[ https://issues.apache.org/jira/browse/SPARK-6469?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Christophe PRÉAUD updated SPARK-6469: ------------------------------------- Description: According to the [Spark YARN doc page|http://spark.apache.org/docs/latest/running-on-yarn.html#important-notes], Spark executors will use the local directories configured for YARN, not {{spark.local.dir}} which should be ignored. It should be noted though that in yarn-client mode, though the executors will indeed use the local directories configured for YARN, the driver will not, because it is not running on the YARN cluster; the driver in yarn-client will use the local directories defined in {{spark.local.dir}} was: According to the [Spark YARN doc page|http://spark.apache.org/docs/latest/running-on-yarn.html#important-notes], Spark executors will use the local directories configured for YARN, not spark.local.dir which should be ignored. If this works correctly in yarn-cluster mode, I've found out that it is not the case in yarn-client mode. The problem seems to originate in the method [isRunningInYarnContainer|https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/util/Utils.scala#L686]. Indeed, I've checked with a simple application that the {{CONTAINER_ID}} environment variable is correctly set in yarn-cluster mode (to something like {{container_1426666761810_0151_01_000001}}, but not in yarn-client mode. > Local directories configured for YARN are not used in yarn-client mode > ---------------------------------------------------------------------- > > Key: SPARK-6469 > URL: https://issues.apache.org/jira/browse/SPARK-6469 > Project: Spark > Issue Type: Documentation > Components: Spark Core > Reporter: Christophe PRÉAUD > Priority: Minor > Attachments: TestYarnVars.scala > > > According to the [Spark YARN doc > page|http://spark.apache.org/docs/latest/running-on-yarn.html#important-notes], > Spark executors will use the local directories configured for YARN, not > {{spark.local.dir}} which should be ignored. > It should be noted though that in yarn-client mode, though the executors will > indeed use the local directories configured for YARN, the driver will not, > because it is not running on the YARN cluster; the driver in yarn-client will > use the local directories defined in {{spark.local.dir}} -- 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