Log more details

2 files changed, 21 insertions, 11 deletions

diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/Autoscaler.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/Autoscaler.java
index 6e7be8f5672..b7729577bda 100644
--- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/Autoscaler.java
+++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/Autoscaler.java
@@ -20,7 +20,7 @@ import java.util.logging.Logger;
  */
 public class Autoscaler {
 
-    protected final Logger log = Logger.getLogger(this.getClass().getName());
+    private final Logger log = Logger.getLogger(this.getClass().getName());
 
     /** What cost difference factor is worth a reallocation? */
     private static final double costDifferenceWorthReallocation = 0.1;
@@ -71,11 +71,10 @@ public class Autoscaler {
 
         ClusterTimeseries clusterTimeseries = new ClusterTimeseries(cluster, clusterNodes, metricsDb, nodeRepository);
 
-        Optional<Double> cpuLoad    = clusterTimeseries.averageLoad(Resource.cpu);
-        Optional<Double> memoryLoad = clusterTimeseries.averageLoad(Resource.memory);
-        Optional<Double> diskLoad   = clusterTimeseries.averageLoad(Resource.disk);
+        Optional<Double> cpuLoad    = clusterTimeseries.averageLoad(Resource.cpu, cluster);
+        Optional<Double> memoryLoad = clusterTimeseries.averageLoad(Resource.memory, cluster);
+        Optional<Double> diskLoad   = clusterTimeseries.averageLoad(Resource.disk, cluster);
         if (cpuLoad.isEmpty() || memoryLoad.isEmpty() || diskLoad.isEmpty()) {
-            log.fine(() -> "Missing average load - Advice.none  " + cluster.toString());
             return Advice.none();
         }
         var target = ResourceTarget.idealLoad(cpuLoad.get(), memoryLoad.get(), diskLoad.get(), currentAllocation);
@@ -83,11 +82,11 @@ public class Autoscaler {
         Optional<AllocatableClusterResources> bestAllocation =
                 allocationOptimizer.findBestAllocation(target, currentAllocation, limits, exclusive);
         if (bestAllocation.isEmpty()) {
-            log.fine(() -> "bestAllocation.isEmpty - Advice.dontScale " + cluster.toString());
+            log.fine(() -> "bestAllocation.isEmpty: Advice.dontScale for " + cluster.toString());
             return Advice.dontScale();
         }
         if (similar(bestAllocation.get(), currentAllocation)) {
-            log.fine(() -> "Current allocation similar - Advice.dontScale " + cluster.toString());
+            log.fine(() -> "Current allocation similar: Advice.dontScale for " + cluster.toString());
             return Advice.dontScale();
         }
         return Advice.scaleTo(bestAllocation.get().toAdvertisedClusterResources());
diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/ClusterTimeseries.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/ClusterTimeseries.java
index 2123ecd0224..bb91b77dce5 100644
--- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/ClusterTimeseries.java
+++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/ClusterTimeseries.java
@@ -11,6 +11,7 @@ import java.util.HashMap;
 import java.util.List;
 import java.util.Map;
 import java.util.Optional;
+import java.util.logging.Logger;
 import java.util.stream.Collectors;
 
 /**
@@ -20,6 +21,8 @@ import java.util.stream.Collectors;
  */
 public class ClusterTimeseries {
 
+    private static final Logger log = Logger.getLogger(ClusterTimeseries.class.getName());
+
     private final List<Node> clusterNodes;
     private final Map<String, Instant> startTimePerNode;
 
@@ -64,9 +67,9 @@ public class ClusterTimeseries {
 
     /**
      * Returns the average load of this resource in the measurement window,
-     * or empty if we are not in a position to make decisions from these measurements at this time.
+     * or empty if we do not have a reliable measurement across the cluster nodes.
      */
-    public Optional<Double> averageLoad(Resource resource) {
+    public Optional<Double> averageLoad(Resource resource, Cluster cluster) {
         ClusterSpec.Type clusterType = clusterNodes.get(0).allocation().get().membership().cluster().type();
 
         List<NodeTimeseries> currentMeasurements = filterStale(nodeTimeseries, startTimePerNode);
@@ -74,8 +77,16 @@ public class ClusterTimeseries {
         // Require a total number of measurements scaling with the number of nodes,
         // but don't require that we have at least that many from every node
         int measurementCount = currentMeasurements.stream().mapToInt(m -> m.size()).sum();
-        if (measurementCount / clusterNodes.size() < Autoscaler.minimumMeasurementsPerNode(clusterType)) return Optional.empty();
-        if (currentMeasurements.size() != clusterNodes.size()) return Optional.empty();
+        if (measurementCount / clusterNodes.size() < Autoscaler.minimumMeasurementsPerNode(clusterType)) {
+            log.fine(() -> "Too few measurements per node for " + cluster.toString() + ": measurementCount " + measurementCount +
+                           " (" + nodeTimeseries.stream().mapToInt(m -> m.size()).sum() + " before filtering");
+            return Optional.empty();
+        }
+        if (currentMeasurements.size() != clusterNodes.size()) {
+            log.fine(() -> "Mssing measurements from some nodes for " + cluster.toString() + ": Has from " + currentMeasurements.size() +
+                           "but need " + clusterNodes.size() + "(before filtering: " +  nodeTimeseries.size() + ")");
+            return Optional.empty();
+        }
 
         double measurementSum = currentMeasurements.stream().flatMap(m -> m.asList().stream()).mapToDouble(m -> value(resource, m)).sum();
         return Optional.of(measurementSum / measurementCount);