Github user mindprince commented on the pull request:
https://github.com/apache/spark/pull/5233#issuecomment-87160289
Hi @sryza
We faced an issue where not doing this was causing incorrect accounting of
the queue's AM Resources.
As you may be aware that the YARN fair scheduler has a queue property
called `maxAMShare` - which limits the fraction of the queue's fair share that
can be used to run application masters.
There is a variable called `amResourceUsage` which tracks the usage of AMs
in a queue.
[This is the
block](https://github.com/apache/hadoop/blob/release-2.6.0/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/FSAppAttempt.java#L525-L529)
in `FSAppAttempt.java` which increases `amResourceUsage`:
```
if (getLiveContainers().size() == 1 && !getUnmanagedAM()) {
getQueue().addAMResourceUsage(container.getResource());
setAmRunning(true);
}
```
[The following
block](https://github.com/apache/hadoop/blob/release-2.6.0/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/FSLeafQueue.java#L91-L96)
in `FSLeafQueue.java` decreases `amResourceUsage`:
```
public boolean removeApp(FSAppAttempt app) {
if (runnableApps.remove(app)) {
// Update AM resource usage
if (app.isAmRunning() && app.getAMResource() != null) {
Resources.subtractFrom(amResourceUsage, app.getAMResource());
}
.
.
```
So one issue (in my opinion) on YARN's side is that it is using
`getLiveContainers().size()` to detect whether a container is an AM or not.
Now to main issue - the spark AM goes down without releasing the
containers. So, this is what I observed stepping through the code:
```
Spark AM starts: request 1024MB
That is allocated.
numLiveContainers for this app = 1.
amResourceUsage for queue = 1024MB
Spark executor starts: request1 6144MB
This is allocated.
numLiveContainers for this app = 2.
amResourceUsage is not changed.
Spark Command requests another container: request2 6144MB
Assume for some reason this is not allocated. (Insufficient resources in
the cluster.)
Now, I see that a container with memory = 1024MB is completed. (This is the
AM container.)
numLiveContainers = 1
A container with memory = 6144MB is completed.
numLiveContainers = 0
The request2 for 6144MB which was unsuccessful earlier is now successful.
numLiveContainers = 1
amResourceUsage = 1024+6144 (This shouldn't have happened but happens
because live containers = 1.)
Now app is complete and removeApp is called.
It sets amResourceUsage to 7168 - 1024 = 6144. Because the AM for spark app
took 1024.
```
After a couple of spark applications are launched, the cluster becomes
deadlocked - because no more AMs are admitted in the queue as the
amResourceUsage is incorrectly too high.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
