Maxim Khutornenko created AURORA-686:
----------------------------------------
Summary: Job updates may fail due to exceeding role quota
Key: AURORA-686
URL: https://issues.apache.org/jira/browse/AURORA-686
Project: Aurora
Issue Type: Story
Components: Scheduler
Reporter: Maxim Khutornenko
Current way of checking job quota during in-flight updates (i.e. within
addInstance transaction) may lead to failed updates and inferior user
experience. Since we are tracking quota at the role level but the update lock
applied at the job level, there is always a possibility to exceed the allowed
quota for long running updates.
This is especially a problem with the server side-driven process where a
resumed update will restart in a potentially quite different quota environment
(i.e. due to other jobs created while the update was paused).
Possible solutions:
- per job quota tracking - requires significant refactoring;
- hierarchical locking (e.g. add role lock in addition to job lock) - limits
update concurrency per role;
- front-loaded consumption (e.g. add additional job consumption during
startJobUpdate and re-evaluate on update completion/termination) - will require
persisting front-loaded value within job update schema but may be the way to go
given current quota implementation.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
