I ran the HdfsWordCount example using this command: spark-submit run-example \ --conf spark.streaming.dynamicAllocation.enabled=true \ --conf spark.executor.instances=0 \ --conf spark.dynamicAllocation.enabled=false \ --conf spark.master=yarn \ --conf spark.submit.deployMode=client \ org.apache.spark.examples.streaming.HdfsWordCount /foo
I tried it on both Spark 2.1.1 (through HDP 2.6) and Spark 2.2.0 (through Google Dataproc 1.2), and I get the same message repeatedly that Spark cannot allocate any executors. 17/08/22 19:34:57 INFO org.spark_project.jetty.util.log: Logging initialized @1694ms 17/08/22 19:34:57 INFO org.spark_project.jetty.server.Server: jetty-9.3.z-SNAPSHOT 17/08/22 19:34:57 INFO org.spark_project.jetty.server.Server: Started @1756ms 17/08/22 19:34:57 INFO org.spark_project.jetty.server.AbstractConnector: Started ServerConnector@578782d6{HTTP/1.1,[http/1.1]}{0.0.0.0:4040} 17/08/22 19:34:58 INFO com.google.cloud.hadoop.fs.gcs.GoogleHadoopFileSystemBase: GHFS version: 1.6.1-hadoop2 17/08/22 19:34:58 INFO org.apache.hadoop.yarn.client.RMProxy: Connecting to ResourceManager at hadoop-m/10.240.1.92:8032 17/08/22 19:35:00 INFO org.apache.hadoop.yarn.client.api.impl.YarnClientImpl: Submitted application application_1503036971561_0022 17/08/22 19:35:04 WARN org.apache.spark.streaming.StreamingContext: Dynamic Allocation is enabled for this application. Enabling Dynamic allocation for Spark Streaming applications can cause data loss if Write Ahead Log is not enabled for non-replayable sources like Flume. See the programming guide for details on how to enable the Write Ahead Log. 17/08/22 19:35:21 WARN org.apache.spark.scheduler.cluster.YarnScheduler: Initial job has not accepted any resources; check your cluster UI to ensure that workers are registered and have sufficient resources 17/08/22 19:35:36 WARN org.apache.spark.scheduler.cluster.YarnScheduler: Initial job has not accepted any resources; check your cluster UI to ensure that workers are registered and have sufficient resources 17/08/22 19:35:51 WARN org.apache.spark.scheduler.cluster.YarnScheduler: Initial job has not accepted any resources; check your cluster UI to ensure that workers are registered and have sufficient resources I confirmed that the YARN cluster has enough memory for dozens of executors, and verified that the application allocates executors when using Core's spark.dynamicAllocation.enabled=true, and leaving spark.streaming.dynamicAllocation.enabled=false. Is streaming dynamic allocation actually supported? Sean Owen suggested it might have been experimental: https://issues.apache.org/jira/browse/SPARK-21792.