20457: Don't displace locked containers with queued containers.

[arvados.git] / lib / dispatchcloud / scheduler / run_queue.go

diff --git a/lib/dispatchcloud/scheduler/run_queue.go b/lib/dispatchcloud/scheduler/run_queue.go
index e6b1b06a2ca46e84d33d87e9dc1b86ddd3108d11..8f4c2e083da584a869c7b8c34bb8d69434ad1584 100644 (file)
--- a/lib/dispatchcloud/scheduler/run_queue.go
+++ b/lib/dispatchcloud/scheduler/run_queue.go
@@ -22,7 +22,19 @@ func (sch *Scheduler) runQueue() {
                sorted = append(sorted, ent)
        }
        sort.Slice(sorted, func(i, j int) bool {
-               if pi, pj := sorted[i].Container.Priority, sorted[j].Container.Priority; pi != pj {
+               ilocked := sorted[i].Container.State == arvados.ContainerStateLocked
+               jlocked := sorted[j].Container.State == arvados.ContainerStateLocked
+               if ilocked != jlocked {
+                       // Give precedence to containers that we have
+                       // already locked, even if higher-priority
+                       // containers have since arrived in the
+                       // queue. This avoids undesirable queue churn
+                       // effects including extra lock/unlock cycles
+                       // and bringing up new instances and quickly
+                       // shutting them down to make room for
+                       // different instance sizes.
+                       return ilocked
+               } else if pi, pj := sorted[i].Container.Priority, sorted[j].Container.Priority; pi != pj {
                        return pi > pj
                } else {
                        // When containers have identical priority,