Compare commits

...
Author SHA1 Message Date
mesh-admin d2d171f2d2 Merge pull request 'Compose a module's Go service as a process the host runs (hq issue 213, 1 of 2)' (#252) from fix/issue-213-the-controller-is-a-process into main 2026-10-03 23:40:38 +00:00
mesh-admin 73fa64ea68 Merge pull request 'A TypeScript bundle installs its module's own packages before it is compiled (hq ADR 0198 §4)' (#255) from feat/a-bundle-installs-its-own-packages into main 2026-10-03 23:20:57 +00:00
jochen 1a13dbeb17 A TypeScript bundle installs its module's own packages before it is compiled
A bundle could import only what the toolchain image carried: the compiler and the bundler resolve an import from the module's directory and then the toolchain's node_modules, and nothing ever put anything in the first. So a module needing a database driver (pg, mongodb, mssql) could not be a bundle, and kept a container whose recipe installed it (hq ADR 0198 §4: the backend's own driver inside the bundle).

Now, when a module's package.json depends on anything beyond the SDK, the build installs its production dependencies into the module's directory, in the toolchain image, before the compile: npm ci from the lockfile when there is one, npm install from the ranges otherwise, the mesh's registry for the SDK's scope and the public one for the rest, install scripts off. esbuild then inlines them. A module depending only on the SDK runs exactly the commands it did before.

The SDK stays the toolchain's (hq issue 212): it is taken out of what is installed and any copy something pulls in is removed, so every import of it resolves past the module's node_modules to the one the toolchain carries; a module's own range never shadows it. npm's verified download cache is a named volume; nothing installed is kept between builds. Without a registry, a scoped package is refused rather than resolved on the public registry.
2026-10-04 01:17:49 +02:00
jochen e11caecdad Let two controllers overlap safely while one hands over to the other (hq issue 213)
The controller's machine moves it from the container to a process by
starting the process first and removing the container once the process
is up (mesh-host's `replaces`). For that moment two controllers share the
store and the bus. Checked what each does:

- the seat's verbs: a queue group per seat, each call answered once. Safe.
- the controller's consumers on CONTROL and EVENTS: push consumers with
  no delivery group, so the second bind is refused with "consumer is
  already bound" and serve exited. The process would restart for ever,
  the host would never see it up, and the container would never go. The
  second controller now stands by and binds when the first lets go
  (tested on a real bus; fails without the change).
- plans: read, changed and saved whole by the 30s timer, by build
  outcomes, by a merge and by `plans stop`. Two timers would each ask a
  tier the other had just asked. Working the plans now takes a
  session-level advisory lock on the inventory: the timer skips while
  another holds it, the other paths wait for it. Build asks happen only
  inside plan work and are covered by the same lock.
2026-10-04 01:11:26 +02:00
jochen 7bb9e55d0b Compose a module's Go service as a process the host runs (hq issue 213)
The controller is to be declared as a Go bundle run by a process instead of
an image (novox/hq issue 213, ADR 0188 §1, §3). The composer could not
express that honestly yet:

- a module declaring tools had every bundle served by the node's runtime,
  so the controller's own binary would have been launched a second time as
  an MCP child; a bundle one of the module's resources runs is now served
  only when it says `loads`
- a module's accounts went after the mesh-computed files, so secrets owned
  by the account a process runs as were refused on the first apply; a
  module's `user` resources now go first
- `prepares` derived its step only from a container; a process is now
  prepared by the same program with `prepare` as a run-once process
- a process may say what it `replaces` (a resource of its module it no
  longer declares), prefixed as the host records it, so the host keeps the
  old one running until the process is (needs mesh-host's `replaces`)

This lands before the controller's manifest uses any of it: the running
controller composes its own declaration, so the code that fills the new
shape must be live first.
2026-10-04 01:11:26 +02:00
mesh-admin 00037608ae Merge pull request 'The forge's tests compose its code as a bundle the node's runtime serves (hq ADR 0198, to-be 38 WP4c)' (#254) from feat/0198-waves-2-3-the-forges-code-is-a-bundle into main 2026-10-03 23:01:17 +00:00
jochen cea59428b1 The forge's tests compose its code as a bundle the node's runtime serves (hq ADR 0198)
gitea's own code moves out of its runtime container (mesh-catalog, to-be 38 WP4c waves 2-3), so the three tests that composed the forge from the catalogue beside this checkout resolve its build as the code bundle, compose it beside the node's runtime, and read the forge's address from the words the runtime hands the module rather than from a sidecar's env.
2026-10-04 00:50:06 +02:00
mesh-admin 5c832f2d19 Merge pull request 'Compose a process's environment as a container's' (#250) from feat/a-process-env-is-composed-like-a-containers into main 2026-10-03 22:29:11 +00:00
jochen cdebb7d1a5 Compose a process's environment as a container's
A module's own code moving out of its container (novox/hq to-be 38 WP4c)
becomes a process on the machine, and still has to be told what its
container was: the port this machine gave the module and where the
foundation's seats are. ${port:…} and ${seat:…} were filled only in a
file's content and a container's env, so in a process's env they reached
the machine as literals, and the modules that moved first (mesh-catalog
#245) wrote their run-once steps a 0600 env file instead. A process's env
now takes the same resolution and the same refusals; ${dir:…} and
${access:…} already did, and a bundle's env (ADR 0192) already resolves
${dir:…} and ${port:…}.
2026-10-04 00:27:42 +02:00
mesh-admin 6a803ea5b3 Merge pull request 'Issue 219: an older build request never replaces a newer one's artifact' (#249) from fix/issue-219-an-older-build-never-replaces-a-newer into main 2026-10-03 22:23:31 +00:00
jochen 9745c1ab31 An older build request never replaces a newer one's artifact
Builds of one module in flight together finish in any order, and the mesh
took whatever it heard last as what the module is: RegisterModule overwrote
the module's manifest unconditionally, and Held/BuiltAgainst/ReadRepositories
ordered builds by when they were recorded. A postgres build asked before the
mesh-tools runtime fix finished after the one asked after it, and the next
push deployed the stale image (novox/hq issue 219).

A build is now ordered by when it was asked, read from the build-<nanos> id
the controller writes: build.asked and module.built_asked (migration 0055).
A registration from an earlier request than the module's current one is
recorded and refused as superseded. A plan takes as its outcome only a build
asked at or after its own ask, so an earlier plan's leftover build cannot
settle a later plan. Ids of any other shape keep the old order.
2026-10-04 00:22:11 +02:00
mesh-admin c0c3c3fed4 Merge pull request 'A TypeScript bundle is one file per entrypoint, bundled in the toolchain (hq ADR 0193); a toolchain follows the SDK it stands on (hq issue 212)' (#247) from feat/a-typescript-bundle-is-one-file into main 2026-10-03 21:50:44 +00:00
mesh-admin c1449fffe9 Merge pull request 'Issue only the recorded holder a seat held once for the mesh (hq issue 218)' (#248) from fix/issue-218-only-the-holder-serves-a-mesh-seat into main 2026-10-03 21:30:50 +00:00
jochen f873c97db5 Issue only the recorded holder a seat held once for the mesh (novox/hq issue 218)
A module claiming a mesh-scoped seat was granted and issued the seat's subjects on every machine
it runs on, so the store's verbs answered from whichever postgres replied first. Where the mesh
records the seat's holder, only that (node, module) is now issued it; the module's own tools are
untouched everywhere.
2026-10-03 23:30:27 +02:00
jochen b0b3d87fe2 Run the toolchain's esbuild as itself: npm installs its native binary in place of the script 2026-10-03 23:26:08 +02:00
jochen ba189e6943 A TypeScript bundle is one file per entrypoint, bundled in the toolchain (hq ADR 0193); a toolchain follows the SDK it stands on (hq issue 212)
Every served bundle is its own process now, so each carries its own copy of what it imports: after
the compile and the launchers, the toolchain image's esbuild bundles every entrypoint in place and
every launcher under its own name into one ES module file, the SDK inlined, require provided to
inlined CommonJS, the launcher's shebang kept and its mode 0755. The toolchain's node_modules is
copied only for packages an artifact names external. An image without the bundler is refused by
name. Issue 212: build.on already passes a published package by its exact version and plans the
toolchain after it; tests say so.
2026-10-03 23:24:08 +02:00
33 changed files with 1650 additions and 89 deletions
+14 -2
View File
@@ -148,6 +148,11 @@ func buildFrom(result link.BuildResult) inventory.Build {
// rebuild the graph rather than a list of names.
Path: result.Path,
}
// When it was asked, which is what orders it against another build of the same module
// (novox/hq 04-ISSUES/219) — not when it was heard.
if asked, ok := link.BuildAskedAt(result.ID); ok {
kept.Asked = asked
}
for _, ref := range result.Against {
kept.Against = append(kept.Against, catalogue.Recorded(ref))
}
@@ -409,7 +414,7 @@ func buildOne(ctx context.Context, source buildSource, path, ref string, wait ti
// Correlated by something the control plane makes, not by the module's name: two builds of one
// module can be in flight, and the second answer is not the first one's.
request := link.BuildRequest{
ID: fmt.Sprintf("%s-%d", "build", time.Now().UnixNano()),
ID: link.NewBuildID(time.Now()),
Repository: repository,
Path: path,
Ref: ref,
@@ -526,6 +531,9 @@ func takeIn(ctx context.Context, inv *inventory.Inventory, result link.BuildResu
BuiltFrom: result.Commit, Head: result.Commit,
// What it stood on, so registration can judge a built manifest's base (to-be 38 WP2.4).
Against: kept.Against,
// When it was asked, so an older request heard later does not replace a newer one
// (novox/hq 04-ISSUES/219).
Asked: kept.Asked,
}
if result.Source != nil && result.Source.Seat != "" {
recorded.Repository, recorded.Seat = result.Source.Repository, result.Source.Seat
@@ -544,6 +552,10 @@ func takeIn(ctx context.Context, inv *inventory.Inventory, result link.BuildResu
result.On, result.Repository, short(result.Commit), err)
}
if err := inv.RegisterModule(ctx, manifest, recorded); err != nil {
if errors.Is(err, inventory.ErrSuperseded) {
return manifest, kept, fmt.Errorf("%s built %s (%s), recorded and not registered: %w",
result.On, manifest.Module, short(result.Commit), err)
}
return manifest, kept, err
}
return manifest, kept, nil
@@ -573,7 +585,7 @@ func buildAndShow(ctx context.Context, source buildSource, path, ref string, wai
defer ask.Close()
result, err := ask.Submit(ctx, link.BuildRequest{
ID: fmt.Sprintf("%s-%d", "build", time.Now().UnixNano()),
ID: link.NewBuildID(time.Now()),
Repository: repository, Path: path, Ref: ref,
Held: heldBy(ctx), Seats: seatBases(ctx),
}, wait)
+50
View File
@@ -2,9 +2,12 @@ package main
import (
"encoding/json"
"errors"
"strings"
"testing"
"time"
"github.com/novox/mesh-controller/internal/inventory"
"github.com/novox/mesh-controller/internal/link"
)
@@ -100,3 +103,50 @@ func TestABuildAtACommitKeepsTheBranchTheModuleFollows(t *testing.T) {
t.Errorf("a module first built at a commit follows %q, want the default branch", src.Ref)
}
}
// novox/hq 04-ISSUES/219: an older request heard after a newer one is recorded and not registered,
// so a push sends what the newer request built.
func TestAnOlderBuildHeardLaterDoesNotReplaceTheNewer(t *testing.T) {
open := aMesh(t)
ctx := t.Context()
older := time.Date(2026, 10, 3, 21, 33, 45, 0, time.UTC)
newer := time.Date(2026, 10, 3, 21, 51, 57, 0, time.UTC)
result := func(asked time.Time, image string) link.BuildResult {
manifest, _ := json.Marshal(map[string]any{"module": "postgres", "version": image})
return link.BuildResult{ID: link.NewBuildID(asked), Repository: "http://forge.internal:20000/novox/mesh-catalog.git",
Path: "modules/postgres", Ref: "main", On: "anchor", Commit: "efff5415", Manifest: manifest,
Source: &link.SourceOnSeat{Seat: "git", Repository: "novox/mesh-catalog"}}
}
if _, _, err := takeIn(ctx, open.inventory, result(newer, "4bcd5f73")); err != nil {
t.Fatal(err)
}
_, _, err := takeIn(ctx, open.inventory, result(older, "0ab07fa9"))
if !errors.Is(err, inventory.ErrSuperseded) {
t.Fatalf("the older request's outcome was taken in as current: %v", err)
}
shelf, err := open.inventory.Catalogue(ctx)
if err != nil {
t.Fatal(err)
}
if got := shelf["postgres"].Version; got != "4bcd5f73" {
t.Errorf("postgres is %q; want the newer request's 4bcd5f73", got)
}
if builds, _ := open.inventory.Builds(ctx, "postgres", 5); len(builds) != 2 {
t.Errorf("the late build was not recorded: %v", builds)
}
}
func TestABuildIDSaysWhenItWasAsked(t *testing.T) {
at := time.Date(2026, 10, 3, 21, 51, 57, 392539762, time.UTC)
if got, ok := link.BuildAskedAt(link.NewBuildID(at)); !ok || !got.Equal(at) {
t.Errorf("read back %v %v; want %v", got, ok, at)
}
if got, ok := link.BuildAskedAt("build-1791064317392539762"); !ok || got.Format(time.TimeOnly) != "21:51:57" {
t.Errorf("the incident's id reads as %v %v", got, ok)
}
for _, id := range []string{"b-1", "build-2", "build-", "build-x", ""} {
if _, ok := link.BuildAskedAt(id); ok {
t.Errorf("%q read as a request time", id)
}
}
}
+13 -3
View File
@@ -8,6 +8,7 @@ package main
import (
"context"
"errors"
"flag"
"fmt"
"os"
@@ -253,25 +254,34 @@ func parseAround(set *flag.FlagSet, args []string) ([]string, error) {
// became of a build nobody was watching.
func (b builds) Built(ctx context.Context, result link.BuildResult) error {
manifest, _, err := takeIn(ctx, b.inv, result)
// When it was asked, so a plan takes as its outcome only a build asked for it or after it
// (novox/hq 04-ISSUES/219). Zero when the id does not say.
asked, _ := link.BuildAskedAt(result.ID)
switch {
case err != nil && result.Failed != "":
fmt.Printf("%s: %v\n", result.ID, err)
if result.Module != "" {
planBuilt(ctx, b.open, result.Module, result.Commit, result.Failed)
planBuilt(ctx, b.open, result.Module, result.Commit, result.Failed, asked)
} else {
planFailedBuild(ctx, b.open, result)
}
return nil
case errors.Is(err, inventory.ErrSuperseded):
// Not a failure: the module is already at what a later request built. A plan that asked
// before that later request is answered by it; one that asked after it ignores this.
fmt.Printf("%s: %v\n", result.ID, err)
planBuilt(ctx, b.open, manifest.Module, result.Commit, "", asked)
return nil
case err != nil:
fmt.Printf("%s: heard and recorded, and not registered: %v\n", result.ID, err)
if manifest.Module != "" {
planBuilt(ctx, b.open, manifest.Module, result.Commit, err.Error())
planBuilt(ctx, b.open, manifest.Module, result.Commit, err.Error(), asked)
}
return nil
}
fmt.Printf("%s: %s %s registered, built on %s from %s\n",
result.ID, manifest.Module, manifest.Version, result.On, short(result.Commit))
saysWhenThePolicyActs(ctx, b.inv, manifest.Module)
planBuilt(ctx, b.open, manifest.Module, result.Commit, "")
planBuilt(ctx, b.open, manifest.Module, result.Commit, "", asked)
return nil
}
@@ -0,0 +1,47 @@
package main
import (
"testing"
"time"
"github.com/novox/mesh-controller/internal/inventory"
)
// novox/hq issue 213: for the moment a machine hands its controller over, the container and the
// process both run the plan timer on one store. Only the one holding the plans moves them; the other
// leaves them alone, and moves them once they are let go.
func TestAControllerLeavesThePlansToTheOneHoldingThem(t *testing.T) {
open := aMesh(t)
ctx := t.Context()
now := time.Now().UTC()
// Every tier done: the next step is the plan's last, and needs nothing but the store.
plan := inventory.Plan{ID: "plan-213", Repository: "r", Commit: "abc", Created: now, Updated: now,
State: inventory.PlanRolling, Tier: 1, Tiers: [][]string{{"app"}},
Modules: map[string]*inventory.PlanModule{"app": {State: "built"}}}
if err := open.inventory.SavePlan(ctx, plan); err != nil {
t.Fatal(err)
}
// The other controller: its own connections to the same store, holding the plans.
other, err := inventory.Open(ctx)
if err != nil {
t.Fatal(err)
}
t.Cleanup(other.Close)
release, err := other.HoldPlans(ctx, false)
if err != nil {
t.Fatal(err)
}
t.Cleanup(release) // before the close above: a pool waits for a connection still held
advancePlans(ctx, open)
if p, err := open.inventory.PlanByID(ctx, "plan-213"); err != nil || !p.Open() {
t.Fatalf("a controller moved a plan another held: %+v %v", p, err)
}
release()
advancePlans(ctx, open)
if p, err := open.inventory.PlanByID(ctx, "plan-213"); err != nil || p.State != inventory.PlanDone {
t.Fatalf("the plan did not move once it was let go: %+v %v", p, err)
}
}
+64 -3
View File
@@ -2,6 +2,7 @@ package main
import (
"context"
"errors"
"flag"
"fmt"
"sort"
@@ -288,8 +289,23 @@ func askTier(ctx context.Context, inv *inventory.Inventory, p *inventory.Plan) e
// planBuilt marks a module built (or failed) in every open plan whose current tier holds it, and
// advances what that completes. Called from the daemon's take-in of every outcome.
func planBuilt(ctx context.Context, open *stores, module, commit, failed string) {
//
// **Only a build asked at or after the plan's ask is its outcome** (novox/hq 04-ISSUES/219). Two
// plans a few minutes apart both ask for a module; the earlier plan's build, finishing late, is not
// the later plan's answer — it stood on the bases from before the later plan's merge, and taking it
// would send machines, and the next tier, what the later merge replaced. asked is zero when the
// build's request time is not known, and such an outcome is taken as before.
func planBuilt(ctx context.Context, open *stores, module, commit, failed string, asked time.Time) {
inv := open.inventory
// One controller works the plans at a time (novox/hq issue 213); an outcome waits its turn rather
// than write over what the holder is about to save. Not taken, it is still in the build records,
// which the holder settles the plan from (issue 214).
release, err := inv.HoldPlans(ctx, true)
if err != nil {
fmt.Printf("plans: %s's outcome is left to the build records: %v\n", module, err)
return
}
defer release()
plans, err := inv.OpenPlans(ctx)
if err != nil {
fmt.Printf("plans: cannot read them: %v\n", err)
@@ -315,6 +331,9 @@ func planBuilt(ctx context.Context, open *stores, module, commit, failed string)
state = &inventory.PlanModule{}
p.Modules[module] = state
}
if askedBefore(asked, state.AskedAt) {
continue
}
if failed != "" {
state.State = "failed"
state.Why = failed
@@ -333,13 +352,30 @@ func planBuilt(ctx context.Context, open *stores, module, commit, failed string)
fmt.Printf("%s: %s; the tiers after it are not asked\n", p.ID, p.Note)
}
}
advancePlans(ctx, open)
advanceHeld(ctx, open)
}
// advancePlans moves every open plan as far as the facts allow: a tier whose modules are all built
// and whose gates are applied gives way to the next; the last tier done is the plan done. Called
// after every outcome and on a timer, so a plan waiting on a machine's report moves when it comes.
//
// **One controller at a time** (novox/hq issue 213). A plan is read, changed and saved whole; two
// controllers — the old and the new while a machine hands its controller over — would each ask a
// tier the other had just asked. Taken without waiting: whoever holds the plans is moving them.
func advancePlans(ctx context.Context, open *stores) {
release, err := open.inventory.HoldPlans(ctx, false)
if err != nil {
if !errors.Is(err, inventory.ErrPlansBusy) {
fmt.Printf("plans: cannot hold them: %v\n", err)
}
return
}
defer release()
advanceHeld(ctx, open)
}
// advanceHeld is advancePlans for a caller already holding the plans.
func advanceHeld(ctx context.Context, open *stores) {
inv := open.inventory
plans, err := inv.OpenPlans(ctx)
if err != nil {
@@ -549,7 +585,8 @@ func planFailedBuild(ctx context.Context, open *stores, result link.BuildResult)
}
for _, e := range entries {
if repositoryMatches(e.Source.Repository, result.Repository) && e.Source.Path == result.Path {
planBuilt(ctx, open, e.Manifest.Module, result.Commit, result.Failed)
asked, _ := link.BuildAskedAt(result.ID)
planBuilt(ctx, open, e.Manifest.Module, result.Commit, result.Failed, asked)
return
}
}
@@ -666,6 +703,19 @@ func plansCommand(ctx context.Context, args []string) error {
}
p.State = inventory.PlanFailed
p.Note = "stopped by hand at tier " + fmt.Sprint(p.Tier)
release, err := inv.HoldPlans(ctx, true)
if err != nil {
return err
}
defer release()
if p, err = inv.PlanByID(ctx, positionals[1]); err != nil {
return err
}
if !p.Open() {
return fmt.Errorf("%s is already %s", p.ID, p.State)
}
p.State = inventory.PlanFailed
p.Note = "stopped by hand at tier " + fmt.Sprint(p.Tier)
if err := inv.SavePlan(ctx, p); err != nil {
return err
}
@@ -791,6 +841,11 @@ func settleFromRecords(p *inventory.Plan, tier []string, recorded map[string][]i
if b.At.Before(*s.AskedAt) {
break
}
// Recorded after the ask and asked before it: an earlier ask's late outcome, not this
// one's (novox/hq 04-ISSUES/219).
if askedBefore(b.Asked, s.AskedAt) {
continue
}
outcome = &b
}
if outcome == nil {
@@ -812,3 +867,9 @@ func settleFromRecords(p *inventory.Plan, tier []string, recorded map[string][]i
}
return changed
}
// askedBefore is whether a build asked at asked was asked before a plan asked for its module — and
// so is not that plan's outcome (novox/hq 04-ISSUES/219). False when either time is not known.
func askedBefore(asked time.Time, planAsked *time.Time) bool {
return !asked.IsZero() && planAsked != nil && asked.Before(*planAsked)
}
+18
View File
@@ -155,6 +155,24 @@ func TestAPlanSettlesAnAskedBuildFromTheRecords(t *testing.T) {
t.Errorf("an ask with no record after it was settled: %+v", s)
}
// novox/hq 04-ISSUES/219: a build recorded after the ask but asked before it — an earlier
// plan's late outcome — is not this ask's, built or failed.
r := inventory.Plan{ID: "plan-3", Tiers: [][]string{{"postgres"}},
Modules: map[string]*inventory.PlanModule{"postgres": {State: "asked", AskedAt: &asked}}}
late := map[string][]inventory.Build{"postgres": {
{ID: "build-old", Commit: "efff5415", Asked: asked.Add(-18 * time.Minute), At: asked.Add(12 * time.Minute)},
}}
if settleFromRecords(&r, r.Tiers[0], late) || r.Modules["postgres"].State != "asked" {
t.Errorf("an earlier ask's late outcome settled this ask: %+v", r.Modules["postgres"])
}
// Newest heard first: the earlier ask's late outcome, then this ask's own, heard before it.
late["postgres"] = append(late["postgres"], inventory.Build{ID: "build-mine", Commit: "4bcd5f73",
Asked: asked.Add(time.Second), At: asked.Add(5 * time.Minute)})
if !settleFromRecords(&r, r.Tiers[0], late) || r.Modules["postgres"].State != "built" ||
r.Modules["postgres"].Commit != "4bcd5f73" {
t.Errorf("this ask's own outcome, heard before the earlier ask's, did not settle it: %+v", r.Modules["postgres"])
}
// A failure recorded after the ask fails the plan, as hearing it would have.
q := inventory.Plan{ID: "plan-2", Tiers: [][]string{{"x"}},
Modules: map[string]*inventory.PlanModule{"x": {State: "asked", AskedAt: &asked}}}
+7
View File
@@ -315,6 +315,13 @@ func (f following) SourceMoved(ctx context.Context, m link.SourceMoved) error {
for _, e := range moved {
movedNames = append(movedNames, e.Manifest.Module)
}
// Written and its first tier asked as one act on the plans (novox/hq issue 213): a timer on
// another controller reading it between the two would ask the tier again.
release, err := inv.HoldPlans(ctx, true)
if err != nil {
return notNow(err)
}
defer release()
plan := planOfMerge(m, movedNames, edges)
if hasCycle(plan.Tiers, edges) {
fmt.Printf(" the last tier depends on itself: %s — built together, in no order\n",
+106 -3
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, seatBases, say)
made, err := one(ctx, run, publish, manifest.Module, within, workspace, commit, credentials, a, args, held, npmrcPath, npmrc, 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, seats map[string]string,
held map[string]string, npmrc string, registry Npmrc, seats map[string]string,
say func(step, format string, args ...any)) (catalogue.Built, error) {
switch a.Kind {
@@ -582,6 +582,11 @@ 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 {
@@ -593,6 +598,12 @@ func one(ctx context.Context, run Runner, publish Publisher,
if err != nil {
return catalogue.Built{}, fmt.Errorf("%s: writing %s's launchers failed: %w", module, a.Name, err)
}
if chain.Bundler != "" {
say("bundle", "bundling each entrypoint into one file")
if compiled, err = bundled(ctx, run, tree, chain, base, a, launchers); err != nil {
return catalogue.Built{}, fmt.Errorf("%s: bundling %s failed: %w", module, a.Name, err)
}
}
say("bundle", "compiled, packing")
body, err := pack(compiled)
if err != nil {
@@ -974,7 +985,7 @@ func compile(ctx context.Context, run Runner, tree string, chain Toolchain,
if _, err := run(ctx, tree, "docker", invocation...); err != nil {
return "", err
}
if chain.Dependencies != "" {
if chain.Dependencies != "" && chain.Bundler == "" {
// **What the bundle runs with, from the image it was compiled in** (Toolchain.Dependencies).
// A second run in the same image rather than a shell wrapped around the compiler: the
// compile line stays a plain command a reader can run by hand, and the copy is one more
@@ -1224,3 +1235,95 @@ func writeLaunchers(root string, chain Toolchain, a catalogue.Artifact) (map[str
}
return out, nil
}
// bundledSuffix is where a bundle's one-file output is written, beside what the compiler wrote.
const bundledSuffix = ".bundled"
// bundled makes every entrypoint and every launcher of a compiled bundle ONE file, in the toolchain
// image's bundler, and answers the directory to pack (novox/hq ADR 0193).
//
// **What a launched bundle runs is what it imports, and nothing else.** Every served bundle is its
// own process, so it carries its own copy of the SDK and its own dependencies inlined — the
// toolchain's whole node_modules no longer travels in every bundle. An entrypoint a process runs by
// name (`node daemon/index.js`) is bundled in place under its own name; a launcher keeps its name
// and its first line, and stays executable. A package the bundler cannot inline is named by the
// artifact (`external`), kept as an import, and only then is the toolchain's runtime directory
// copied beside the files. CommonJS inlined into an ES module still finds `require`.
func bundled(ctx context.Context, run Runner, tree string, chain Toolchain, base string,
a catalogue.Artifact, launchers map[string]string) (string, error) {
const within = "/app/modules/module"
out, final := Out(a.Name), Out(a.Name)+bundledSuffix
if err := os.RemoveAll(filepath.Join(tree, final)); err != nil {
return "", err
}
if err := os.MkdirAll(filepath.Join(tree, final), 0o755); err != nil {
return "", err
}
common := []string{"--bundle", "--platform=node", "--format=esm", "--target=node22",
"--outbase=" + out, "--outdir=" + final, "--log-level=warning",
"--banner:js=import { createRequire as __meshRequire } from 'node:module'; const require = __meshRequire(import.meta.url);"}
for _, x := range a.External {
common = append(common, "--external:"+x)
}
var plain []string
for _, e := range a.Entrypoints {
if strings.HasSuffix(e, ".js") {
plain = append(plain, out+"/"+e)
}
}
var launch []string
for _, l := range sortedValues(launchers) {
launch = append(launch, out+"/"+l)
}
// Refused by name in an image that predates the bundler, as the dependencies copy is: a bundle
// packed without it would carry nothing it imports. Run as itself: npm installs esbuild's native
// binary in place of its script, which `node` cannot run.
guard := `test -x "$0" || { echo "the toolchain image carries no bundler at $0: it predates one-file bundles, rebuild mesh-tools first" >&2; exit 1; }; exec "$0" "$@"`
step := func(entries []string, extra ...string) error {
if len(entries) == 0 {
return nil
}
invocation := []string{"run", "--rm", "--volume", tree + ":" + within, "--workdir", within, base,
"sh", "-c", guard, chain.Bundler}
invocation = append(invocation, entries...)
invocation = append(invocation, common...)
invocation = append(invocation, extra...)
_, err := run(ctx, tree, "docker", invocation...)
return err
}
if err := step(plain); err != nil {
return "", err
}
if err := step(launch, "--out-extension:.js=.mjs"); err != nil {
return "", err
}
// Plain `.js` output is an ES module; said once, as the runtime directory used to say it.
if err := os.WriteFile(filepath.Join(tree, final, "package.json"), []byte(`{"type":"module","private":true}`+"\n"), 0o644); err != nil {
return "", err
}
for _, l := range launchers {
path := filepath.Join(tree, final, filepath.FromSlash(l))
if _, err := os.Stat(path); err == nil {
if err := os.Chmod(path, 0o755); err != nil {
return "", err
}
}
}
if len(a.External) > 0 && chain.Dependencies != "" {
copying := []string{"run", "--rm", "--volume", tree + ":" + within, "--workdir", within, base,
"sh", "-c", `cp -a "$0/node_modules" "$1/"`, chain.Dependencies, final}
if _, err := run(ctx, tree, "docker", copying...); err != nil {
return "", fmt.Errorf("copying the packages %s keeps external: %w", a.Name, err)
}
}
return filepath.Join(tree, final), nil
}
func sortedValues(m map[string]string) []string {
out := make([]string, 0, len(m))
for _, v := range m {
out = append(out, v)
}
sort.Strings(out)
return out
}
+38 -14
View File
@@ -83,25 +83,49 @@ func TestABundleIsCompiledAndPackedWithNoDockerfile(t *testing.T) {
t.Fatalf("the bundle was not pinned: %v", got.Manifest.Resources[0])
}
// **And what it runs with, from the image it was compiled in** (novox/hq to-be 38 WP3). A
// second run in the same toolchain image copies the toolchain's runtime directory — the
// `"type": "module"` package.json and the pruned node_modules — into the output's root, and
// refuses by name when the image carries none rather than packing a bundle that starts nowhere.
var copied string
// **One file per entrypoint and launcher, in the toolchain's bundler** (novox/hq ADR 0193). A
// second run in the same toolchain image bundles each into the artifact's bundled output, the SDK
// inlined, refusing by name in an image that predates the bundler; and the toolchain's
// node_modules is no longer copied into a bundle that keeps nothing external.
var bundling []string
for _, line := range r.ran {
if strings.HasPrefix(line, "docker run") && strings.Contains(line, "/app/runtime") {
copied = line
if strings.HasPrefix(line, "docker run") && strings.Contains(line, "esbuild") {
bundling = append(bundling, line)
}
}
if copied == "" {
t.Fatalf("the bundle's dependencies were not copied in after the compile:\n%s", strings.Join(r.ran, "\n"))
if len(bundling) != 2 {
t.Fatalf("want one bundling run for the entrypoints and one for the launchers:\n%s", strings.Join(r.ran, "\n"))
}
if !strings.Contains(copied, "mesh-tools/build@sha256:") || !strings.Contains(copied, "predates") ||
!strings.Contains(copied, Out("code")) {
t.Fatalf("the copy does not run in the same toolchain, refuse an older image by name, or land in the artifact's output: %s", copied)
for _, want := range []string{"mesh-tools/build@sha256:", "predates one-file bundles", "--bundle", "--format=esm",
"--platform=node", "--outdir=" + Out("code") + ".bundled", Out("code") + "/index.js"} {
if !strings.Contains(bundling[0], want) {
t.Errorf("the entrypoints' bundling lacks %q: %s", want, bundling[0])
}
}
if strings.Index(strings.Join(r.ran, "\n"), "--outDir") > strings.Index(strings.Join(r.ran, "\n"), "/app/runtime") {
t.Fatal("the dependencies were copied before the compile wrote its output")
if !strings.Contains(bundling[1], Out("code")+"/index.serve.mjs") || !strings.Contains(bundling[1], "--out-extension:.js=.mjs") {
t.Errorf("the launcher is not bundled under its own name: %s", bundling[1])
}
if strings.Contains(strings.Join(r.ran, "\n"), "/app/runtime") {
t.Errorf("the toolchain's node_modules was copied into a bundle that keeps nothing external:\n%s", strings.Join(r.ran, "\n"))
}
if strings.Index(strings.Join(r.ran, "\n"), "--outDir") > strings.Index(strings.Join(r.ran, "\n"), "esbuild") {
t.Fatal("the bundler ran before the compile wrote its output")
}
}
// A bundle naming packages it keeps external is bundled with them as imports, and carries the
// toolchain's node_modules for them — the one case it still does.
func TestABundleKeepingAPackageExternalCarriesTheToolchainsModules(t *testing.T) {
manifest := strings.Replace(aBundle, `"entrypoints":["index.js"]`, `"entrypoints":["index.js"],"external":["sharp"]`, 1)
r, workspace := aRepository(t, manifest, map[string]string{"index.ts": "console.log(1)"})
held := map[string]string{"mesh-tools/build": "registry.invalid/mesh-tools/build@sha256:" + strings.Repeat("b", 64)}
if _, err := Build(context.Background(), compiling{r}.run, r,
"https://forge.invalid/greeter.git", "", "", workspace, held, Npmrc{}, GitCredential{}, nil); err != nil {
t.Fatal(err)
}
all := strings.Join(r.ran, "\n")
if !strings.Contains(all, "--external:sharp") || !strings.Contains(all, "/app/runtime") {
t.Errorf("an external package was not kept as an import with the toolchain's modules beside it:\n%s", all)
}
}
+147
View File
@@ -0,0 +1,147 @@
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
@@ -0,0 +1,131 @@
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)
}
}
+20
View File
@@ -200,3 +200,23 @@ func TestWhatABuildReadIsTheRepositoriesItsRecipesName(t *testing.T) {
t.Fatal("a module whose recipes name no other repository read one")
}
}
// novox/hq 04-ISSUES/212: a toolchain stands on the SDK's published package, and is built with the
// exact version the mesh published — an argument that changes when the SDK does, so a rebuild after
// a release never reuses an install of the version before it.
func TestAPackageTheMeshPublishedIsPassedByItsExactVersion(t *testing.T) {
manifest := catalogue.Manifest{
Module: "mesh-tools",
Build: &catalogue.Build{
On: []catalogue.BuildsOn{{Arg: "MESH_SDK", Module: "mesh-sdk", Artifact: "lib"}},
},
}
held := map[string]string{"mesh-sdk/lib": "@novox/mesh-sdk@0.1.6"}
args, resolved, err := standingOn(context.Background(), manifest, held, noMirror)
if err != nil {
t.Fatal(err)
}
if fmt.Sprint(args) != "[--build-arg MESH_SDK=@novox/mesh-sdk@0.1.6]" || fmt.Sprint(resolved) != "[@novox/mesh-sdk@0.1.6]" {
t.Errorf("the package was passed as %v, recorded as %v", args, resolved)
}
}
+13 -2
View File
@@ -67,13 +67,23 @@ 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 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 —
// 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 —
// 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
// that starts nowhere: the image predates this and must be rebuilt first.
//
// *Since the bundler (below):* copied only for a bundle that names packages it keeps external,
// which cannot be inlined; a bundle with none carries no node_modules at all.
Dependencies string
// Bundler is the bundler inside the toolchain image that makes each compiled entrypoint and each
// launcher ONE self-contained file (novox/hq ADR 0193): every served bundle is its own process
// now, so each carries its own copy of what it imports — the SDK included — and nothing else.
// A bundle shrinks from the toolchain's whole node_modules to the code it runs. Empty for a
// language whose build is already one file.
Bundler string
// SystemStamp is the variable this language's linker fills with the artifact's declared system,
// for a language whose binaries are pinned to one at link time (novox/hq ADR 0005).
//
@@ -139,6 +149,7 @@ var toolchains = []Toolchain{
Unit: UnitSources,
SourceExt: ".ts",
Dependencies: "/app/runtime",
Bundler: "/app/node_modules/esbuild/bin/esbuild",
},
{
Language: "go",
@@ -0,0 +1,166 @@
package catalogue
import (
"encoding/json"
"strings"
"testing"
)
// novox/hq issue 213 (ADR 0188 §1, §3): a module's own Go service is a bundle the host runs as a
// process, not an image. Each test holds one thing that had to change in the composer for the
// controller to be declared that way.
const aServiceDigest = "sha256:dddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddd"
// aServiceModule is the controller's shape in miniature: it answers tools of its own, its code is a
// Go bundle a process runs as an account it declares, its secrets belong to that account, it
// prepares its state, and its process replaces the container it used to run as.
func aServiceModule(t *testing.T) Manifest {
t.Helper()
raw := `{
"module": "svc", "version": "1", "tools": ["status"], "prepares": true,
"own-secrets": {"store": "${dir:state}/store"},
"secrets-owner": "svc",
"resources": [
{"id": "state", "type": "directory", "mode": "0700", "place": "mesh", "owner": "svc"},
{"id": "service", "type": "process", "name": "svc", "artifact": "code",
"run": ["./svc", "serve"], "user": "svc", "replaces": ["server"],
"env": {"SVC_STORE_FILE": "${dir:state}/store", "SVC_STORE_PORT": "${seat:mesh-store:5432}"}},
{"id": "account", "type": "user", "name": "svc", "shell": "/usr/bin/nologin", "home": "/var/lib/svc"}
],
"build": {"artifacts": [{"name": "code", "kind": "bundle", "language": "go", "system": "arch",
"from": "cmd/svc", "binary": "svc"}]}
}`
m, err := ParseManifest([]byte(raw))
if err != nil {
t.Fatalf("the service's manifest is refused: %v", err)
}
resolved, err := m.Resolve([]Built{{Name: "code", Kind: ArtifactBundle,
Reference: ArtifactStoreScheme + "svc/code@" + aServiceDigest, Digest: aServiceDigest}})
if err != nil {
t.Fatal(err)
}
return resolved
}
func composeTheService(t *testing.T, with Rendering) []map[string]any {
t.Helper()
with.Needed = map[string]map[string]string{"svc": {"store": "sealed-store"}}
with.ArtifactStore = "anchor.internal:5100"
out, err := Resolution{Node: "anchor", Modules: []Manifest{aServiceModule(t)}}.Declaration(with)
if err != nil {
t.Fatalf("the service does not compose: %v", err)
}
return out
}
func indexOf(out []map[string]any, id string) int {
for i, r := range out {
if r["id"] == id {
return i
}
}
return -1
}
// A module that declares tools has every bundle served by the node's runtime unless it says
// otherwise — and the controller declares the verbs it answers as tools. Its service bundle is run
// by its own process; launched a second time by the runtime it would be a second controller
// pretending to be an MCP server.
func TestABundleItsOwnProcessRunsIsNotServedByTheRuntime(t *testing.T) {
m := aServiceModule(t)
if len(m.Bundles) != 1 {
t.Fatalf("the service's bundle was not kept: %+v", m.Bundles)
}
if loads := m.Bundles[0].Loads; len(loads) != 0 {
t.Fatalf("the runtime would launch the service's own bundle as tools: %v", loads)
}
// And a bundle no resource runs still is served, as a module declaring tools always had it.
tools := Manifest{Module: "t", Version: "1", Tools: []string{"x"},
Build: &Build{Artifacts: []Artifact{{Name: "tools", Kind: ArtifactBundle, Language: "go",
System: "arch", From: "cmd/t"}}}}
resolved, err := tools.Resolve([]Built{{Name: "tools", Kind: ArtifactBundle,
Reference: ArtifactStoreScheme + "t/tools@" + aServiceDigest, Digest: aServiceDigest}})
if err != nil {
t.Fatal(err)
}
if loads := resolved.Bundles[0].Loads; len(loads) != 1 || loads[0] != "t" {
t.Fatalf("a tools bundle nothing runs is no longer served: %v", loads)
}
}
// The account is created before anything is given to it. Its secrets are mesh-computed and so
// placed before the module's own resources; given to a user the machine did not have yet, they were
// refused on the first apply and the process started without them.
func TestAModulesAccountComesBeforeWhatBelongsToIt(t *testing.T) {
out := composeTheService(t, Rendering{})
account, secret := indexOf(out, "svc.account"), indexOf(out, "svc."+NeedID("store"))
if account < 0 || secret < 0 {
t.Fatalf("the account or the secret is missing: %v", out)
}
if account > secret {
t.Fatalf("the secret owned by svc is written before svc exists: account at %d, secret at %d",
account, secret)
}
if owner := out[secret]["owner"]; owner != "svc" {
t.Errorf("the secret belongs to %v, not the account its process runs as", owner)
}
}
// The process is the module's program; its preparation is the same program asked to prepare, as a
// step before it — with the same account and environment, and handing nothing over.
func TestAProcessIsPreparedByItsOwnProgram(t *testing.T) {
out := composeTheService(t, Rendering{})
step, process := indexOf(out, "svc.service-prepare"), indexOf(out, "svc.service")
if step < 0 || process < 0 || step > process {
t.Fatalf("the preparation is not a step before the process (%d, %d): %v", step, process, out)
}
s := out[step]
if s["type"] != "process" || s["run-once"] != true || s["name"] != "svc-prepare" {
t.Errorf("the preparation is not a run-once process: %v", s)
}
if run, _ := json.Marshal(s["run"]); string(run) != `["./svc","prepare"]` {
t.Errorf("the preparation runs %s", run)
}
if s["user"] != "svc" || s["source"] != out[process]["source"] {
t.Errorf("the preparation does not run the same bundle as the same account: %v", s)
}
if env, _ := s["env"].(map[string]any); env["SVC_STORE_FILE"] == nil {
t.Errorf("the preparation is not given the process's environment: %v", s["env"])
}
if _, has := s["replaces"]; has {
t.Errorf("the preparation would hand over what the process replaces: %v", s)
}
if _, has := s["args"]; has {
t.Errorf("the preparation carries a container's args: %v", s)
}
}
// What the process replaces is named as the host recorded it, `<module>.<id>`; unprefixed, the host
// matches nothing and removes the container first, as before.
func TestWhatAProcessReplacesIsNamedAsTheHostRecordedIt(t *testing.T) {
out := composeTheService(t, Rendering{})
p := out[indexOf(out, "svc.service")]
if got, _ := json.Marshal(p["replaces"]); string(got) != `["svc.server"]` {
t.Fatalf("the process replaces %s", got)
}
}
func TestWhatReplacesMayNameIsRefusedNearItsAuthor(t *testing.T) {
for what, resource := range map[string]string{
"a container": `{"id":"c","type":"container","name":"c","image":"x@` + aServiceDigest + `","replaces":["old"]}`,
"a step": `{"id":"p","type":"process","name":"p","run":["./p"],"run-once":true,"replaces":["old"]}`,
"something declared": `{"id":"p","type":"process","name":"p","run":["./p"],"replaces":["p"]}`,
"another module's": `{"id":"p","type":"process","name":"p","run":["./p"],"replaces":["other.old"]}`,
"not a list": `{"id":"p","type":"process","name":"p","run":["./p"],"replaces":"old"}`,
} {
raw := `{"module":"m","version":"1","resources":[` + resource + `]}`
if _, err := ParseManifest([]byte(raw)); err == nil || !strings.Contains(err.Error(), "replace") {
t.Errorf("replaces on %s was accepted: %v", what, err)
}
}
ok := `{"module":"m","version":"1","resources":[{"id":"p","type":"process","name":"p","run":["./p"],"replaces":["old"]}]}`
if _, err := ParseManifest([]byte(ok)); err != nil {
t.Errorf("a process replacing what its module no longer declares was refused: %v", err)
}
}
+25 -7
View File
@@ -77,6 +77,7 @@ func (m Manifest) Resolve(built []Built) (Manifest, error) {
// build compare equal.
out.Bundles = nil
if m.Build != nil {
run := runByAResource(m)
for _, a := range m.Build.Artifacts {
if a.Kind != ArtifactBundle {
continue
@@ -84,8 +85,13 @@ func (m Manifest) Resolve(built []Built) (Manifest, error) {
made := by[a.Name]
// What the runtime loads: what the artifact said, else every entrypoint of a module
// that declares tools, else nothing (the field's own rule; see Artifact.Loads).
//
// **Never, unasked, a bundle one of the module's own resources runs** (novox/hq issue 213).
// A process the host runs is the module's service, not its tools: the controller declares
// the verbs it answers as `tools` and serves them itself, and its bundle would otherwise
// have been launched a second time by the node's runtime, as an MCP child it is not.
loads := append([]string(nil), a.Loads...)
if a.Loads == nil && len(m.Tools) > 0 {
if a.Loads == nil && len(m.Tools) > 0 && !run[a.Name] {
loads = append([]string(nil), a.Entrypoints...)
// A bundle compiled to a binary has no entrypoints: the binary is what it is, and what
// the runtime starts to serve it (novox/hq ADR 0193). So a Go tools bundle is served
@@ -214,6 +220,11 @@ func (b *Build) problems(module string) []string {
// A bundle's source is the module's own directory by definition, and what it needs to say
// is which compiler — because the mesh chooses that, and cannot choose for a module that
// has not said.
if len(a.External) > 0 && (a.Kind != ArtifactBundle || a.Language != "typescript") {
problems = append(problems, fmt.Sprintf(
"%s: %q names packages it keeps external, and only a TypeScript bundle is bundled into "+
"one file with some kept out (novox/hq ADR 0193)", module, a.Name))
}
if len(a.Env) > 0 && a.Kind != ArtifactBundle {
problems = append(problems, fmt.Sprintf(
"%s: %q is a %q and says what it is given (env). Only a bundle the node's runtime "+
@@ -448,6 +459,18 @@ func BinaryOf(a Artifact) string {
return a.Name
}
// runByAResource is the artifacts one of a module's own resources names — a process that runs it,
// a step, an archive that unpacks it — by name.
func runByAResource(m Manifest) map[string]bool {
named := map[string]bool{}
for _, r := range m.Resources {
if a, ok := r["artifact"].(string); ok && a != "" {
named[a] = true
}
}
return named
}
// undeliveredBundles says which of a module's bundles nothing would ever put on a machine (novox/hq
// 04-ISSUES/216). A bundle reaches a machine three ways: the node's runtime serves it (it says
// `loads`, or its module declares `tools`), a resource names it (a process, a step, an archive), or
@@ -458,12 +481,7 @@ func undeliveredBundles(m Manifest) []string {
if m.Build == nil || m.Module == RuntimeModule {
return nil
}
named := map[string]bool{}
for _, r := range m.Resources {
if a, ok := r["artifact"].(string); ok && a != "" {
named[a] = true
}
}
named := runByAResource(m)
var problems []string
for _, a := range m.Build.Artifacts {
if a.Kind != ArtifactBundle || named[a.Name] || len(a.Loads) > 0 || len(m.Tools) > 0 {
+52 -4
View File
@@ -693,7 +693,15 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
// Now, and not before: a module whose resources are computed replaces them wholesale, and
// merging earlier would throw away the files it still needs.
resources = append(append([]map[string]any{}, first...), resources...)
//
// **Except the module's own accounts, which go before even those** (novox/hq issue 213). What
// the mesh computes may belong to one: a module whose code runs as an account it declares has
// its secrets written owned by that account, and a file given to a user the machine does not
// have yet fails — so on the first apply the secrets were refused, the process started without
// them, and the second apply healed it, which is the fault the paragraph above describes.
// An account depends on nothing the mesh computes.
accounts, rest := accountsFirst(resources)
resources = append(append(accounts, first...), rest...)
// No container is given the mesh's names (novox/hq ADR 0148). It used to be: every
// container got the whole roster as `--add-host` entries at creation, and a name that
@@ -872,6 +880,12 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
if renamed := reflectsRenamed(m.Module, resource["reload-on"]); renamed != nil {
copied["reload-on"] = renamed
}
// And what a process replaces (novox/hq issue 213): a resource of this module's that it
// no longer declares, named as the host recorded it, or the host hands nothing over and
// removes it first.
if renamed := reflectsRenamed(m.Module, resource["replaces"]); renamed != nil {
copied["replaces"] = renamed
}
// **What reads one of this module's own secrets is restarted when it changes** (novox/hq
// issue 203, issue 206). A credential is re-issued by the mesh, and a container that
// mounted the old file keeps the old one open: the build machine ran for an hour on a
@@ -2010,12 +2024,18 @@ func preparationTarget(m Manifest) string {
return ""
}
for _, r := range m.Resources {
if fmt.Sprint(r["type"]) != "container" || !ownArtifact(r, m.Module) {
// A container, or a process the host runs from a bundle the module built (novox/hq issue
// 213): the same program in the same context, hosted as a unit rather than a container.
kind := fmt.Sprint(r["type"])
if (kind != "container" && kind != "process") || !ownArtifact(r, m.Module) {
continue
}
if once, _ := r["run-once"].(bool); once {
continue
}
if r["schedule"] != nil {
continue
}
return fmt.Sprint(r["id"])
}
return ""
@@ -2029,7 +2049,10 @@ func ownArtifact(resource map[string]any, module string) bool {
return true
}
image, _ := resource["image"].(string)
return strings.HasPrefix(image, ArtifactStoreScheme+module+"/")
// A process or an archive carries what was built as its source (novox/hq issue 213).
source, _ := resource["source"].(string)
return strings.HasPrefix(image, ArtifactStoreScheme+module+"/") ||
strings.HasPrefix(source, ArtifactStoreScheme+module+"/")
}
// prepared is the module's own resource as the step that prepares its state: the same image, the same
@@ -2051,7 +2074,19 @@ func prepared(from map[string]any) map[string]any {
step["id"] = fmt.Sprint(from["id"]) + "-prepare"
step["name"] = fmt.Sprint(from["name"]) + "-prepare"
step["run-once"] = true
step["args"] = []any{PreparationArgument}
if fmt.Sprint(from["type"]) == "process" {
// A process says its whole command: the program, then its arguments. The step is the same
// program asked to prepare (novox/hq issue 213). It replaces nothing — what the process
// replaces is handed over to the process, never to the step that runs before it — and a
// step is not restarted, it runs again when what it reads changed, which `restart-on` says.
run := stringsIn(from["run"])
if len(run) > 0 {
step["run"] = []any{run[0], PreparationArgument}
}
delete(step, "replaces")
} else {
step["args"] = []any{PreparationArgument}
}
delete(step, "ports")
delete(step, "ip")
delete(step, "schedule")
@@ -2196,3 +2231,16 @@ func withRestartOn(have any, add []string) []any {
}
return out
}
// accountsFirst splits a module's resources into its accounts and everything else, each in the order
// written.
func accountsFirst(resources []map[string]any) (accounts, rest []map[string]any) {
for _, r := range resources {
if fmt.Sprint(r["type"]) == "user" {
accounts = append(accounts, r)
continue
}
rest = append(rest, r)
}
return accounts, rest
}
+23 -17
View File
@@ -1,6 +1,7 @@
package catalogue
import (
"encoding/json"
"fmt"
"os"
"reflect"
@@ -170,7 +171,7 @@ func TestTheForgeHoldsTheNpmAndGitSeats(t *testing.T) {
// **And the forge's own address follows it**, composed from the manifest in the catalogue beside
// this checkout (novox/hq 04-ISSUES/088).
//
// The forge is reached a third way that neither test above covers: by its own sidecar, over the
// The forge is reached a third way that neither test above covers: by its own code, over the
// machine's loopback, told where to go in its environment. The `2999:3000` mapping that lets the
// forge go on binding 3000 does nothing for a caller dialling the machine — so a literal there is
// wrong on every node whose assignment differs, and wrong for a second reason on a node given the
@@ -178,13 +179,13 @@ func TestTheForgeHoldsTheNpmAndGitSeats(t *testing.T) {
// in an `env` at all is a declaration, not a manifest.
func TestTheForgesOwnAddressFollowsThePortTheNodeGaveIt(t *testing.T) {
forge, err := catalogueManifest(t, "gitea").Resolve([]Built{{
Name: "runtime", Kind: ArtifactImage,
Reference: "registry.example/gitea-runtime@sha256:" + strings.Repeat("a", 64),
Name: "code", Kind: ArtifactBundle,
Reference: ArtifactStoreScheme + "gitea/code/blobs/" + bundleDigest, Digest: bundleDigest,
}})
if err != nil {
t.Fatalf("the forge's manifest does not resolve against its own build: %v", err)
}
r := Resolution{Node: "anchor", Modules: []Manifest{forge}, Needs: []Needed{
r := Resolution{Node: "anchor", Modules: []Manifest{forge, theRuntime(t)}, Needs: []Needed{
{Name: "postgres-database", For: "gitea", From: "anchor", At: "127.0.0.1",
Serves: map[string]any{"port": float64(5432)}, Sealed: "sealed-db"},
{Name: "route", For: "gitea", From: "anchor"},
@@ -194,8 +195,8 @@ func TestTheForgesOwnAddressFollowsThePortTheNodeGaveIt(t *testing.T) {
// The number this node was given for the forge — the one the machine it is about to run on
// already publishes.
out, err := r.Declaration(Rendering{
Needed: map[string]map[string]string{"gitea": {"broker": "sealed-broker"}},
out, err := r.Declaration(Rendering{ArtifactStore: "anchor.internal:5101",
Needed: map[string]map[string]string{RuntimeModule: {"broker": "sealed-broker"}},
Given: map[string]map[int]int{"gitea": {3000: 2999}},
})
if err != nil {
@@ -210,14 +211,19 @@ func TestTheForgesOwnAddressFollowsThePortTheNodeGaveIt(t *testing.T) {
if published := fmt.Sprint(server["ports"]); !strings.Contains(published, "2999:3000") {
t.Fatalf("the forge is not published on the port this node gave it: %v", server["ports"])
}
runtime := fileNamed(out, "gitea.runtime")
// The forge's own code runs in the node's runtime (novox/hq ADR 0198), given its words there.
runtime := fileNamed(out, RuntimeModule+"."+RuntimeProcessID())
if runtime == nil {
t.Fatalf("the forge's sidecar is not in the declaration: %v", out)
t.Fatalf("the node's runtime is not in the declaration: %v", ids(out))
}
env, _ := runtime["env"].(map[string]any)
if env["MESH_GITEA_URL"] != "http://127.0.0.1:2999" {
t.Fatalf("the forge's sidecar dials %v while the machine publishes the forge on 2999 — "+
"whatever reads it dials a dead port", env["MESH_GITEA_URL"])
env, _ := runtime["env"].(map[string]string)
var given map[string]map[string]string
if err := json.Unmarshal([]byte(env[RuntimeToolEnv]), &given); err != nil {
t.Fatalf("the runtime's %s is not JSON: %q", RuntimeToolEnv, env[RuntimeToolEnv])
}
if given["gitea"]["MESH_GITEA_URL"] != "http://127.0.0.1:2999" {
t.Fatalf("the forge's code dials %v while the machine publishes the forge on 2999 — "+
"whatever reads it dials a dead port", given["gitea"]["MESH_GITEA_URL"])
}
}
@@ -229,13 +235,13 @@ func declaredGiteaSsh(t *testing.T, given map[int]int) map[string]any {
t.Helper()
forge := catalogueManifest(t, "gitea")
resolved, err := forge.Resolve([]Built{{
Name: "runtime", Kind: ArtifactImage,
Reference: "registry.example/gitea-runtime@sha256:" + strings.Repeat("a", 64),
Name: "code", Kind: ArtifactBundle,
Reference: ArtifactStoreScheme + "gitea/code/blobs/" + bundleDigest, Digest: bundleDigest,
}})
if err != nil {
t.Fatalf("the forge's manifest does not resolve against its own build: %v", err)
}
r := Resolution{Node: "anchor", Modules: []Manifest{resolved}, Needs: []Needed{
r := Resolution{Node: "anchor", Modules: []Manifest{resolved, theRuntime(t)}, Needs: []Needed{
{Name: "postgres-database", For: "gitea", From: "anchor", At: "127.0.0.1",
Serves: map[string]any{"port": float64(5432)}, Sealed: "sealed-db"},
{Name: "route", For: "gitea", From: "anchor"},
@@ -246,8 +252,8 @@ func declaredGiteaSsh(t *testing.T, given map[int]int) map[string]any {
for k, v := range given {
givenPorts[k] = v
}
out, err := r.Declaration(Rendering{
Needed: map[string]map[string]string{"gitea": {"broker": "sealed-broker"}},
out, err := r.Declaration(Rendering{ArtifactStore: "anchor.internal:5101",
Needed: map[string]map[string]string{RuntimeModule: {"broker": "sealed-broker"}},
Ports: map[string]map[int]int{"gitea": givenPorts},
Given: map[string]map[int]int{"gitea": given},
})
+53 -1
View File
@@ -757,6 +757,11 @@ type Artifact struct {
// list twice. A module declaring no tools has nothing the runtime loads, whatever it compiles.
Loads []string `json:"loads,omitempty"`
// External are packages a TypeScript bundle keeps as imports rather than inlining — a native
// addon, a package that reads its own files — and so carries the toolchain's node_modules for
// (novox/hq ADR 0193). Absent for nearly every bundle, which is then one file per entrypoint.
External []string `json:"external,omitempty"`
// Env is what a tools bundle is given on a machine (novox/hq ADR 0192): words and their values,
// paths and constants composed with ${dir:…} and ${port:…} exactly as a container's environment
// is, never a secret's content. The node's runtime hands it to this bundle and to no other.
@@ -1510,6 +1515,53 @@ func ParseManifest(raw []byte) (Manifest, error) {
"program that reads what the mesh delivered and reconciles",
m.Module, r["id"]))
}
// **What a process replaces is something the module no longer declares** (novox/hq issue 213).
// The host keeps it running until the process is, then removes it: so it is named by the id the
// module used to give it, it is never a resource the module still declares — that would be
// applied and removed by one declaration — and only a process that stays up has anything to
// hand over to. Said here, near the author, as the host would refuse it far away.
ids := map[string]bool{}
for _, r := range m.Resources {
ids[fmt.Sprint(r["id"])] = true
}
for _, r := range m.Resources {
raw, present := r["replaces"]
if !present {
continue
}
if fmt.Sprint(r["type"]) != "process" {
problems = append(problems, fmt.Sprintf(
"%s: %v says what it replaces, and only a process does", m.Module, r["id"]))
continue
}
if once, _ := r["run-once"].(bool); once || r["schedule"] != nil {
problems = append(problems, fmt.Sprintf(
"%s: %v replaces something and runs once or on a schedule — only a process that stays "+
"up is there a moment later to hand over to", m.Module, r["id"]))
}
list, ok := raw.([]any)
if !ok {
problems = append(problems, fmt.Sprintf(
"%s: %v replaces %v; replaces is a list of the ids this module no longer declares",
m.Module, r["id"], raw))
continue
}
for _, item := range list {
id, ok := item.(string)
switch {
case !ok || strings.TrimSpace(id) == "":
problems = append(problems, fmt.Sprintf(
"%s: %v replaces %v, which is not an id", m.Module, r["id"], item))
case strings.Contains(id, "."):
problems = append(problems, fmt.Sprintf(
"%s: %v replaces %q; a process replaces only a resource of its own module, named "+
"by its own id", m.Module, r["id"], id))
case ids[id]:
problems = append(problems, fmt.Sprintf(
"%s: %v replaces %q, which this module still declares", m.Module, r["id"], id))
}
}
}
// **A module that prepares its state must have code the mesh can run** (novox/hq ADR 0135). The
// preparation is the module's own program in its preparation mode, so it is derived from the
// resource that runs that program — and a module declaring none has asked for something the mesh
@@ -1517,7 +1569,7 @@ func ParseManifest(raw []byte) (Manifest, error) {
// quietly prepares nothing.
if m.Prepares && preparationTarget(m) == "" {
problems = append(problems, fmt.Sprintf(
"%s says it prepares its state, and declares no container running an artifact it built — "+
"%s says it prepares its state, and declares no container or process running an artifact it built — "+
"the preparation is this module's own program, so there has to be one for the mesh to "+
"run it in", m.Module))
}
+10 -5
View File
@@ -31,7 +31,7 @@ import (
//
// So a module asks. `${port:8080}` is "the machine-side port you gave me for the 8080 I said I
// listen on", and the module writes that where it would otherwise have written a literal — in a
// file's content, or in a value of a container's `env`.
// file's content, or in a value of a container's or a process's `env`.
//
// **The environment is filled by the control plane, exactly as a bound value is.** A port is not
// secret — the mesh holds it in the clear — so there is nothing for the host to be the only
@@ -64,7 +64,12 @@ func portsUsed(content string) []int {
}
// portInto replaces a resource's ${port:…} placeholders with what this machine assigned — in a
// file's content, and in a value of a container's environment.
// file's content, and in a value of a container's or a process's environment.
//
// **A process's environment is a container's** (novox/hq to-be 38 WP4c). A module's code moving out
// of its container becomes a process on the machine and still has to be told what the container
// was told; filled for one kind and not the other, the literal reached the process and was read as
// a port, and the modules that moved first wrote their run-once steps a 0600 env file instead.
//
// A port the module did not say it listens on is refused, for the same reason a binding's unknown
// key is: the module is asking about something it never declared, and the answer would be a guess.
@@ -84,7 +89,7 @@ func portInto(resource map[string]any, module string, listens []Listening, with
}
resource["content"] = filled
case "container":
case "container", "process":
env, ok := resource["env"].(map[string]any)
if !ok {
return nil
@@ -106,8 +111,8 @@ func portInto(resource map[string]any, module string, listens []Listening, with
continue
}
value, err := portsFilledInto(written,
fmt.Sprintf("%s's container %s sets %s to something that",
module, resource["name"], key), module, listens, with)
fmt.Sprintf("%s's %s %s sets %s to something that",
module, resource["type"], resource["name"], key), module, listens, with)
if err != nil {
return err
}
+80
View File
@@ -0,0 +1,80 @@
package catalogue
import (
"strings"
"testing"
)
// **A process's environment is composed as a container's is** (novox/hq to-be 38 WP4c).
//
// A module's code moving out of its container becomes a process on the machine, and what its
// container's environment asked for — the port this machine gave the module, the place it put the
// module's directory — it still has to be told. Filled for a container and not for a process, the
// literal `${port:8080}` reached the process as its environment and was read as a port; the modules
// that moved first wrote their run-once steps an env file instead.
func processModule(env map[string]any) Manifest {
return Manifest{
Module: "showcase",
Listens: []Listening{{Port: 8080, From: FromMesh}},
Resources: []map[string]any{
{"id": "data", "type": "directory", "mode": "0700"},
{"id": "setup", "type": "process", "name": "showcase-setup", "run-once": true,
"run": []any{"/usr/bin/showcase", "setup"}, "env": env},
},
}
}
func TestAProcessIsToldItsPortAndItsPlaceInItsEnvironment(t *testing.T) {
env := map[string]any{
"SHOWCASE_URL": "http://127.0.0.1:${port:8080}",
"SHOWCASE_DATA": "${dir:data}/objects",
"SHOWCASE_DB": "127.0.0.1:${seat:mesh-store:5432}",
"GREETING": "hello",
}
out, err := Resolution{Node: "anchor", Modules: []Manifest{processModule(env)}}.Declaration(Rendering{
Ports: map[string]map[int]int{"showcase": {8080: 21000}},
Seats: map[string]map[int]int{"mesh-store": {5432: 6852}},
})
if err != nil {
t.Fatalf("a process asking for its port and its place does not compose: %v", err)
}
setup := fileNamed(out, "showcase.setup")
if setup == nil {
t.Fatalf("the process is not in the declaration: %v", out)
}
got, _ := setup["env"].(map[string]any)
for key, want := range map[string]string{
"SHOWCASE_URL": "http://127.0.0.1:21000",
"SHOWCASE_DATA": "/var/lib/showcase/data/objects",
"SHOWCASE_DB": "127.0.0.1:6852",
"GREETING": "hello",
} {
if got[key] != want {
t.Errorf("the process is told %s=%v, want %q", key, got[key], want)
}
}
if env["SHOWCASE_URL"] != "http://127.0.0.1:${port:8080}" {
t.Fatalf("composing for one machine edited the module's own manifest: %v", env)
}
}
// An unknown reference in a process's environment is refused as a container's is, naming the
// process and the variable — left alone, it would reach the machine as a literal.
func TestAProcessAskingAboutAnUndeclaredPortIsRefused(t *testing.T) {
env := map[string]any{"SHOWCASE_URL": "http://127.0.0.1:${port:9999}"}
_, err := Resolution{Node: "anchor", Modules: []Manifest{processModule(env)}}.Declaration(Rendering{})
if err == nil {
t.Fatal("a process was told a port its module never said it listens on")
}
for _, said := range []string{"showcase-setup", "SHOWCASE_URL", "${port:9999}", "8080"} {
if !strings.Contains(err.Error(), said) {
t.Errorf("the refusal does not say %q: %v", said, err)
}
}
env = map[string]any{"SHOWCASE_DATA": "${dir:date}/objects"}
if _, err := (Resolution{Node: "anchor", Modules: []Manifest{processModule(env)}}).Declaration(Rendering{}); err == nil ||
!strings.Contains(err.Error(), "${dir:date}") {
t.Fatalf("a process naming no directory of its module was not refused: %v", err)
}
}
+5 -5
View File
@@ -43,8 +43,8 @@ import (
var ofSeat = regexp.MustCompile(`\$\{seat:([a-z0-9][a-z0-9-]*):([0-9]+)\}`)
// seatInto replaces a resource's ${seat:…} placeholders with where this machine put each seat's
// holder — in a file's content, and in a value of a container's environment. The same two places
// portInto fills, for the same reason: they are where a process reads a number from.
// holder — in a file's content, and in a value of a container's or a process's environment. The
// same places portInto fills, for the same reason: they are where a program reads a number from.
func seatInto(resource map[string]any, module string, with Rendering) error {
switch fmt.Sprint(resource["type"]) {
case "file":
@@ -58,7 +58,7 @@ func seatInto(resource map[string]any, module string, with Rendering) error {
}
resource["content"] = filled
case "container":
case "container", "process":
env, ok := resource["env"].(map[string]any)
if !ok {
return nil
@@ -78,8 +78,8 @@ func seatInto(resource map[string]any, module string, with Rendering) error {
continue
}
value, err := seatsFilledInto(written,
fmt.Sprintf("%s's container %s sets %s to something that",
module, resource["name"], key), with)
fmt.Sprintf("%s's %s %s sets %s to something that",
module, resource["type"], resource["name"], key), with)
if err != nil {
return err
}
+43 -14
View File
@@ -42,9 +42,29 @@ type Build struct {
// Failed is the builder's own words, empty when it worked.
Failed string
Made []Artifact
At time.Time
// Asked is when the build was requested, zero when that is not known (an id of another shape,
// or a build recorded before the mesh kept it). **What orders one build of a module against
// another** (novox/hq 04-ISSUES/219): builds in flight together finish in any order, and the
// one asked last stood on the newest bases.
Asked time.Time
// At is when the outcome was recorded — when it finished, not when it was asked.
At time.Time
}
// AskedOrAt is when the build was asked, or when it was recorded when that is not known — the
// order the mesh had before it kept the request time.
func (b Build) AskedOrAt() time.Time {
if !b.Asked.IsZero() {
return b.Asked
}
return b.At
}
// newestRequestFirst is the ordering every "what a module currently is" question uses: the newest
// request wins, whenever it finished (novox/hq 04-ISSUES/219). A build whose request time is not
// known is placed at the moment it was recorded, which is the rule that held before.
const newestRequestFirst = `coalesce(asked, at) desc, at desc`
// ReadRepository is a repository a build read source from besides the module's own.
type ReadRepository struct {
Repository string `json:"repository"`
@@ -83,13 +103,17 @@ func (i *Inventory) RecordBuild(ctx context.Context, b Build) error {
if b.Module != "" {
module = &b.Module
}
var asked *time.Time
if !b.Asked.IsZero() {
asked = &b.Asked
}
_, err = i.store.Pool().Exec(ctx,
`insert into build (id, repository, ref, module, commit_hash, built_on, failed, made,
source_path, manifest, built_against, built_contexts)
values ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12)
source_path, manifest, built_against, built_contexts, asked)
values ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13)
on conflict (id) do nothing`,
b.ID, b.Repository, b.Ref, module, b.Commit, b.On, b.Failed, made,
b.Path, manifestOrNil(b.Manifest), against, read)
b.Path, manifestOrNil(b.Manifest), against, read, asked)
return err
}
@@ -102,11 +126,11 @@ func (i *Inventory) Builds(ctx context.Context, module string, limit int) ([]Bui
if limit <= 0 {
limit = 20
}
query := `select id, repository, ref, coalesce(module,''), commit_hash, built_on, failed, made, at
query := `select id, repository, ref, coalesce(module,''), commit_hash, built_on, failed, made, asked, at
from build order by at desc limit $1`
args := []any{limit}
if module != "" {
query = `select id, repository, ref, coalesce(module,''), commit_hash, built_on, failed, made, at
query = `select id, repository, ref, coalesce(module,''), commit_hash, built_on, failed, made, asked, at
from build where module = $2 order by at desc limit $1`
args = append(args, module)
}
@@ -121,10 +145,14 @@ func (i *Inventory) Builds(ctx context.Context, module string, limit int) ([]Bui
for rows.Next() {
var b Build
var made []byte
var asked *time.Time
if err := rows.Scan(&b.ID, &b.Repository, &b.Ref, &b.Module, &b.Commit,
&b.On, &b.Failed, &made, &b.At); err != nil {
&b.On, &b.Failed, &made, &asked, &b.At); err != nil {
return nil, err
}
if asked != nil {
b.Asked = *asked
}
if err := json.Unmarshal(made, &b.Made); err != nil {
return nil, err
}
@@ -135,8 +163,9 @@ func (i *Inventory) Builds(ctx context.Context, module string, limit int) ([]Bui
// Held is every artifact this mesh has built, keyed "<module>/<artifact>".
//
// **The newest successful build of each module wins**, which is the same rule the rest of the mesh
// uses for what a module currently is. A module rebuilt to something broken and then rebuilt again
// **The successful build of each module asked last wins**, which is the same rule the rest of the
// mesh uses for what a module currently is — asked last, not finished last (novox/hq
// 04-ISSUES/219): an older request that finishes later stood on older bases. A module rebuilt to something broken and then rebuilt again
// is at the second one; a module whose last build failed is at the last one that worked, because a
// failure published nothing and the thing it published before is still what exists.
//
@@ -147,7 +176,7 @@ func (i *Inventory) Held(ctx context.Context) (map[string]string, error) {
`select distinct on (module) module, made
from build
where module is not null and module <> '' and failed = ''
order by module, at desc`)
order by module, `+newestRequestFirst)
if err != nil {
return nil, err
}
@@ -185,7 +214,7 @@ func (i *Inventory) BuiltAgainst(ctx context.Context) (map[string][]string, erro
`select distinct on (module) module, built_against
from build
where module is not null and module <> '' and failed = ''
order by module, at desc`)
order by module, `+newestRequestFirst)
if err != nil {
return nil, err
}
@@ -223,7 +252,7 @@ func (i *Inventory) ReadRepositories(ctx context.Context) (map[string][]ReadRepo
`select distinct on (module) module, built_contexts
from build
where module is not null and module <> '' and failed = ''
order by module, at desc`)
order by module, `+newestRequestFirst)
if err != nil {
return nil, err
}
@@ -274,7 +303,7 @@ func manifestOrNil(raw []byte) any {
// follows when it decides whether to announce at all.
//
// One row per module and commit: a module built twice at the same commit is one fact, and the
// latest row is the one whose artifacts are current.
// row asked last is the one whose artifacts are current (novox/hq 04-ISSUES/219).
func (i *Inventory) Announceable(ctx context.Context) ([]Build, error) {
rows, err := i.store.Pool().Query(ctx,
`select distinct on (module, commit_hash)
@@ -282,7 +311,7 @@ func (i *Inventory) Announceable(ctx context.Context) ([]Build, error) {
source_path, manifest, built_against, at
from build
where failed = '' and module is not null and module <> '' and commit_hash <> ''
order by module, commit_hash, at desc`)
order by module, commit_hash, `+newestRequestFirst)
if err != nil {
return nil, err
}
+39 -1
View File
@@ -49,6 +49,12 @@ func (i *Inventory) BusRecords(ctx context.Context) (broker.Records, error) {
}
}
// Who holds each seat held once for the mesh, where the mesh recorded it (novox/hq issue 218).
holdings, err := i.Holdings(ctx)
if err != nil {
return broker.Records{}, fmt.Errorf("cannot read who holds the mesh's seats: %w", err)
}
out := broker.Records{Assigned: map[string][]broker.Declared{}, People: map[string][]string{},
Interchangeable: map[string]bool{}}
for _, n := range nodes {
@@ -72,7 +78,9 @@ func (i *Inventory) BusRecords(ctx context.Context) (broker.Records, error) {
"%s is assigned to %s and is not in the catalogue, so what it may say cannot "+
"be derived", module, n.Name)
}
out.Assigned[n.Name] = append(out.Assigned[n.Name], declaredFor(m, seats))
d := declaredFor(m, seats)
d.Holds = heldHere(d.Holds, holdings, n.Name, module)
out.Assigned[n.Name] = append(out.Assigned[n.Name], d)
if m.Instances == catalogue.InstancesInterchangeable {
out.Interchangeable[m.Module] = true
}
@@ -183,3 +191,33 @@ func (i *Inventory) NodesWithALiveToken(ctx context.Context) ([]string, error) {
}
return out, rows.Err()
}
// heldHere keeps of what a module claims only the seats it holds on this machine (novox/hq issue 218).
// A seat held once per machine is held by every assignment that claims it. A seat held once for the
// mesh is held by one assignment: where the mesh recorded who holds it, a claim on any other machine
// grants nothing and issues nothing — or the module would serve the role's verbs from a machine that
// is not the role's, and a question to the mesh's store would be answered from the wrong database. A
// mesh seat with no holder on record is left as it was derived.
func heldHere(claimed []broker.Seat, holdings []catalogue.Held, node, module string) []broker.Seat {
recorded := map[string][]catalogue.Held{}
for _, h := range holdings {
if h.Scope == catalogue.ScopeMesh {
recorded[h.Claim] = append(recorded[h.Claim], h)
}
}
var out []broker.Seat
for _, s := range claimed {
holders, onRecord := recorded[s.Name]
if s.Scope != catalogue.ScopeMesh || !onRecord {
out = append(out, s)
continue
}
for _, h := range holders {
if h.Node == node && h.Module == module {
out = append(out, s)
break
}
}
}
return out
}
+38 -6
View File
@@ -17,6 +17,10 @@ import (
// ErrNoSuchModule is what the mesh says about a module it has never been told about.
var ErrNoSuchModule = errors.New("no module of that name")
// ErrSuperseded is a registration from a build asked before the one the module is already at
// (novox/hq 04-ISSUES/219). The build is recorded; what the module is does not change.
var ErrSuperseded = errors.New("a build asked later is already what the module is")
// ErrStillAssigned is why a module cannot be forgotten.
//
// Its own error because it is not a fault: it means a machine is running that module now, and
@@ -47,6 +51,10 @@ type Source struct {
// itself no longer carries its build (novox/hq to-be 38 WP2.4). Empty for a manifest handed over
// by hand, which carries its `build.on` itself.
Against []string
// Asked is when the build this manifest came from was requested (novox/hq 04-ISSUES/219). Zero
// is a manifest handed over by hand, or a build whose request time is not known: either is
// taken as asked at the moment it is registered.
Asked time.Time
}
// Current reports whether what the mesh holds is what the source last had.
@@ -97,13 +105,23 @@ func (i *Inventory) RegisterModule(ctx context.Context, m catalogue.Manifest, fr
return err
}
asked := from.Asked
if asked.IsZero() {
asked = time.Now()
}
// A module registered without provenance keeps whatever it had. Handing over a manifest by
// hand is a legitimate way to fix something in a hurry, and it should not silently erase the
// record of where the module normally comes from — which is the only thing that would say,
// afterwards, that the machine is running something nobody can rebuild.
_, err = i.store.Pool().Exec(ctx,
`insert into module (name, manifest, version, source, source_path, source_seat, ref, built_from, source_head)
values ($1, $2, nullif($3,''), nullif($4,''), $7, $8, nullif($5,''), nullif($6,''), nullif($6,''))
//
// **An older request never replaces a newer one** (novox/hq 04-ISSUES/219). Builds of one
// module in flight together finish in any order, and each stood on the bases the mesh held when
// it was asked; the one asked later is what the module is, whichever is heard last. An outcome
// of an earlier request is kept in the build records and changes nothing here.
tag, err := i.store.Pool().Exec(ctx,
`insert into module (name, manifest, version, source, source_path, source_seat, ref, built_from, source_head, built_asked)
values ($1, $2, nullif($3,''), nullif($4,''), $7, $8, nullif($5,''), nullif($6,''), nullif($6,''), $9)
on conflict (name) do update set
manifest = excluded.manifest,
version = excluded.version,
@@ -115,9 +133,23 @@ func (i *Inventory) RegisterModule(ctx context.Context, m catalogue.Manifest, fr
else excluded.source_seat end,
ref = coalesce(excluded.ref, module.ref),
built_from = coalesce(excluded.built_from, module.built_from),
source_head = coalesce(excluded.built_from, module.source_head)`,
m.Module, raw, m.Version, from.Repository, from.Ref, from.BuiltFrom, from.Path, from.Seat)
return err
source_head = coalesce(excluded.built_from, module.source_head),
built_asked = excluded.built_asked
where module.built_asked is null or module.built_asked <= excluded.built_asked`,
m.Module, raw, m.Version, from.Repository, from.Ref, from.BuiltFrom, from.Path, from.Seat, asked)
if err != nil {
return err
}
if tag.RowsAffected() == 0 {
var current time.Time
if err := i.store.Pool().QueryRow(ctx,
`select built_asked from module where name = $1`, m.Module).Scan(&current); err != nil {
return err
}
return fmt.Errorf("%w: %s is at a build asked %s, and this one was asked %s",
ErrSuperseded, m.Module, current.UTC().Format(time.RFC3339), asked.UTC().Format(time.RFC3339))
}
return nil
}
// registeredInThatShape is whether the catalogue already holds this module as a tools container on
+20
View File
@@ -100,3 +100,23 @@ func TestABundleStandsOnTheToolchainItIsCompiledIn(t *testing.T) {
}
}
}
// novox/hq 04-ISSUES/212: a toolchain standing on the SDK's package is planned after the SDK, so a
// release of the SDK rebuilds the toolchain, and every bundle compiled in it after that.
func TestAToolchainStandingOnTheSDKFollowsIt(t *testing.T) {
entries := []Entry{
{Manifest: catalogue.Manifest{Module: "mesh-sdk"}},
{Manifest: catalogue.Manifest{Module: "mesh-tools", Build: &catalogue.Build{
On: []catalogue.BuildsOn{{Arg: "MESH_SDK", Module: "mesh-sdk", Artifact: "lib"}}}}},
}
edges := dependenciesOf(entries, nil, nil)
found := false
for _, e := range edges {
if e.From == "mesh-tools" && e.To == "mesh-sdk" {
found = true
}
}
if !found {
t.Errorf("no edge from the toolchain to the SDK: %v", edges)
}
}
+32
View File
@@ -0,0 +1,32 @@
package inventory
import (
"testing"
"github.com/novox/mesh-controller/internal/broker"
"github.com/novox/mesh-controller/internal/catalogue"
)
// novox/hq issue 218: a seat held once for the mesh is granted and issued only to the holder on record;
// a node seat to every machine's claimant; a mesh seat with no holder on record as derived.
func TestOnlyTheRecordedHolderHoldsAMeshSeat(t *testing.T) {
claimed := []broker.Seat{
{Name: "mesh-store", Scope: catalogue.ScopeMesh},
{Name: "node-packet-filter", Scope: catalogue.ScopeNode},
{Name: "unrecorded", Scope: catalogue.ScopeMesh},
}
holdings := []catalogue.Held{{Claim: "mesh-store", Scope: catalogue.ScopeMesh, Node: "control", Module: "postgres"}}
names := func(ss []broker.Seat) (out []string) {
for _, s := range ss {
out = append(out, s.Name)
}
return
}
if got := names(heldHere(claimed, holdings, "control", "postgres")); len(got) != 3 {
t.Errorf("the holder lost a seat: %v", got)
}
got := names(heldHere(claimed, holdings, "other", "postgres"))
if len(got) != 2 || got[0] != "node-packet-filter" || got[1] != "unrecorded" {
t.Errorf("a claimant on another machine holds %v; want the node seat and the unrecorded one, not the store", got)
}
}
+59
View File
@@ -87,3 +87,62 @@ func (i *Inventory) tryHold(ctx context.Context, sorted []string) (func(), strin
}
return release, "", nil
}
// ErrPlansBusy is the plans held by another act — on a machine replacing its controller, the other
// controller — for longer than a caller waits, or at all for one that does not wait.
var ErrPlansBusy = errors.New("another controller is working the plans")
// HoldPlans makes working the plans one act at a time, across every controller on the store
// (novox/hq issue 213). A plan is read, changed and written whole; two controllers doing that at
// once — the old and the new for the moment a machine hands its controller over, or a controller
// and a person's `plans stop` — each act on what the other has not saved yet: a tier asked twice,
// an outcome written over. A session-level advisory lock on one connection, released by the
// returned function and by the session ending, so a controller that dies holding it holds nothing.
//
// wait false gives ErrPlansBusy at once when another holds them — the timer's way: the holder is
// moving the plans already. wait true looks again every HoldPoll for up to HoldWaitFor — an
// outcome's or a merge's way, which must be written.
func (i *Inventory) HoldPlans(ctx context.Context, wait bool) (func(), error) {
deadline := time.Now().Add(HoldWaitFor)
for {
release, took, err := i.tryLock(ctx, "mesh-plans")
if err != nil || took {
return release, err
}
if !wait || time.Now().After(deadline) {
return nil, ErrPlansBusy
}
select {
case <-ctx.Done():
return nil, ctx.Err()
case <-time.After(HoldPoll):
}
}
}
// tryLock takes one named advisory lock on a connection of its own, or gives the connection back.
func (i *Inventory) tryLock(ctx context.Context, key string) (func(), bool, error) {
conn, err := i.store.Pool().Acquire(ctx)
if err != nil {
return nil, false, err
}
var once sync.Once
release := func() {
once.Do(func() {
if _, err := conn.Exec(context.WithoutCancel(ctx), `select pg_advisory_unlock_all()`); err != nil {
_ = conn.Conn().Close(context.WithoutCancel(ctx))
}
conn.Release()
})
}
var took bool
if err := conn.QueryRow(ctx, `select pg_try_advisory_lock(hashtext($1)::bigint)`, key).Scan(&took); err != nil {
release()
return nil, false, err
}
if !took {
release()
return nil, false, nil
}
return release, true, nil
}
+38
View File
@@ -0,0 +1,38 @@
package inventory
import (
"errors"
"testing"
"time"
)
// novox/hq issue 213: while a machine hands its controller over from the container to the process,
// two controllers run on one store for a moment. Working the plans is one act at a time across them.
func TestThePlansAreWorkedByOneControllerAtATime(t *testing.T) {
first := ForTest(t)
// A second controller: its own connections to the same store.
second, err := Open(t.Context())
if err != nil {
t.Fatal(err)
}
t.Cleanup(second.Close)
release, err := first.HoldPlans(t.Context(), false)
if err != nil {
t.Fatalf("the plans could not be held when nobody held them: %v", err)
}
t.Cleanup(release) // a pool closing waits for a connection still held; release is idempotent
if _, err := second.HoldPlans(t.Context(), false); !errors.Is(err, ErrPlansBusy) {
t.Fatalf("a second controller held the plans while the first did: %v", err)
}
// A waiter gets them once they are let go.
was := HoldPoll
HoldPoll = 10 * time.Millisecond
defer func() { HoldPoll = was }()
go func() { time.Sleep(50 * time.Millisecond); release() }()
again, err := second.HoldPlans(t.Context(), true)
if err != nil {
t.Fatalf("a waiting controller never got the plans once they were let go: %v", err)
}
again()
}
@@ -0,0 +1,23 @@
-- A build is ordered by when it was asked, not when it finished (novox/hq 04-ISSUES/219).
--
-- Two builds of one module can be in flight together — two merge plans a few minutes apart, each
-- asking for everything standing on what it changed — and they finish in any order. Each build
-- stands on the bases the mesh held when it was *asked*, so the one asked later is the newer one.
-- The mesh ordered builds by `at`, which is when the outcome was recorded, and registered whatever
-- it heard last: an older request that took longer replaced a newer one as what the module is, and
-- the next push sent machines an image built on a base the mesh had already replaced.
--
-- `build.asked` is when the build was requested, read from the correlation id the controller wrote
-- (`build-<unix nanoseconds>`). Nullable: an id of any other shape says no request time, and such a
-- build is placed where it was recorded, which is the order the mesh had before this.
alter table build add column asked timestamptz;
update build
set asked = to_timestamp((substring(id from '^build-([0-9]{19})$'))::numeric / 1000000000)
where id ~ '^build-[0-9]{19}$';
-- `module.built_asked` is when the build the module's registered manifest came from was asked, so
-- a later-heard outcome of an earlier request is recorded and not registered. A manifest handed over
-- by hand is a request made when it is handed over. Null for a module registered before this was
-- kept: its next registration, whichever it is, sets it.
alter table module add column built_asked timestamptz;
+124
View File
@@ -0,0 +1,124 @@
package inventory
import (
"context"
"errors"
"testing"
"time"
"github.com/novox/mesh-controller/internal/catalogue"
)
// novox/hq 04-ISSUES/219: two builds of one module in flight together, the one asked first heard
// last. The newer request stood on the newer base; the older one's late outcome is recorded and is
// not what the module is.
func postgresBuild(id string, asked time.Time, image string) Build {
b := aBuild(id, "postgres", "")
b.Asked = asked
b.Against = []string{"mesh-tools/runtime@sha256:" + id}
b.Made = []Artifact{{Name: "server", Kind: "image", Reference: "postgres@sha256:" + image}}
return b
}
func TestAnOlderRequestFinishingLaterIsNotWhatTheModuleHolds(t *testing.T) {
inv := fresh(t)
ctx := context.Background()
older := time.Date(2026, 10, 3, 21, 33, 45, 0, time.UTC)
newer := time.Date(2026, 10, 3, 21, 51, 57, 0, time.UTC)
// The newer request finishes first, the older one last — recorded in that order.
if err := inv.RecordBuild(ctx, postgresBuild("newer", newer, "4bcd5f73")); err != nil {
t.Fatal(err)
}
if err := inv.RecordBuild(ctx, postgresBuild("older", older, "0ab07fa9")); err != nil {
t.Fatal(err)
}
held, err := inv.Held(ctx)
if err != nil {
t.Fatal(err)
}
if got := held["postgres/server"]; got != "postgres@sha256:4bcd5f73" {
t.Errorf("postgres holds %q; want the newer request's image 4bcd5f73", got)
}
against, err := inv.BuiltAgainst(ctx)
if err != nil {
t.Fatal(err)
}
if got := against["postgres"]; len(got) != 1 || got[0] != "mesh-tools/runtime@sha256:newer" {
t.Errorf("postgres stands on %v; want what the newer request stood on", got)
}
// Both are still recorded, the late one first as what happened lately.
builds, err := inv.Builds(ctx, "postgres", 5)
if err != nil {
t.Fatal(err)
}
if len(builds) != 2 || builds[0].ID != "older" || !builds[0].Asked.Equal(older) {
t.Fatalf("both builds, newest heard first, with when they were asked: %+v", builds)
}
}
func TestABuildWithNoKnownRequestTimeIsOrderedByWhenItWasRecorded(t *testing.T) {
// What the mesh did before it kept the request time, so a row from before still answers.
inv := fresh(t)
ctx := context.Background()
for _, id := range []string{"first", "second"} {
b := aBuild(id, "shell", "")
b.Made = []Artifact{{Name: "config", Kind: "archive", Reference: "…/" + id}}
if err := inv.RecordBuild(ctx, b); err != nil {
t.Fatal(err)
}
}
held, err := inv.Held(ctx)
if err != nil {
t.Fatal(err)
}
if got := held["shell/config"]; got != "…/second" {
t.Errorf("shell holds %q; want the one recorded last", got)
}
}
func TestARegistrationFromAnOlderRequestDoesNotReplaceANewerOne(t *testing.T) {
inv := fresh(t)
ctx := context.Background()
older := time.Date(2026, 10, 3, 21, 33, 45, 0, time.UTC)
newer := time.Date(2026, 10, 3, 21, 51, 57, 0, time.UTC)
from := func(asked time.Time) Source {
return Source{Repository: "novox/mesh-catalog", Seat: "git", Path: "modules/postgres",
BuiltFrom: "efff5415", Asked: asked}
}
if err := inv.RegisterModule(ctx, catalogue.Manifest{Module: "postgres", Version: "fixed"}, from(newer)); err != nil {
t.Fatal(err)
}
err := inv.RegisterModule(ctx, catalogue.Manifest{Module: "postgres", Version: "stale"}, from(older))
if !errors.Is(err, ErrSuperseded) {
t.Fatalf("an older request's registration was not refused as superseded: %v", err)
}
shelf, err := inv.Catalogue(ctx)
if err != nil {
t.Fatal(err)
}
if got := shelf["postgres"].Version; got != "fixed" {
t.Fatalf("postgres is %q; want the newer request's manifest", got)
}
// A later request, and a manifest handed over by hand — asked when it is handed over — both
// replace it as before.
if err := inv.RegisterModule(ctx, catalogue.Manifest{Module: "postgres", Version: "later"},
from(newer.Add(time.Minute))); err != nil {
t.Fatal(err)
}
if err := inv.RegisterModule(ctx, catalogue.Manifest{Module: "postgres", Version: "by-hand"}, Source{}); err != nil {
t.Fatal(err)
}
if shelf, _ := inv.Catalogue(ctx); shelf["postgres"].Version != "by-hand" {
t.Fatalf("postgres is %q; want the manifest handed over by hand", shelf["postgres"].Version)
}
src, err := inv.SourceOf(ctx, "postgres")
if err != nil || src.Repository != "novox/mesh-catalog" {
t.Fatalf("a hand registration erased the provenance: %+v %v", src, err)
}
}
+29
View File
@@ -2,6 +2,9 @@ package link
import (
"encoding/json"
"strconv"
"strings"
"time"
)
// Asking a machine to build a module, and hearing what came out.
@@ -17,6 +20,32 @@ import (
// holds no opinion about what they contain, and a host that also built things would be a host
// with a container runtime requirement and a git dependency (novox/hq ADR 0005).
// NewBuildID is the correlation for a build asked at that moment: `build-<unix nanoseconds>`.
//
// **The id carries when the build was asked, and that is read back** (novox/hq 04-ISSUES/219). Builds
// of one module can be in flight together and finish in any order; what a module currently is must
// be the newest *request's* outcome, not the last one heard, and the id is the one thing every
// outcome echoes whichever builder answered it. One place writes the shape and one reads it.
func NewBuildID(asked time.Time) string {
return "build-" + strconv.FormatInt(asked.UnixNano(), 10)
}
// BuildAskedAt is when the build with this id was asked, as NewBuildID wrote it. False for an id
// of any other shape — one written before this was read, or by hand — whose request time the mesh
// does not know.
func BuildAskedAt(id string) (time.Time, bool) {
digits, ok := strings.CutPrefix(id, "build-")
if !ok || digits == "" {
return time.Time{}, false
}
nanos, err := strconv.ParseInt(digits, 10, 64)
// A number too small to be a moment this mesh could have asked at is a name, not a time.
if err != nil || nanos < time.Date(2020, 1, 1, 0, 0, 0, 0, time.UTC).UnixNano() {
return time.Time{}, false
}
return time.Unix(0, nanos).UTC(), true
}
// BuildRequest is one module to build.
type BuildRequest struct {
// ID correlates the answer with the asking. Not the module name: two builds of one module can
+54 -2
View File
@@ -5,6 +5,7 @@ import (
"encoding/json"
"errors"
"fmt"
"log"
"strings"
"time"
@@ -78,10 +79,15 @@ func (n *natsInbound) Receive(ctx context.Context, act func(context.Context, Con
// than creating one here: the consumer is an object with a configuration — ack policy, ack
// wait, redelivery — and a client that creates its own would be a second opinion about it.
control := make(chan *nats.Msg, Prefetch)
said, err := js.ChanSubscribe("", control, nats.Bind("CONTROL", broker.ControllerName))
said, err := standingBy(ctx, log.Default(), "CONTROL", func() (*nats.Subscription, error) {
return js.ChanSubscribe("", control, nats.Bind("CONTROL", broker.ControllerName))
})
if err != nil {
return fmt.Errorf("subscribing to what nodes say: %w", err)
}
if said == nil {
return nil // stopped while standing by
}
defer func() { _ = said.Unsubscribe() }()
// Heartbeats, on core NATS and off any stream (design 25 §3). Their own subscription because
@@ -97,10 +103,15 @@ func (n *natsInbound) Receive(ctx context.Context, act func(context.Context, Con
var events chan *nats.Msg
if len(n.follows) > 0 {
events = make(chan *nats.Msg, Prefetch)
followed, err := js.ChanSubscribe("", events, nats.Bind("EVENTS", broker.ControllerName))
followed, err := standingBy(ctx, log.Default(), "EVENTS", func() (*nats.Subscription, error) {
return js.ChanSubscribe("", events, nats.Bind("EVENTS", broker.ControllerName))
})
if err != nil {
return fmt.Errorf("subscribing to what the catalogue says: %w", err)
}
if followed == nil {
return nil
}
defer func() { _ = followed.Unsubscribe() }()
}
@@ -324,3 +335,44 @@ func (m *natsControl) forget() {
type replyAddressed struct {
ReplyTo string `json:"reply_to,omitempty"`
}
// StandbyPoll is how often a controller standing by looks again for its consumers. A variable so a
// test need not wait.
var StandbyPoll = 2 * time.Second
// standingBy binds one of the controller's consumers, waiting while another controller holds it.
//
// **Two controllers, one consumer** (novox/hq issue 213). The controller's consumers are push
// consumers with no delivery group, so the server lets one subscription bind each — on purpose:
// two would each act on every message (issue 146). When a machine hands its controller over from
// the container to the process, the host starts the process first and removes the container only
// once the process is up; the process then finds the consumers bound. Exiting on that would never
// be up, so the container would never go. It stands by instead — the seat's verbs are already
// served from a queue group, and the plans wait on their lock — and binds as soon as the other lets
// go. Nil and no error is ctx ending while it waited.
func standingBy(ctx context.Context, logger interface{ Printf(string, ...any) }, stream string,
bind func() (*nats.Subscription, error)) (*nats.Subscription, error) {
said := false
for {
sub, err := bind()
if err == nil {
if said {
logger.Printf("took the controller's consumer on %s: the controller that held it let go", stream)
}
return sub, nil
}
if !strings.Contains(err.Error(), "already bound") {
return nil, err
}
if !said {
logger.Printf("another controller holds the controller's consumer on %s; standing by "+
"until it lets go", stream)
said = true
}
select {
case <-ctx.Done():
return nil, nil
case <-time.After(StandbyPoll):
}
}
}
+69
View File
@@ -0,0 +1,69 @@
package link
import (
"context"
"encoding/json"
"os"
"testing"
"time"
"github.com/novox/mesh-controller/internal/broker"
)
// novox/hq issue 213: while a machine hands its controller over, the new controller (the process)
// is started while the old one (the container) still holds the controller's consumers. It must not
// exit — the host would read that as a replacement that did not come up and never remove the
// container — and must not act on what the old one is handed. It stands by, and takes the consumers
// when the old one lets go.
func TestNatsASecondControllerStandsByAndTakesOverWhenTheFirstLetsGo(t *testing.T) {
js := aBus(t)
was := StandbyPoll
StandbyPoll = 50 * time.Millisecond
defer func() { StandbyPoll = was }()
old := &counted{}
_, stopOld := servingOn(t, js, old)
eventually(t, "the first controller binding its consumer", func() bool {
info, err := js.Context().ConsumerInfo("CONTROL", broker.ControllerName)
return err == nil && info.PushBound
})
// The new one, on a connection of its own as the process would have.
second, err := broker.Dial(os.Getenv("MESH_TEST_NATS"))
if err != nil {
t.Fatal(err)
}
t.Cleanup(second.Close)
fresh := &counted{}
s := &Server{inbound: Nats(second), bus: OverNATS{Conn: second.Conn(), JS: second.Context()},
listener: fresh, log: quiet()}
ctx, stopNew := context.WithCancel(context.Background())
defer stopNew()
ended := make(chan error, 1)
go func() { ended <- s.Serve(ctx) }()
select {
case err := <-ended:
t.Fatalf("the second controller stopped instead of standing by: %v", err)
case <-time.After(500 * time.Millisecond):
}
report := func(declared string) {
body, _ := json.Marshal(Report{Node: "anchor", Declared: declared, Applied: []string{"store"}})
if _, err := js.Context().Publish(ReportSubject("anchor"), body); err != nil {
t.Fatal(err)
}
}
report("d1")
eventually(t, "the holding controller hearing the report", func() bool { return old.count() == 1 })
if fresh.count() != 0 {
t.Fatal("the controller standing by acted on a report the holder was handed")
}
stopOld()
report("d2")
eventually(t, "the second controller taking over once the first let go", func() bool { return fresh.count() == 1 })
if old.count() != 1 {
t.Errorf("the first controller heard %d reports", old.count())
}
}