[
https://issues.apache.org/jira/browse/YARN-4665?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15128768#comment-15128768
]
Jason Lowe commented on YARN-4665:
----------------------------------
As Rohith points out, I believe the intent of YARN-514 is to handle this mostly
on the client side. Technically until the application gets into the ACCEPTED
state (or later) the submission process has not completed, and the YARN client
code will not return back to the application until that has completed.
Failures besides ApplicationNotFoundException should be automatically retried
by the retry proxy later, and the code snippet above should handle the case
where the new RM doesn't understand the application being queried because it
failed to make it into the state store.
Could you elaborate one or more specific scenarios that aren't handled
properly? I'm curious to know the sequence of events that occurs during the
failover that makes the user code believe the application submission was
successful (i.e.: the YARN client submit code returned successfully to the
caller), and how that same error doesn't occur if the call was made synchronous.
> Asynch submit can lose application submissions
> ----------------------------------------------
>
> Key: YARN-4665
> URL: https://issues.apache.org/jira/browse/YARN-4665
> Project: Hadoop YARN
> Issue Type: Bug
> Affects Versions: 2.1.0-beta
> Reporter: Daniel Templeton
> Assignee: Daniel Templeton
>
> The change introduced in YARN-514 opens up a hole into which applications can
> fall and be lost. Prior to YARN-514, the {{submitApplication()}} call did
> not complete until the application state was persisted to the state store.
> After YARN-514, the {{submitApplication()}} call is asynchronous, with the
> application state being saved later.
> If the state store is slow or unresponsive, it may be that an application's
> state may not be persisted for quite a while. During that time, if the RM
> fails (over), all applications that have not yet been persisted to the state
> store will be lost. If the active RM loses ZK connectivity, a significant
> number of job submissions can pile up before the ZK connection times out,
> resulting in a large pile of client failures when it finally does.
> This issue is inherent in the design of YARN-514. I see three solutions:
> 1. Add a WAL to the state store. HBase does it, so we know how to do it. It
> seems like a heavy solution to the original problem, however. It's certainly
> not a trivial change.
> 2. Revert YARN-514 and update the RPC layer to allow a connection to be
> parked if it's doing something that may take a while. This is a generally
> useful feature but could be a deep rabbit hole.
> 3. Revert YARN-514 and add back-pressure to the job submission. For example,
> we set a maximum number of threads that can simultaneously be assigned to
> handle job submissions. When that threshold is reached, new job submissions
> get a try-again-later response. This is also a generally useful feature and
> should be a fairly constrained set of changes.
> I think the third option is the most approachable. It's the smallest change,
> and it adds useful behavior beyond solving the original issue.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
