[IGNITE-21722] Sql. NPE in correlated nested loop join

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.0
• Component/s: sql
• Labels: calcite2-required ignite-3
• Assignee: Pavel Pereslegin
• Reporter: Konstantin Orlov
• Created: 08/Mar/24 17:34
• Updated: 02/Oct/24 18:56
• Resolved: 29/Mar/24 07:39

Description

See stack trace below:

Caused by: java.lang.NullPointerException
	at org.apache.ignite.internal.sql.engine.exec.rel.CorrelatedNestedLoopJoinNode.join(CorrelatedNestedLoopJoinNode.java:362)
	at org.apache.ignite.internal.sql.engine.exec.rel.CorrelatedNestedLoopJoinNode.lambda$onRequest$1(CorrelatedNestedLoopJoinNode.java:274)
	at org.apache.ignite.internal.sql.engine.exec.ExecutionContext.lambda$execute$0(ExecutionContext.java:320)
	... 4 more

To reproduce the issue, run TPC-H q21 with sf=0.1 on 3-node cluster

Update

The following sequence occurs:

1. CorrelatedNestedLoopJoinNode pushes last row to downstream (FilterNode). See joinType == JoinRelType.LEFT branch in join() method.
2. FilterNode receives last requested row and tries to request next batch from the upstream (CorrelatedNestedLoopJoinNode).
3. CorrelatedNestedLoopJoinNode is in IDLE state (which is incorrect), so it submits join() to the execution queue.
4. CorrelatedNestedLoopJoinNode resets rightInBuf, changes state to FILLING_LEFT and requests rows from leftsource().
5. Previously submitted join() starts executing and crashes with NullPointerException, because rightInBuf is null.

Issue Links

• blocks: IGNITE-21579 Sql. Investigate performance bottlenecks in sql engine
• is related to: IGNITE-23345 Calcite engine. Correlated nested loop LEFT join failure
• links to: GitHub Pull Request #3504