Use pull consumers (#2140)

* Pull consumers

* Pull consumers

* Only nuke consumers if they are push consumers

* Clean up old consumers

* Better error handling

* Update comments

setup/config/config_jetstream.go

@@ -2,8 +2,6 @@ package config
 
 import (
 	"fmt"
-
-	"github.com/nats-io/nats.go"
 )
 
 type JetStream struct {
@@ -25,8 +23,8 @@ func (c *JetStream) TopicFor(name string) string {
 	return fmt.Sprintf("%s%s", c.TopicPrefix, name)
 }
 
-func (c *JetStream) Durable(name string) nats.SubOpt {
-	return nats.Durable(c.TopicFor(name))
+func (c *JetStream) Durable(name string) string {
+	return c.TopicFor(name)
 }
 
 func (c *JetStream) Defaults(generate bool) {

setup/jetstream/helpers.go

@@ -1,12 +1,81 @@
 package jetstream
 
-import "github.com/nats-io/nats.go"
+import (
+	"context"
+	"fmt"
 
-func WithJetStreamMessage(msg *nats.Msg, f func(msg *nats.Msg) bool) {
-	_ = msg.InProgress()
-	if f(msg) {
-		_ = msg.Ack()
-	} else {
-		_ = msg.Nak()
+	"github.com/nats-io/nats.go"
+	"github.com/sirupsen/logrus"
+)
+
+func JetStreamConsumer(
+	ctx context.Context, js nats.JetStreamContext, subj, durable string,
+	f func(ctx context.Context, msg *nats.Msg) bool,
+	opts ...nats.SubOpt,
+) error {
+	defer func() {
+		// If there are existing consumers from before they were pull
+		// consumers, we need to clean up the old push consumers. However,
+		// in order to not affect the interest-based policies, we need to
+		// do this *after* creating the new pull consumers, which have
+		// "Pull" suffixed to their name.
+		if _, err := js.ConsumerInfo(subj, durable); err == nil {
+			if err := js.DeleteConsumer(subj, durable); err != nil {
+				logrus.WithContext(ctx).Warnf("Failed to clean up old consumer %q", durable)
+			}
+		}
+	}()
+
+	name := durable + "Pull"
+	sub, err := js.PullSubscribe(subj, name, opts...)
+	if err != nil {
+		return fmt.Errorf("nats.SubscribeSync: %w", err)
 	}
+	go func() {
+		for {
+			// The context behaviour here is surprising — we supply a context
+			// so that we can interrupt the fetch if we want, but NATS will still
+			// enforce its own deadline (roughly 5 seconds by default). Therefore
+			// it is our responsibility to check whether our context expired or
+			// not when a context error is returned. Footguns. Footguns everywhere.
+			msgs, err := sub.Fetch(1, nats.Context(ctx))
+			if err != nil {
+				if err == context.Canceled || err == context.DeadlineExceeded {
+					// Work out whether it was the JetStream context that expired
+					// or whether it was our supplied context.
+					select {
+					case <-ctx.Done():
+						// The supplied context expired, so we want to stop the
+						// consumer altogether.
+						return
+					default:
+						// The JetStream context expired, so the fetch probably
+						// just timed out and we should try again.
+						continue
+					}
+				} else {
+					// Something else went wrong, so we'll panic.
+					logrus.WithContext(ctx).WithField("subject", subj).Fatal(err)
+				}
+			}
+			if len(msgs) < 1 {
+				continue
+			}
+			msg := msgs[0]
+			if err = msg.InProgress(); err != nil {
+				logrus.WithContext(ctx).WithField("subject", subj).Warn(fmt.Errorf("msg.InProgress: %w", err))
+				continue
+			}
+			if f(ctx, msg) {
+				if err = msg.Ack(); err != nil {
+					logrus.WithContext(ctx).WithField("subject", subj).Warn(fmt.Errorf("msg.Ack: %w", err))
+				}
+			} else {
+				if err = msg.Nak(); err != nil {
+					logrus.WithContext(ctx).WithField("subject", subj).Warn(fmt.Errorf("msg.Nak: %w", err))
+				}
+			}
+		}
+	}()
+	return nil
 }