Kuai Yu created GOBBLIN-661:
-------------------------------
Summary: Prevent jobs resubmission after manager failure
Key: GOBBLIN-661
URL: https://issues.apache.org/jira/browse/GOBBLIN-661
Project: Apache Gobblin
Issue Type: Improvement
Reporter: Kuai Yu
Assignee: Kuai Yu
In gobblin cluster, if manager failed and relaunched, all the jobs persisted in
the job catalog will be relaunched. This can cause a few issues:
1) Scalability issue: because the unfinished job might be submitted at
different point of time, now if all of them are submitted at the same time, it
can cause a performance issue.
2) Waste effort: because the unfinished job now needs to be deleted, we have to
kill the existing running job, and resubmit.
In this change, we improve both 1) and 2)
1) In taskdriver mode, we will delete the job spec once we submit to Helix,
because we believe Helix is durable and all the jobs submitted wont' be lost,
so that we can safely delete the job specs. Next reboot manager won't see those
deleted job spec, thus no resubmission is needed.
2) In taskdriver mode, we will cleanup Helix running jobs. If it is a planning
job, we won't delete it. Instead we just let it run to the end.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
