Maryann Xue created SPARK-25914:
-----------------------------------
Summary: Separate projection from grouping and aggregate in
logical Aggregate
Key: SPARK-25914
URL: https://issues.apache.org/jira/browse/SPARK-25914
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 2.4.0
Reporter: Maryann Xue
Currently the Spark SQL logical Aggregate has two expression fields:
{{groupingExpressions}} and {{aggregateExpressions}}, in which
{{aggregateExpressions}} is actually the result expressions, or in other words,
the project list in the SELECT clause.
This would cause an exception while processing the following query:
SELECT concat('x', concat(a, 's'))
FROM testData2
GROUP BY concat(a, 's')
After optimization, the query becomes:
SELECT concat('x', a, 's')
FROM testData2
GROUP BY concat(a, 's')
The optimization rule {{CombineConcats}} optimizes the expressions by
flattening "concat" and causes the query to fail since the expression
{{concat('x', a, 's')}} in the SELECT clause is neither referencing a grouping
expression nor a aggregate expression.
The problem is that we try to mix two operations in one operator, and worse, in
one field: the group-and-aggregate operation and the project operation. There
are two ways to solve this problem:
1. Break the two operations into two logical operators, which means a group-by
query can usually to be mapped into a Project-over-Aggregate pattern.
2. Break the two operations into multiple fields in the Aggregate operator, the
same way we do for physical aggregate classes (e.g., {{HashAggregateExec}}, or
{{SortAggregateExec}}). Thus, {{groupingExpressions}} would still be the
expressions from the GROUP BY clause (as before), but {{aggregateExpressions}}
would contain aggregate functions only, and {{resultExpressions}} would be the
project list in the SELECT clause holding references to either
{{groupingExpressions}} or {{aggregateExpressions}}.
I would say option 1 is even clearer, but it would be more likely to break the
pattern matching in existing optimization rules and thus require more changes
in the compiler. So we'd probably wanna go with option 2. That said, I suggest
we achieve this goal through two iterative steps:
Phase 1: Keep the current fields of logical Aggregate as
{{groupingExpressions}} and {{aggregateExpressions}}, but change the semantics
of {{aggregateExpressions}} by replacing the grouping expressions with
corresponding references to expressions in {{groupingExpressions}}. The
aggregate expressions in {{aggregateExpressions}} will remain the same.
Phase 2: Add {{resultExpressions}} for the project list, and keep only
aggregate expressions in {{aggregateExpressions}}.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
