[
https://issues.apache.org/jira/browse/YARN-1547?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15218681#comment-15218681
]
Vinod Kumar Vavilapalli commented on YARN-1547:
-----------------------------------------------
Looked at the updated document, thanks [~giovanni.fumarola] and [~subru]! Few
comments
bq. Keeps tracks of all the requests
In addition to the sliding window, we should have an upper limit on all
requests possible from any single application in its entire life-time. The
sliding window avoids unreasonable spikes while the across-the-lifetime limit
avoids slow abuse of unreasonable amount of resources.
bq. We can apply an additional control to check if the list-size of a single
request overflows a specific threshold.
+1, we must. To me this is an equally pressing concern compared to the spikes
in list-sizes.
It will be good if we can implement the enforcers / limiters in such a way that
we can easily reuse them as the APIs themselves change over time from lists to
maps etc.
Overall, you have covered significant ground in the latest doc. We should get
going with code etc.
Thanks for taking care of this!
> Prevent DoS of ApplicationMasterProtocol by putting in limits
> -------------------------------------------------------------
>
> Key: YARN-1547
> URL: https://issues.apache.org/jira/browse/YARN-1547
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Vinod Kumar Vavilapalli
> Assignee: Giovanni Matteo Fumarola
> Attachments: YARN-1547.pdf, YARN-1547.v0.pdf
>
>
> Points of DoS in ApplicationMasterProtocol
> - Host and trackingURL in RegisterApplicationMasterRequest
> - Diagnostics, final trackingURL in FinishApplicationMasterRequest
> - Unlimited number of resourceAsks, containersToBeReleased and
> resourceBlacklistRequest in AllocateRequest
> -- Unbounded number of priorities and/or resourceRequests in each ask.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
