[SPARK-4194] Exceptions thrown during SparkContext or SparkEnv construction might lead to resource leaks or corrupted global state

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.2.0, 1.3.0
• Fix Version/s: 1.4.0
• Component/s: Spark Core
• Assignee: Marcelo Masiero Vanzin
• Reporter: Josh Rosen
• Created: 02/Nov/14 06:28
• Updated: 16/Apr/15 09:49
• Resolved: 16/Apr/15 09:48

Description

The SparkContext and SparkEnv constructors instantiate a bunch of objects that may need to be cleaned up after they're no longer needed. If an exception is thrown during SparkContext or SparkEnv construction (e.g. due to a bad configuration setting), then objects created earlier in the constructor may not be properly cleaned up.

This is unlikely to cause problems for batch jobs submitted through spark-submit, since failure to construct SparkContext will probably cause the JVM to exit, but it is a potentially serious issue in interactive environments where a user might attempt to create SparkContext with some configuration, fail due to an error, and re-attempt the creation with new settings. In this case, resources from the previous creation attempt might not have been cleaned up and could lead to confusing errors (especially if the old, leaked resources share global state with the new SparkContext).

Issue Links

• blocks: SPARK-4180 SparkContext constructor should throw exception if another SparkContext is already running
• links to: [Github] Pull Request #5335 (vanzin)

Sub-Tasks

• Clean up SparkContext after DAGScheduler initialization errors: Resolved