This past week we had quite few issues with users not being able to run Spark jobs running in YARN Cluster mode. Particularly a team that was on tight schedule used to get errors like this all the time:
java.io.IOException: Failed to send RPC 8277242275361198650 to datanode-055: java.nio.channels.ClosedChannelException
Mostly accompanied by error messages like:
org.apache.spark.SparkException: Error sending message [message = Heartbeat(9,[Lscala.Tuple2;@e47ba81,BlockManagerId(9, datanode-50 , 43381))]
ERROR Executor: Exit as unable to send heartbeats to driver more than 60 times
These errors basically mean the connection between Spark driver and executors are broken, mainly because executor is killed. This could happen because of a number of reasons:
- We realized this happens a lot more often when our cluster is too busy and has hit maximum usage. What it means is that executors are accepted to DataNodes, but they fail to acquire enough memory on the datanode and therefore get killed.
- Metaspace attempts to grow beyond the executor(JVM) memory limits, resulting in loss of executors.The best way to stop this error from appearing is to set below properties when launching Spark-Shell or submitting application using spark-submit:spark.driver.extraJavaOptions = -XX:ReservedCodeCacheSize=100M
spark.executor.extraJavaOptions = -XX:ReservedCodeCacheSize=100M
Please note that depending on your project and code, you may need to increase the values mentioned above.
- Network is slow for whatever reason. In our case, this was caused by a change in DNS which resulted in turning off caching.This case could be fixed by adjusting spark.executor.heartbeatInterval and spark.network.timeout. Default values for these 2 parameters are 10s and 120s. You can adjust these 2 values based on how your network, the only point to consider here is that the later property, spark.network.timeout, should be greater than the first one.
If none of what mentioned above helps your situation, then it is something you need to take to your cluster’s administrator. There could be something wrong with the datanodes where executors are sent to that admins are not aware of.