More detailed autoscaling status

2 files changed, 34 insertions, 29 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 c4f11ee76d0..d2c943794fe 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
@@ -60,6 +60,7 @@ public class Autoscaler {
     }
 
     private Advice autoscale(Cluster cluster, List<Node> clusterNodes, Limits limits, boolean exclusive) {
+        ClusterSpec.Type clusterType = clusterNodes.get(0).allocation().get().membership().cluster().type();
         if (unstable(clusterNodes, nodeRepository))
             return Advice.none("Cluster change in progress");
 
@@ -68,13 +69,21 @@ public class Autoscaler {
 
         ClusterTimeseries clusterTimeseries = new ClusterTimeseries(cluster, clusterNodes, metricsDb, nodeRepository);
 
-        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())
-            return Advice.none("Collecting more data before making new scaling decisions");
+        int measurementsPerNode = clusterTimeseries.measurementsPerNode();
+        if  (measurementsPerNode < minimumMeasurementsPerNode(clusterType))
+            return Advice.none("Collecting more data before making new scaling decisions" +
+                               ": Has " + measurementsPerNode + " data points per node");
 
-        var target = ResourceTarget.idealLoad(cpuLoad.get(), memoryLoad.get(), diskLoad.get(), currentAllocation);
+        int nodesMeasured = clusterTimeseries.nodesMeasured();
+        if (nodesMeasured != clusterNodes.size())
+            return Advice.none("Collecting more data before making new scaling decisions" +
+                               ": Has measurements from " + nodesMeasured + " but need from " + clusterNodes.size());
+
+        double cpuLoad    = clusterTimeseries.averageLoad(Resource.cpu);
+        double memoryLoad = clusterTimeseries.averageLoad(Resource.memory);
+        double diskLoad   = clusterTimeseries.averageLoad(Resource.disk);
+
+        var target = ResourceTarget.idealLoad(cpuLoad, memoryLoad, diskLoad, currentAllocation);
 
         Optional<AllocatableClusterResources> bestAllocation =
                 allocationOptimizer.findBestAllocation(target, currentAllocation, limits, exclusive);
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 3c93e7ee7f6..e325e797ca5 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
@@ -10,8 +10,6 @@ import java.time.Instant;
 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;
 
 /**
@@ -21,10 +19,7 @@ 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;
 
     /** The measurements for all hosts in this snapshot */
     private final List<NodeTimeseries> nodeTimeseries;
@@ -32,9 +27,10 @@ public class ClusterTimeseries {
     public ClusterTimeseries(Cluster cluster, List<Node> clusterNodes, MetricsDb db, NodeRepository nodeRepository) {
         this.clusterNodes = clusterNodes;
         ClusterSpec.Type clusterType = clusterNodes.get(0).allocation().get().membership().cluster().type();
-        this.nodeTimeseries = db.getNodeTimeseries(nodeRepository.clock().instant().minus(Autoscaler.scalingWindow(clusterType)),
-                                                   clusterNodes.stream().map(Node::hostname).collect(Collectors.toSet()));
-        this.startTimePerNode = metricStartTimes(cluster, clusterNodes, nodeRepository);
+        var allTimeseries = db.getNodeTimeseries(nodeRepository.clock().instant().minus(Autoscaler.scalingWindow(clusterType)),
+                                                 clusterNodes.stream().map(Node::hostname).collect(Collectors.toSet()));
+        Map<String, Instant> startTimePerNode = metricStartTimes(cluster, clusterNodes, allTimeseries, nodeRepository);
+        nodeTimeseries = filterStale(allTimeseries, startTimePerNode);
     }
 
     /**
@@ -43,6 +39,7 @@ public class ClusterTimeseries {
      */
     private Map<String, Instant> metricStartTimes(Cluster cluster,
                                                   List<Node> clusterNodes,
+                                                  List<NodeTimeseries> nodeTimeseries,
                                                   NodeRepository nodeRepository) {
         Map<String, Instant> startTimePerHost = new HashMap<>();
         if ( ! cluster.scalingEvents().isEmpty()) {
@@ -65,23 +62,22 @@ public class ClusterTimeseries {
         return startTimePerHost;
     }
 
-    /**
-     * Returns the average load of this resource in the measurement window,
-     * or empty if we do not have a reliable measurement across the cluster nodes.
-     */
-    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);
+    /** Returns the average number of measurements per node */
+    public int measurementsPerNode() {
+        int measurementCount = nodeTimeseries.stream().mapToInt(m -> m.size()).sum();
+        return measurementCount / clusterNodes.size();
+    }
 
-        // 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();
+    /** Returns the number of nodes measured in this */
+    public int nodesMeasured() {
+        return nodeTimeseries.size();
+    }
 
-        double measurementSum = currentMeasurements.stream().flatMap(m -> m.asList().stream()).mapToDouble(m -> value(resource, m)).sum();
-        return Optional.of(measurementSum / measurementCount);
+    /** Returns the average load of this resource in this */
+    public double averageLoad(Resource resource) {
+        int measurementCount = nodeTimeseries.stream().mapToInt(m -> m.size()).sum();
+        double measurementSum = nodeTimeseries.stream().flatMap(m -> m.asList().stream()).mapToDouble(m -> value(resource, m)).sum();
+        return measurementSum / measurementCount;
     }
 
     private double value(Resource resource, MetricSnapshot snapshot) {