We are running a local mode Spark application that runs a spark job every 5 minutes using a singleton SparkContext.
- We are using Spark 220.127.116.11.4.3.0-227
- The application is long running and "sleeps" in-between the jobs
- We are using SparkContext.getOrCreate
- We are running spark in "local[*]" mode
- "spark.worker.cleanup.enabled" is set to "true"
- The application is written in Scala
- On failure we are invoking the Spark Context "stop" method in order to get a healthy SparkContext for the next job.
The "spark.local.dir" directory is filling up over time, and we eventually get "java.io.IOException: No space left on device".
We found an old Jira ticket mentioning the issue (https://issues.apache.org/jira/browse/SPARK-7439), but it seems it was closed with the motivation "the dirs should already be cleaned up on JVM exit".
Are we closing the spark context here ? Usually a ".close()" call is done, the JVM should be able to clean up those directories .
We are using a singleton spark context (SparkContext.getOrCreate). When the business logic fails we call ".stop()" (close is not available) on it to make sure a new one is created for the next run.