
"run-once": true, "while-stopped": []any{"store"}},
"A maintenance window is for a recurring step",
},
{
"naming a container this module does not declare",
map[string]any{"id": "collect", "type": "container", "name": "c",
"image": "registry@sha256:" + strings.Repeat("a", 64),
"schedule": "30 3 * * *", "while-stopped": []any{"the-broker"}},
"could quiesce a neighbour could stop the mesh",
},
{
"naming itself",
map[string]any{"id": "collect", "type": "container", "name": "c",
"image": "registry@sha256:" + strings.Repeat("a", 64),
"schedule": "30 3 * * *", "while-stopped": []any{"collect"}},
"naming itself",
},
{
"written as something that is not a list",
map[string]any{"id": "collect", "type": "container", "name": "c",
"image": "registry@sha256:" + strings.Repeat("a", 64),
"schedule": "30 3 * * *", "while-stopped": "store"},
"a list of this module's container ids",
},
} {
_, err := ParseManifest(aStoreManifest(c.step))
if err == nil {
t.Errorf("%s was accepted", c.name)
continue
}
if !strings.Contains(err.Error(), c.says) {
t.Errorf("%s: the refusal does not say %q:\n%v", c.name, c.says, err)
}
}
}
+32 -7
View File
@@ -78,12 +78,11 @@ func (i *Inventory) RegisterModule(ctx context.Context, m catalogue.Manifest, fr
// (novox/hq ADR 0175, to-be 38 WP2.4): a module serving its tools from a container built on the
// runtime's image. Refused at registration, by name, for a module that is new to the catalogue
// or that was registered in another shape — the mechanism that keeps the old pattern from
// returning by habit. Before the runtime exists the pattern is accepted as it always was.
//
// *Since 2026-10-04 (to-be 38 WP4b's last step):* refused for **every** module. While some
// thirty modules still stood in that shape, one already registered so was rebuilt without
// complaint, so the pipeline kept running while each moved; every module has moved since, and
// the exception would only let one move back.
// returning by habit. **Not refused for a module already registered in that shape**: the
// catalogue holds some thirty of them the day the runtime arrives, each moves to a bundle in
// its own change (to-be 38 WP4 onward), and a gate that refused every rebuild of every unmoved
// module in the meantime would stop the whole pipeline to make a point the record already makes.
// Before the runtime exists the pattern is accepted as it always was.
if m.Module != catalogue.RuntimeModule {
if why := catalogue.ToolContainerOnTheRuntime(m, from.Against); why != "" {
runtime, err := i.hasModule(ctx, catalogue.RuntimeModule)
@@ -91,7 +90,13 @@ func (i *Inventory) RegisterModule(ctx context.Context, m catalogue.Manifest, fr
return err
}
if runtime {
return fmt.Errorf("%s is not registered: %s", m.Module, why)
already, err := i.registeredInThatShape(ctx, m.Module)
if err != nil {
return err
}
if !already {
return fmt.Errorf("%s is not registered: %s", m.Module, why)
}
}
}
}
@@ -147,6 +152,26 @@ func (i *Inventory) RegisterModule(ctx context.Context, m catalogue.Manifest, fr
return nil
}
// registeredInThatShape is whether the catalogue already holds this module as a tools container on
// the runtime's image — judged from the manifest it holds and what that module's newest build stood
// on, the same two things the gate judges a new registration by. False for a module the catalogue
// does not hold.
func (i *Inventory) registeredInThatShape(ctx context.Context, name string) (bool, error) {
held, err := i.Catalogue(ctx)
if err != nil {
return false, err
}
stored, has := held[name]
if !has {
return false, nil
}
against, err := i.BuiltAgainst(ctx)
if err != nil {
return false, err
}
return catalogue.ToolContainerOnTheRuntime(stored, against[name]) != "", nil
}
// hasModule is whether the catalogue holds a module of that name.
func (i *Inventory) hasModule(ctx context.Context, name string) (bool, error) {
var one int
+8 -8
View File
@@ -688,9 +688,9 @@ func TestRegisteringWithoutProvenanceKeepsTheSeat(t *testing.T) {
// Once the node's tool runtime is in the catalogue, a module serving its tools from a container
// built on the runtime's image is refused at registration, naming the record (novox/hq ADR 0175,
// to-be 38 WP2.4) — for every module, since every module has moved (WP4b's last step; WP3's
// amendment let one already standing in that shape be rebuilt while each moved). Before the
// runtime, it is accepted as it always was — so a
// to-be 38 WP2.4) — for a module new to the catalogue or one that had moved away from it; a module
// already standing in that shape is rebuilt as before, so the catalogue's pipeline keeps running
// while each moves (WP3's amendment). Before the runtime, it is accepted as it always was — so a
// mesh converts in the order the design says and nothing is refused before there is anything to
// move to.
func TestAToolContainerIsRefusedOnceTheRuntimeIsRegistered(t *testing.T) {
@@ -715,11 +715,11 @@ func TestAToolContainerIsRefusedOnceTheRuntimeIsRegistered(t *testing.T) {
t.Fatal(err)
}
// **A module already registered in that shape is refused too** (WP4b's last step): every module
// has moved, and a rebuild in the old shape is one moving back.
if err := inv.RegisterModule(ctx, filter, Source{Repository: "/r", Against: stoodOn}); err == nil ||
!strings.Contains(err.Error(), "ADR 0175") {
t.Fatalf("a rebuild of a module in the old pattern was registered beside the runtime: %v", err)
// **A module already registered in that shape is rebuilt without complaint** (to-be 38 WP2.4 as
// amended by WP3): some thirty of them stand the day the runtime arrives, and each moves in its
// own change. The gate is against the pattern spreading, not against the pipeline running.
if err := inv.RegisterModule(ctx, filter, Source{Repository: "/r", Against: stoodOn}); err != nil {
t.Fatalf("a rebuild of a module that already had the pattern was refused: %v", err)
}
// A module new to the catalogue in that shape is refused, naming the record.
newcomer := filter
-242
View File
@@ -1,242 +0,0 @@
package inventory
import (
"context"
"encoding/json"
"strings"
)
// What the artifact store keeps, and what it may let go (novox/hq ADR 0189, issue 108).
//
// The store has never collected anything: every build pushes another layer set and nothing has
// ever removed one. The registry's own answer — collect what no tag names — is wrong here, because
// the mesh pushes each artifact under one moving tag and pins machines by digest, so every build
// but the newest is untagged and some machine may still be running it.
//
// **So the mesh decides, from its own records, and it never has to look in the store to do it.**
// It has never put anything there it did not record, which means every digest it could remove is
// already in a build row. A digest the mesh did not record making is therefore never named here —
// not as a safety margin but as the rule restated, and it is what keeps the sweep away from the
// images genesis pushed before any record existed (04-ISSUES/102, F4).
// KeptBuilds is how many successful builds of each module keep their artifacts, counting the
// newest. The newest is what the mesh hands a machine now; the four behind it are how far back a
// release that turns out wrong can be taken.
const KeptBuilds = 5
// ToCollect is every artifact the mesh made, no longer keeps, and has not already collected.
//
// Three reasons an artifact stays, and nothing else is a reason:
//
// - **a definition names it** — the reference appears in a module's recorded manifest, which is
// what the mesh would hand a machine now. No age limit: this is the floor;
// - **the mesh can still go back to it** — it is an artifact of one of the KeptBuilds most
// recent successful builds of its module;
// - it was already collected, in which case there is nothing left to do.
//
// Returned in a stated order so two runs over the same records ask for the same things in the
// same sequence, which is what makes a failed sweep safe to simply run again.
func (i *Inventory) ToCollect(ctx context.Context) ([]string, error) {
keep, err := i.keptReferences(ctx)
if err != nil {
return nil, err
}
rows, err := i.store.Pool().Query(ctx,
// Every artifact of every successful build, oldest first, minus what has already been
// collected. A failed build published nothing, so it names nothing to remove.
`select b.made
from build b
where b.failed = '' and b.module is not null and b.module <> ''
order by b.at asc, b.id asc`)
if err != nil {
return nil, err
}
defer rows.Close()
collected, err := i.alreadyCollected(ctx)
if err != nil {
return nil, err
}
seen := map[string]bool{}
var out []string
for rows.Next() {
var raw []byte
if err := rows.Scan(&raw); err != nil {
return nil, err
}
var made []Artifact
if err := json.Unmarshal(raw, &made); err != nil {
// One unreadable record must not stop the rest being collected — and an artifact this
// row named is simply not offered, which errs toward keeping.
continue
}
for _, a := range made {
if a.Reference == "" || keep[a.Reference] || collected[a.Reference] || seen[a.Reference] {
continue
}
seen[a.Reference] = true
out = append(out, a.Reference)
}
}
return out, rows.Err()
}
// keptReferences is every artifact reference the mesh still keeps, for either of the two reasons.
func (i *Inventory) keptReferences(ctx context.Context) (map[string]bool, error) {
keep := map[string]bool{}
// **Whatever a definition the mesh holds names.** Read as text rather than by walking the
// resource shapes: a reference may be a container's image, a bundle's source, or a field some
// later kind of resource grows, and what matters is only whether the mesh could hand this
// string to a machine. A manifest that mentions it is a manifest that might.
manifests, err := i.store.Pool().Query(ctx, `select manifest::text from module where manifest is not null`)
if err != nil {
return nil, err
}
defer manifests.Close()
var named []string
for manifests.Next() {
var text string
if err := manifests.Scan(&text); err != nil {
return nil, err
}
named = append(named, text)
}
if err := manifests.Err(); err != nil {
return nil, err
}
// The KeptBuilds most recent successful builds of each module, whole.
recent, err := i.store.Pool().Query(ctx,
`select made from (
select made, row_number() over (partition by module order by at desc, id desc) as back
from build
where failed = '' and module is not null and module <> ''
) ranked where back <= $1`, KeptBuilds)
if err != nil {
return nil, err
}
defer recent.Close()
for recent.Next() {
var raw []byte
if err := recent.Scan(&raw); err != nil {
return nil, err
}
var made []Artifact
if err := json.Unmarshal(raw, &made); err != nil {
continue
}
for _, a := range made {
if a.Reference != "" {
keep[a.Reference] = true
}
}
}
if err := recent.Err(); err != nil {
return nil, err
}
// And anything a manifest mentions. Done after the recent set so the scan runs over the
// candidates rather than over every reference ever recorded: a manifest holds a reference
// composed with the store's address or kept bare, so the search is for the digest within it.
if len(named) > 0 {
all, err := i.everyReferenceMade(ctx)
if err != nil {
return nil, err
}
for _, reference := range all {
if keep[reference] {
continue
}
digest := digestIn(reference)
if digest == "" {
// Not something the store holds by digest; nothing here can speak for it, so it
// is kept rather than guessed about.
keep[reference] = true
continue
}
for _, text := range named {
if strings.Contains(text, digest) {
keep[reference] = true
break
}
}
}
}
return keep, nil
}
// everyReferenceMade is every artifact reference any successful build recorded.
func (i *Inventory) everyReferenceMade(ctx context.Context) ([]string, error) {
rows, err := i.store.Pool().Query(ctx,
`select made from build where failed = '' and module is not null and module <> ''`)
if err != nil {
return nil, err
}
defer rows.Close()
seen := map[string]bool{}
var out []string
for rows.Next() {
var raw []byte
if err := rows.Scan(&raw); err != nil {
return nil, err
}
var made []Artifact
if err := json.Unmarshal(raw, &made); err != nil {
continue
}
for _, a := range made {
if a.Reference == "" || seen[a.Reference] {
continue
}
seen[a.Reference] = true
out = append(out, a.Reference)
}
}
return out, rows.Err()
}
// digestIn is the `sha256:<hex>` a reference names, empty when it names none.
func digestIn(reference string) string {
for _, marker := range []string{"@sha256:", "/sha256:"} {
if _, after, ok := strings.Cut(reference, marker); ok {
return "sha256:" + after
}
}
return ""
}
// alreadyCollected is what the store has already been asked to let go.
func (i *Inventory) alreadyCollected(ctx context.Context) (map[string]bool, error) {
rows, err := i.store.Pool().Query(ctx, `select reference from artifact_collected`)
if err != nil {
return nil, err
}
defer rows.Close()
out := map[string]bool{}
for rows.Next() {
var reference string
if err := rows.Scan(&reference); err != nil {
return nil, err
}
out[reference] = true
}
return out, rows.Err()
}
// MarkCollected records that the store no longer holds these.
//
// **A store that answered "not found" is recorded too.** The outcome wanted is that the artifact
// is gone, and it is; retrying it every sweep for ever is the failure this table exists to
// prevent. Only a store that could not be reached, or refused, leaves a reference unmarked — and
// then the next sweep asks again, which is what should happen.
func (i *Inventory) MarkCollected(ctx context.Context, references []string) error {
for _, reference := range references {
if _, err := i.store.Pool().Exec(ctx,
`insert into artifact_collected (reference) values ($1) on conflict (reference) do nothing`,
reference); err != nil {
return err
}
}
return nil
}
-147
View File
@@ -1,147 +0,0 @@
package inventory
import (
"context"
"fmt"
"testing"
"github.com/novox/mesh-controller/internal/catalogue"
)
// What the store keeps, and what it may let go (novox/hq ADR 0189, issue 108).
//
// The store has collected nothing since it was raised, and the registry's own answer — collect
// what no tag names — would delete images machines are running, because the mesh pushes under one
// moving tag and pins by digest. So the rule is the mesh's, read from its own records, and these
// are the three reasons an artifact stays and the one reason it goes.
// ref is an artifact reference as the mesh records one.
func ref(module, artifact string, n int) string {
return fmt.Sprintf("%s%s/%s@sha256:%064x", catalogue.ArtifactStoreScheme, module, artifact, n)
}
// built records one successful build of a module publishing one image.
func built(t *testing.T, inv *Inventory, id, module string, n int) string {
t.Helper()
reference := ref(module, "app", n)
b := aBuild(id, module, "")
b.Made = []Artifact{{Name: "app", Kind: "image", Reference: reference}}
if err := inv.RecordBuild(context.Background(), b); err != nil {
t.Fatal(err)
}
return reference
}
func TestTheStoreKeepsTheRecentBuildsAndLetsGoOfTheRest(t *testing.T) {
inv := fresh(t)
ctx := context.Background()
// Eight builds of one module, oldest first. Five are kept — the newest, and the four a
// release that turns out wrong can be taken back to.
var made []string
for i := 1; i <= 8; i++ {
made = append(made, built(t, inv, fmt.Sprintf("b%02d", i), "web", i))
}
go_, err := inv.ToCollect(ctx)
if err != nil {
t.Fatal(err)
}
want := made[:3] // the three oldest
if len(go_) != len(want) {
t.Fatalf("offered %v to collect; want the %d oldest of %d", go_, len(want), len(made))
}
for i := range want {
if go_[i] != want[i] {
t.Fatalf("offered %v; want %v — and in that order, so a failed sweep is safe to run again",
go_, want)
}
}
}
func TestADefinitionNamingAnArtifactKeepsItHoweverOldItIs(t *testing.T) {
// The floor: no age limit. A module recorded at an older commit still names what the mesh
// would hand a machine now, and that is what must not be collected out from under it.
inv := fresh(t)
ctx := context.Background()
var made []string
for i := 1; i <= 8; i++ {
made = append(made, built(t, inv, fmt.Sprintf("b%02d", i), "web", i))
}
oldest := made[0]
// A definition the mesh holds, whose container runs that oldest image.
m := catalogue.Manifest{Module: "web", Version: "1", Resources: []map[string]any{{
"id": "app", "type": "container", "name": "web", "image": oldest,
}}}
if err := inv.RegisterModule(ctx, m, Source{Repository: "https://forge.invalid/web.git"}); err != nil {
t.Fatal(err)
}
go_, err := inv.ToCollect(ctx)
if err != nil {
t.Fatal(err)
}
for _, reference := range go_ {
if reference == oldest {
t.Fatalf("the mesh offered to collect %s, which a definition it holds names", oldest)
}
}
if len(go_) != 2 {
t.Fatalf("offered %v; want the two oldest that nothing names", go_)
}
}
func TestWhatHasBeenCollectedIsNotOfferedAgain(t *testing.T) {
// Without this the sweep reissues a delete for every artifact it has ever collected, every
// time it runs, for ever — a number of requests that grows with the mesh's whole history.
inv := fresh(t)
ctx := context.Background()
for i := 1; i <= 7; i++ {
built(t, inv, fmt.Sprintf("b%02d", i), "web", i)
}
first, err := inv.ToCollect(ctx)
if err != nil {
t.Fatal(err)
}
if len(first) != 2 {
t.Fatalf("offered %v, want two", first)
}
if err := inv.MarkCollected(ctx, first); err != nil {
t.Fatal(err)
}
again, err := inv.ToCollect(ctx)
if err != nil {
t.Fatal(err)
}
if len(again) != 0 {
t.Fatalf("offered %v again after collecting it", again)
}
}
func TestAFailedBuildNamesNothingToCollectAndEachModuleIsCountedOnItsOwn(t *testing.T) {
inv := fresh(t)
ctx := context.Background()
// A failed build published nothing, so it is neither kept nor collected — and it must not
// count against the module's five.
for i := 1; i <= 6; i++ {
built(t, inv, fmt.Sprintf("w%02d", i), "web", i)
}
if err := inv.RecordBuild(ctx, aBuild("w99", "web", "the recipe would not build")); err != nil {
t.Fatal(err)
}
// And a second module with three builds keeps all three: five each, not five between them.
for i := 1; i <= 3; i++ {
built(t, inv, fmt.Sprintf("d%02d", i), "db", 100+i)
}
go_, err := inv.ToCollect(ctx)
if err != nil {
t.Fatal(err)
}
if len(go_) != 1 || go_[0] != ref("web", "app", 1) {
t.Fatalf("offered %v; want only web's oldest — db's three are all within its five", go_)
}
}
@@ -1,23 +0,0 @@
-- What the artifact store no longer keeps (novox/hq ADR 0189, issue 108).
--
-- The mesh removes from its store only what it put there and can account for: every digest it
-- could remove is already in a build record, so the sweep reads its own records rather than
-- enumerating the store. What it does not get from those records is whether it has already
-- removed something -- `build.made` says what that build published, for ever, which is history
-- and not an index of what is on disk.
--
-- Without this the sweep would reissue a delete for every artifact it has ever collected, every
-- time it runs, and each one would answer 404 -- a number of requests that grows with the mesh's
-- whole history and never shrinks.
--
-- Keyed by the reference as the mesh records it (`artifact-store://<module>/<artifact>@sha256:…`),
-- because that is the identity the record uses everywhere else. Not a foreign key to build: two
-- builds can publish the same digest (the same source built twice produces the same bytes), and
-- what is collected is the artifact, not the attempt that made it.
create table artifact_collected (
reference text primary key,
-- When the store answered. Kept so a reader of an old build record can tell "this artifact is
-- gone" from "this artifact was never there", which are different kinds of surprise.
at timestamptz not null default now()
);