[
https://issues.apache.org/jira/browse/SPARK-2641?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Kanwaljit Singh updated SPARK-2641:
-----------------------------------
Description:
When running spark-submit in Yarn cluster mode, we provide properties file
using --properties-file option.
spark.executor.instances=5
spark.executor.memory=2120m
spark.executor.cores=3
The running job picks up the cores and memory, but not the correct instances.
I think the issue is here in org.apache.spark.deploy.SparkSubmitArguments:
// Use properties file as fallback for values which have a direct analog to
// arguments in this script.
master =
Option(master).getOrElse(defaultProperties.get("spark.master").orNull)
executorMemory = Option(executorMemory)
.getOrElse(defaultProperties.get("spark.executor.memory").orNull)
executorCores = Option(executorCores)
.getOrElse(defaultProperties.get("spark.executor.cores").orNull)
totalExecutorCores = Option(totalExecutorCores)
.getOrElse(defaultProperties.get("spark.cores.max").orNull)
name =
Option(name).getOrElse(defaultProperties.get("spark.app.name").orNull)
jars = Option(jars).getOrElse(defaultProperties.get("spark.jars").orNull)
Along with these defaults, we should also set default for instances:
numExecutors=Option(numExecutors).getOrElse(defaultProperties.get("spark.executor.instances").orNull)
> Spark submit doesn't pick up executor instances from properties file
> --------------------------------------------------------------------
>
> Key: SPARK-2641
> URL: https://issues.apache.org/jira/browse/SPARK-2641
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
> Affects Versions: 1.0.0
> Reporter: Kanwaljit Singh
>
> When running spark-submit in Yarn cluster mode, we provide properties file
> using --properties-file option.
> spark.executor.instances=5
> spark.executor.memory=2120m
> spark.executor.cores=3
> The running job picks up the cores and memory, but not the correct instances.
> I think the issue is here in org.apache.spark.deploy.SparkSubmitArguments:
> // Use properties file as fallback for values which have a direct analog to
> // arguments in this script.
> master =
> Option(master).getOrElse(defaultProperties.get("spark.master").orNull)
> executorMemory = Option(executorMemory)
> .getOrElse(defaultProperties.get("spark.executor.memory").orNull)
> executorCores = Option(executorCores)
> .getOrElse(defaultProperties.get("spark.executor.cores").orNull)
> totalExecutorCores = Option(totalExecutorCores)
> .getOrElse(defaultProperties.get("spark.cores.max").orNull)
> name =
> Option(name).getOrElse(defaultProperties.get("spark.app.name").orNull)
> jars = Option(jars).getOrElse(defaultProperties.get("spark.jars").orNull)
> Along with these defaults, we should also set default for instances:
> numExecutors=Option(numExecutors).getOrElse(defaultProperties.get("spark.executor.instances").orNull)
--
This message was sent by Atlassian JIRA
(v6.2#6252)
