[
https://issues.apache.org/jira/browse/HDFS-13923?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16621294#comment-16621294
]
Konstantin Shvachko commented on HDFS-13923:
--------------------------------------------
Hey [~csun], some thoughts.
* If we want to turn off observer reads globally we can simply kill
ObserverNode (or transition it to SBN) and restart it (transition back) when we
are ready to enable observer reads.
* We can configure client with {{ConfiguredProxyProvider}} in order to make
client not ever go to Observer. (Just as you mentioned).
* Here you are talking, I believe, about a way of telling the client
dynamically not to use ObserverNode. Do you?
I don't see how configuration parameter can help with that. If the client is
cached, then whatever value of a config variable stored in it during the
initial instantiation will still be there, even if you updated the config file.
So it looks more like you want a flag somewhere in ORPP, which tells the
invoker to call ActiveNN on reads no matter what, and the flag should be
updatable dynamically.
Not sure I like the idea, because it sounds like it will require adding a new
FileSystem method turning of observer reads, which exposes implementation
details with Observer reads to the user space. I would rather use options like
{{fs.hdfs.impl.disable.cache = true}} or spawn a different JVM.
> Add a configuration to turn on/off observer reads
> -------------------------------------------------
>
> Key: HDFS-13923
> URL: https://issues.apache.org/jira/browse/HDFS-13923
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Chao Sun
> Priority: Major
>
> In some situations having a config to turn on/off observer reads dynamically
> may be useful. For instance, some applications may be sensitive for data
> freshness and want to always reach directly to active NN. In a more complex
> scenario, services such as Presto may want to apply observer reads for
> different types of queries. In this case, simply change
> {{dfs.client.failover.proxy.provider.<nameservice>}} may not be enough, since
> with FileSystem cache (which is usually turned on) will ignore the change and
> still use the same FileSystem object.
> Here I'm proposing to add a flag in {{HdfsClientConfigKeys}}, such as
> {{dfs.client.observer.reads.enabled}}, that can be used to dynamically turn
> on/off observer reads. The FileSystem cache key should also take account of
> this flag in its {{hashCode}} and {{equals}} impl, so that different
> FileSystem objects will be used depending on the flag.
>
> cc [~shv], [~xkrogen], [~vagarychen], [~zero45] for discussion.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
