[
https://issues.apache.org/jira/browse/SPARK-11720?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15004206#comment-15004206
]
Xiangrui Meng edited comment on SPARK-11720 at 11/13/15 4:25 PM:
-----------------------------------------------------------------
Computing average in the normal way should be sufficient. The precision issue
is different from moment aggregation. Using m' = m + delta/n won't help. We
need https://en.wikipedia.org/wiki/Kahan_summation_algorithm or an equivalent
algorithm to be very accurate, but it seems unnecessary in Spark.
was (Author: mengxr):
Computing average in the normal way should be sufficient. The precision issue
is different from moment aggregation. Using m_{k+1} = m_{k} + delta/n won't
help. We need https://en.wikipedia.org/wiki/Kahan_summation_algorithm or an
equivalent algorithm to be very accurate, but it seems unnecessary in Spark.
> Return Double.NaN instead of null for Mean and Average when count = 0
> ---------------------------------------------------------------------
>
> Key: SPARK-11720
> URL: https://issues.apache.org/jira/browse/SPARK-11720
> Project: Spark
> Issue Type: Sub-task
> Reporter: Jihong MA
>
> change the default behavior of mean in case of count = 0 from null to
> Double.NaN, to make it inline with all other univariate stats function.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
