shards: only trigger rescan on .zoekt files changing

Any write to the index dir triggered a scan. This means on busy
instances we are constantly rescanning, leading to an
over-representation in CPU profiles around watch. The events are
normally writes to our temporary files. By only considering events for
.zoekt files (which is what scan reads) we can avoid the constant scan
calls.

Just in case we also introduce a re-scan every minute in case we miss an
event. There is error handling around this, but I thought it is just
more reliable to call scan every once in a while.

Note: this doesn't represent significant CPU use, but it does muddy the
CPU profiler output. So this makes it easier to understand trends in our
continuous cpu profiling.

Test Plan: CI

shards/watcher.go

@@ -216,21 +216,38 @@ func (s *DirectoryWatcher) watch() error {
 	signal := make(chan struct{}, 1)
 
 	go func() {
+		notify := func() {
+			select {
+			case signal <- struct{}{}:
+			default:
+			}
+		}
+
+		ticker := time.NewTicker(time.Minute)
+
 		for {
 			select {
-			case <-watcher.Events:
-				select {
-				case signal <- struct{}{}:
-				default:
+			case event := <-watcher.Events:
+				// Only notify if a file we read in has changed. This is important to
+				// avoid all the events writing to temporary files.
+				if strings.HasSuffix(event.Name, ".zoekt") {
+					notify()
 				}
+
+			case <-ticker.C:
+				// Periodically just double check the disk
+				notify()
+
 			case err := <-watcher.Errors:
 				// Ignore ErrEventOverflow since we rely on the presence of events so
 				// safe to ignore.
 				if err != nil && err != fsnotify.ErrEventOverflow {
 					log.Println("watcher error:", err)
 				}
+
 			case <-s.quit:
 				watcher.Close()
+				ticker.Stop()
 				close(signal)
 				return
 			}