[BugFix] Restore dedup pushdown when combined with WHERE clause (#5482) (#5488)

ryan-gh-bot · RyanL1997 · web-flow · commit e5de8f6a77a3 · 2026-06-08T09:55:24.000-07:00
* [BugFix] Restore dedup pushdown when combined with WHERE clause (#5482) Run PPLSimplifyDedupRule before FilterMergeRule in the HEP optimizer so the bucket-non-null filter PPL emits for dedup is matched as-is. With the previous order, an upstream user where filter sat adjacent to the bucket-non-null filter; FilterMergeRule fired first and merged them into a conjunction that no longer satisfied PPLSimplifyDedupRule's operand predicate, defeating dedup pushdown to the shard. Use sequential addRuleInstance phases for explicit ordering rather than addRuleCollection, which is documented as non-deterministic in firing order. Adds two regression tests in CalcitePPLDedupTest: one that asserts LogicalDedup is produced under the fixed order, and one that pins the buggy behavior under the swapped order. Signed-off-by: ryan-gh-bot <ryan-gh-bot@users.noreply.github.com> * [BugFix] Drop issue-link reference from regression-test JavaDoc (#5488) Per maintainer review feedback, the regression-test JavaDoc for testDedupAfterWhereProducesLogicalDedup mentioned the originating issue URL. The remaining JavaDoc paragraphs already describe the bug shape and the rule-ordering invariant, so the explicit issue link is unnecessary noise. Signed-off-by: ryan-gh-bot <ryan-gh-bot@users.noreply.github.com> * [BugFix] Make dedup simplify operand order-independent (#5488) Address review feedback on #5488: extend mayBeFilterFromBucketNonNull to accept the merged conjunction shape FilterMergeRule produces, so PPLSimplifyDedupRule fires regardless of whether FilterMergeRule has already merged the user where clause into the bucket-non-null filter. PPLSimplifyDedupRule.apply now splits the bottom filter into IS_NOT_NULL conjuncts on partition keys (absorbed into LogicalDedup semantics) and any remaining conjuncts (preserved as a separate filter below the new LogicalDedup), so a user predicate that was folded in is no longer dropped. With the operand predicate order-independent, the HEP rule order is no longer a load-bearing invariant. Revert the addRuleCollection -> addRuleInstance change in CalciteToolsHelper.HEP_PROGRAM that the previous patch introduced. Replace the regression test that pinned the buggy rule order with one that asserts the user-visible contract: with where preceding dedup, a LogicalDedup is produced and the user predicate is preserved regardless of which order FilterMergeRule and PPLSimplifyDedupRule fire. Signed-off-by: ryan-gh-bot <ryan-gh-bot@users.noreply.github.com> * Address review comments on #5488 Per @penghuo review: PlanUtils.java - Revert mayBeFilterFromBucketNonNull to the original ternary form; drop the early-return refactor (no behavior change, just cleaner). - Drop the !rexCall.getOperands().isEmpty() guard before .get(0): IS NOT NULL is always unary in Calcite, so the check is dead. - Trim the JavaDoc to the essentials (un-merged vs merged-AND shape; concrete partition-key match happens in PPLSimplifyDedupRule#apply). - Promote isNotNullOnRef from package-private to public so the dedup rule can reuse it from a different package. PPLSimplifyDedupRule.java - isNotNullOnPartitionKey now delegates the IS NOT NULL($ref) structural check to PlanUtils.isNotNullOnRef and adds the partition-key index check on top. CalciteExplainIT.java - Add testDedupAfterWherePushDown: an end-to-end regression that runs the shape `... | where <pred> | dedup <field>` and asserts (a) LogicalDedup appears in the explain output (PPLSimplifyDedupRule fired even after FilterMergeRule had a chance to merge the two filters), and (b) EnumerableWindow does NOT appear (the in-memory ROW_NUMBER fallback the bug caused is gone). Signed-off-by: Jialiang Liang <jiallian@amazon.com> * Push user where filter into scan when blocking dedup pushdown PPLSimplifyDedupRule correctly produces Dedup -> Filter(user where) -> Scan when a `where` precedes `dedup`. The Filter between Dedup and Scan blocks DedupPushdownRule's strict Dedup -> Project -> Scan operand chain, so Volcano falls back to PPLDedupConvertRule and the plan ends up with an in-memory ROW_NUMBER window instead of the pushed-down composite + top_hits aggregation. Add a WITH_FILTER operand variant to DedupPushdownRule that matches Dedup -> Filter -> Scan, pushes the filter into the scan, then runs the standard apply() on the resulting Dedup -> Project -> Scan shape. Signed-off-by: Jialiang Liang <jiallian@amazon.com> * Bail when filter is only partially pushable pushDownFilter returns a Filter (not a CalciteLogicalIndexScan) when the predicate analyzer can only partially push the condition. The previous cast would have thrown ClassCastException in that case. Use an instanceof-pattern check so the rule bails out cleanly and leaves the plan untouched, letting other rules handle the residual. Also drop a stale issue-link reference from a test comment. Signed-off-by: Jialiang Liang <jiallian@amazon.com> * Apply spotless formatting to dedup unit test Reflowed the JavaDoc on testDedupAfterWhereProducesLogicalDedupWithProductionHepProgram to match Google Java Format's preferred line break, fixing the spotlessJavaCheck violation that failed the unit-test matrix on CI. Signed-off-by: Jialiang Liang <jiallian@amazon.com> --------- Signed-off-by: ryan-gh-bot <ryan-gh-bot@users.noreply.github.com> Signed-off-by: Jialiang Liang <jiallian@amazon.com> Co-authored-by: ryan-gh-bot <ryan-gh-bot@users.noreply.github.com> Co-authored-by: Jialiang Liang <jiallian@amazon.com>
diff --git a/core/src/main/java/org/opensearch/sql/calcite/plan/rule/PPLSimplifyDedupRule.java b/core/src/main/java/org/opensearch/sql/calcite/plan/rule/PPLSimplifyDedupRule.java
@@ -6,11 +6,14 @@
 package org.opensearch.sql.calcite.plan.rule;
 
 import java.util.ArrayList;
+import java.util.HashSet;
 import java.util.List;
+import java.util.Set;
 import java.util.function.Predicate;
 import java.util.stream.Collectors;
 import javax.annotation.Nullable;
 import org.apache.calcite.plan.RelOptRuleCall;
+import org.apache.calcite.plan.RelOptUtil;
 import org.apache.calcite.plan.RelRule;
 import org.apache.calcite.rel.RelCollation;
 import org.apache.calcite.rel.RelCollations;
@@ -22,6 +25,7 @@
 import org.apache.calcite.rex.RexInputRef;
 import org.apache.calcite.rex.RexLiteral;
 import org.apache.calcite.rex.RexNode;
+import org.apache.calcite.rex.RexUtil;
 import org.apache.calcite.rex.RexWindow;
 import org.apache.calcite.sql.SqlKind;
 import org.apache.calcite.sql.type.SqlTypeName;
@@ -115,8 +119,41 @@ protected void apply(
 
     RelCollation inputCollation = extractCollationFromWindow(windows.get(0));
 
+    // Split the bucket-non-null filter into two parts:
+    //   1) IS_NOT_NULL conjuncts on a partition key — these are the bucket-non-null guards PPL
+    //      emits as part of the dedup pattern; LogicalDedup absorbs their semantics.
+    //   2) Everything else — for example, a user `where` predicate that FilterMergeRule may
+    //      have folded into the same conjunction, or a user IS_NOT_NULL filter on a non-partition
+    //      column. These must be preserved as a separate filter below the new LogicalDedup so
+    //      user-visible behavior is unchanged regardless of whether FilterMergeRule fired.
+    Set<Integer> partitionKeyIndices = new HashSet<>();
+    for (RexNode key : dedupColumns) {
+      if (key instanceof RexInputRef ref) {
+        partitionKeyIndices.add(ref.getIndex());
+      }
+    }
+    List<RexNode> bucketNonNullConjuncts = new ArrayList<>();
+    List<RexNode> remainingConjuncts = new ArrayList<>();
+    for (RexNode conjunct : RelOptUtil.conjunctions(bucketNonNullFilter.getCondition())) {
+      if (isNotNullOnPartitionKey(conjunct, partitionKeyIndices)) {
+        bucketNonNullConjuncts.add(conjunct);
+      } else {
+        remainingConjuncts.add(conjunct);
+      }
+    }
+    // Defensive: if no IS_NOT_NULL conjunct on a partition key is present, this filter is not
+    // actually a bucket-non-null filter — bail out without transforming. The loose operand
+    // predicate may have matched on an unrelated AND that happens to contain an IS_NOT_NULL on
+    // some other ref.
+    if (bucketNonNullConjuncts.isEmpty()) {
+      return;
+    }
+
     RelBuilder relBuilder = call.builder();
     relBuilder.push(bucketNonNullFilter.getInput());
+    if (!remainingConjuncts.isEmpty()) {
+      relBuilder.filter(RexUtil.composeConjunction(relBuilder.getRexBuilder(), remainingConjuncts));
+    }
     List<Pair<RexNode, String>> targetProjections =
         projectWithWindow.getNamedProjects().stream()
             .filter(p -> !p.getKey().isA(SqlKind.ROW_NUMBER))
@@ -134,6 +171,14 @@ protected void apply(
     call.transformTo(relBuilder.build());
   }
 
+  private static boolean isNotNullOnPartitionKey(RexNode rex, Set<Integer> partitionKeyIndices) {
+    if (!PlanUtils.isNotNullOnRef(rex)) {
+      return false;
+    }
+    RexInputRef ref = (RexInputRef) ((RexCall) rex).getOperands().get(0);
+    return partitionKeyIndices.contains(ref.getIndex());
+  }
+
   private static @Nullable RelCollation extractCollationFromWindow(RexWindow window) {
     if (window.orderKeys.isEmpty()) {
       return null;
diff --git a/core/src/main/java/org/opensearch/sql/calcite/utils/PlanUtils.java b/core/src/main/java/org/opensearch/sql/calcite/utils/PlanUtils.java
@@ -788,15 +788,21 @@ static Mapping mapping(List<RexNode> rexNodes, RelDataType schema) {
     return Mappings.target(getSelectColumns(rexNodes), schema.getFieldCount());
   }
 
+  /**
+   * Accepts the un-merged {@code IS NOT NULL($ref)} shape and the merged-{@code AND} shape that
+   * {@link org.apache.calcite.rel.rules.FilterMergeRule} produces when a user {@code where}
+   * precedes the dedup. The concrete partition-key match — and the split-out of any user predicate
+   * folded into the AND — happens in {@link PPLSimplifyDedupRule#apply}.
+   */
   static boolean mayBeFilterFromBucketNonNull(LogicalFilter filter) {
     RexNode condition = filter.getCondition();
     return isNotNullOnRef(condition)
         || (condition instanceof RexCall rexCall
             && rexCall.getOperator().equals(SqlStdOperatorTable.AND)
-            && rexCall.getOperands().stream().allMatch(PlanUtils::isNotNullOnRef));
+            && rexCall.getOperands().stream().anyMatch(PlanUtils::isNotNullOnRef));
   }
 
-  private static boolean isNotNullOnRef(RexNode rex) {
+  public static boolean isNotNullOnRef(RexNode rex) {
     return rex instanceof RexCall rexCall
         && rexCall.isA(SqlKind.IS_NOT_NULL)
         && rexCall.getOperands().get(0) instanceof RexInputRef;
diff --git a/integ-test/src/test/java/org/opensearch/sql/calcite/remote/CalciteExplainIT.java b/integ-test/src/test/java/org/opensearch/sql/calcite/remote/CalciteExplainIT.java
@@ -2153,6 +2153,29 @@ public void testTransposeExplain() throws IOException {
                 + "|  transpose 4 column_name='column_names'"));
   }
 
+  /**
+   * With a user {@code where} clause preceding {@code dedup}, the physical plan must push both the
+   * filter and the dedup-as-aggregation into the OpenSearch scan, not fall back to an in-memory
+   * {@code ROW_NUMBER} window above a row-fetching scan.
+   */
+  @Test
+  public void testDedupAfterWherePushDown() throws IOException {
+    enabledOnlyWhenPushdownIsEnabled();
+    String result =
+        explainQueryToString(
+            "source=opensearch-sql_test_index_account | where age > 25 | dedup gender");
+    assertTrue(
+        "Expected user where filter pushed down to the scan:\n" + result,
+        result.contains("FILTER->>($8, 25)"));
+    assertTrue(
+        "Expected dedup pushed down as AGGREGATION (composite + top_hits):\n" + result,
+        result.contains("AGGREGATION->"));
+    assertFalse(
+        "Unexpected EnumerableWindow — dedup fell back to the in-memory ROW_NUMBER form:\n"
+            + result,
+        result.contains("EnumerableWindow"));
+  }
+
   public void testComplexDedup() throws IOException {
     enabledOnlyWhenPushdownIsEnabled();
     String expected = loadExpectedPlan("explain_dedup_complex1.yaml");
diff --git a/opensearch/src/main/java/org/opensearch/sql/opensearch/planner/rules/DedupPushdownRule.java b/opensearch/src/main/java/org/opensearch/sql/opensearch/planner/rules/DedupPushdownRule.java
@@ -17,6 +17,7 @@
 import org.apache.calcite.rel.RelCollations;
 import org.apache.calcite.rel.RelFieldCollation;
 import org.apache.calcite.rel.logical.LogicalAggregate;
+import org.apache.calcite.rel.logical.LogicalFilter;
 import org.apache.calcite.rel.logical.LogicalProject;
 import org.apache.calcite.rel.rules.SubstitutionRule;
 import org.apache.calcite.rex.RexInputRef;
@@ -47,6 +48,24 @@ protected DedupPushdownRule(Config config) {
   @Override
   protected void onMatchImpl(RelOptRuleCall call) {
     final LogicalDedup logicalDedup = call.rel(0);
+    if (call.rels[1] instanceof LogicalFilter filter) {
+      // Push the filter into the scan, then synthesize an identity project so the standard
+      // apply() can run on the resulting Dedup → Project → Scan shape. If the filter is only
+      // partially pushable, pushDownFilter returns a Filter wrapping the residual condition over
+      // the new scan; we can't strip a residual filter without breaking semantics, so bail.
+      final CalciteLogicalIndexScan scan = call.rel(2);
+      if (!(scan.pushDownFilter(filter) instanceof CalciteLogicalIndexScan newScan)) {
+        return;
+      }
+      RelBuilder relBuilder = call.builder();
+      relBuilder.push(newScan);
+      // force=true so the identity project is materialized (apply() requires a LogicalProject)
+      relBuilder.project(
+          relBuilder.fields(), newScan.getRowType().getFieldNames(), /* force= */ true);
+      LogicalProject identityProject = (LogicalProject) relBuilder.build();
+      apply(call, logicalDedup, identityProject, newScan);
+      return;
+    }
     final LogicalProject projectWithExpr = call.rel(1);
     final CalciteLogicalIndexScan scan = call.rel(2);
     apply(call, logicalDedup, projectWithExpr, scan);
@@ -226,6 +245,27 @@ public interface Config extends OpenSearchRuleConfig {
                                                 .predicate(Config::tableScanChecker)
                                                 .noInputs())));
 
+    // +- LogicalDedup
+    //    +- LogicalFilter      (e.g. a `where` predicate preserved below dedup by
+    //    +- CalciteLogicalIndexScan   PPLSimplifyDedupRule when an upstream `where` is folded
+    //                                 into the bucket-non-null filter)
+    Config WITH_FILTER =
+        ImmutableDedupPushdownRule.Config.builder()
+            .build()
+            .withDescription("Dedup-to-Aggregate-WithFilter")
+            .withOperandSupplier(
+                b0 ->
+                    b0.operand(LogicalDedup.class)
+                        .predicate(dedup -> !dedup.getKeepEmpty())
+                        .oneInput(
+                            b1 ->
+                                b1.operand(LogicalFilter.class)
+                                    .oneInput(
+                                        b2 ->
+                                            b2.operand(CalciteLogicalIndexScan.class)
+                                                .predicate(Config::tableScanChecker)
+                                                .noInputs())));
+
     /**
      * Project must be not pushed since the name of expression would lose after project pushed. E.g.
      * in query "eval new_a = a + 1 | dedup b", the "new_a" will lose.
diff --git a/opensearch/src/main/java/org/opensearch/sql/opensearch/planner/rules/OpenSearchIndexRules.java b/opensearch/src/main/java/org/opensearch/sql/opensearch/planner/rules/OpenSearchIndexRules.java
@@ -50,6 +50,8 @@ public class OpenSearchIndexRules {
       SortIndexScanRule.Config.DEFAULT.toRule();
   private static final DedupPushdownRule DEDUP_PUSH_DOWN =
       DedupPushdownRule.Config.DEFAULT.toRule();
+  private static final DedupPushdownRule DEDUP_PUSH_DOWN_WITH_FILTER =
+      DedupPushdownRule.Config.WITH_FILTER.toRule();
   private static final SortProjectExprTransposeRule SORT_PROJECT_EXPR_TRANSPOSE =
       SortProjectExprTransposeRule.Config.DEFAULT.toRule();
   private static final ExpandCollationOnProjectExprRule EXPAND_COLLATION_ON_PROJECT_EXPR =
@@ -75,6 +77,7 @@ public class OpenSearchIndexRules {
           LIMIT_INDEX_SCAN,
           SORT_INDEX_SCAN,
           DEDUP_PUSH_DOWN,
+          DEDUP_PUSH_DOWN_WITH_FILTER,
           SORT_PROJECT_EXPR_TRANSPOSE,
           SORT_AGGREGATION_METRICS_RULE,
           RARE_TOP_PUSH_DOWN,
diff --git a/ppl/src/test/java/org/opensearch/sql/ppl/calcite/CalcitePPLAbstractTest.java b/ppl/src/test/java/org/opensearch/sql/ppl/calcite/CalcitePPLAbstractTest.java
@@ -110,6 +110,20 @@ public RelNode getRelNode(String ppl) {
     return root;
   }
 
+  /**
+   * Get the root RelNode of the given PPL query without merging adjacent filters. Useful for
+   * regression tests that need to exercise rule ordering against the un-merged shape that PPL
+   * actually emits to the production HEP optimizer.
+   */
+  public RelNode getRelNodeRaw(String ppl) {
+    CalcitePlanContext context = createBuilderContext();
+    Query query = (Query) plan(pplParser, ppl);
+    planTransformer.analyze(query.getPlan(), context);
+    RelNode root = context.relBuilder.build();
+    System.out.println(root.explain());
+    return root;
+  }
+
   private RelNode mergeAdjacentFilters(RelNode relNode) {
     HepProgram program =
         new HepProgramBuilder().addRuleInstance(FilterMergeRule.Config.DEFAULT.toRule()).build();
diff --git a/ppl/src/test/java/org/opensearch/sql/ppl/calcite/CalcitePPLDedupTest.java b/ppl/src/test/java/org/opensearch/sql/ppl/calcite/CalcitePPLDedupTest.java
@@ -5,9 +5,15 @@
 
 package org.opensearch.sql.ppl.calcite;
 
+import org.apache.calcite.plan.hep.HepPlanner;
+import org.apache.calcite.plan.hep.HepProgram;
+import org.apache.calcite.plan.hep.HepProgramBuilder;
 import org.apache.calcite.rel.RelNode;
+import org.apache.calcite.rel.rules.FilterMergeRule;
 import org.apache.calcite.test.CalciteAssert;
+import org.junit.Assert;
 import org.junit.Test;
+import org.opensearch.sql.calcite.plan.rule.PPLSimplifyDedupRule;
 
 public class CalcitePPLDedupTest extends CalcitePPLAbstractTest {
 
@@ -353,4 +359,129 @@ public void testSortFieldProjectedAwayBeforeDedup() {
             + "            LogicalTableScan(table=[[scott, EMP]])\n";
     verifyLogical(root, expectedLogical);
   }
+
+  /**
+   * When a user {@code where} precedes {@code dedup}, the user filter sits adjacent to the
+   * bucket-non-null filter that PPL emits for the dedup pattern. {@link PPLSimplifyDedupRule} must
+   * fold the dedup pattern into a {@link org.opensearch.sql.calcite.plan.rel.LogicalDedup} so that
+   * {@code DedupPushdownRule} can match it; otherwise dedup falls through to the in-memory {@code
+   * ROW_NUMBER} window form, defeating dedup pushdown to the shard. The user predicate must be
+   * preserved as a separate filter below the new {@code LogicalDedup}.
+   */
+  @Test
+  public void testDedupAfterWhereProducesLogicalDedup() {
+    String ppl = "source=EMP | where SAL > 1000 | dedup DEPTNO";
+    RelNode raw = getRelNodeRaw(ppl);
+
+    // Sanity: the un-merged plan has both the bucket-non-null filter and the user where filter
+    // adjacent to each other above the scan — exactly the shape that triggered the original bug.
+    String rawExplain = raw.explain();
+    Assert.assertTrue(
+        "Raw plan should contain the bucket-non-null filter:\n" + rawExplain,
+        rawExplain.contains("IS NOT NULL"));
+    Assert.assertTrue(
+        "Raw plan should contain the user where filter:\n" + rawExplain,
+        rawExplain.contains(">($5, 1000)"));
+    Assert.assertTrue(
+        "Raw plan should contain ROW_NUMBER prior to simplification:\n" + rawExplain,
+        rawExplain.contains("ROW_NUMBER"));
+
+    // Apply rules in the order PPLSimplifyDedupRule -> FilterMergeRule.
+    HepProgram program =
+        new HepProgramBuilder()
+            .addRuleInstance(PPLSimplifyDedupRule.DEDUP_SIMPLIFY_RULE)
+            .addRuleInstance(FilterMergeRule.Config.DEFAULT.toRule())
+            .build();
+    RelNode optimized = runHepPlanner(raw, program);
+
+    String optimizedExplain = optimized.explain();
+    Assert.assertTrue(
+        "Optimized plan should contain LogicalDedup so DedupPushdownRule can match it:\n"
+            + optimizedExplain,
+        optimizedExplain.contains("LogicalDedup"));
+    Assert.assertFalse(
+        "Optimized plan should not retain ROW_NUMBER after simplification:\n" + optimizedExplain,
+        optimizedExplain.contains("ROW_NUMBER"));
+    Assert.assertTrue(
+        "User where predicate must be preserved as a filter below LogicalDedup:\n"
+            + optimizedExplain,
+        optimizedExplain.contains(">($5, 1000)"));
+  }
+
+  /**
+   * Companion to {@link #testDedupAfterWhereProducesLogicalDedup} that pins the user-visible
+   * contract: with a {@code where} preceding {@code dedup}, a {@code LogicalDedup} must be produced
+   * regardless of the order in which {@link FilterMergeRule} and {@link PPLSimplifyDedupRule} fire.
+   *
+   * <p>The simplify rule's bucket-non-null operand predicate is order-independent — it accepts both
+   * a pure {@code IS NOT NULL} and an {@code AND} that contains an {@code IS NOT NULL} on a
+   * partition key — so {@code FilterMergeRule} firing first no longer disables dedup pushdown. This
+   * guards against re-introducing the original regression by reordering, removing, or adding rules
+   * to {@code CalciteToolsHelper#HEP_PROGRAM}.
+   */
+  @Test
+  public void testDedupAfterWhereProducesLogicalDedupRegardlessOfRuleOrder() {
+    String ppl = "source=EMP | where SAL > 1000 | dedup DEPTNO";
+    RelNode raw = getRelNodeRaw(ppl);
+
+    HepProgram program =
+        new HepProgramBuilder()
+            .addRuleInstance(FilterMergeRule.Config.DEFAULT.toRule())
+            .addRuleInstance(PPLSimplifyDedupRule.DEDUP_SIMPLIFY_RULE)
+            .build();
+    RelNode optimized = runHepPlanner(raw, program);
+
+    String optimizedExplain = optimized.explain();
+    Assert.assertTrue(
+        "Even with FilterMergeRule firing first, PPLSimplifyDedupRule must still produce"
+            + " LogicalDedup so DedupPushdownRule can match it:\n"
+            + optimizedExplain,
+        optimizedExplain.contains("LogicalDedup"));
+    Assert.assertFalse(
+        "ROW_NUMBER window form should be removed after simplification:\n" + optimizedExplain,
+        optimizedExplain.contains("ROW_NUMBER"));
+    Assert.assertTrue(
+        "User where predicate must be preserved as a filter below LogicalDedup, even when"
+            + " FilterMergeRule fired first and folded it into the bucket-non-null filter:\n"
+            + optimizedExplain,
+        optimizedExplain.contains(">($5, 1000)"));
+  }
+
+  /**
+   * Mirrors the exact shape of {@code CalciteToolsHelper.HEP_PROGRAM} used in production ({@code
+   * addRuleCollection(List.of(FilterMergeRule, PPLSimplifyDedupRule))}). The two sequenced-{@code
+   * addRuleInstance} tests above prove the rule fires under either ordering, but production runs
+   * both rules in a single collection instruction. This test pins that exact shape to catch any
+   * addRuleCollection-vs-addRuleInstance traversal differences.
+   */
+  @Test
+  public void testDedupAfterWhereProducesLogicalDedupWithProductionHepProgram() {
+    String ppl = "source=EMP | where SAL > 1000 | dedup DEPTNO";
+    RelNode raw = getRelNodeRaw(ppl);
+
+    HepProgram program =
+        new HepProgramBuilder()
+            .addRuleCollection(
+                java.util.List.of(
+                    FilterMergeRule.Config.DEFAULT.toRule(),
+                    PPLSimplifyDedupRule.DEDUP_SIMPLIFY_RULE))
+            .build();
+    RelNode optimized = runHepPlanner(raw, program);
+
+    String optimizedExplain = optimized.explain();
+    Assert.assertTrue(
+        "Production HEP_PROGRAM (addRuleCollection) must still produce LogicalDedup:\n"
+            + optimizedExplain,
+        optimizedExplain.contains("LogicalDedup"));
+    Assert.assertFalse(
+        "ROW_NUMBER window form should be removed under production HEP_PROGRAM:\n"
+            + optimizedExplain,
+        optimizedExplain.contains("ROW_NUMBER"));
+  }
+
+  private static RelNode runHepPlanner(RelNode root, HepProgram program) {
+    HepPlanner planner = new HepPlanner(program);
+    planner.setRoot(root);
+    return planner.findBestExp();
+  }
 }
