Affects Version/s: 2.2.2, 2.3.0, 2.3.1
Component/s: Structured Streaming
Java(TM) SE Runtime Environment (build 1.8.0_112-b15)
Java HotSpot(TM) 64-Bit Server VM (build 25.112-b15, mixed mode)
- Spark 2.3.1
- Java 1.8.0 (112)
- Standalone Cluster Manager
- 3 Nodes, 1 Executor per node.
Spark 2.3.0 introduced a new mechanism for caching Kafka consumers (https://issues.apache.org/jira/browse/SPARK-23623?page=com.atlassian.jira.plugin.system.issuetabpanels%3Aall-tabpanel) via KafkaDataConsumer.acquire.
It seems that there are situations (I've been trying to debug it, haven't been able to find the root cause as of yet) where cached consumers remain "in use" throughout the life time of the task and are never released. This can be identified by the following line of the stack trace:
Which points to:
Meaning the existing consumer created for that `TopicPartition` is still in use for some reason. The weird thing is that you can see this for very old tasks which have already finished successfully.
I've traced down this leak using file leak detector, attaching it to the running Executor JVM process. I've emitted the list of open file descriptors which you can find here, and you can see that the majority of them are epoll FD used by Kafka Consumers, indicating that they aren't closing.