Sean Mackrory created HDFS-11077:
------------------------------------
Summary: RequestHedgingProxyProvider does not behave correctly
with more than 3+ NameNodes
Key: HDFS-11077
URL: https://issues.apache.org/jira/browse/HDFS-11077
Project: Hadoop HDFS
Issue Type: Bug
Reporter: Sean Mackrory
RequestHedgingProxyProvider will send its initial request to all NameNodes. It
is meant to then detect the active, and send subsequent requests to the active
until it fails over.
It does this by remember the last NameNode to fail, but unless the set of
NameNodes minus the last to fail is a set of size 1, it will continue to use a
thread pool and send multiple requests at a time. If there are 3 or more
NameNodes, however, that means a thread pool and multiple requests will always
be sent.
The correct logic in the case of more than 2 NameNodes is probably to use the
last known active and then when it fails, to reset the state, query all of them
in parallell, and then use the one the one that looks like the active.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-dev-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-dev-h...@hadoop.apache.org
