Let a process consume its module's work queue, acknowledging, delaying or terminating each message (issue 1190)

go/process/queue.go

@@ -0,0 +1,127 @@
+package process
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/nats-io/nats.go/jetstream"
+)
+
+// Message is one piece of work this process's queue consumer delivered, with how many times the bus has
+// delivered it. Nothing acts on its behalf: the handler decides its fate exactly once, by calling Ack,
+// NakWithDelay or Term (mesh-tools node-tools/internal/bus/seat.go's seat work Consume, with the same
+// three outcomes, offered here explicitly instead of by a handler's return value, since the caller
+// chooses its own delay before a message comes back).
+type Message struct {
+	Subject string
+	Body    []byte
+	Headers map[string]string
+	// Delivered is how many times the bus has handed this message over, this time included.
+	Delivered uint64
+
+	ack  func() error
+	nak  func(time.Duration) error
+	term func() error
+}
+
+// Ack acknowledges the message: the bus never delivers it again.
+func (m *Message) Ack() error { return m.ack() }
+
+// NakWithDelay asks for the message to be delivered again no sooner than delay, its delivery count
+// raised by one.
+func (m *Message) NakWithDelay(delay time.Duration) error { return m.nak(delay) }
+
+// Term terminates the message: the bus never delivers it again, and does not count it a dead letter the
+// way running out of deliveries does.
+func (m *Message) Term() error { return m.term() }
+
+// ProcessID is this account's own process id, the last segment of its user name <node>.<module>.<id>
+// (mesh-controller internal/broker, Principal.Username for KindProcess).
+func (a Account) ProcessID() string {
+	prefix := a.Node + "." + a.Module + "."
+	if !strings.HasPrefix(a.User, prefix) {
+		return ""
+	}
+	return strings.TrimPrefix(a.User, prefix)
+}
+
+// QueueConsumerName is the one durable consumer a process's bus.queues names it on its work-queue stream
+// (novox/hq ADR 0436 clauses 1, 2, 6; mesh-controller internal/broker/modstreams.go QueueConsumer):
+// <node>_<module>_<process id>.
+func (a Account) QueueConsumerName() string { return a.Node + "_" + a.Module + "_" + a.ProcessID() }
+
+// errNoMessage is one pull that fetched nothing within its wait: not an error to the caller, just a
+// reason to pull again.
+var errNoMessage = errors.New("no message this pull")
+
+// Consume binds this process's one durable consumer on stream (QueueConsumerName) and pulls from it, one
+// message at a time, handing each to handle until ctx ends: it then stops pulling and returns ctx.Err(),
+// never acknowledging a message still in handle's hands — only handle decides a message's fate, by
+// calling Ack, NakWithDelay or Term on it.
+func (c *Conn) Consume(ctx context.Context, stream string, handle func(*Message)) error {
+	js, err := jetstream.New(c.nc)
+	if err != nil {
+		return err
+	}
+	name := c.account.QueueConsumerName()
+	cons, err := js.Consumer(ctx, stream, name)
+	if err != nil {
+		return fmt.Errorf("binding %s's queue consumer %s on %s: %w", c.account.Module, name, stream, err)
+	}
+	return runQueue(ctx, fetchOne(cons), handle)
+}
+
+// fetchOne is one pull of at most one message, non-blocking past ctx: Fetch itself does not take a
+// context, so this selects on its message channel and ctx together, and returns promptly once ctx ends
+// even while a fetch is still outstanding.
+func fetchOne(cons jetstream.Consumer) func(ctx context.Context) (*Message, error) {
+	return func(ctx context.Context) (*Message, error) {
+		batch, err := cons.Fetch(1, jetstream.FetchMaxWait(5*time.Second))
+		if err != nil {
+			return nil, err
+		}
+		select {
+		case msg, ok := <-batch.Messages():
+			if !ok {
+				return nil, errNoMessage
+			}
+			headers := map[string]string{}
+			for k := range msg.Headers() {
+				headers[k] = msg.Headers().Get(k)
+			}
+			var delivered uint64 = 1
+			if meta, err := msg.Metadata(); err == nil && meta != nil {
+				delivered = meta.NumDelivered
+			}
+			return &Message{Subject: msg.Subject(), Body: append([]byte(nil), msg.Data()...), Headers: headers,
+				Delivered: delivered, ack: msg.Ack, nak: msg.NakWithDelay, term: msg.Term}, nil
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		}
+	}
+}
+
+// runQueue is Consume's loop, apart from the bus so a unit test can drive it with a fake next
+// (novox/hq issue 1190 AC4): it checks ctx before every pull, and stops as soon as next or the check
+// says ctx has ended, without touching the message past handing it to handle.
+func runQueue(ctx context.Context, next func(context.Context) (*Message, error), handle func(*Message)) error {
+	for {
+		if err := ctx.Err(); err != nil {
+			return err
+		}
+		msg, err := next(ctx)
+		if err != nil {
+			if ctx.Err() != nil {
+				return ctx.Err()
+			}
+			if errors.Is(err, errNoMessage) {
+				continue
+			}
+			return err
+		}
+		handle(msg)
+	}
+}