[IGNITE-13443] distributed left joins errored with "java.lang.ClassCastException"

• Type: Bug
• Status: Open
• Priority: Blocker
• Resolution: Unresolved
• Affects Version/s: 2.8.1
• Component/s: h2-limitation, sql
• Ignite Flags: Docs Required, Release Notes Required
• Assignee: Unassigned
• Reporter: Isaac Zhu
• Created: 13/Sep/20 04:57
• Updated: 21/Sep/20 14:40

Description

This is actually an H2 1.4.197 Bug

When using distributed joins, SQL like this (table A and table B are not collocated, so distributed join is used here):

     select * from table A left join table B on A.id = B.id left join table C on B.id = C.id where ...

has chance to error out at the second left join with a "java.lang.ClassCastException" when the first left join had an EmptyCursor

 

This is because, for distributed join, H2 JoinBatch is used - it seems JoinBatch is not used for collocated join.

In JoinBatch line 292:

https://github.com/h2database/h2database/blob/c8a861bb1a3f04967ec40cb5e3336535c43af5fb/h2/src/main/org/h2/table/JoinBatch.java#L292

It increases the filter ID to let the next fetchCurrent call resolve the "EMPTY_CURSOR" to a NullRow when a "JoinFilter.find" is not called.

Then, in the "find" call, the resolved NullRow will be updated to some cursor from the future - because only EmptyCursor is skipped in the find(), a NullRow will be updated again and the bitwise state goes wrong

This then causes the exception in "fetchCurrent" line:  358

https://github.com/h2database/h2database/blob/c8a861bb1a3f04967ec40cb5e3336535c43af5fb/h2/src/main/org/h2/table/JoinBatch.java#L358

The "x" is a Future, so it can't be cast to a Cursor - the state is wrong, the "isFuture" returns false while it is a Future.

 

I did a local fix in our server as a workaround in org.h2.table.JoinBatch, the changed code is attached.