From 3a1650a1d5b22d50a45b125d6fcd41965b3b98a2 Mon Sep 17 00:00:00 2001
From: Valeriy Khakhutskyy <1292899+valeriy42@users.noreply.github.com>
Date: Wed, 3 Sep 2025 14:45:05 +0200
Subject: [PATCH] [ML] Fix double-counting of inference memory in the
 assignment rebalancer (#133919)

The static method TrainedModelAssignmentRebalancer.getNodeFreeMemoryExcludingPerNodeOverheadAndNativeInference was used to subtract load.getAssignedNativeInferenceMemory() from load.getFreeMemoryExcludingPerNodeOverhead(). However, in NodeLoad.getFreeMemoryExcludingPerNodeOverhead(), native inference memory was already subtracted as part of the getAssignedJobMemoryExcludingPerNodeOverhead() calculation.

This led to double-counting of the native inference memory. Avoiding this double-counting allows us to remove the private method getNodeFreeMemoryExcludingPerNodeOverheadAndNativeInference() entirely.
---
 docs/changelog/133919.yaml                                | 5 +++++
 .../assignment/TrainedModelAssignmentRebalancer.java      | 8 +-------
 2 files changed, 6 insertions(+), 7 deletions(-)
 create mode 100644 docs/changelog/133919.yaml

diff --git a/docs/changelog/133919.yaml b/docs/changelog/133919.yaml
new file mode 100644
index 0000000000000..34c3ecd3ebe57
--- /dev/null
+++ b/docs/changelog/133919.yaml
@@ -0,0 +1,5 @@
+pr: 133919
+summary: Fix double-counting of inference memory in the assignment rebalancer
+area: Machine Learning
+type: bug
+issues: []
diff --git a/x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/assignment/TrainedModelAssignmentRebalancer.java b/x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/assignment/TrainedModelAssignmentRebalancer.java
index f523b4b086f35..90f86dbc243f0 100644
--- a/x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/assignment/TrainedModelAssignmentRebalancer.java
+++ b/x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/assignment/TrainedModelAssignmentRebalancer.java
@@ -298,9 +298,7 @@ private Map<List<String>, List<AssignmentPlan.Node>> createNodesByZoneMap() {
                         nodes.add(
                             new AssignmentPlan.Node(
                                 discoveryNode.getId(),
-                                // We subtract native inference memory as the planner expects available memory for
-                                // native inference including current assignments.
-                                getNodeFreeMemoryExcludingPerNodeOverheadAndNativeInference(load),
+                                load.getFreeMemoryExcludingPerNodeOverhead(),
                                 MlProcessors.get(discoveryNode, allocatedProcessorsScale).roundUp()
                             )
                         );
@@ -317,10 +315,6 @@ private Map<List<String>, List<AssignmentPlan.Node>> createNodesByZoneMap() {
         }));
     }
 
-    private static long getNodeFreeMemoryExcludingPerNodeOverheadAndNativeInference(NodeLoad load) {
-        return load.getFreeMemoryExcludingPerNodeOverhead() - load.getAssignedNativeInferenceMemory();
-    }
-
     private TrainedModelAssignmentMetadata.Builder buildAssignmentsFromPlan(AssignmentPlan assignmentPlan) {
         TrainedModelAssignmentMetadata.Builder builder = TrainedModelAssignmentMetadata.Builder.empty();
         for (AssignmentPlan.Deployment deployment : assignmentPlan.deployments()) {