Nick Young created SPARK-60861:
----------------------------------
Summary: Adding a window expression changes HAVING name resolution
for a same-named grouping column
Key: SPARK-60861
URL: https://issues.apache.org/jira/browse/SPARK-60861
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 5.0.0
Reporter: Nick Young
When an input grouping column and a SELECT-list alias have the same name,
adding a window expression changes which name HAVING resolves. HAVING should
continue to prefer the input grouping column. The changed binding silently
filters the wrong groups.
Minimal reproduction using two synthetic rows:
{code:sql}
-- Returns 10: HAVING refers to the grouping key t.k.
SELECT sum(v) AS k
FROM VALUES (1, 10), (2, 20) AS t(k, v)
GROUP BY k
HAVING k = 1;
-- Returns no rows; expected (10, 1).
SELECT sum(v) AS k, row_number() OVER (ORDER BY sum(v)) AS rn
FROM VALUES (1, 10), (2, 20) AS t(k, v)
GROUP BY k
HAVING k = 1;
-- Qualifying the grouping key returns the expected (10, 1).
SELECT sum(v) AS k, row_number() OVER (ORDER BY sum(v)) AS rn
FROM VALUES (1, 10), (2, 20) AS t(k, v)
GROUP BY k
HAVING t.k = 1;
{code}
The analyzed plan for the failing query applies the filter to the BIGINT
aggregate alias sum(v), rather than the INT grouping key. The control queries
filter the input grouping key.
The public Analyzer's ExtractWindowExpressions rule converts unresolved HAVING
into a Filter over the extracted Aggregate. This allows SELECT aliases to
shadow grouping columns that would otherwise take precedence.
Reproduced with ANSI mode enabled and disabled against the HAVING analyzer code
from Apache Spark master 523596fdc5495bb191a0af903622355999a62eaa.
SQLQueryTestSuite execution and analyzed-plan output confirm the different
bindings and results.
A candidate fix and regression coverage are available at:
https://github.com/apache/spark/pull/60076
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
