Force bq.enter()/bq.leave() to be called if the scheduler is idle

Once the final workers shut down, bq.enter()/bq.leave() won't end up
getting called. This means that they only get removed from the scheduler
the next time any RPC is called (e.g., when workers start to come online
again). This unfortunately causes Prometheus metrics to be incorrect in
the meantime.

Author: EdSchouten

cmd/bb_scheduler/main.go

@@ -141,7 +141,25 @@ func main() {
 			actionRouter,
 			executeAuthorizer,
 			modifyDrainsAuthorizer,
-			killOperationsAuthorizer)
+			killOperationsAuthorizer,
+		)
+
+		// Force periodic cleanups of stale workers. This also
+		// happens automatically when RPCs occur, but that's not
+		// sufficient to ensure Prometheus metrics are updated
+		// if the final workers disappear.
+		dependenciesGroup.Go(func(ctx context.Context, siblingsGroup, dependenciesGroup program.Group) error {
+			t := time.NewTicker(time.Minute)
+			for {
+				select {
+				case <-t.C:
+					buildQueue.ForceCleanup()
+				case <-ctx.Done():
+					t.Stop()
+					return nil
+				}
+			}
+		})
 
 		// Create predeclared platform queues.
 		for _, platformQueue := range configuration.PredeclaredPlatformQueues {

pkg/scheduler/in_memory_build_queue.go

@@ -1258,6 +1258,15 @@ func (bq *InMemoryBuildQueue) leave() {
 	bq.lock.Unlock()
 }
 
+// ForceCleanup forcefully runs any pending cleanup tasks. This method
+// can be invoked periodically to ensure that workers are removed, even
+// if no other RPC traffic occurs. This ensures that Prometheus metrics
+// report the correct values.
+func (bq *InMemoryBuildQueue) ForceCleanup() {
+	bq.enter(bq.clock.Now())
+	bq.leave()
+}
+
 // getIdleSynchronizeResponse returns a synchronization response that
 // explicitly instructs a worker to return to the idle state.
 func (bq *InMemoryBuildQueue) getIdleSynchronizeResponse() *remoteworker.SynchronizeResponse {