[
https://issues.apache.org/jira/browse/SPARK-9971?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14697868#comment-14697868
]
Davies Liu commented on SPARK-9971:
-----------------------------------
We had a long discussion about how to support NaN, but didn't find a good one
from other databases to follow (they all handle them in different way from each
other), then we came out a plan, see
https://issues.apache.org/jira/browse/SPARK-9079.
For 4 (in NaN in aggregation), we have not decided yet. The current behavior is
not consistent across functions
{code}
>>> sf = sqlContext.createDataFrame([(1.0,), (float('nan'),)], ['f'])
>>> sf.selectExpr('min(f)', 'max(f)', 'sum(f)', 'avg(f)', 'count(f)').show()
+-------+-------+-------+-------+---------+
|'min(f)|'max(f)|'sum(f)|'avg(f)|'count(f)|
+-------+-------+-------+-------+---------+
| 1.0| NaN| NaN| NaN| 2|
+-------+-------+-------+-------+---------+
{code}
cc [~yhuai] [~rxin] [~joshrosen]
> MaxFunction not working correctly with columns containing Double.NaN
> --------------------------------------------------------------------
>
> Key: SPARK-9971
> URL: https://issues.apache.org/jira/browse/SPARK-9971
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 1.4.1
> Reporter: Frank Rosner
> Priority: Minor
>
> h4. Problem Description
> When using the {{max}} function on a {{DoubleType}} column that contains
> {{Double.NaN}} values, the returned maximum value will be {{Double.NaN}}.
> This is because it compares all values with the running maximum. However, {{x
> < Double.NaN}} will always lead false for all {{x: Double}}, so will {{x >
> Double.NaN}}.
> h4. How to Reproduce
> {code}
> import org.apache.spark.sql.{SQLContext, Row}
> import org.apache.spark.sql.types._
> val sql = new SQLContext(sc)
> val rdd = sc.makeRDD(List(Row(Double.NaN), Row(-10d), Row(0d)))
> val dataFrame = sql.createDataFrame(rdd, StructType(List(
> StructField("col", DoubleType, false)
> )))
> dataFrame.select(max("col")).first
> // returns org.apache.spark.sql.Row = [NaN]
> {code}
> h4. Solution
> The {{max}} and {{min}} functions should ignore NaN values, as they are not
> numbers. If a column contains only NaN values, then the maximum and minimum
> is not defined.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
