[
https://issues.apache.org/jira/browse/CASSANDRA-21091?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18112018#comment-18112018
]
Amal Sebastian commented on CASSANDRA-21091:
--------------------------------------------
Dug into this and found the issue. RequestsSize reads the shared
endpoint/global reserve counter, and pre-v5 dispatch hits that on every request
so it's accurate there. V5 reuses AbstractMessageHandler (shared with internode
messaging), which only touches the reserve once a connection's local queue
budget (1MiB by default) is exceeded - normal traffic never gets close to that,
so the gauge just sits at 0.
Fixed by adding hooks that fire on every acquire/release regardless of path,
feeding a dedicated counter in ClientMetrics from both v4 and v5. Backpressure
logic untouched.
PR: [https://github.com/apache/cassandra/pull/5120]
> Missing metrics with protocol v5
> --------------------------------
>
> Key: CASSANDRA-21091
> URL: https://issues.apache.org/jira/browse/CASSANDRA-21091
> Project: Apache Cassandra
> Issue Type: Bug
> Reporter: Andy Salnikov
> Priority: Normal
>
> I am testing our client-side setup switching from protocol v4 to v5 and
> monitoring Cassandra metrics. What I see is that one particular metric is
> missing (or is 0) on server side when I switch to v5, metric is type=Client
> name=RequestSize.
> Server side is Cassandra 5.0.6 with Java 17. Client side is DataStax Python
> driver latest version. When I configure client to use protocol v4, then the
> above metric is generating reasonable values. When I switch protocol to v5,
> the metric shows zeros.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
