[
https://issues.apache.org/jira/browse/SPARK-45846?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18048947#comment-18048947
]
Dipanshu Pandey edited comment on SPARK-45846 at 1/4/26 10:45 AM:
------------------------------------------------------------------
[~manisethu] I have successfully replicated this issue. The NAAJ (Null-Aware
Anti Join) optimization
ignores the spark.sql.autoBroadcastJoinThreshold configuration, forcing a
broadcast join
even when the threshold is set to -1 (broadcast disabled).
Reproduction Steps:
1. Create tables with nullable columns
2. Set spark.sql.autoBroadcastJoinThreshold = -1
3. Set spark.sql.optimizeNullAwareAntiJoin = true
4. Execute a NOT IN subquery
Example Query:
{code:java}
CREATE OR REPLACE TEMP VIEW t1 AS SELECT * FROM VALUES (1), (2), (null) AS t(a);
CREATE OR REPLACE TEMP VIEW t2 AS SELECT * FROM VALUES (1), (null) AS t(b);
SET spark.sql.autoBroadcastJoinThreshold = -1;
SET spark.sql.optimizeNullAwareAntiJoin = true;
SELECT * FROM t1 WHERE a NOT IN (SELECT b FROM t2);
{code}
Observed Behavior:
The query plan still shows BroadcastHashJoin despite threshold = -1:
{code:java}
+- BroadcastHashJoin [a#9], [b#11], LeftAnti, BuildRight, true
:- LocalTableScan [a#9]
+- BroadcastExchange HashedRelationBroadcastMode(...)
+- LocalTableScan [b#11]{code}
Expected Behavior:
When autoBroadcastJoinThreshold = -1, the query should fall back to a
non-broadcast
join strategy (e.g., BroadcastNestedLoopJoin or SortMergeJoin).
Root Cause:
The NAAJ pattern matching in SparkStrategies.scala (line 331-333)
unconditionally
creates a BroadcastHashJoinExec without checking canBroadcastBySize().
I will be working on a fix for this issue.
was (Author: JIRAUSER311986):
I have successfully replicated this issue. The NAAJ (Null-Aware Anti Join)
optimization
ignores the spark.sql.autoBroadcastJoinThreshold configuration, forcing a
broadcast join
even when the threshold is set to -1 (broadcast disabled).
Reproduction Steps:
1. Create tables with nullable columns
2. Set spark.sql.autoBroadcastJoinThreshold = -1
3. Set spark.sql.optimizeNullAwareAntiJoin = true
4. Execute a NOT IN subquery
Example Query:
{code:java}
CREATE OR REPLACE TEMP VIEW t1 AS SELECT * FROM VALUES (1), (2), (null) AS t(a);
CREATE OR REPLACE TEMP VIEW t2 AS SELECT * FROM VALUES (1), (null) AS t(b);
SET spark.sql.autoBroadcastJoinThreshold = -1;
SET spark.sql.optimizeNullAwareAntiJoin = true;
SELECT * FROM t1 WHERE a NOT IN (SELECT b FROM t2);
{code}
Observed Behavior:
The query plan still shows BroadcastHashJoin despite threshold = -1:
{code:java}
+- BroadcastHashJoin [a#9], [b#11], LeftAnti, BuildRight, true
:- LocalTableScan [a#9]
+- BroadcastExchange HashedRelationBroadcastMode(...)
+- LocalTableScan [b#11]{code}
Expected Behavior:
When autoBroadcastJoinThreshold = -1, the query should fall back to a
non-broadcast
join strategy (e.g., BroadcastNestedLoopJoin or SortMergeJoin).
Root Cause:
The NAAJ pattern matching in SparkStrategies.scala (line 331-333)
unconditionally
creates a BroadcastHashJoinExec without checking canBroadcastBySize().
I will be working on a fix for this issue.
> spark.sql.optimizeNullAwareAntiJoin should respect
> spark.sql.autoBroadcastJoinThreshold
> ---------------------------------------------------------------------------------------
>
> Key: SPARK-45846
> URL: https://issues.apache.org/jira/browse/SPARK-45846
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 4.0.0
> Reporter: Chao Sun
> Priority: Major
>
> Normally broadcast join can be disabled when users set
> {{spark.sql.autoBroadcastJoinThreshold}} to -1. However this doesn't apply to
> {{spark.sql.optimizeNullAwareAntiJoin}}:
> {code}
> case j @ ExtractSingleColumnNullAwareAntiJoin(leftKeys, rightKeys) =>
> Seq(joins.BroadcastHashJoinExec(leftKeys, rightKeys, LeftAnti,
> BuildRight,
> None, planLater(j.left), planLater(j.right), isNullAwareAntiJoin =
> true))
> {code}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
