Chenhao Li created SPARK-42775:
----------------------------------
Summary: approx_percentile produces wrong results for large
decimals.
Key: SPARK-42775
URL: https://issues.apache.org/jira/browse/SPARK-42775
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 3.3.0, 3.2.0, 3.1.0, 3.0.0, 2.4.0, 2.3.0, 2.2.0, 2.1.0,
3.4.0
Reporter: Chenhao Li
In the `approx_percentile` expression, Spark casts decimal to double to update
the aggregation state
([ApproximatePercentile.scala#L181|https://github.com/apache/spark/blob/933dc0c42f0caf74aaa077fd4f2c2e7208452b9b/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/aggregate/ApproximatePercentile.scala#L181])
and casts the result double back to decimal
([ApproximatePercentile.scala#L206|https://github.com/apache/spark/blob/933dc0c42f0caf74aaa077fd4f2c2e7208452b9b/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/aggregate/ApproximatePercentile.scala#L206]).
The precision loss in the casts can make the result decimal out of its
precision range. This can lead to the following counter-intuitive results:
{code:sql}
spark-sql> select approx_percentile(col, 0.5) from values (9999999999999999999)
as tab(col);
NULL
spark-sql> select approx_percentile(col, 0.5) is null from values
(9999999999999999999) as tab(col);
false
spark-sql> select cast(approx_percentile(col, 0.5) as string) from values
(9999999999999999999) as tab(col);
10000000000000000000
spark-sql> desc select approx_percentile(col, 0.5) from values
(9999999999999999999) as tab(col);
approx_percentile(col, 0.5, 10000) decimal(19,0)
{code}
The result is actually not null, so the second query returns false. The first
query returns null because the result cannot fit into {{decimal(19, 0)}}.
A suggested fix is to use `Decimal.changePrecision` here to ensure the result
fits, and really returns a null or throws an exception when the result doesn't
fit.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
