Merge branch 'main' into ccs-test-discovery

Author: richard-dennehy

docs/changelog/131990.yaml

@@ -0,0 +1,6 @@
+pr: 131990
+summary: Prevent the trained model deployment memory estimation from double-counting
+  allocations
+area: Machine Learning
+type: bug
+issues: []

x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/assignment/TrainedModelAssignmentRebalancer.java

@@ -140,9 +140,13 @@ private static void copyAssignments(
             for (Map.Entry<AssignmentPlan.Node, Integer> assignment : nodeAssignments.entrySet()) {
                 AssignmentPlan.Node originalNode = originalNodeById.get(assignment.getKey().id());
                 dest.assignModelToNode(m, originalNode, assignment.getValue());
-                // As the node has all its available memory we need to manually account memory of models with
-                // current allocations.
-                dest.accountMemory(m, originalNode);
+                if (m.currentAllocationsByNodeId().containsKey(originalNode.id())) {
+                    // TODO (#101612) requiredMemory should be calculated by the AssignmentPlan.Builder
+                    // As the node has all its available memory we need to manually account memory of models with
+                    // current allocations.
+                    long requiredMemory = m.estimateMemoryUsageBytes(m.currentAllocationsByNodeId().get(originalNode.id()));
+                    dest.accountMemory(m, originalNode, requiredMemory);
+                }
             }
         }
     }

x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/assignment/planning/AbstractPreserveAllocations.java

@@ -70,7 +70,7 @@ Deployment modifyModelPreservingPreviousAssignments(Deployment m) {
     AssignmentPlan mergePreservedAllocations(AssignmentPlan assignmentPlan) {
         // As the model/node objects the assignment plan are the modified ones,
         // they will not match the models/nodes members we have in this class.
-        // Therefore, we build a lookup table based on the ids, so we can merge the plan
+        // Therefore, we build a lookup table based on the ids so we can merge the plan
         // with its preserved allocations.
         final Map<Tuple<String, String>, Integer> plannedAssignmentsByDeploymentNodeIdPair = new HashMap<>();
         for (Deployment d : assignmentPlan.deployments()) {
@@ -85,6 +85,7 @@ AssignmentPlan mergePreservedAllocations(AssignmentPlan assignmentPlan) {
 
         AssignmentPlan.Builder mergedPlanBuilder = AssignmentPlan.builder(nodes, deployments);
         for (Node n : nodes) {
+            // TODO (#101612) Should the first loop happen in the builder constructor?
             for (Deployment deploymentAllocationsToPreserve : deployments) {
 
                 // if the model m is already allocated on the node n and I want to preserve this allocation

x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/assignment/planning/AssignmentPlan.java

@@ -424,7 +424,8 @@ public Builder assignModelToNode(Deployment deployment, Node node, int allocatio
             if (allocations <= 0) {
                 return this;
             }
-            if (requiredMemory > remainingNodeMemory.get(node)) {
+            if (/*isAlreadyAssigned(deployment, node) == false
+                &&*/ requiredMemory > remainingNodeMemory.get(node)) {
                 throw new IllegalArgumentException(
                     "not enough memory on node ["
                         + node.id()
@@ -470,14 +471,13 @@ private static int getCurrentAllocations(Deployment m, Node n) {
         }
 
         public void accountMemory(Deployment m, Node n) {
-            if (m.currentAllocationsByNodeId().containsKey(n.id())) {
-                int allocations = m.currentAllocationsByNodeId().get(n.id());
-                long requiredMemory = m.estimateMemoryUsageBytes(allocations);
-                accountMemory(m, n, requiredMemory);
-            }
+            // TODO (#101612) remove or refactor unused method
+            long requiredMemory = getDeploymentMemoryRequirement(m, n, getCurrentAllocations(m, n));
+            accountMemory(m, n, requiredMemory);
         }
 
-        private void accountMemory(Deployment m, Node n, long requiredMemory) {
+        public void accountMemory(Deployment m, Node n, long requiredMemory) {
+            // TODO (#101612) computation of required memory should be done internally
             remainingNodeMemory.computeIfPresent(n, (k, v) -> v - requiredMemory);
             if (remainingNodeMemory.containsKey(n) && remainingNodeMemory.get(n) < 0) {
                 throw new IllegalArgumentException("not enough memory on node [" + n.id() + "] to assign model [" + m.deploymentId() + "]");

x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/assignment/planning/RandomizedAssignmentRounding.java

@@ -310,6 +310,8 @@ private void unassignOversizedModels(Node n) {
         private AssignmentPlan toPlan() {
             AssignmentPlan.Builder builder = AssignmentPlan.builder(nodes, deployments);
             for (Map.Entry<Tuple<AssignmentPlan.Deployment, Node>, Integer> assignment : tryAssigningRemainingCores().entrySet()) {
+                // TODO (#101612) The model should be assigned to the node only when it is possible. This means, that canAssign should be
+                // integrated into the assignModelToNode.
                 if (builder.canAssign(assignment.getKey().v1(), assignment.getKey().v2(), assignment.getValue())) {
                     builder.assignModelToNode(assignment.getKey().v1(), assignment.getKey().v2(), assignment.getValue());
                 }

x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/assignment/planning/ZoneAwareAssignmentPlanner.java

@@ -200,9 +200,15 @@ private AssignmentPlan swapOriginalModelsInPlan(
             for (Map.Entry<Node, Integer> assignment : nodeAssignments.entrySet()) {
                 Node originalNode = originalNodeById.get(assignment.getKey().id());
                 planBuilder.assignModelToNode(originalDeployment, originalNode, assignment.getValue());
-                // As the node has all its available memory we need to manually account memory of models with
-                // current allocations.
-                planBuilder.accountMemory(originalDeployment, originalNode);
+                if (originalDeployment.currentAllocationsByNodeId().containsKey(originalNode.id())) {
+                    // TODO (#101612) requiredMemory should be calculated by the AssignmentPlan.Builder
+                    // As the node has all its available memory we need to manually account memory of models with
+                    // current allocations.
+                    long requiredMemory = originalDeployment.estimateMemoryUsageBytes(
+                        originalDeployment.currentAllocationsByNodeId().get(originalNode.id())
+                    );
+                    planBuilder.accountMemory(m, originalNode, requiredMemory);
+                }
             }
         }
         return planBuilder.build();