Use popOne in JS API routed request workers (#6355)

This updates the JS API workers to use `popOne`, so that if the queue
stacks up at all, the recovery is fairer across workers, rather than
single workers stealing the entire queue.

This was extracted from #6342.

Signed-off-by: Neil Twigg <[email protected]>

server/jetstream_api.go

@@ -922,8 +922,10 @@ func (s *Server) processJSAPIRoutedRequests() {
 	for {
 		select {
 		case <-queue.ch:
-			reqs := queue.pop()
-			for _, r := range reqs {
+			// Only pop one item at a time here, otherwise if the system is recovering
+			// from queue buildup, then one worker will pull off all the tasks and the
+			// others will be starved of work.
+			for r, ok := queue.popOne(); ok && r != nil; r, ok = queue.popOne() {
 				client.pa = r.pa
 				start := time.Now()
 				r.jsub.icb(r.sub, client, r.acc, r.subject, r.reply, r.msg)
@@ -932,7 +934,6 @@ func (s *Server) processJSAPIRoutedRequests() {
 				}
 				atomic.AddInt64(&js.apiInflight, -1)
 			}
-			queue.recycle(&reqs)
 		case <-s.quitCh:
 			return
 		}

server/jetstream_cluster_4_test.go

@@ -5265,3 +5265,69 @@ func TestJetStreamClusterOnlyPublishAdvisoriesWhenInterest(t *testing.T) {
 	// it should succeed.
 	require_True(t, s1.publishAdvisory(s1.GlobalAccount(), subj, "test"))
 }
+
+func TestJetStreamClusterRoutedAPIRecoverPerformance(t *testing.T) {
+	c := createJetStreamClusterExplicit(t, "R3S", 3)
+	defer c.shutdown()
+
+	nc, _ := jsClientConnect(t, c.randomNonLeader())
+	defer nc.Close()
+
+	// We only run 16 JetStream API workers.
+	mp := runtime.GOMAXPROCS(0)
+	if mp > 16 {
+		mp = 16
+	}
+
+	leader := c.leader()
+	ljs := leader.js.Load()
+
+	// Take the JS lock, which allows the JS API queue to build up.
+	ljs.mu.Lock()
+	defer ljs.mu.Unlock()
+
+	count := JSDefaultRequestQueueLimit - 1
+	ch := make(chan *nats.Msg, count)
+
+	inbox := nc.NewRespInbox()
+	_, err := nc.ChanSubscribe(inbox, ch)
+	require_NoError(t, err)
+
+	// To ensure a fair starting line, we need to submit as many tasks as
+	// there are JS workers whilst holding the JS lock. This will ensure that
+	// each JS API worker is properly wedged.
+	msg := &nats.Msg{
+		Subject: fmt.Sprintf(JSApiConsumerInfoT, "Doesnt", "Exist"),
+		Reply:   "no_one_here",
+	}
+	for i := 0; i < mp; i++ {
+		require_NoError(t, nc.PublishMsg(msg))
+	}
+
+	// Then we want to submit a fixed number of tasks, big enough to fill
+	// the queue, so that we can measure them.
+	msg = &nats.Msg{
+		Subject: fmt.Sprintf(JSApiConsumerInfoT, "Doesnt", "Exist"),
+		Reply:   inbox,
+	}
+	for i := 0; i < count; i++ {
+		require_NoError(t, nc.PublishMsg(msg))
+	}
+	checkFor(t, 5*time.Second, 25*time.Millisecond, func() error {
+		if queued := leader.jsAPIRoutedReqs.len(); queued != count {
+			return fmt.Errorf("expected %d queued requests, got %d", count, queued)
+		}
+		return nil
+	})
+
+	// Now we're going to release the lock and start timing. The workers
+	// will now race to clear the queues and we'll wait to see how long
+	// it takes for them all to respond.
+	start := time.Now()
+	ljs.mu.Unlock()
+	for i := 0; i < count; i++ {
+		<-ch
+	}
+	ljs.mu.Lock()
+	t.Logf("Took %s to clear %d items", time.Since(start), count)
+}