[Ml] Validate tree feature index is within range (elastic#52460)

This changes the tree validation code to ensure no node in the tree has a
feature index that is beyond the bounds of the feature_names array.
Specifically this handles the situation where the C++ emits a tree containing
a single node and an empty feature_names list. This is valid tree used to
centre the data in the ensemble but the validation code would reject this
as feature_names is empty. This meant a broken workflow as you cannot GET
the model and PUT it back

Author: davidkyle

client/rest-high-level/src/test/java/org/elasticsearch/client/ml/inference/trainedmodel/tree/TreeTests.java

@@ -61,11 +61,11 @@ public static Tree createRandom() {
     }
 
     public static Tree buildRandomTree(List<String> featureNames, int depth, TargetType targetType) {
-        int numFeatures = featureNames.size();
+        int maxFeatureIndex = featureNames.size() -1;
         Tree.Builder builder = Tree.builder();
         builder.setFeatureNames(featureNames);
 
-        TreeNode.Builder node = builder.addJunction(0, randomInt(numFeatures), true, randomDouble());
+        TreeNode.Builder node = builder.addJunction(0, randomInt(maxFeatureIndex), true, randomDouble());
         List<Integer> childNodes = Arrays.asList(node.getLeftChild(), node.getRightChild());
 
         for (int i = 0; i < depth -1; i++) {
@@ -76,7 +76,7 @@ public static Tree buildRandomTree(List<String> featureNames, int depth, TargetT
                     builder.addLeaf(nodeId, randomDouble());
                 } else {
                     TreeNode.Builder childNode =
-                        builder.addJunction(nodeId, randomInt(numFeatures), true, randomDouble());
+                        builder.addJunction(nodeId, randomInt(maxFeatureIndex), true, randomDouble());
                     nextNodes.add(childNode.getLeftChild());
                     nextNodes.add(childNode.getRightChild());
                 }

x-pack/plugin/core/src/main/java/org/elasticsearch/xpack/core/ml/inference/trainedmodel/tree/Tree.java

@@ -253,8 +253,10 @@ public static Builder builder() {
 
     @Override
     public void validate() {
-        if (featureNames.isEmpty()) {
-            throw ExceptionsHelper.badRequestException("[{}] must not be empty for tree model", FEATURE_NAMES.getPreferredName());
+        int maxFeatureIndex = maxFeatureIndex();
+        if (maxFeatureIndex >= featureNames.size()) {
+            throw ExceptionsHelper.badRequestException("feature index [{}] is out of bounds for the [{}] array",
+                    maxFeatureIndex, FEATURE_NAMES.getPreferredName());
         }
         checkTargetType();
         detectMissingNodes();
@@ -267,6 +269,23 @@ public long estimatedNumOperations() {
         return (long)Math.ceil(Math.log(nodes.size())) + featureNames.size();
     }
 
+    /**
+     * The highest index of a feature used any of the nodes.
+     * If no nodes use a feature return -1. This can only happen
+     * if the tree contains a single leaf node.
+     *
+     * @return The max or -1
+     */
+    int maxFeatureIndex() {
+        int maxFeatureIndex = -1;
+
+        for (TreeNode node : nodes) {
+            maxFeatureIndex = Math.max(maxFeatureIndex, node.getSplitFeature());
+        }
+
+        return maxFeatureIndex;
+    }
+
     private void checkTargetType() {
         if (this.classificationLabels != null && this.targetType != TargetType.CLASSIFICATION) {
             throw ExceptionsHelper.badRequestException(

x-pack/plugin/core/src/test/java/org/elasticsearch/xpack/core/ml/inference/trainedmodel/tree/TreeTests.java

@@ -29,6 +29,7 @@
 import java.util.stream.IntStream;
 
 import static org.hamcrest.Matchers.closeTo;
+import static org.hamcrest.Matchers.containsString;
 import static org.hamcrest.Matchers.equalTo;
 
 
@@ -72,10 +73,10 @@ public static Tree createRandom() {
 
     public static Tree buildRandomTree(List<String> featureNames, int depth) {
         Tree.Builder builder = Tree.builder();
-        int numFeatures = featureNames.size() - 1;
+        int maxFeatureIndex = featureNames.size() - 1;
         builder.setFeatureNames(featureNames);
 
-        TreeNode.Builder node = builder.addJunction(0, randomInt(numFeatures), true, randomDouble());
+        TreeNode.Builder node = builder.addJunction(0, randomInt(maxFeatureIndex), true, randomDouble());
         List<Integer> childNodes = Arrays.asList(node.getLeftChild(), node.getRightChild());
 
         for (int i = 0; i < depth -1; i++) {
@@ -86,7 +87,7 @@ public static Tree buildRandomTree(List<String> featureNames, int depth) {
                     builder.addLeaf(nodeId, randomDouble());
                 } else {
                     TreeNode.Builder childNode =
-                        builder.addJunction(nodeId, randomInt(numFeatures), true, randomDouble());
+                        builder.addJunction(nodeId, randomInt(maxFeatureIndex), true, randomDouble());
                     nextNodes.add(childNode.getLeftChild());
                     nextNodes.add(childNode.getRightChild());
                 }
@@ -339,26 +340,83 @@ public void testTreeWithTargetTypeAndLabelsMismatch() {
         assertThat(ex.getMessage(), equalTo(msg));
     }
 
-    public void testTreeWithEmptyFeatureNames() {
-        String msg = "[feature_names] must not be empty for tree model";
-        ElasticsearchException ex = expectThrows(ElasticsearchException.class, () -> {
-            Tree.builder()
-                .setRoot(TreeNode.builder(0)
-                    .setLeftChild(1)
-                    .setSplitFeature(1)
-                    .setThreshold(randomDouble()))
-                .setFeatureNames(Collections.emptyList())
-                .build()
-                .validate();
-        });
-        assertThat(ex.getMessage(), equalTo(msg));
-    }
-
     public void testOperationsEstimations() {
         Tree tree = buildRandomTree(Arrays.asList("foo", "bar", "baz"), 5);
         assertThat(tree.estimatedNumOperations(), equalTo(7L));
     }
 
+    public void testMaxFeatureIndex() {
+
+        int numFeatures = randomIntBetween(1, 15);
+        // We need a tree where every feature is used, choose a depth big enough to
+        // accommodate those non-leave nodes (leaf nodes don't have a feature index)
+        int depth = (int) Math.ceil(Math.log(numFeatures +1) / Math.log(2)) + 1;
+        List<String> featureNames = new ArrayList<>(numFeatures);
+        for (int i=0; i<numFeatures; i++) {
+            featureNames.add("feature" + i);
+        }
+
+        Tree.Builder builder = Tree.builder().setFeatureNames(featureNames);
+
+        // build a tree using feature indices 0..numFeatures -1
+        int featureIndex = 0;
+        TreeNode.Builder node = builder.addJunction(0, featureIndex++, true, randomDouble());
+        List<Integer> childNodes = Arrays.asList(node.getLeftChild(), node.getRightChild());
+
+        for (int i = 0; i < depth -1; i++) {
+            List<Integer> nextNodes = new ArrayList<>();
+            for (int nodeId : childNodes) {
+                if (i == depth -2) {
+                    builder.addLeaf(nodeId, randomDouble());
+                } else {
+                    TreeNode.Builder childNode =
+                            builder.addJunction(nodeId, featureIndex++ % numFeatures, true, randomDouble());
+                    nextNodes.add(childNode.getLeftChild());
+                    nextNodes.add(childNode.getRightChild());
+                }
+            }
+            childNodes = nextNodes;
+        }
+
+        Tree tree = builder.build();
+
+        assertEquals(numFeatures, tree.maxFeatureIndex() +1);
+    }
+
+    public void testMaxFeatureIndexSingleNodeTree() {
+        Tree tree = Tree.builder()
+                .setRoot(TreeNode.builder(0).setLeafValue(10.0))
+                .setFeatureNames(Collections.emptyList())
+                .build();
+
+        assertEquals(-1, tree.maxFeatureIndex());
+    }
+
+    public void testValidateGivenMissingFeatures() {
+        List<String> featureNames = Arrays.asList("foo", "bar", "baz");
+
+        // build a tree referencing a feature at index 3 which is not in the featureNames list
+        Tree.Builder builder = Tree.builder().setFeatureNames(featureNames);
+        builder.addJunction(0, 0, true, randomDouble());
+        builder.addJunction(1, 1, true, randomDouble());
+        builder.addJunction(2, 3, true, randomDouble());
+        builder.addLeaf(3, randomDouble());
+        builder.addLeaf(4, randomDouble());
+        builder.addLeaf(5, randomDouble());
+        builder.addLeaf(6, randomDouble());
+
+        ElasticsearchStatusException e = expectThrows(ElasticsearchStatusException.class, () -> builder.build().validate());
+        assertThat(e.getDetailedMessage(), containsString("feature index [3] is out of bounds for the [feature_names] array"));
+    }
+
+    public void testValidateGivenTreeWithNoFeatures() {
+        Tree.builder()
+                .setRoot(TreeNode.builder(0).setLeafValue(10.0))
+                .setFeatureNames(Collections.emptyList())
+                .build()
+                .validate();
+    }
+
     private static Map<String, Object> zipObjMap(List<String> keys, List<? extends Object> values) {
         return IntStream.range(0, keys.size()).boxed().collect(Collectors.toMap(keys::get, values::get));
     }

x-pack/plugin/ml/qa/ml-with-security/build.gradle

@@ -137,7 +137,7 @@ integTest.runner {
     'ml/inference_crud/Test get given missing trained model',
     'ml/inference_crud/Test get given expression without matches and allow_no_match is false',
     'ml/inference_crud/Test put ensemble with empty models',
-    'ml/inference_crud/Test put ensemble with tree where tree has empty feature-names',
+    'ml/inference_crud/Test put ensemble with tree where tree has out of bounds feature_names index',
     'ml/inference_crud/Test put model with empty input.field_names',
     'ml/inference_stats_crud/Test get stats given missing trained model',
     'ml/inference_stats_crud/Test get stats given expression without matches and allow_no_match is false',

x-pack/plugin/src/test/resources/rest-api-spec/test/ml/inference_crud.yml

@@ -333,6 +333,40 @@ setup:
   - match: { count: 1 }
   - match: { trained_model_configs.0.model_id: "lang_ident_model_1" }
 ---
+"Test put ensemble with single node and empty feature_names":
+
+  - do:
+      ml.put_trained_model:
+        model_id: "ensemble_tree_empty_feature_names"
+        body: >
+          {
+            "input": {
+              "field_names": "fieldy_mc_fieldname"
+            },
+            "definition": {
+              "trained_model": {
+                "ensemble": {
+                  "feature_names": [],
+                  "trained_models": [
+                     {
+                        "tree": {
+                           "feature_names": [],
+                           "tree_structure": [
+                             {
+                               "node_index": 0,
+                               "decision_type": "lte",
+                               "leaf_value": 12.0,
+                               "default_left": true
+                             }]
+                        }
+                     }
+                  ]
+                }
+              }
+            }
+          }
+
+---
 "Test put ensemble with empty models":
   - do:
       catch: /\[trained_models\] must not be empty/
@@ -353,11 +387,11 @@ setup:
             }
           }
 ---
-"Test put ensemble with tree where tree has empty feature-names":
+"Test put ensemble with tree where tree has out of bounds feature_names index":
   - do:
-      catch: /\[feature_names\] must not be empty/
+      catch: /feature index \[1\] is out of bounds for the \[feature_names\] array/
       ml.put_trained_model:
-        model_id: "ensemble_tree_missing_feature_names"
+        model_id: "ensemble_tree_out_of_bounds_feature_names_index"
         body: >
           {
             "input": {
@@ -374,7 +408,7 @@ setup:
                            "tree_structure": [
                              {
                                "node_index": 0,
-                               "split_feature": 0,
+                               "split_feature": 1,
                                "split_gain": 12.0,
                                "threshold": 10.0,
                                "decision_type": "lte",