[
https://issues.apache.org/jira/browse/GOBBLIN-1728?focusedWorklogId=818542&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-818542
]
ASF GitHub Bot logged work on GOBBLIN-1728:
-------------------------------------------
Author: ASF GitHub Bot
Created on: 19/Oct/22 17:28
Start Date: 19/Oct/22 17:28
Worklog Time Spent: 10m
Work Description: homatthew commented on code in PR #3586:
URL: https://github.com/apache/gobblin/pull/3586#discussion_r999752146
##########
gobblin-yarn/src/main/java/org/apache/gobblin/yarn/YarnService.java:
##########
@@ -474,11 +477,21 @@ public synchronized void
requestTargetNumberOfContainers(YarnContainerRequestBun
for (Map.Entry<String, Integer> entry :
yarnContainerRequestBundle.getHelixTagContainerCountMap().entrySet()) {
String currentHelixTag = entry.getKey();
int desiredContainerCount = entry.getValue();
+ Resource resourceForHelixTag =
yarnContainerRequestBundle.getHelixTagResourceMap().get(currentHelixTag);
+
// Calculate requested container count based on adding allocated count
and outstanding ContainerRequests in Yarn
- int requestedContainerCount =
allocatedContainerCountMap.getOrDefault(currentHelixTag, 0)
- +
getMatchingRequestsCount(yarnContainerRequestBundle.getHelixTagResourceMap().get(currentHelixTag));
- for(; requestedContainerCount < desiredContainerCount;
requestedContainerCount++) {
- requestContainer(Optional.absent(),
yarnContainerRequestBundle.getHelixTagResourceMap().get(currentHelixTag));
+ allocatedContainerCountMap.putIfAbsent(currentHelixTag, new
AtomicInteger(0));
Review Comment:
The main logical change is changing the value type from `Integer` to
`AtomicInteger`. This prevents potential desync between our
allocatedContainerCountMap (HelixTag -> numCountainers) and the containerMap
(containerId -> Container POJO) due to race condition.
I suspect the incorrect value for allocated container map is one of the
reasons we saw incorrect behavior with shrink + allocating. This edge case
would be generally rare and that's why we only start to see issues after a
large number of days and the issue is resolved after a restart.
Issue Time Tracking
-------------------
Worklog Id: (was: 818542)
Time Spent: 0.5h (was: 20m)
> Yarn Service requests too many containers due to improper calculation
> ---------------------------------------------------------------------
>
> Key: GOBBLIN-1728
> URL: https://issues.apache.org/jira/browse/GOBBLIN-1728
> Project: Apache Gobblin
> Issue Type: New Feature
> Reporter: Matthew Ho
> Priority: Major
> Time Spent: 0.5h
> Remaining Estimate: 0h
>
> Yarn Service is responsible for calculating the number of instances based on
> the helix tasks. Yarn service tracks the number of instances by asking Yarn
> for the number of resource requests and the number of allocated containers.
>
> It uses this count to determine if it should ask for more containers or
> shrink the number of containers. This calculation is currently done
> improperly and we continue to request containers when we have enough
> requested.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
