Peter Toth created SPARK-61549:
----------------------------------
Summary: Estimate the size of a single join from its left side
Key: SPARK-61549
URL: https://issues.apache.org/jira/browse/SPARK-61549
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 5.0.0
Reporter: Peter Toth
The size estimate of a logical join of type {{LeftSingle}} is the product of
the sizes of its two sides. {{SizeInBytesOnlyStatsPlanVisitor.visitJoin}} falls
back to {{default}} for it, which multiplies the sizes of the children. CBO's
{{JoinEstimation}} does not support {{LeftSingle}} either, so it falls back to
the same product.
A single join outputs each row of its left side once, with the columns of the
right side added. They are null when no row matches, and the join fails on a
second match. So its row count is the left side's.
{{RewriteCorrelatedScalarSubquery}} plans a single join for a correlated scalar
subquery that the optimizer cannot prove to return one row
({{spark.sql.optimizer.scalarSubqueryUseSingleJoin}}, true by default). The
joins above it see that side as far bigger than it is, and can miss a broadcast.
Estimating it from its left side, as SPARK-61534 does for an existence join,
would fix it.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]