Builder: count only deliverable asks ahead, drop the old message of an ask asked again, and read a seat's queue (issue 1945)
mesh/delivery merged: source.main moved to novox/mesh-delivery@bcb5aafdbe80
mesh/check pass 3e8bd9b0def2 on bcb5aafdbe80 | unit 168 run, 472 kept | integration 178 run, 538 kept | composition passed on shanks

mesh-delivery #51's re-ask never fired for mesh-build-node.4e7b118e: messages the
SEAT_NODE_BUILD_GO workers had given up on stay on the work queue, and each counted
as waiting ahead, so the ask was never asked again (note 62). A message ahead now
waits only when no node wrote its step and no ask of its kind asked after it was
taken; the old message of an ask asked again is deleted from its stream; and
builder_seat_queue (answered by the builder through its queues state, as only its
account may read the stream) shows each message and whether it still waits.
This commit is contained in:
2026-10-26 01:19:30 +01:00
parent b842b9422d
commit 3e8bd9b0de
10 changed files with 686 additions and 77 deletions
@@ -5,19 +5,26 @@
package main
import (
"context"
"encoding/json"
"fmt"
"os"
"strings"
"sync"
"time"
stdio "git.novox.be/novox/mesh-sdk/go"
"git.novox.be/novox/mesh-sdk/go/delivery"
"mesh-delivery-builder/internal/build"
)
// Tools are the builder's two tools: stuck reads, askAgain writes the reasks value.
func Tools(r build.StuckReader, askAgain func(key string, value []byte) error, now func() time.Time) []stdio.Tool {
// queueAsker writes a queues value and waits for the builder's answer to it.
type queueAsker func(key string, value []byte) (json.RawMessage, error)
// Tools are the builder's tools: stuck reads, askAgain writes the reasks value, queue writes a queues value and waits for
// the builder's read.
func Tools(r build.StuckReader, askAgain func(key string, value []byte) error, queue queueAsker, now func() time.Time) []stdio.Tool {
str := func(args map[string]any, k string) string { s, _ := args[k].(string); return strings.TrimSpace(s) }
return []stdio.Tool{
{Name: build.StuckTool,
@@ -48,6 +55,78 @@ func Tools(r build.StuckReader, askAgain func(key string, value []byte) error, n
return map[string]any{"key": key, "asked": true,
"next": "the builder answers in " + build.ReAsksState + "." + key + "; " + build.StuckTool + " shows the answer and the new ask"}, nil
}},
{Name: build.SeatQueueTool,
Description: "What waits on a build seat's stream (SEAT_NODE_BUILD_<LANGUAGE>), as the builder reads it: the stream's message count and " +
"sequences, and per kind (delivery, check, recheck) its messages oldest first, at most limit each, with sequence, ask id, when asked " +
"and put, age, the state of the step a build node wrote for it, and whether the builder counts it as still waiting or as given up " +
"on (its worker delivered it its most times) or taken; and the ask asked last that a node took. The workers' delivery counts are " +
"not readable by the builder (nats.nats_streams shows them). Read by the builder, which alone may read the stream; read-only.",
Input: map[string]any{"type": "object", "properties": map[string]any{
"language": map[string]any{"type": "string", "description": "go, node or image (default go)"},
"seat": map[string]any{"type": "string", "description": "or the seat: node-build-go, node-build-node, node-build-image"},
"limit": map[string]any{"type": "integer", "description": fmt.Sprintf("messages read of each kind (default %d, at most %d)", build.DefaultQueueLimit, build.MaxQueueLimit)}}},
Run: func(args map[string]any) (any, error) {
language := build.LanguageOf(str(args, "language"))
if language == "" {
language = build.LanguageOf(str(args, "seat"))
}
if language == "" {
language = delivery.AskGo
}
limit, _ := args["limit"].(float64)
key, value, err := build.NewQueueAsk(language, int(limit), stdio.CallerOf(args), now())
if err != nil {
return nil, err
}
return queue(key, value)
}},
}
}
// askQueue writes the queues value at key and waits, on a watch of the key, at most bound for the builder's answer.
func askQueue(s stdio.KeptState, bound time.Duration) queueAsker {
return func(key string, value []byte) (json.RawMessage, error) {
answered := make(chan json.RawMessage, 1)
var rev uint64
var mu sync.Mutex
if err := s.Watch(key, func(c stdio.StateChange) error {
mu.Lock()
mine := rev
mu.Unlock()
if c.Op != "put" || mine == 0 || c.Revision <= mine {
return nil
}
if a, err := build.DecodeQueueAsk(c.Value); err == nil && !a.Answered.IsZero() {
select {
case answered <- c.Value:
default:
}
}
return nil
}); err != nil {
return nil, fmt.Errorf("watching %s.%s: %w", build.QueuesState, key, err)
}
r, err := s.Put(key, json.RawMessage(value))
if err != nil {
return nil, fmt.Errorf("the read could not be written to %s: %w", build.QueuesBucket, err)
}
mu.Lock()
rev = r
mu.Unlock()
// The watch may have heard the answer before the revision was known: read it once too.
if e, err := s.Get(key); err == nil && e != nil && e.Revision > r {
if a, err := build.DecodeQueueAsk(e.Value); err == nil && !a.Answered.IsZero() {
return e.Value, nil
}
}
wait, cancel := context.WithTimeout(context.Background(), bound) // a context's bound, no timer (ADR 0437 §8)
defer cancel()
select {
case v := <-answered:
return v, nil
case <-wait.Done():
return nil, fmt.Errorf("the builder did not answer %s.%s within %s: it is not running, or cannot read %s", build.QueuesState, key, bound, build.QueuesBucket)
}
}
}
@@ -92,7 +171,7 @@ func main() {
_, err := reasks.Put(key, json.RawMessage(value))
return err
}
if err := stdio.Serve("", Tools(r, put, time.Now)); err != nil {
if err := stdio.Serve("", Tools(r, put, askQueue(stdio.State(build.QueuesState), 25*time.Second), time.Now)); err != nil {
fmt.Fprintf(os.Stderr, "[mesh-delivery-builder-tools] %v\n", err)
os.Exit(1)
}
@@ -4,6 +4,9 @@ import (
"context"
"errors"
"fmt"
"sort"
"strings"
"time"
"github.com/nats-io/nats.go/jetstream"
@@ -36,22 +39,29 @@ func (h heldOnTheBus) Gone(ctx context.Context, a delivery.StepAsk) (string, err
return seen.Why(), nil
}
// seen reads what the bus holds of the ask; with later, a step of its build module started after the ask was put too,
// when the ask waits first in line.
// maxOnSubject bounds the messages of one subject a read walks (each a read by sequence, next by subject), and maxSteps
// the steps it reads whole to find the last ask taken: past them the read says it cannot tell, and the ask is waited for.
const (
maxOnSubject = 5000
maxSteps = 200
)
// seen reads what the bus holds of the ask; with later, the ask asked last that a build node took too (AskSeen.Last), when
// the ask is on its stream.
func (h heldOnTheBus) seen(ctx context.Context, a delivery.StepAsk, later bool) (build.AskSeen, error) {
var seen build.AskSeen
seen := build.AskSeen{Kind: a.Kind, At: build.AskedAt(a.ID)}
module, err := delivery.BuildModuleOf(a.Language)
if err != nil {
return seen, err
}
pattern, err := delivery.AskStepWatch(a.ID)
if err != nil {
return seen, err
}
steps, err := h.js.KeyValue(ctx, delivery.StepsBucket(module))
if err != nil {
return seen, fmt.Errorf("the bucket %s: %w", delivery.StepsBucket(module), err)
}
pattern, err := delivery.AskStepWatch(a.ID)
if err != nil {
return seen, err
}
keys, err := steps.ListKeysFiltered(ctx, pattern)
if err != nil {
return seen, err
@@ -61,67 +71,236 @@ func (h heldOnTheBus) seen(ctx context.Context, a delivery.StepAsk, later bool)
seen.Step = true // a node took it
return seen, nil
}
subject, err := delivery.AskSubject(a.Language, a.Kind)
if err != nil {
return seen, err
}
// The seat's stream by its name (delivery.AskStream), as the broker grants the builder its information and its
// messages by sequence and nothing wider (issue 1910).
name, err := delivery.AskStream(a.Language)
if err != nil {
return seen, err
}
stream, name, subject, err := h.stream(ctx, a.Language, a.Kind)
seen.Stream = name
stream, err := h.js.Stream(ctx, name)
if errors.Is(err, jetstream.ErrStreamNotFound) {
return seen, nil
}
if err != nil {
return seen, err
}
// A seat's stream is a work queue, which takes no second consumer over its subjects: its messages are read by sequence
// instead, each a direct read that consumes nothing.
info, err := stream.Info(ctx)
var ahead []build.Queued
err = onSubject(ctx, stream, subject, func(q build.Queued) bool {
if q.Ask == a.ID {
seen.OnStream, seen.PutAt = true, q.PutAt
return false
}
ahead = append(ahead, q)
return true
})
if err != nil || !seen.OnStream || !later {
return seen, err
}
taken, err := stepKeys(ctx, steps)
if err != nil {
return seen, err
}
for seq := info.State.FirstSeq; seq != 0 && seq <= info.State.LastSeq; seq++ {
m, err := stream.GetMsg(ctx, seq)
for i := range ahead {
ahead[i].Step = stepState(ctx, steps, taken[ahead[i].Ask])
}
seen.Ahead = ahead
since := seen.At
if since.IsZero() {
since = seen.PutAt
}
for _, q := range ahead {
if q.At.Before(since) {
since = q.At
}
}
seen.Last = lastTaken(ctx, steps, taken, a.Kind, since)
return seen, nil
}
// stream is the build seat's stream of language, its name and the subject of kind on it.
func (h heldOnTheBus) stream(ctx context.Context, language, kind string) (jetstream.Stream, string, string, error) {
subject, err := delivery.AskSubject(language, kind)
if err != nil {
return nil, "", "", err
}
// The seat's stream by its name (delivery.AskStream), as the broker grants the builder its information and its
// messages by sequence and nothing wider (issue 1910).
name, err := delivery.AskStream(language)
if err != nil {
return nil, "", "", err
}
stream, err := h.js.Stream(ctx, name)
return stream, name, subject, err
}
// onSubject hands each message of subject on stream to each, in order, until each answers false or the subject's
// messages end. A seat's stream is a work queue, which takes no second consumer over its subjects: its messages are read
// by sequence instead, each the next of the subject (one read per message, none for the sequences taken between), and
// none consumed.
func onSubject(ctx context.Context, stream jetstream.Stream, subject string, each func(build.Queued) bool) error {
for seq, n := uint64(1), 0; ; n++ {
if n >= maxOnSubject {
return fmt.Errorf("%s holds more than %d messages before the one asked after: not read further", subject, maxOnSubject)
}
m, err := stream.GetMsg(ctx, seq, jetstream.WithGetMsgSubject(subject))
if errors.Is(err, jetstream.ErrMsgNotFound) {
continue // taken and acknowledged, or removed
return nil
}
if err != nil {
return seen, err
return err
}
if m.Subject != subject {
continue
q := build.Queued{Seq: m.Sequence, Ask: m.Header.Get(jetstream.MsgIDHeader), PutAt: m.Time}
if _, kind, err := delivery.ParseAskSubject(m.Subject); err == nil {
q.Kind = kind
}
if m.Header.Get(jetstream.MsgIDHeader) == a.ID {
seen.OnStream, seen.PutAt = true, m.Time
break
if q.At = build.AskedAt(q.Ask); q.At.IsZero() {
q.At = q.PutAt
}
seen.Ahead++
if !each(q) {
return nil
}
seq = m.Sequence + 1
}
if !later || !seen.OnStream || seen.Ahead > 0 {
return seen, nil
}
// First in line and not taken: did a build node take any ask put after it? Its stream then no longer delivers it.
}
// stepKeys are the build module's step keys by ask id: `<node>.<ask id>` (a node's own key, such as its pause, has no
// ask id and is left out).
func stepKeys(ctx context.Context, steps jetstream.KeyValue) (map[string][]string, error) {
all, err := steps.ListKeys(ctx)
if err != nil {
return seen, err
if errors.Is(err, jetstream.ErrNoKeysFound) {
return map[string][]string{}, nil
}
return nil, err
}
defer all.Stop()
out := map[string][]string{}
for k := range all.Keys() {
if _, id, ok := strings.Cut(k, "."); ok && !build.AskedAt(id).IsZero() {
out[id] = append(out[id], k)
}
}
return out, nil
}
// stepState is the state of the first of keys a node wrote, "" for none.
func stepState(ctx context.Context, steps jetstream.KeyValue, keys []string) string {
for _, k := range keys {
e, err := steps.Get(ctx, k)
if err != nil {
continue
}
s, err := delivery.DecodeStep(e.Value())
if err != nil || !s.Started.After(seen.PutAt) {
continue // not a step (a node's pause key), or one started before the ask was put
if s, err := delivery.DecodeStep(e.Value()); err == nil {
return s.State
}
seen.Later = &s
break
}
return seen, nil
return ""
}
// lastTaken is the ask asked last after since that a build node took, of a kind whose taking proves kind passed
// (build.Proves), read newest first and at most maxSteps of them; nil for none. A step's tool is "<kind> ask"
// (mesh-sdk buildstep).
func lastTaken(ctx context.Context, steps jetstream.KeyValue, taken map[string][]string, kind string, since time.Time) *build.Taken {
var ids []string
for id := range taken {
if build.AskedAt(id).After(since) {
ids = append(ids, id)
}
}
sort.Slice(ids, func(i, j int) bool { return build.AskedAt(ids[i]).After(build.AskedAt(ids[j])) })
read := 0
for _, id := range ids {
for _, k := range taken[id] {
if read >= maxSteps {
return nil
}
read++
e, err := steps.Get(ctx, k)
if err != nil {
continue
}
s, err := delivery.DecodeStep(e.Value())
if err != nil {
continue
}
stepKind, ok := strings.CutSuffix(s.Tool, " ask")
if !ok || !build.Proves(stepKind, kind) {
continue
}
return &build.Taken{Ask: id, Kind: stepKind, At: build.AskedAt(id), Node: s.Node, Started: s.Started}
}
}
return nil
}
// Drop is build.Dropper's Drop: it removes the ask's message from its build seat's stream, found by its Nats-Msg-Id on its
// subject, when the stream still holds it. The builder's account must be granted $JS.API.STREAM.MSG.DELETE.<stream>.
func (h heldOnTheBus) Drop(ctx context.Context, a delivery.StepAsk) error {
stream, name, subject, err := h.stream(ctx, a.Language, a.Kind)
if errors.Is(err, jetstream.ErrStreamNotFound) {
return nil
}
if err != nil {
return err
}
var seq uint64
if err := onSubject(ctx, stream, subject, func(q build.Queued) bool {
if q.Ask == a.ID {
seq = q.Seq
return false
}
return true
}); err != nil || seq == 0 {
return err
}
if err := stream.DeleteMsg(ctx, seq); err != nil && !errors.Is(err, jetstream.ErrMsgNotFound) {
return fmt.Errorf("deleting message %d of %s (is $JS.API.STREAM.MSG.DELETE.%s granted to the builder?): %w", seq, name, name, err)
}
return nil
}
// Queue is what the tool builder_seat_queue answers (build.SeatQueue): the messages of language's build seat's stream, of
// every kind, oldest first and at most limit of each kind, each with its step's state and whether it still waits.
func (h heldOnTheBus) Queue(ctx context.Context, language string, limit int) (build.SeatQueue, error) {
q := build.SeatQueue{Language: language, Note: build.QueueNote}
module, err := delivery.BuildModuleOf(language)
if err != nil {
return q, err
}
stream, name, _, err := h.stream(ctx, language, delivery.AskDelivery)
q.Stream = name
if errors.Is(err, jetstream.ErrStreamNotFound) {
q.Note = name + " does not exist"
return q, nil
}
if err != nil {
return q, err
}
info, err := stream.Info(ctx)
if err != nil {
return q, err
}
q.Messages, q.FirstSeq, q.LastSeq = info.State.Msgs, info.State.FirstSeq, info.State.LastSeq
steps, err := h.js.KeyValue(ctx, delivery.StepsBucket(module))
if err != nil {
return q, fmt.Errorf("the bucket %s: %w", delivery.StepsBucket(module), err)
}
taken, err := stepKeys(ctx, steps)
if err != nil {
return q, err
}
for _, kind := range delivery.AskKinds {
subject, _ := delivery.AskSubject(language, kind)
var ahead []build.Queued
if err := onSubject(ctx, stream, subject, func(m build.Queued) bool {
ahead = append(ahead, m)
return len(ahead) < limit
}); err != nil {
return q, err
}
if len(ahead) == 0 {
continue
}
for i := range ahead {
ahead[i].Step = stepState(ctx, steps, taken[ahead[i].Ask])
}
seen := build.AskSeen{Kind: kind, Ahead: ahead, Last: lastTaken(ctx, steps, taken, kind, ahead[0].At)}
q.Kinds = append(q.Kinds, build.NewKindQueue(kind, seen, time.Now()))
}
return q, nil
}
@@ -24,6 +24,8 @@ import (
"git.novox.be/novox/mesh-sdk/go/buildstep"
"git.novox.be/novox/mesh-sdk/go/delivery"
"mesh-delivery-builder/internal/build"
)
// busConf is the shape of the composed accounts file (mesh-controller internal/broker): one account with JetStream, the
@@ -822,6 +824,34 @@ func TestARestartedBuilderPutsNoSecondAsk(t *testing.T) {
t.Fatalf("asks put = %+v within 20 s, want the two the key names", asked())
}
}
// 1945 note 62, on the real bus: the stream read as builder_seat_queue reads it shows the Go ask waiting; the ask can
// still be answered (nothing was taken since); and a message dropped as an ask asked again is gone from the stream.
held := heldOnTheBus{b.js}
goAsk := delivery.StepAsk{ID: v.Asks[delivery.AskGo], Language: delivery.AskGo, Kind: delivery.AskDelivery}
stale := delivery.StepAsk{ID: "mvo53jcm70b5f7", Language: delivery.AskGo, Kind: delivery.AskDelivery}
subject, _ := delivery.AskSubject(delivery.AskGo, delivery.AskDelivery)
if _, err := b.js.Publish(b.ctx, subject, []byte("{}"), jetstream.WithMsgID(stale.ID)); err != nil {
t.Fatal(err)
}
queued := func() []build.QueuedState {
q, err := held.Queue(b.ctx, delivery.AskGo, 10)
if err != nil || len(q.Kinds) != 1 || q.Kinds[0].Kind != delivery.AskDelivery {
t.Fatalf("the Go seat's queue = %+v, %v; want its delivery asks", q, err)
}
return q.Kinds[0].Messages
}
if m := queued(); len(m) != 2 || m[0].Ask != goAsk.ID || !m[0].Waits || m[1].Ask != stale.ID {
t.Fatalf("the Go delivery queue = %+v, want the key's ask waiting, then the stale one", m)
}
if why, err := held.Gone(b.ctx, goAsk); why != "" || err != nil {
t.Errorf("Gone(the waiting ask) = %q, %v; want it answerable", why, err)
}
if err := held.Drop(b.ctx, stale); err != nil {
t.Fatal(err)
}
if m := queued(); len(m) != 1 || m[0].Ask != goAsk.ID {
t.Fatalf("after the drop the Go delivery queue = %+v, want the key's ask alone", m)
}
b.stop()
b.start(t, "n1", nil) // the same builder, started again, while both asks wait
b.fake.open()
@@ -176,7 +176,7 @@ func run(ctx context.Context, getenv func(string) string, logOut io.Writer) erro
// revision (the gate's source state). The source's revisions are watched with their history (a direct get of one revision
// is not what a read grants): a revision heard after the candidate that names it builds that candidate then. Where it
// cannot be read the trunk's keys are built and the candidates wait (issue 1945).
errs := make(chan error, 4)
errs := make(chan error, 6)
trunks := &trunkSource{ctx: ctx, js: js, heads: &sourceHeads{}, moved: b.TrunkMoved, errs: errs, logf: logger.Printf}
trunks.start()
b.Trunk = trunks.head
@@ -288,6 +288,16 @@ func run(ctx context.Context, getenv func(string) string, logOut io.Writer) erro
go func() { errs <- follow(ctx, rw, eachReAsk) }()
}
}
// A person reads a build seat's stream with builder_seat_queue, which writes the queues state (issue 1945 note 62): the
// builder alone may read the stream, so it answers there.
if kv := process.OptionalBucket(ctx, js, build.QueuesBucket, "builder_seat_queue is not answered until the builder starts again", logger.Printf); kv != nil {
eachRead := func(e jetstream.KeyValueEntry) { queueRead(ctx, heldOnTheBus{js}, kv, e, logger.Printf) }
if qw, err := settle(ctx, kv, eachRead); err != nil {
process.NotWatched(logger.Printf, build.QueuesBucket, err, "builder_seat_queue is not answered until the builder starts again")
} else {
go func() { errs <- follow(ctx, qw, eachRead) }()
}
}
logger.Printf("started on %s", cred.Node)
go func() { errs <- follow(ctx, w, each) }()
select {
@@ -0,0 +1,49 @@
package main
import (
"context"
"encoding/json"
"errors"
"time"
"github.com/nats-io/nats.go/jetstream"
"mesh-delivery-builder/internal/build"
)
// queueReader is the part of the builder a queues value reads through: heldOnTheBus.
type queueReader interface {
Queue(ctx context.Context, language string, limit int) (build.SeatQueue, error)
}
// queueRead answers one value of the queues state (the tool builder_seat_queue, issue 1945 note 62): a value not yet
// answered is read now, from the stream the builder's account may read, and written back by compare-and-set at its
// revision with what was read or why it could not be. A value answered already (this builder's own write, or another
// instance's) or deleted is left.
func queueRead(ctx context.Context, r queueReader, kv jetstream.KeyValue, e jetstream.KeyValueEntry, logf func(string, ...any)) {
if e.Operation() != jetstream.KeyValuePut {
return
}
a, err := build.DecodeQueueAsk(e.Value())
if err != nil {
logf("%s.%s holds what is not a queue read, left: %v", build.QueuesState, e.Key(), err)
return
}
if !a.Answered.IsZero() {
return
}
read, cancel := context.WithTimeout(ctx, 20*time.Second)
q, err := r.Queue(read, e.Key(), max(1, min(a.Limit, build.MaxQueueLimit)))
cancel()
if err != nil {
a.Error = err.Error()
}
a.Queue, a.Answered = &q, time.Now().UTC()
data, err := json.Marshal(a)
if err != nil {
return
}
if _, err := kv.Update(ctx, e.Key(), data, e.Revision()); err != nil && !errors.Is(err, jetstream.ErrKeyExists) && !errors.Is(err, jetstream.ErrKeyRevisionMismatch) {
logf("%s.%s: the read could not be written: %v", build.QueuesState, e.Key(), err)
}
}