(doris) 01/01: [fix](local shuffle) fix NLJ COREDUMP: don't insert build-side BROADCAST for serial NLJ

[hu...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

huajianlan pushed a commit to branch fe_local_shuffle
in repository https://gitbox.apache.org/repos/asf/doris.git

commit d632320f8f6f5927ab3db501e80777b40fd587ad
Author: 924060929 <la...@selectdb.com>
AuthorDate: Wed Apr 1 12:23:14 2026 +0800

    [fix](local shuffle) fix NLJ COREDUMP: don't insert build-side BROADCAST for serial NLJ
    
    When NLJ is serial (RIGHT_OUTER/FULL_OUTER etc.), the probe pipeline has 1 task.
    FE was inserting a BROADCAST local exchange on the build side for pooling scan,
    which inflated the build pipeline's num_tasks to _num_instances. This caused
    instance 1+ to create build tasks without corresponding probe tasks, leaving
    source_deps empty and crashing in set_ready_to_read().
    
    BE-native avoids this via the num_tasks_of_parent()<=1 check in _add_local_exchange.
    Fix: set buildSideRequire to noRequire() for serial NLJ to match BE-native behavior.
---
 .../java/org/apache/doris/planner/NestedLoopJoinNode.java     | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/fe/fe-core/src/main/java/org/apache/doris/planner/NestedLoopJoinNode.java b/fe/fe-core/src/main/java/org/apache/doris/planner/NestedLoopJoinNode.java
index 8bcb837e6da..4f3ecdea388 100644
--- a/fe/fe-core/src/main/java/org/apache/doris/planner/NestedLoopJoinNode.java
+++ b/fe/fe-core/src/main/java/org/apache/doris/planner/NestedLoopJoinNode.java
@@ -196,11 +196,14 @@ public class NestedLoopJoinNode extends JoinNodeBase {
         } else if (isSerialOperator()) {
             // RIGHT_OUTER/RIGHT_SEMI/RIGHT_ANTI/FULL_OUTER: probe side must be serial (1 task)
             // to avoid duplicate unmatched rows from build side. No exchange needed for probe.
-            // BE: NestedLoopJoinProbeOperatorX returns NOOP for these join types.
+            // Build side: noRequire() regardless of pooling scan. The probe pipeline has 1 task
+            // (serial), so inserting BROADCAST on the build side would inflate build pipeline's
+            // num_tasks to _num_instances while probe stays at 1. This causes instance 1+ to
+            // create build tasks without corresponding probe tasks, leaving source_deps empty
+            // and crashing in set_ready_to_read(). BE-native avoids this via the
+            // num_tasks_of_parent()<=1 check in _add_local_exchange.
             probeSideRequire = LocalExchangeTypeRequire.noRequire();
-            buildSideRequire = childUsePoolingScan
-                    ? LocalExchangeTypeRequire.requireBroadcast()
-                    : LocalExchangeTypeRequire.noRequire();
+            buildSideRequire = LocalExchangeTypeRequire.noRequire();
             outputType = LocalExchangeType.NOOP;
         } else if (childUsePoolingScan) {
             probeSideRequire = LocalExchangeTypeRequire.requireAdaptivePassthrough();


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org