[FLINK-4150] Problem with Blobstore in Yarn HA setting on recovery after cluster shutdown

• Type: Bug
• Status: Closed
• Priority: Blocker
• Resolution: Fixed
• Fix Version/s: 1.1.0
• Component/s: Runtime / Coordination
• Assignee: Ufuk Celebi
• Reporter: Stefan Richter
• Created: 04/Jul/16 16:13
• Updated: 28/Feb/19 13:30
• Resolved: 25/Jul/16 13:21

Description

Submitting a job in Yarn with HA can lead to the following exception:

org.apache.flink.streaming.runtime.tasks.StreamTaskException: Cannot load user class: org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumer09
ClassLoader info: URL ClassLoader:
    file: '/tmp/blobStore-ccec0f4a-3e07-455f-945b-4fcd08f5bac1/cache/blob_7fafffe9595cd06aff213b81b5da7b1682e1d6b0' (invalid JAR: zip file is empty)
Class not resolvable through given classloader.
	at org.apache.flink.streaming.api.graph.StreamConfig.getStreamOperator(StreamConfig.java:207)
	at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:222)
	at org.apache.flink.runtime.taskmanager.Task.run(Task.java:588)
	at java.lang.Thread.run(Thread.java:745)

Some job information, including the Blob ids, are stored in Zookeeper. The actual Blobs are stored in a dedicated BlobStore, if the recovery mode is set to Zookeeper. This BlobStore is typically located in a FS like HDFS. When the cluster is shut down, the path for the BlobStore is deleted. When the cluster is then restarted, recovering jobs cannot restore because it's Blob ids stored in Zookeeper now point to deleted files.

Issue Links

• is duplicated by: FLINK-4182 HA recovery not working properly under ApplicationMaster failures.
• relates to: FLINK-4166 Generate automatic different namespaces in Zookeeper for Flink applications
• links to: GitHub Pull Request #2256