[
https://issues.apache.org/jira/browse/HDFS-5369?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14041497#comment-14041497
]
Andrew Wang commented on HDFS-5369:
-----------------------------------
Hey [~eddyxu], 30s sounds okay to me, maybe even a bit longer than that (i.e. 1
or 2 min). [~kihwal] might be able to make a quick comment about this, since he
mentioned tight job SLAs in HADOOP-8088. HADOOP-8088 also mentions handling
error codes indicative of a transient error differently, so let's keep that in
mind here too.
Would also still be good to squish the stack trace if possible too, since it's
not very useful.
> Support negative caching of user-group mapping
> ----------------------------------------------
>
> Key: HDFS-5369
> URL: https://issues.apache.org/jira/browse/HDFS-5369
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: namenode
> Affects Versions: 2.2.0
> Reporter: Andrew Wang
>
> We've seen a situation at a couple of our customers where interactions from
> an unknown user leads to a high-rate of group mapping calls. In one case,
> this was happening at a rate of 450 calls per second with the shell-based
> group mapping, enough to severely impact overall namenode performance and
> also leading to large amounts of log spam (prints a stack trace each time).
> Let's consider negative caching of group mapping, as well as quashing the
> rate of this log message.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
