[
https://issues.apache.org/jira/browse/PHOENIX-6671?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Lars Hofhansl updated PHOENIX-6671:
-----------------------------------
Description:
See PHOENIX-6501, PHOENIX-6458, and HBASE-26812.
HBase's ShortCircuit Connection are fundamentally broken in HBase 2. We might
be able to fix it there, but with all the work the RPC handlers perform now
(closing scanning, resolving current user, etc), I doubt we'll get that 100%
right. HBase 3 has removed this functionality.
Even with HBase, which does not have the async protobuf code, I could hardly
see any performance improvement from circumventing the RPC stack in case the
target of a Get or Scan is local. Even in the most ideal conditions where
everything is local, there was improvement outside of noise.
I suggest we do not use ShortCircuited Connections in Phoenix 5+.
was:
See PHOENIX-6501, PHOENIX-6458, and HBASE-26812.
HBase's ShortCircuit Connection are fundamentally broken in HBase 2. We might
be able to fix it there, but with all the work the RPC handlers perform now
(closing scanning, resolving current user, etc), I doubt we'll get that 100%
right. HBase 3 has removed this functionality.
Even with HBase, which does not have the async protobuf code, I could hardly
see any performance improvement from circumventing the RPC stack in case the
target of a Get or Scan is local. Even in the most ideal conditions where
everything is local, there was improvement outside of noise.
I suggest we do not use ShortCircuited Connections.
> Avoid ShortCirtuation Coprocessor Connection with HBase 2.x
> -----------------------------------------------------------
>
> Key: PHOENIX-6671
> URL: https://issues.apache.org/jira/browse/PHOENIX-6671
> Project: Phoenix
> Issue Type: Bug
> Reporter: Lars Hofhansl
> Priority: Major
>
> See PHOENIX-6501, PHOENIX-6458, and HBASE-26812.
> HBase's ShortCircuit Connection are fundamentally broken in HBase 2. We might
> be able to fix it there, but with all the work the RPC handlers perform now
> (closing scanning, resolving current user, etc), I doubt we'll get that 100%
> right. HBase 3 has removed this functionality.
> Even with HBase, which does not have the async protobuf code, I could hardly
> see any performance improvement from circumventing the RPC stack in case the
> target of a Get or Scan is local. Even in the most ideal conditions where
> everything is local, there was improvement outside of noise.
> I suggest we do not use ShortCircuited Connections in Phoenix 5+.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
