[SPARK-42937] Join with subquery in condition can fail with wholestage codegen and adaptive execution disabled

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.3.2, 3.4.0
• Fix Version/s: 3.3.3, 3.4.1
• Component/s: SQL
• Assignee: Bruce Robbins
• Reporter: Bruce Robbins
• Created: 27/Mar/23 19:45
• Updated: 28/Mar/23 12:41
• Resolved: 28/Mar/23 12:41

Description

The below left outer join gets an error:

create or replace temp view v1 as
select * from values
(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1),
(2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2),
(3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1)
as v1(key, value1, value2, value3, value4, value5, value6, value7, value8, value9, value10);

create or replace temp view v2 as
select * from values
(1, 2),
(3, 8),
(7, 9)
as v2(a, b);

create or replace temp view v3 as
select * from values
(3),
(8)
as v3(col1);

set spark.sql.codegen.maxFields=10; -- let's make maxFields 10 instead of 100
set spark.sql.adaptive.enabled=false;

select *
from v1
left outer join v2
on key = a
and key in (select col1 from v3);

The join fails during predicate codegen:

23/03/27 12:24:12 WARN Predicate: Expr codegen error and falling back to interpreter mode
java.lang.IllegalArgumentException: requirement failed: input[0, int, false] IN subquery#34 has not finished
	at scala.Predef$.require(Predef.scala:281)
	at org.apache.spark.sql.execution.InSubqueryExec.prepareResult(subquery.scala:144)
	at org.apache.spark.sql.execution.InSubqueryExec.doGenCode(subquery.scala:156)
	at org.apache.spark.sql.catalyst.expressions.Expression.$anonfun$genCode$3(Expression.scala:201)
	at scala.Option.getOrElse(Option.scala:189)
	at org.apache.spark.sql.catalyst.expressions.Expression.genCode(Expression.scala:196)
	at org.apache.spark.sql.catalyst.expressions.codegen.CodegenContext.$anonfun$generateExpressions$2(CodeGenerator.scala:1278)
	at scala.collection.immutable.List.map(List.scala:293)
	at org.apache.spark.sql.catalyst.expressions.codegen.CodegenContext.generateExpressions(CodeGenerator.scala:1278)
	at org.apache.spark.sql.catalyst.expressions.codegen.GeneratePredicate$.create(GeneratePredicate.scala:41)
	at org.apache.spark.sql.catalyst.expressions.codegen.GeneratePredicate$.generate(GeneratePredicate.scala:33)
	at org.apache.spark.sql.catalyst.expressions.Predicate$.createCodeGeneratedObject(predicates.scala:73)
	at org.apache.spark.sql.catalyst.expressions.Predicate$.createCodeGeneratedObject(predicates.scala:70)
	at org.apache.spark.sql.catalyst.expressions.CodeGeneratorWithInterpretedFallback.createObject(CodeGeneratorWithInterpretedFallback.scala:51)
	at org.apache.spark.sql.catalyst.expressions.Predicate$.create(predicates.scala:86)
	at org.apache.spark.sql.execution.joins.HashJoin.boundCondition(HashJoin.scala:146)
	at org.apache.spark.sql.execution.joins.HashJoin.boundCondition$(HashJoin.scala:140)
	at org.apache.spark.sql.execution.joins.BroadcastHashJoinExec.boundCondition$lzycompute(BroadcastHashJoinExec.scala:40)
	at org.apache.spark.sql.execution.joins.BroadcastHashJoinExec.boundCondition(BroadcastHashJoinExec.scala:40)

It fails again after fallback to interpreter mode:

23/03/27 12:24:12 ERROR Executor: Exception in task 2.0 in stage 2.0 (TID 7)
java.lang.IllegalArgumentException: requirement failed: input[0, int, false] IN subquery#34 has not finished
	at scala.Predef$.require(Predef.scala:281)
	at org.apache.spark.sql.execution.InSubqueryExec.prepareResult(subquery.scala:144)
	at org.apache.spark.sql.execution.InSubqueryExec.eval(subquery.scala:151)
	at org.apache.spark.sql.catalyst.expressions.InterpretedPredicate.eval(predicates.scala:52)
	at org.apache.spark.sql.execution.joins.HashJoin.$anonfun$boundCondition$2(HashJoin.scala:146)
	at org.apache.spark.sql.execution.joins.HashJoin.$anonfun$boundCondition$2$adapted(HashJoin.scala:146)
	at org.apache.spark.sql.execution.joins.HashJoin.$anonfun$outerJoin$1(HashJoin.scala:205)

Both the predicate codegen and the evaluation fail for the same reason: PlanSubqueries creates InSubqueryExec with shouldBroadcast=false. The driver waits for the subquery to finish, but it's the executor that uses the results of the subquery (for predicate codegen or evaluation). Because shouldBroadcast is set to false, the result is stored in a transient field (InSubqueryExec#result), so the result of the subquery is not serialized when the InSubqueryExec instance is sent to the executor.

When wholestage codegen is enabled, the predicate codegen happens on the driver, so the subquery's result is available. When adaptive execution is enabled, PlanAdaptiveSubqueries always sets shouldBroadcast=true, so the subquery's result is available on the executor, if needed.

Issue Links

• links to: [Github] Pull Request #40569 (bersprockets)