[
https://issues.apache.org/jira/browse/SPARK-51666?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Weichen Xu updated SPARK-51666:
-------------------------------
Description:
Fix sparkStageCompleted executorRunTime metric calculation:
In case of when a spark task uses multiple CPU’s, the CPU seconds should
capture the total execution seconds across all CPU’s. i.e. if a stage set
cpus-of-task to be 48, if we used 10 seconds on each CPU, the total CPU seconds
for that stage should be 10 seconds X 1 Tasks X 48 CPU = 480 CPU-seconds. If
another task only used 1 CPU then its total CPU seconds is 10 seconds X 1 CPU =
10 CPU-Seconds.
*This is very important fix since spark introduces stage level scheduling (so
that different stage tasks are configured with different number of CPUs) ,
without this fix, in stage level scheduling case, the metric calculation is
wrong.*
was:
Fix sparkStageCompleted executorRunTime metric calculation:
In case of when a spark task uses multiple CPU’s, the CPU seconds should
capture the total execution seconds across all CPU’s. i.e. if a stage set
cpus-of-task to be 48, if we used 10 seconds on each CPU, the total CPU seconds
for that stage should be 10 seconds X 1 Tasks X 48 CPU = 480 CPU-seconds. If
another task only used 1 CPU then its total CPU seconds is 10 seconds X 1 CPU =
10 CPU-Seconds.
*This is very important fix since spark introduces stage level scheduling (so
that different stage tasks are configured with different number of CPUs) ,
without this fix, the data pipeline revenue calculation spreads DBUs evenly
across these tasks.*
> Fix sparkStageCompleted executorRunTime metric calculation
> -----------------------------------------------------------
>
> Key: SPARK-51666
> URL: https://issues.apache.org/jira/browse/SPARK-51666
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
> Affects Versions: 4.1.0
> Reporter: Weichen Xu
> Priority: Major
> Labels: pull-request-available
>
> Fix sparkStageCompleted executorRunTime metric calculation:
> In case of when a spark task uses multiple CPU’s, the CPU seconds should
> capture the total execution seconds across all CPU’s. i.e. if a stage set
> cpus-of-task to be 48, if we used 10 seconds on each CPU, the total CPU
> seconds for that stage should be 10 seconds X 1 Tasks X 48 CPU = 480
> CPU-seconds. If another task only used 1 CPU then its total CPU seconds is 10
> seconds X 1 CPU = 10 CPU-Seconds.
> *This is very important fix since spark introduces stage level scheduling (so
> that different stage tasks are configured with different number of CPUs) ,
> without this fix, in stage level scheduling case, the metric calculation is
> wrong.*
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
