[MAPREDUCE-5863] Killing task attempts while speculation is enabled can cause the job to fail

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.4.0
• Assignee: Unassigned
• Reporter: Mingzhe Hao
• Created: 27/Apr/14 04:22
• Updated: 28/Apr/14 02:01

Description

There could be a race condition when a T_ADD_SPEC_ATTEMPT is being fired, the task gets succeeded and then killed by the client. In that case, the task state changes from SUCCEEDED to SCHEDULED, and then task gets a T_ADD_SPEC_ATTEMPT event, which is invalid for SCHEDULED state.

1. Task is running.
2. Speculator fires a T_ADD_SPEC_ATTEMPT
3. Before task receives T_ADD_SPEC_ATTEMPT, it succeeds
4. Succeeded TA receives TA_KILL from client. Now the task is at SCHEDULED state.
5. Task receives T_ADD_SPEC_ATTEMPT, since this is an unexpected event, the job fails.