Details
Critical Description
Run spark-shell like this:
spark-shell --conf spark.shuffle.service.enabled=true \ --conf spark.dynamicAllocation.enabled=true \ --conf spark.dynamicAllocation.minExecutors=1 \ --conf spark.dynamicAllocation.maxExecutors=20 \ --conf spark.dynamicAllocation.executorIdleTimeout=10 \ --conf spark.dynamicAllocation.schedulerBacklogTimeout=5 \ --conf spark.dynamicAllocation.sustainedSchedulerBacklogTimeout=5
Then run this simple test:
scala> val verySmallRdd = sc.parallelize(1 to 10, 10).map { i =>
| if (i % 2 == 0) { Thread.sleep(30 * 1000); i } else 0
| }
verySmallRdd: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[1] at map at <console>:21
scala> verySmallRdd.collect()
When Spark starts ramping down the number of allocated executors, it will hit an assert in YarnAllocator.scala:
assert(targetNumExecutors >= 0, "Allocator killed more executors than are allocated!")
This assert will cause the akka backend to die, but not the AM itself. So the app will be in a zombie-like state, where the driver is alive but can't talk to the AM. Sadness ensues.
I have a working fix, just need to add unit tests. Stay tuned.
Thanks to wypoon for finding the problem, and for the test case.
Attachments
Issue Links
- links to
