[jira] [Commented] (PHOENIX-6687) The region server hosting the SYSTEM.CATALOG fails to serve any metadata requests as default handler pool threads are exhausted.

["Istvan Toth (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/PHOENIX-6687?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17523014#comment-17523014 ] 

Istvan Toth commented on PHOENIX-6687:
--------------------------------------

Looking at the code, this only happens when an old (pre 5.1/4.15) client tries to connect, is that correct ?

> The region server hosting the SYSTEM.CATALOG fails to serve any metadata requests as default handler pool  threads are exhausted.
> ---------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: PHOENIX-6687
>                 URL: https://issues.apache.org/jira/browse/PHOENIX-6687
>             Project: Phoenix
>          Issue Type: Bug
>          Components: core
>    Affects Versions: 4.16.1, 5.2.0
>            Reporter: Jacob Isaac
>            Priority: Major
>             Fix For: 4.17.0, 5.2.0
>
>         Attachments: stacktraces.txt
>
>
> When SYSTEM.CATALOG region server is restarted and the server is experiencing heavy metadata call volume.
> The stack traces indicate that all the default handler pool threads are waiting for the CQSI.init thread to finish initializing.
> The CQSI.init thread itself cannot proceed since it cannot complete the second RPC call (org.apache.phoenix.query.ConnectionQueryServicesImpl.checkClientServerCompatibility) due to thread starvation.
> For e.g
> The following [code|https://github.com/apache/phoenix/blob/3cff97087d79b85e282fca4ac69ddf499fb1f40f/phoenix-core/src/main/java/org/apache/phoenix/coprocessor/MetaDataEndpointImpl.java#L661] turned the getTable(..) into needing an additional server-to-server RPC call when initializing a PhoenixConnection (CQSI.init) for the first time on the JVM. 
> It is well-known that server-to-server RPC calls are prone to deadlocking due to thread pool exhaustion.



--
This message was sent by Atlassian Jira
(v8.20.1#820001)