[STORM-112] Race condition between Topology Kill and Worker Timeout can crash supervisor

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: storm-core
• Assignee: Unassigned
• Reporter: James Xu
• Created: 14/Dec/13 16:04
• Updated: 09/Oct/15 00:19

Description

Recently during testing on a single node cluster we saw a supervisor crash when a topology was killed. The supervisor came back up and recovered, so it was not that big of a deal, but when we dug into it, it appears that there is a race.
https://github.com/nathanmarz/storm/issues/656

When a topology is killed the local assignments are reset, and then stormconf.ser is deleted right away. But at the same time sync-process may already be running with old state indicating that a worker timed out and needs to be relaunched. launch-worker then tries to read in the topology conf which was deleted and crashes.

The following is a sanitized version of the supervisor log that shows this happening.
https://gist.github.com/revans2/6282830