Name no controller verb in hints and drop its handover and paging reads, since the controller ended (issue 1126 B19)

Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
This commit is contained in:
2026-11-02 11:23:14 +01:00
parent 49f9269376
commit 4f636b7080
36 changed files with 171 additions and 416 deletions
+20 -58
View File
@@ -32,9 +32,7 @@ import (
"github.com/novox/mesh-tools/node-tools/internal/wire"
)
// RequestTimeout is how long a call waits for its answer — what modules already expect. The
// controller answers its verbs well inside it (its AnswerWithin, ten seconds), with a call id while a
// verb still runs, so a caller is never left guessing past this (novox/hq issue 265).
// RequestTimeout is how long a call waits for its answer — what modules already expect.
const RequestTimeout = 30 * time.Second
// Credential is a broker credential as the mesh delivers it (novox/hq ADR 0120).
@@ -344,7 +342,7 @@ func (c *Conn) Credential() Credential { return c.cred }
// Follow reads a module's membership on this machine once and follows it live, so its tools are
// served where the mesh issued them (ADR 0175). It is read from the grants part's state memberships, its key
// `<node>.<module>` (novox/hq ADR 0571 clause 7 and 10, mesh-sdk buscompose.GrantsMembershipsBucket), and nowhere else:
// the controller's assignments stream is not read and its subject is not heard.
// the old assignments stream is not read and its subject is not heard.
func (c *Conn) Follow(module string) {
c.mu.Lock()
if c.node == "" {
@@ -790,83 +788,47 @@ func (c *Conn) Ask(key string, body any, on string) (Answered, error) {
// account publish with its own name alone; asked again on the subject that names no caller only when the
// first reached nobody — nothing serves it yet, or the bus does not grant it yet — so no call is done twice.
if called := CalledSubject(subject, c.cred.User); called != "" {
answered, handingOver, fallBack, err := c.askCalled(called, data)
answered, fallBack, err := c.askCalled(called, data)
if !fallBack {
if !handingOver {
return answered, err
}
time.Sleep(HandoverPause)
answered, again, _, err := c.askCalled(called, data)
if err != nil && again {
return Answered{}, fmt.Errorf("asked twice, %s apart, and both times refused as a handover: %w", HandoverPause, err)
}
return answered, err
}
}
answered, handingOver, err := c.askOnce(subject, data)
if !handingOver {
return answered, err
}
// **Refused because the controller is handing over: asked once more** (novox/hq issue 289). The
// controller being replaced did nothing and said so; the one after it answers the same call. Once:
// a second refusal is a controller that is not coming back, and the caller is told.
time.Sleep(HandoverPause)
answered, again, err := c.askOnce(subject, data)
if err != nil && again {
return Answered{}, fmt.Errorf("asked twice, %s apart, and both times refused as a handover: %w", HandoverPause, err)
}
return answered, err
return c.askOnce(subject, data)
}
// HandoverPause is how long a call refused as a handover waits before it is asked once more: the
// supervisor's restart of the controller, which stops the one being replaced and starts the next. A
// variable so a test need not wait.
var HandoverPause = 3 * time.Second
// RetryHandingOver is the mark the controller puts on a call it refused because it is being replaced
// (mesh-controller internal/link, `"retry": "handing-over"`).
const RetryHandingOver = "handing-over"
// ErrNoResponders is the bus's own refusal: nothing subscribes to the subject asked. It is told apart
// from a tool's own error by errors.Is, never by its text (a tool may say "503" itself, issue 1529).
var ErrNoResponders = errors.New("503 no responders")
// askOnce asks once, and says whether the answer was a refusal marked to be asked again.
func (c *Conn) askOnce(subject string, data []byte) (Answered, bool, error) {
// askOnce asks once.
func (c *Conn) askOnce(subject string, data []byte) (Answered, error) {
msg, err := c.nc.Request(subject, data, RequestTimeout)
if err != nil {
if errors.Is(err, nats.ErrNoResponders) {
return Answered{}, false, ErrNoResponders
return Answered{}, ErrNoResponders
}
if errors.Is(err, nats.ErrTimeout) {
return Answered{}, false, errors.New("timeout")
return Answered{}, errors.New("timeout")
}
return Answered{}, false, err
return Answered{}, err
}
return c.answerOf(subject, msg.Data)
return answerOf(msg.Data)
}
// answerOf reads one answer, asking the rest of a paged one on the subject it was asked on, and says whether it
// was a refusal marked to be asked again.
func (c *Conn) answerOf(subject string, data []byte) (Answered, bool, error) {
// An answer larger than one message comes in parts (pages.go, novox/hq issue 314).
body, err := c.whole(subject, data)
if err != nil {
return Answered{}, false, err
}
// answerOf reads one answer.
func answerOf(data []byte) (Answered, error) {
var r struct {
Result json.RawMessage `json:"result"`
Error string `json:"error"`
Node string `json:"node"`
Retry string `json:"retry"`
}
if err := json.Unmarshal(body, &r); err != nil {
return Answered{}, false, err
if err := json.Unmarshal(data, &r); err != nil {
return Answered{}, err
}
if r.Error != "" {
return Answered{}, r.Retry == RetryHandingOver, errors.New(r.Error)
return Answered{}, errors.New(r.Error)
}
return Answered{Result: r.Result, Node: r.Node}, false, nil
return Answered{Result: r.Result, Node: r.Node}, nil
}
// PublishAs emits an event as a module: published into JetStream and awaited, de-duplicated by its
@@ -1070,7 +1032,7 @@ func (c *Conn) AskAs(module, key string, body any) (Answered, error) {
}
// EventsStream is where every module's events land, and ConsumerOf the durable consumer the
// controller makes for a module on a machine: the same names the module's own runtime bound
// mesh makes for a module on a machine: the same names the module's own runtime bound
// (node-tools/src/broker-nats.ts), so moving the module into the node's runtime neither loses an
// event nor sees one twice.
const EventsStream = "EVENTS"
@@ -1085,7 +1047,7 @@ func ConsumerOf(node, module string) string {
// NakDelay is how long an event a handler failed waits before it is offered again: a transient cause
// gets another attempt, a permanent one exhausts the consumer's max-deliver rather than spinning. The
// consumer then gives the event up, and the controller keeps it in DEAD_LETTERS until a person delivers
// consumer then gives the event up, and the mesh keeps it in DEAD_LETTERS until a person delivers
// it again — under `mesh.again.<consumer>.…`, from which keyOf reads the same key — or drops it
// (novox/hq issue 330).
var NakDelay = 5 * time.Second
@@ -1099,7 +1061,7 @@ var EventHeartbeat = 10 * time.Second
// ConsumeAs reads a module's durable consumer on EVENTS, and each one on a module's stream its
// membership names (novox/hq ADR 0363 §8), and hands each event to deliver one at a time (novox/hq
// ADR 0198): acknowledged when deliver returns nil, negatively acknowledged after NakDelay when it
// returns an error, terminated when it is not an event at all. The consumers are the controller's to
// returns an error, terminated when it is not an event at all. The consumers are the mesh's to
// create; this binds to them and never makes one. Runs until stopped.
func (c *Conn) ConsumeAs(module string, deliver func(Envelope) error) (func(), error) {
durable := ConsumerOf(c.node, module)
@@ -1109,7 +1071,7 @@ func (c *Conn) ConsumeAs(module string, deliver func(Envelope) error) (func(), e
}
h := newHearing(c, js, module, deliver)
// Bound by name, never created and never matched against a subject: the consumer and its
// filters are the controller's (design 29 §3), exactly as the module's own runtime bound it.
// filters are the mesh's (design 29 §3), exactly as the module's own runtime bound it.
events, _, doing, err := h.bind(EventsStream, durable, nil)
if err != nil {
return nil, fmt.Errorf("%s %s's consumer %s: %w", doing, module, durable, err)