v0.3.3.

cmd/ts3monolith/main.go

@@ -52,15 +52,31 @@ func main() {
 func collect(ctx context.Context, cfg config.Config, tq *ts3.Query, rm *rrd.Manager) {
 	run := func() {
 		c, cancel := context.WithTimeout(ctx, cfg.TSQueryTimeout)
-		defer cancel()
 		n, err := tq.CountOnline(c)
+		cancel()
 		if err != nil {
 			log.Printf("stats query failed: %v", err)
-			return
 		}
-		c2, cancel2 := context.WithTimeout(ctx, 5*time.Second)
-		defer cancel2()
-		if err := rm.Update(c2, n); err != nil {
+		// Treat "query failed" and "query succeeded but nobody is online"
+		// the same way: repeat the last known sample instead of writing 0.
+		// This matches the original PHP app's collector exactly (it never
+		// distinguished the two cases either). At a 1-minute sampling
+		// interval a real but brief empty-server moment plots as a sharp
+		// spike down to 0 and back up, which reads as broken/jagged far
+		// more than it reads as useful information -- so, deliberately,
+		// a true "server went empty for a while" no longer shows as 0 in
+		// the graph, same trade-off the original app made.
+		if n == 0 {
+			c2, cancel2 := context.WithTimeout(ctx, 2*time.Second)
+			last, lastErr := rm.Last(c2)
+			cancel2()
+			if lastErr == nil {
+				n = last
+			}
+		}
+		c3, cancel3 := context.WithTimeout(ctx, 5*time.Second)
+		defer cancel3()
+		if err := rm.Update(c3, n); err != nil {
 			log.Printf("rrd update failed: %v", err)
 		}
 	}