You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@datafusion.apache.org by gi...@apache.org on 2026/05/30 04:27:31 UTC
(datafusion) 01/01: fix(physical-plan): make HashJoinExec dynamic filter pushdown idempotent (#22523)
This is an automated email from the ASF dual-hosted git repository.
github-merge-queue[bot] pushed a commit to branch gh-readonly-queue/main/pr-22523-b6d4c252634f6ab94b1363c751d7bbd101d86dd2
in repository https://gitbox.apache.org/repos/asf/datafusion.git
commit 2e7b8e1b2d3cd800823b3813f2f7efdde0d21af8
Author: Xuanyi Li <sh...@outlook.com>
AuthorDate: Fri May 29 21:27:10 2026 -0700
fix(physical-plan): make HashJoinExec dynamic filter pushdown idempotent (#22523)
## Which issue does this PR close?
Related to https://github.com/apache/datafusion-ballista/issues/1359
## Rationale
Ballista's Adaptive Query Execution (AQE) planner re-invokes
DataFusion's full `PhysicalOptimizer` chain after every completed stage.
`FilterPushdown::new_post_optimization()` is not idempotent on plans
containing `HashJoinExec`.
In the `Post` phase, `HashJoinExec::gather_filters_for_pushdown`
unconditionally creates a new `DynamicFilterPhysicalExpr` and installs
it on the probe-side child via `with_self_filter`. After pass 1 the join
already carries a `dynamic_filter: Some(...)`, and the shared
`Arc<DynamicFilterPhysicalExpr>` is already wired into the probe-side
scan's predicate. On pass 2 a *second* dynamic filter is created and
ANDed onto the existing predicate, producing `DynamicFilter AND
DynamicFilter`. Each subsequent pass adds another duplicate, compounding
indefinitely in AQE replan loops.
## What changes are included in this PR?
- **Guard in `HashJoinExec::gather_filters_for_pushdown`**: skip
dynamic-filter creation when `self.dynamic_filter.is_some()`, meaning a
previous pass already installed one. The existing `Arc` remains valid
and correctly wired into the probe-side scan.
- **Comment** explaining why the guard is needed (AQE replan context).
- **Test** `post_phase_is_idempotent_on_hash_join` in
`tests/physical_optimizer/filter_pushdown.rs`: builds a `HashJoinExec`,
runs `FilterPushdown::new_post_optimization()` twice, and asserts
structural equality via `get_plan_string`.
## Are these changes tested?
Yes. The new test fails without the fix (plan strings diverge due to
duplicated dynamic filter predicates) and passes with it.
## Are there any user-facing changes?
No. Dynamic filter pushdown is an internal optimization; the idempotence
guard only affects re-optimization scenarios (AQE).
---------
Co-authored-by: Claude Opus 4.7 (1M context) <no...@anthropic.com>
---
.../tests/physical_optimizer/filter_pushdown.rs | 39 ++++++++++++++++++++++
.../physical-plan/src/joins/hash_join/exec.rs | 8 ++++-
2 files changed, 46 insertions(+), 1 deletion(-)
diff --git a/datafusion/core/tests/physical_optimizer/filter_pushdown.rs b/datafusion/core/tests/physical_optimizer/filter_pushdown.rs
index b420326596..909b80cada 100644
--- a/datafusion/core/tests/physical_optimizer/filter_pushdown.rs
+++ b/datafusion/core/tests/physical_optimizer/filter_pushdown.rs
@@ -3269,3 +3269,42 @@ fn test_filter_pushdown_through_sort_with_projection() {
"
);
}
+
+/// `FilterPushdown::new_post_optimization()` must be idempotent. When applied
+/// to a HashJoinExec, the rule installs a dynamic filter on the probe-side
+/// scan; before the fix in `HashJoinExec::gather_filters_for_pushdown`, every
+/// invocation created a *new* `DynamicFilterPhysicalExpr` and ANDed it onto
+/// the probe side's existing predicate, producing
+/// `DynamicFilter AND DynamicFilter AND ...` after N passes.
+///
+/// AQE (datafusion-ballista#1359) re-runs the optimizer chain after every
+/// completed stage, so this would compound indefinitely without the guard.
+#[test]
+fn post_phase_is_idempotent_on_hash_join() {
+ use crate::physical_optimizer::test_utils::{hash_join_exec, parquet_exec, schema};
+ use datafusion_common::JoinType;
+ use datafusion_physical_expr::expressions::Column;
+ use datafusion_physical_optimizer::filter_pushdown::FilterPushdown;
+ use datafusion_physical_plan::get_plan_string;
+ use datafusion_physical_plan::joins::utils::JoinOn;
+
+ let s = schema();
+ let left = parquet_exec(Arc::clone(&s));
+ let right = parquet_exec(Arc::clone(&s));
+ let join_on: JoinOn = vec![(
+ Arc::new(Column::new_with_schema("a", &left.schema()).unwrap()),
+ Arc::new(Column::new_with_schema("a", &right.schema()).unwrap()),
+ )];
+ let plan = hash_join_exec(left, right, join_on, None, &JoinType::Inner).unwrap();
+
+ let config = ConfigOptions::new();
+ let rule = FilterPushdown::new_post_optimization();
+ let once = rule.optimize(plan, &config).unwrap();
+ let twice = rule.optimize(Arc::clone(&once), &config).unwrap();
+
+ assert_eq!(
+ get_plan_string(&once),
+ get_plan_string(&twice),
+ "second invocation of FilterPushdown::new_post_optimization mutated the plan",
+ );
+}
diff --git a/datafusion/physical-plan/src/joins/hash_join/exec.rs b/datafusion/physical-plan/src/joins/hash_join/exec.rs
index f7391feb29..03387c316b 100644
--- a/datafusion/physical-plan/src/joins/hash_join/exec.rs
+++ b/datafusion/physical-plan/src/joins/hash_join/exec.rs
@@ -1635,8 +1635,14 @@ impl ExecutionPlan for HashJoinExec {
ChildFilterDescription::all_unsupported(&parent_filters)
};
- // Add dynamic filters in Post phase if enabled
+ // Add dynamic filters in Post phase if enabled. Skip when this join
+ // already carries a dynamic filter from a previous pass — the shared
+ // `Arc<DynamicFilterPhysicalExpr>` is still wired into the probe-side
+ // scan's predicate, and re-creating it would AND a fresh duplicate
+ // onto every Post-phase invocation (apache/datafusion-ballista#1359
+ // surfaces this in AQE replan loops).
if phase == FilterPushdownPhase::Post
+ && self.dynamic_filter.is_none()
&& self.allow_join_dynamic_filter_pushdown(config)
{
// Add actual dynamic filter to right side (probe side)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@datafusion.apache.org
For additional commands, e-mail: commits-help@datafusion.apache.org