fix: preserve hash join dynamic filters across recursive CTE iterations

datafusion/physical-expr/src/expressions/dynamic_filters/mod.rs

@@ -29,6 +29,7 @@ use datafusion_common::{
     tree_node::{Transformed, TransformedResult, TreeNode},
 };
 
+use crate::expressions::lit;
 use datafusion_expr::ColumnarValue;
 use datafusion_expr_common::dyn_eq::DynHash;
 
@@ -343,6 +344,24 @@ impl DynamicFilterPhysicalExpr {
         Ok(())
     }
 
+    /// Reset this dynamic filter so its producer can publish a new expression
+    /// during a later execution of the same plan.
+    pub fn reset(&self) {
+        let mut current = self.inner.write();
+        let new_generation = current.generation + 1;
+        *current = Inner {
+            expression_id: current.expression_id,
+            generation: new_generation,
+            expr: lit(true),
+            is_complete: false,
+        };
+        drop(current);
+
+        let _ = self.state_watch.send(FilterState::InProgress {
+            generation: new_generation,
+        });
+    }
+
     /// Mark this dynamic filter as complete and broadcast to all waiters.
     ///
     /// This signals that all expected updates have been received.
@@ -1032,6 +1051,21 @@ mod test {
         dynamic_filter.wait_complete().await;
     }
 
+    #[test]
+    fn test_reset_reopens_completed_filter() {
+        let dynamic_filter =
+            DynamicFilterPhysicalExpr::new(vec![], lit(false) as Arc<dyn PhysicalExpr>);
+        dynamic_filter.mark_complete();
+        let generation_before_reset = dynamic_filter.inner.read().generation;
+
+        dynamic_filter.reset();
+
+        let inner = dynamic_filter.inner.read();
+        assert!(!inner.is_complete);
+        assert!(inner.generation > generation_before_reset);
+        assert_eq!(inner.expr.to_string(), "true");
+    }
+
     #[test]
     fn test_with_new_children_independence() {
         // Create a schema with columns a, b, c, d

datafusion/physical-plan/src/aggregates/mod.rs

@@ -2617,6 +2617,19 @@ impl ExecutionPlan for AggregateExec {
         )
     }
 
+    fn reset_state(self: Arc<Self>) -> Result<Arc<dyn ExecutionPlan>> {
+        if let Some(dynamic_filter) = &self.dynamic_filter {
+            dynamic_filter.filter.reset();
+            for accumulator_filter in &dynamic_filter.accumulator_dyn_filter_info {
+                *accumulator_filter.shared_bound.lock() = ScalarValue::Null;
+            }
+        }
+
+        let mut aggregate = (*self).clone();
+        aggregate.metrics = ExecutionPlanMetricsSet::new();
+        Ok(Arc::new(aggregate))
+    }
+
     fn execute(
         &self,
         partition: usize,
@@ -9814,4 +9827,58 @@ mod tests {
         assert!(agg.set_dynamic_filter(df).is_err());
         Ok(())
     }
+
+    #[test]
+    fn test_aggregate_exec_reset_state() -> Result<()> {
+        let schema = Arc::new(Schema::new(vec![Field::new("a", DataType::Int64, false)]));
+        let child = Arc::new(EmptyExec::new(Arc::clone(&schema)));
+
+        let agg = AggregateExec::try_new(
+            AggregateMode::Partial,
+            PhysicalGroupBy::new_single(vec![]),
+            vec![Arc::new(
+                AggregateExprBuilder::new(min_udaf(), vec![col("a", &schema)?])
+                    .schema(Arc::clone(&schema))
+                    .alias("min_a")
+                    .build()?,
+            )],
+            vec![None],
+            child,
+            Arc::clone(&schema),
+        )?;
+
+        let df = Arc::new(DynamicFilterPhysicalExpr::new(
+            vec![col("a", &schema)?],
+            lit(false),
+        ));
+        let agg = Arc::new(agg.set_dynamic_filter(Arc::clone(&df))?);
+
+        // Simulate an iteration setting a shared bound and completing the filter
+        if let Some(dyn_filter) = &agg.dynamic_filter {
+            for info in &dyn_filter.accumulator_dyn_filter_info {
+                *info.shared_bound.lock() = ScalarValue::Int64(Some(42));
+            }
+            dyn_filter.filter.mark_complete();
+        }
+
+        // Call reset_state
+        let reset_plan = agg.reset_state()?;
+        let reset_agg = reset_plan
+            .downcast_ref::<AggregateExec>()
+            .expect("should downcast back to AggregateExec");
+
+        let dyn_filter = reset_agg
+            .dynamic_filter
+            .as_ref()
+            .expect("dynamic filter should be retained");
+
+        for info in &dyn_filter.accumulator_dyn_filter_info {
+            assert_eq!(*info.shared_bound.lock(), ScalarValue::Null);
+        }
+
+        let current_expr = dyn_filter.filter.current()?;
+        assert_eq!(format!("{current_expr}"), format!("{}", lit(true)));
+
+        Ok(())
+    }
 }

datafusion/physical-plan/src/execution_plan.rs

@@ -1996,7 +1996,9 @@ pub fn check_not_null_constraints(
 ///
 /// While this function enables plan reuse, it does not allow the same plan to be executed if it (OR):
 ///
-/// * uses dynamic filters,
+/// * uses dynamic filters and the original plan (or another reset copy) is still
+///   executing, because operators such as `HashJoinExec` and `AggregateExec`
+///   reset their shared dynamic filter in place,
 /// * represents a recursive query.
 ///
 pub fn reset_plan_states(plan: Arc<dyn ExecutionPlan>) -> Result<Arc<dyn ExecutionPlan>> {

datafusion/physical-plan/src/joins/hash_join/exec.rs

@@ -568,8 +568,14 @@ impl HashJoinExecBuilder {
     /// Reset task-local runtime state while retaining the immutable prepared build.
     pub fn reset_state(mut self) -> Self {
         self.exec.left_fut = Default::default();
-        self.exec.dynamic_filter = None;
+
+        if let Some(dynamic_filter) = &mut self.exec.dynamic_filter {
+            dynamic_filter.filter.reset();
+            dynamic_filter.build_accumulator = OnceLock::new();
+        }
+
         self.exec.metrics = ExecutionPlanMetricsSet::new();
+
         self
     }
 

datafusion/sqllogictest/test_files/cte.slt

@@ -1727,3 +1727,108 @@ order by 1, 2;
 
 statement ok
 drop table test;
+
+# Dynamic filters produced by a recursive term must be reset between
+# iterations. Otherwise the second execution of this uncorrelated aggregate
+# reuses the first iteration's MAX bound and scans no rows.
+statement ok
+COPY (SELECT CAST(value AS DOUBLE) AS r FROM generate_series(0, 128)) TO 'test_files/scratch/cte/dynamic_filter_recursive.parquet' STORED AS PARQUET;
+
+statement ok
+CREATE EXTERNAL TABLE dynamic_filter_recursive STORED AS PARQUET LOCATION 'test_files/scratch/cte/dynamic_filter_recursive.parquet';
+
+query IR rowsort
+WITH RECURSIVE w(n, m) AS (
+    SELECT 0, CAST(NULL AS DOUBLE)
+    UNION ALL
+    SELECT w.n + 1, x.m
+    FROM w CROSS JOIN (
+        SELECT MAX(r) AS m
+        FROM dynamic_filter_recursive
+        WHERE r % 2 = 0
+    ) x
+    WHERE w.n < 3
+)
+SELECT n, m FROM w;
+----
+0 NULL
+1 128
+2 128
+3 128
+
+statement ok
+DROP TABLE dynamic_filter_recursive;
+
+# Hash join dynamic filters must remain shared with the probe scan across
+# recursive iterations (#26641). Each iteration looks up a different key; the
+# EXPLAIN ANALYZE below checks that the filter is reset and republished, so the
+# probe scan prunes rows on every iteration.
+statement ok
+SET datafusion.execution.target_partitions = 1;
+
+statement ok
+SET datafusion.execution.parquet.pushdown_filters = true;
+
+statement ok
+COPY (SELECT value AS k FROM generate_series(1, 100))
+TO 'test_files/scratch/cte/hash_join_dynamic_filter.parquet' STORED AS PARQUET;
+
+statement ok
+CREATE EXTERNAL TABLE hash_join_dynamic_filter STORED AS PARQUET
+LOCATION 'test_files/scratch/cte/hash_join_dynamic_filter.parquet';
+
+query I
+WITH RECURSIVE w(n) AS (
+    SELECT 1
+    UNION ALL
+    SELECT t.k
+    FROM w JOIN hash_join_dynamic_filter t ON t.k = w.n + 1
+    WHERE w.n < 5
+)
+SELECT * FROM w ORDER BY n;
+----
+1
+2
+3
+4
+5
+
+statement ok
+SET datafusion.explain.analyze_categories = 'rows';
+
+# The probe scan should output four rows and prune the other 396, rather than
+# outputting all 400 rows across the four iterations.
+query TT
+EXPLAIN ANALYZE
+WITH RECURSIVE w(n) AS (
+    SELECT 1
+    UNION ALL
+    SELECT t.k
+    FROM w JOIN hash_join_dynamic_filter t ON t.k = w.n + 1
+    WHERE w.n < 5
+)
+SELECT * FROM w ORDER BY n;
+----
+Plan with Metrics
+01)SortExec: expr=[n@0 ASC NULLS LAST], preserve_partitioning=[false], metrics=[output_rows=5, output_batches=1, spill_count=0, spilled_rows=0]
+02)--RecursiveQueryExec: name=w, is_distinct=false, metrics=[output_rows=5, output_batches=5]
+03)----ProjectionExec: expr=[1 as n], metrics=[output_rows=1, output_batches=1]
+04)------PlaceholderRowExec, metrics=[]
+05)----ProjectionExec: expr=[k@0 as n], metrics=[]
+06)------HashJoinExec: mode=CollectLeft, join_type=Inner, on=[(w.n + Int64(1)@1, k@0)], projection=[k@2], metrics=[]
+07)--------ProjectionExec: expr=[n@0 as n, n@0 + 1 as w.n + Int64(1)], metrics=[]
+08)----------FilterExec: n@0 < 5, metrics=[]
+09)------------WorkTableExec: name=w, metrics=[]
+10)--------DataSourceExec: file_groups={1 group: [[WORKSPACE_ROOT/datafusion/sqllogictest/test_files/scratch/cte/hash_join_dynamic_filter.parquet]]}, projection=[k], output_ordering=[k@0 ASC NULLS LAST], file_type=parquet, predicate=DynamicFilter [ true ], dynamic_rg_pruning=eligible, metrics=[output_rows=4, output_batches=4, files_ranges_pruned_statistics=4 total → 4 matched, row_groups_pruned_statistics=4 total → 4 matched, row_groups_pruned_bloom_filter=4 total → 4 matched, page_index_pages_pruned=4 total → 4 matched, page_index_rows_pruned=400 total → 400 matched, limit_pruned_row_groups=0 total → 0 matched, batches_split=0, file_open_errors=0, file_scan_errors=0, files_opened=4, files_processed=4, num_predicate_creation_errors=0, predicate_evaluation_errors=0, pushdown_rows_matched=4, pushdown_rows_pruned=396, predicate_cache_inner_records=400, predicate_cache_records=4, scan_efficiency_ratio=157.36% (1.14 K/727)]
+
+statement ok
+DROP TABLE hash_join_dynamic_filter;
+
+statement ok
+RESET datafusion.explain.analyze_categories;
+
+statement ok
+SET datafusion.execution.target_partitions = 4;
+
+statement ok
+RESET datafusion.execution.parquet.pushdown_filters;