Files
mesh-controller/internal/broker/jetstream.go
T
mesh-tools 97a2d841c7 Land feature 57's controller work on its release branch
The controller plans config, templates and moved providers' provisions,
hands bus accounts to the vault and judges the new manifest words, on the
feature's SDK and node-engine validator (feature 57, config and provisioning).
2026-10-15 21:27:33 +00:00

515 lines
21 KiB
Go

package broker
import (
"context"
"crypto/sha256"
"crypto/tls"
"crypto/x509"
"encoding/hex"
"errors"
"fmt"
"os"
"strings"
"time"
"github.com/nats-io/nats.go"
"github.com/nats-io/nats.go/jetstream"
)
// The JetStream side of the controller: the one place the mesh's streams and consumers are
// actually created.
//
// Everything that decides *what* they are is pure and lives beside this (streams.go, derived.go).
// This is only the part that talks to a server, kept small on purpose: a bug in a subject filter
// should be findable in a unit test, and only a bug in "did the server accept it" should need one
// running.
// A JetStream is a connection to the bus, as the controller uses it.
type JetStream struct {
conn *nats.Conn
js nats.JetStreamContext
// borrowed is a connection lent by its owner (Borrow): closing it is the owner's.
borrowed bool
// Note is how this says something it decided not to fail over. Nil is silent, which is only
// right for a caller that has no way to report; the controller sets it.
Note func(string, ...any)
}
// note reports without requiring a caller to have set one.
func (j *JetStream) note(format string, args ...any) {
if j.Note != nil {
j.Note(format, args...)
}
}
// ConnectionName is what a connection this process dials says it is, in the server's list of
// connections: the controller's role and what it is doing (novox/hq issue 327). Every connection was
// named `mesh-controller` — the serving controller's, each verb's own process, the build agents' — so the
// server's list could not say which was which. The process sets it once, at its start; an option a
// caller passes to Dial names one connection otherwise.
var ConnectionName = "mesh-controller"
// Dial connects and returns the controller's JetStream handle.
func Dial(url string, opts ...nats.Option) (*JetStream, error) {
// The name first, so a name the caller gives is the one that stands.
opts = append([]nats.Option{nats.Name(ConnectionName), nats.Timeout(10 * time.Second)}, opts...)
// **Pinned, not named.** The bus presents the mesh's own certificate, which names nothing a
// public verifier would accept (design 25 §4: a host pins the server's exact certificate and
// checks nothing else, and so does this). Without this, the first connection failed with
// "certificate is not valid for any names" against a bus that was answering (2026-09-28).
if path := strings.TrimSpace(os.Getenv(CertificateVar)); path != "" {
pinned, err := pinnedTo(path)
if err != nil {
return nil, err
}
opts = append(opts, nats.Secure(pinned))
}
// **Its own inbox, and nothing wider.** Every principal is granted `_INBOX.<its user>.>` and
// no other inbox; the client's default prefix is random, and the server refused the first
// subscription to it (2026-09-28). The user is in the URL, so the prefix follows from it.
if user, _, _ := CredentialIn(url); user != "" {
opts = append(opts, nats.CustomInboxPrefix("_INBOX."+user))
}
// The address in an error is the address alone. The URL carries this controller's password,
// and an error here is written on the assumption it will be logged.
where := BareAddress(url)
conn, err := nats.Connect(url, opts...)
if err != nil {
return nil, fmt.Errorf("connecting to the bus at %s: %w", where, err)
}
js, err := conn.JetStream()
if err != nil {
conn.Close()
return nil, fmt.Errorf("the bus at %s has no JetStream: %w", where, err)
}
return &JetStream{conn: conn, js: js}, nil
}
// pinnedTo is a TLS configuration that accepts exactly the certificate in the file and no other:
// the leaf's SHA-256, compared on every handshake, with the name and the chain deliberately not
// consulted — a self-signed certificate with no names is the ordinary case for a mesh's bus.
func pinnedTo(path string) (*tls.Config, error) {
want, err := FingerprintOf(path)
if err != nil {
return nil, err
}
return PinnedToFingerprint(want), nil
}
// OnConn is the JetStream handle over a connection the caller already holds — the control plane's
// link — for asserting what the bus holds without dialling a second time.
func OnConn(conn *nats.Conn) *JetStream {
js, _ := conn.JetStream()
return &JetStream{conn: conn, js: js}
}
// DialPinned is Dial with the server's certificate pinned by a fingerprint the caller already holds
// — a module or a build machine that was handed one beside its credential, and has no file.
func DialPinned(url, fingerprint string, opts ...nats.Option) (*JetStream, error) {
if strings.TrimSpace(fingerprint) != "" {
opts = append(opts, nats.Secure(PinnedToFingerprint(fingerprint)))
}
return Dial(url, opts...)
}
// PinnedToFingerprint accepts exactly the certificate with this SHA-256 and no other.
func PinnedToFingerprint(want string) *tls.Config {
return &tls.Config{
InsecureSkipVerify: true, //nolint:gosec // replaced by the pin below, which is stricter
MinVersion: tls.VersionTLS12,
VerifyPeerCertificate: func(rawCerts [][]byte, _ [][]*x509.Certificate) error {
if len(rawCerts) == 0 {
return errors.New("the bus presented no certificate")
}
sum := sha256.Sum256(rawCerts[0])
got := "sha256:" + hex.EncodeToString(sum[:])
if got != want {
return fmt.Errorf("the bus presented a certificate this mesh does not know (%s…), expected %s…", got[:23], want[:23])
}
return nil
},
}
}
// Conn is the connection itself, for what the mesh keeps off JetStream on purpose — a heartbeat,
// a tool call — where a lost message is answered by the next one or by a timeout the caller
// already handles (design 25 §3).
func (j *JetStream) Conn() *nats.Conn { return j.conn }
// Context is the JetStream handle, for subscribing to what the consumers above define.
func (j *JetStream) Context() nats.JetStreamContext { return j.js }
func (j *JetStream) Close() {
if j.conn != nil && !j.borrowed {
j.conn.Close()
}
}
// Borrow is the same connection for a caller that will close what it was handed when it is done: its
// Close closes nothing, and the connection stays its owner's (novox/hq issue 327). How the serving
// controller lends its own connection to work that would otherwise dial one of its own.
func Borrow(j *JetStream) *JetStream {
lent := *j
lent.borrowed = true
return &lent
}
// EnsureStream creates the stream if it is absent and brings it to match if it is present.
//
// **Idempotent, because the controller asserts on every start** rather than creating once at
// genesis: a stream somebody deleted, or a mesh raised from a restored backup, has to converge
// rather than run without the guarantee its messages assume.
//
// An update, not a delete and recreate. Recreating would discard every message the stream holds
// and every consumer's position in it — which for CONTROL means the pushes being held through a
// store restart, exactly the guarantee the stream exists for.
func (j *JetStream) EnsureStream(s Stream) error {
want := &nats.StreamConfig{
Name: s.Name,
Subjects: s.Subjects,
Retention: retentionOf(s.Retention),
MaxAge: time.Duration(s.MaxAge) * time.Second,
MaxMsgsPerSubject: int64(s.MaxMsgsPerSubject),
Description: s.Why,
}
want.AllowDirect = s.Direct
want.AllowMsgTTL = s.AllowMsgTTL
if s.Restricted {
want.Metadata = map[string]string{RestrictedMetadata: "true"}
}
if s.MaxBytes > 0 {
want.MaxBytes = s.MaxBytes
}
if s.DiscardNew {
want.Discard = nats.DiscardNew
}
if s.DuplicatesSeconds > 0 {
want.Duplicates = time.Duration(s.DuplicatesSeconds) * time.Second
}
if s.Retention == RetentionLastPerSubject {
// Last-per-subject is a limits stream with one message kept per subject, not a
// retention policy of its own — the state shape, spelled the way the server spells it.
want.Retention = nats.LimitsPolicy
want.MaxMsgsPerSubject = 1
want.MaxAge = 0
}
switch _, err := j.js.StreamInfo(s.Name); {
case err == nil:
if _, err := j.js.UpdateStream(want); err != nil {
return fmt.Errorf("bringing stream %s to match: %w", s.Name, err)
}
return nil
case errors.Is(err, nats.ErrStreamNotFound):
if _, err := j.js.AddStream(want); err != nil {
return fmt.Errorf("creating stream %s: %w", s.Name, err)
}
return nil
default:
return fmt.Errorf("asking about stream %s: %w", s.Name, err)
}
}
// EnsureConsumer creates or updates one durable consumer.
//
// Explicit acknowledgement throughout: a consumer that acknowledges on delivery cannot redeliver
// work its holder died in the middle of, which is the whole difference between a queue and a
// firehose.
func (j *JetStream) EnsureConsumer(c Consumer) error {
want := &nats.ConsumerConfig{
Durable: c.Name,
AckPolicy: nats.AckExplicitPolicy,
AckWait: time.Duration(c.AckWaitSeconds) * time.Second,
MaxDeliver: c.MaxDeliver,
MaxAckPending: c.MaxAckPending,
DeliverGroup: c.Queue,
DeliverSubject: "",
Description: c.Why,
}
switch len(c.Filters) {
case 0:
case 1:
want.FilterSubject = c.Filters[0]
default:
want.FilterSubjects = c.Filters
}
// A queue group needs a delivery subject: a pull consumer has no group, and declaring one
// without the other is refused by the server with a message that does not say which half is
// missing.
if c.Queue != "" || c.Push {
// **Per consumer, which means per stream as well as per name** (novox/hq 04-ISSUES/146).
// A push consumer delivers onto an ordinary subject, and everything subscribed to that
// subject gets a copy. The controller holds a consumer called `controller` on CONTROL and
// another called `controller` on EVENTS, and both were given `_DELIVER.controller` — so the
// one process, holding both subscriptions, acted on every message twice. It enrolled a
// joining machine twice from one request, minting a second credential that replaced the one
// the machine had just been given; the same doubling applied to every report and every
// event the controller follows.
//
// The stream is in the name because the pair is what identifies a consumer — the server
// scopes a durable's name to its stream, and this subject is the only place that scoping
// was dropped. Already within what the controller may subscribe (`_DELIVER.controller.>`),
// so no permission moves.
want.DeliverSubject = DeliverSubjectFor(c)
}
switch have, err := j.js.ConsumerInfo(c.Stream, c.Name); {
case err == nil:
// **The controller owns the worker's shape, type included** (novox/hq issue 206). A holder
// built for a pull worker cannot bind a push one — `cannot pull subscribe to push based
// consumer` — and on 2026-10-03 the build machine rolled before the controller that would
// have redefined its worker, restarted on that for an hour, and nothing could build the
// controller that would have ended it. The server cannot change a consumer's type in place,
// so one of the wrong type is re-made: on a work queue nothing is lost, because what was
// acknowledged is gone from the stream and what was not is delivered again from the start.
// On any other stream a re-made consumer would replay what this one acknowledged (issue
// 156), so there it is said and left, and the person re-makes it knowing the cost.
if havePush, wantPush := have.Config.DeliverSubject != "", want.DeliverSubject != ""; havePush != wantPush {
shape := func(push bool) string {
if push {
return "push"
}
return "pull"
}
info, err := j.js.StreamInfo(c.Stream)
if err != nil {
return fmt.Errorf("asking about stream %s to re-make consumer %s: %w", c.Stream, c.Name, err)
}
if info.Config.Retention != nats.WorkQueuePolicy {
// **A stream that keeps its history is re-made from now on, never from the start.**
// Left for a hand, the hand re-makes it with the server's default — everything the
// stream holds — which on 2026-10-03 replayed every build ask since 1 October and
// re-registered nine modules from the past (novox/hq issue 207). What this consumer
// had not yet acknowledged is lost with it, and said: on a history stream that is
// the smaller cost, and the asks in flight are visible to whoever asked.
j.note("consumer %s on %s changes from %s to %s delivery on a stream that keeps its history: "+
"re-made to deliver from now on, so nothing this one acknowledged comes back (novox/hq issue "+
"207); %d ask(s) it had not acknowledged are not carried over and must be asked again",
c.Name, c.Stream, shape(havePush), shape(wantPush), have.NumPending+uint64(have.NumAckPending))
want.DeliverPolicy = nats.DeliverNewPolicy
} else {
j.note("consumer %s on %s changes from %s to %s delivery: re-made where it left off, nothing "+
"acknowledged comes back and nothing pending is lost (novox/hq issue 206); a holder bound to "+
"the old shape binds again", c.Name, c.Stream, shape(havePush), shape(wantPush))
}
if err := j.js.DeleteConsumer(c.Stream, c.Name); err != nil {
return fmt.Errorf("re-making consumer %s on %s as %s: %w", c.Name, c.Stream, shape(wantPush), err)
}
if _, err := j.js.AddConsumer(c.Stream, want); err != nil {
return fmt.Errorf("re-making consumer %s on %s as %s: %w", c.Name, c.Stream, shape(wantPush), err)
}
return nil
}
// Where an existing consumer starts is its history, not something an assertion may move:
// the server refuses a changed deliver policy outright. Carried across, so asserting twice
// is the no-op a restart depends on.
want.DeliverPolicy = have.Config.DeliverPolicy
want.OptStartSeq = have.Config.OptStartSeq
want.OptStartTime = have.Config.OptStartTime
if _, err := j.js.UpdateConsumer(c.Stream, want); err != nil {
// **A consumer that works is not replaced to make its name tidier**
// (novox/hq 04-ISSUES/156).
//
// The server will not move a push consumer's delivery subject while a subscriber is
// bound to it, and answers `consumer name already in use` — a message about the name,
// for a conflict about the subject. A node is bound to its declaration consumer the
// whole time it is up; that IS a node listening. So when 04-ISSUES/146 put the stream
// into the subject, every node consumer in a running mesh became one this could not
// bring to match, and the control plane crash-looped on the assertion it makes before
// it serves. A fresh mesh showed nothing: nothing was bound.
//
// Kept rather than deleted and re-made. Re-making moves the subject, and a holder may
// not be allowed to subscribe to the new one yet — the wider grant travels in the bus's
// user list, which this same control plane composes and a machine applies minutes
// later. Re-making here would have silenced every machine in the mesh, which is worse
// than the collision it was fixing and harder to undo.
//
// Kept rather than fatal, which is what 146's change intended and did not do: the bare
// subject it replaces still delivers, and it collides only where one holder has two
// consumers of one name. That is the controller's own pair, and the controller is not
// bound to them while it asserts, so those do move. A node has one consumer and nothing
// to collide with.
if have.Config.DeliverSubject != want.DeliverSubject {
j.note("consumer %s on %s still delivers to %q and not %q: %v. It keeps working; "+
"the subject moves on an assertion made while nothing is bound to it",
c.Name, c.Stream, have.Config.DeliverSubject, want.DeliverSubject, err)
return nil
}
return fmt.Errorf("bringing consumer %s on %s to match: %w", c.Name, c.Stream, err)
}
return nil
case errors.Is(err, nats.ErrConsumerNotFound):
if c.FromNow {
want.DeliverPolicy = nats.DeliverNewPolicy
}
if _, err := j.js.AddConsumer(c.Stream, want); err != nil {
return fmt.Errorf("creating consumer %s on %s: %w", c.Name, c.Stream, err)
}
return nil
default:
return fmt.Errorf("asking about consumer %s on %s: %w", c.Name, c.Stream, err)
}
}
// ConsumerState is what a reset says about a consumer, before and after.
type ConsumerState struct {
DeliverPolicy string
Delivered uint64
AckFloor uint64
Pending uint64
AckPending int
}
func stateOf(info *nats.ConsumerInfo) ConsumerState {
policy, _ := info.Config.DeliverPolicy.MarshalJSON()
return ConsumerState{DeliverPolicy: strings.Trim(string(policy), `"`), Delivered: info.Delivered.Stream,
AckFloor: info.AckFloor.Stream, Pending: info.NumPending, AckPending: info.NumAckPending}
}
// ResetConsumer re-makes a consumer to start from now, its configuration otherwise unchanged (novox/hq
// issue 248): what it had not yet delivered or acknowledged is dropped, which is the point — on a stream
// that keeps history, a consumer replaying a week of announcements does nothing anyone wants. Refused on
// a work queue, where what is pending is work nobody else will do.
func (j *JetStream) ResetConsumer(stream, name string) (before, after ConsumerState, err error) {
info, err := j.js.StreamInfo(stream)
if err != nil {
return before, after, fmt.Errorf("asking about stream %s: %w", stream, err)
}
if info.Config.Retention == nats.WorkQueuePolicy {
return before, after, fmt.Errorf("%s is a work queue: what its consumer has pending is work, and a reset would drop it", stream)
}
have, err := j.js.ConsumerInfo(stream, name)
if err != nil {
return before, after, fmt.Errorf("asking about consumer %s on %s: %w", name, stream, err)
}
before = stateOf(have)
want := have.Config
want.DeliverPolicy = nats.DeliverNewPolicy
want.OptStartSeq, want.OptStartTime = 0, nil
if err := j.js.DeleteConsumer(stream, name); err != nil {
return before, after, fmt.Errorf("removing consumer %s on %s: %w", name, stream, err)
}
made, err := j.js.AddConsumer(stream, &want)
if err != nil {
return before, after, fmt.Errorf("re-making consumer %s on %s — it is gone until the controller asserts it at its next start: %w", name, stream, err)
}
return before, stateOf(made), nil
}
func retentionOf(r Retention) nats.RetentionPolicy {
switch r {
case RetentionWorkQueue:
return nats.WorkQueuePolicy
default:
return nats.LimitsPolicy
}
}
// EnsureBucket creates a module's bucket if it is absent and brings its options to match if it is
// present (novox/hq ADR 0201).
//
// **An update, never a delete and recreate**, for the reason a stream is updated: recreating
// discards what the bucket holds, and what a module's state holds is data. The mesh's caps are
// asserted with the owner's options, so a bucket made by hand converges to them.
func (j *JetStream) EnsureBucket(b Bucket) error {
history := b.History
if history == 0 {
history = 1
}
js, err := jetstream.New(j.conn)
if err != nil {
return err
}
ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
defer cancel()
if _, err := js.CreateOrUpdateKeyValue(ctx, jetstream.KeyValueConfig{
Bucket: b.Bucket(),
Description: b.Why(),
History: uint8(history),
TTL: time.Duration(b.TTLSeconds) * time.Second,
MaxValueSize: StateMaxValueBytes,
MaxBytes: StateMaxBytes,
Storage: jetstream.FileStorage,
}); err != nil {
return fmt.Errorf("asserting bucket %s: %w", b.Bucket(), err)
}
return nil
}
// BucketNames is every key-value bucket on the server, the mesh's and anybody else's.
func (j *JetStream) BucketNames() ([]string, error) {
js, err := jetstream.New(j.conn)
if err != nil {
return nil, err
}
ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
defer cancel()
lister := js.KeyValueStoreNames(ctx)
var out []string
for name := range lister.Name() {
out = append(out, name)
}
return out, lister.Error()
}
// EnsureLog creates a module's log if it is absent and brings its configuration to match if it is
// present (novox/hq ADR 0297).
//
// **An update, never a delete and recreate**, for a bucket's reason: recreating discards what the log
// holds, and a log is a module's record. A configuration the server will not change in place (its
// storage, say, on a stream made by hand) is said as this assertion's error and the stream is left as
// it is — never removed to be made again.
func (j *JetStream) EnsureLog(l Log) error {
js, err := jetstream.New(j.conn)
if err != nil {
return err
}
ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
defer cancel()
if _, err := js.CreateOrUpdateStream(ctx, l.Config()); err != nil {
return fmt.Errorf("asserting log %s: %w", l.Stream(), err)
}
return nil
}
// LogStreams is every log stream on the server: every stream whose name starts with LogStreamPrefix.
func (j *JetStream) LogStreams() ([]string, error) {
js, err := jetstream.New(j.conn)
if err != nil {
return nil, err
}
ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
defer cancel()
lister := js.StreamNames(ctx)
var out []string
for name := range lister.Name() {
if strings.HasPrefix(name, LogStreamPrefix) {
out = append(out, name)
}
}
return out, lister.Err()
}
// ConsumerNames is every durable consumer on a stream; none for a stream that does not exist yet.
func (j *JetStream) ConsumerNames(stream string) ([]string, error) {
if _, err := j.js.StreamInfo(stream); errors.Is(err, nats.ErrStreamNotFound) {
return nil, nil
} else if err != nil {
return nil, err
}
var out []string
for name := range j.js.ConsumerNames(stream) {
out = append(out, name)
}
return out, nil
}
// RemoveConsumer deletes one durable consumer (novox/hq ADR 0363 §6): one the composition no longer asks for.
func (j *JetStream) RemoveConsumer(stream, name string) error {
if err := j.js.DeleteConsumer(stream, name); err != nil && !errors.Is(err, nats.ErrConsumerNotFound) {
return err
}
return nil
}