Jonathan Hurley created AMBARI-20712:
----------------------------------------
Summary: Parallel Requests With Intersecting Hosts Don't Block
Correctly
Key: AMBARI-20712
URL: https://issues.apache.org/jira/browse/AMBARI-20712
Project: Ambari
Issue Type: Bug
Affects Versions: 2.5.1
Reporter: Jonathan Hurley
Assignee: Jonathan Hurley
Priority: Blocker
Fix For: 2.5.1
AMBARI-20646 introduced a regression which can be seen when deploying a cluster
via blueprint with Kerberization. The problem is caused by the change which was
made to only pull in the most recent stage in progress per request.
There was logic which prevented concurrently executing request stages to be
scheduled if a prior request had an un-executed stage with a conflicting host.
For example:
- Request 1
-- Stage 1 (IN_PROGRESS)
--- Host 1
--- Host 2
--- Host 3
-- Stage 2 (PENDING)
--- Host 4
- Request 2
-- Stage 1
--- Host 4
In the above scenario, the scheduler was tricked into thinking that Request 2
can run even though Host-4 has a conflict. This is because it was only looking
at the stage in progress.
We can't simply look at all stages in progress since this would cause the
performance bug to manifest again. So, the solution is to have the database
tell us which hosts are blocking from prior requests. This SQL should be very
fast and will only run in the specific cases where there are multiple
concurrent requests (which is typically at blueprint time).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
