[SPARK-33031] scheduler with blacklisting doesn't appear to pick up new executor added

• Type: Bug
• Status: Open
• Priority: Critical
• Resolution: Unresolved
• Affects Version/s: 3.0.0, 3.1.0
• Component/s: Scheduler
• Assignee: Unassigned
• Reporter: Thomas Graves
• Created: 29/Sep/20 19:55
• Updated: 29/Jun/21 19:19

Description

I was running a test with blacklisting  standalone mode and all the executors were initially blacklisted.  Then one of the executors died and we got allocated another one. The scheduler did not appear to pick up the new one and try to schedule on it though.

You can reproduce this by starting a master and slave on a single node, then launch a shell like where you will get multiple executors (in this case I got 3)

$SPARK_HOME/bin/spark-shell --master spark://yourhost:7077 --executor-cores 4 --conf spark.blacklist.enabled=true

From shell run:

import org.apache.spark.TaskContext
val rdd = sc.makeRDD(1 to 1000, 5).mapPartitions { it =>
 val context = TaskContext.get()
 if (context.attemptNumber() < 2) {
 throw new Exception("test attempt num")
 }
 it
}
rdd.collect()

 

Note that I tried both with and without dynamic allocation enabled.

 

You can see screen shot related on https://issues.apache.org/jira/browse/SPARK-33029