Set suggestion to highest oberved over a week

Since suggestions are consumed by humans they should change on
the time scale of human decision making.

2 files changed, 25 insertions, 1 deletions

diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/ScalingSuggestionsMaintainer.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/ScalingSuggestionsMaintainer.java
index d75cfbecee7..be8966068c9 100644
--- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/ScalingSuggestionsMaintainer.java
+++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/ScalingSuggestionsMaintainer.java
@@ -85,7 +85,16 @@ public class ScalingSuggestionsMaintainer extends NodeRepositoryMaintainer {
         Optional<Cluster> cluster = application.cluster(clusterId);
         if (cluster.isEmpty()) return;
         var at = nodeRepository().clock().instant();
-        applications().put(application.with(cluster.get().withSuggested(suggestion.map(s -> new Cluster.Suggestion(s,  at)))), lock);
+        var currentSuggestion = cluster.get().suggestedResources();
+        if (currentSuggestion.isEmpty()
+            || currentSuggestion.get().at().isBefore(at.minus(Duration.ofDays(7)))
+            || suggestion.isPresent() && isHigher(suggestion.get(), currentSuggestion.get().resources()))
+            applications().put(application.with(cluster.get().withSuggested(suggestion.map(s -> new Cluster.Suggestion(s,  at)))), lock);
+    }
+
+    private boolean isHigher(ClusterResources r1, ClusterResources r2) {
+        // Use cost as a measure of "highness" over multiple dimensions
+        return r1.totalResources().cost() > r2.totalResources().cost();
     }
 
     private Map<ClusterSpec.Id, List<Node>> nodesByCluster(List<Node> applicationNodes) {
diff --git a/node-repository/src/test/java/com/yahoo/vespa/hosted/provision/maintenance/ScalingSuggestionsMaintainerTest.java b/node-repository/src/test/java/com/yahoo/vespa/hosted/provision/maintenance/ScalingSuggestionsMaintainerTest.java
index 069759c43f2..5ec29cb9043 100644
--- a/node-repository/src/test/java/com/yahoo/vespa/hosted/provision/maintenance/ScalingSuggestionsMaintainerTest.java
+++ b/node-repository/src/test/java/com/yahoo/vespa/hosted/provision/maintenance/ScalingSuggestionsMaintainerTest.java
@@ -70,6 +70,21 @@ public class ScalingSuggestionsMaintainerTest {
                      tester.nodeRepository().applications().get(app1).get().cluster(cluster1.id()).get().suggestedResources().get().resources().toString());
         assertEquals("8 nodes with [vcpu: 14.7, memory: 4.0 Gb, disk 11.8 Gb, bandwidth: 0.1 Gbps, storage type: remote]",
                      tester.nodeRepository().applications().get(app2).get().cluster(cluster2.id()).get().suggestedResources().get().resources().toString());
+
+        // Utilization goes way down
+        tester.clock().advance(Duration.ofHours(13));
+        addMeasurements(0.10f, 0.10f, 0.10f, 0, 500, app1, tester.nodeRepository(), metricsDb);
+        maintainer.maintain();
+        assertEquals("Suggestion stays at the peak value observed",
+                     "14 nodes with [vcpu: 6.9, memory: 5.1 Gb, disk 15.0 Gb, bandwidth: 0.1 Gbps, storage type: remote]",
+                     tester.nodeRepository().applications().get(app1).get().cluster(cluster1.id()).get().suggestedResources().get().resources().toString());
+        // Utilization is still way down and a week has passed
+        tester.clock().advance(Duration.ofDays(7));
+        addMeasurements(0.10f, 0.10f, 0.10f, 0, 500, app1, tester.nodeRepository(), metricsDb);
+        maintainer.maintain();
+        assertEquals("Peak suggestion has been  outdated",
+                     "6 nodes with [vcpu: 2.0, memory: 4.0 Gb, disk 10.0 Gb, bandwidth: 0.1 Gbps, storage type: remote]",
+                     tester.nodeRepository().applications().get(app1).get().cluster(cluster1.id()).get().suggestedResources().get().resources().toString());
     }
 
     public void addMeasurements(float cpu, float memory, float disk, int generation, int count, ApplicationId applicationId,