hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Derek Dagit (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-256) Increase retention of applications in AppManager without sending more applications to the UI
Date Tue, 04 Dec 2012 00:01:58 GMT

    [ https://issues.apache.org/jira/browse/YARN-256?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13509348#comment-13509348

Derek Dagit commented on YARN-256:

One approach would be to add a new configuration variable to control limit the maximum number
of applications used in the web UI presentation.  With such a configuration variable set to
N applications, the user would see, more or less, the last N applications that were submitted
when browsing.

Implementing the use of such a configuration has challenges:

The RMAppManager has member that implements RMContext for the purpose of concurrent accesses
to pieces of the RM state.  The RMContext interface defines a method getRMApps() that returns
a java.util.concurrent.ConcurrentMap (with ConcurrentHashMap as its implementation) holding
the mapping of ApplicationId to RMApps.

Given that we want to return the newest "N" RMApps, we would need do walk the entire map since
there is no ordering of keys.

A couple of strategies and some cons:

1) Implement a second data structure that maintains the order the ConcurrentHashMap lacks.

- Maintenance of the two separate structures in a concurrent environment could be nasty.

2) Change the map to a data structure that supports fast deletions, updates, and retrieval
while maintaining ordering.

- No provided, concurrent structures exists that have these qualities, so more work.
- Locking would need to be done by the caller.

3) Encapsulate the map behind a set of method calls

- Large scope of code change.

One other thing to note: With the current implementation, there are cases in which we walk
the elements in the map without locking.  ConcurrentHashMap does not guarantee that iterators
remain consistent to changes in structure after the ithe iterators are created.  Practically,
this means that if an RMApp is removed from the ConcurrentHashMap while we are walking it,
then there is a possibility we may crash.  (ConcurrentHashMap will not throw a ConcurrentModificationException.)


When collections are returned from .values(), these Collections are backed by the list.  We
do such a thing in several places currently.

- GetAllApplicationsResponse#getAllApplications()
- RMWebServices#getApps()
- ClientRMService#getQueueInfo()
- AppsList#toDataTableArrays()

I am leaning toward option 3) above, and what I would want to do is something like the following:

- Remove the map from the RMContext

- Add a LinkedHashMap private to the RMContext

- Remove getRMApps() from RMContext() 

- Add to RMContext getRMAppForAppId(ApplicationId) -> returns the desired app

Use a visitor pattern to do proper locking and hide the underlying data structure:

- Add to RMContext acceptVisitor(RMAppsVisitor) -> execute logic on each RMApp with proper

- Provide the RMAppsVisitor interface.

- Change all getRMApps().get(ApplicationId) -> getRMAppForAppId(ApplicationId)

- Where there are walks over map values, pass in loop logic as an RMAppsVisitor to visitRMApps()

At this point I am looking for input, and I would appreciate any comments.

> Increase retention of applications in AppManager without sending more applications to
the UI
> --------------------------------------------------------------------------------------------
>                 Key: YARN-256
>                 URL: https://issues.apache.org/jira/browse/YARN-256
>             Project: Hadoop YARN
>          Issue Type: Improvement
>    Affects Versions: 0.23.5
>            Reporter: Derek Dagit
>            Assignee: Derek Dagit
> In very busy clusters we would like to retain applications longer so that users' links
will not expire too soon.  Very often links to application history expire before they can
be followed.
> Simply increasing max-completed applications has an adverse performance impact on the
applications list in the web UI because it presents the entire list of applications with a
> Therefore, we would like some way to be able to increase the retention of applications
without increasing the number of applications sent to the Web UI.

This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

View raw message