[
https://issues.apache.org/jira/browse/SOLR-14401?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17479098#comment-17479098
]
David Smiley commented on SOLR-14401:
-------------------------------------
I'll try and be more clear for everyone's sake.
http://localhost:8983/solr/admin/metrics?group=core
Today, _every_ request handler has 4 metrics affected by this decision (using
/admin/mbeans as an example):
* "ADMIN./admin/mbeans.distrib.requestTimes" (a histogram)
* "ADMIN./admin/mbeans.distrib.totalTime"
* "ADMIN./admin/mbeans.local.requestTimes" (a histogram)
* "ADMIN./admin/mbeans.local.totalTime"
The choice of the name ".distrib" and ".local." was already made in SOLR-13979;
we are revisiting all this now. It's not 100% obvious but basically,
".distrib." is for requests coming from a client, and that which may be
distributed/sharded out to other requests (if necessary). The ".local." is
only for a sharded request (solr-to-solr). The client identity isn't truly
part of the algorithm but I find it best to explain it this way.
Henceforth what _should_ it be? I had proposed that ".distrib." stays but as
you seemed to indicate in your question, it could be stripped out, which I'm +1
to. I do think it's clearer stripped out. Downside is that there is more
disruption to users who have gotten used to Solr 8's naming choice (okay in
9.0). I think most consumers are unlikely consuming ".local." and so
back-compat is even less of a concern there. For /select and other affected
handlers, it seems we agree on ".shard." instead.
bq. SearchHandler, ExportHandler, StreamHandler, and UpdateHandler
Yeah. ExportHandler subclasses SearchHandler so it'd get this automatically.
I didn't notice StreamHandler at first. UpdateHandler might need special code
for this because it doesn't use the "distrib" param, which is how the
underlying algorithm decides this.
> "distrib" request handler metrics should only be tracked on pertinent handlers
> ------------------------------------------------------------------------------
>
> Key: SOLR-14401
> URL: https://issues.apache.org/jira/browse/SOLR-14401
> Project: Solr
> Issue Type: Improvement
> Components: metrics
> Reporter: David Smiley
> Priority: Major
>
> SOLR-13979 in 8.4 added separate request handler metrics for distributed
> requests. However this was done for _all_ request handlers, even though it's
> mainly SearchHandler (and maybe one or two others?) where a distributed
> request is even possible. I refer to this as "metrics pollution" and it's a
> bad thing. It's more weight per handler (latency load & memory), more weight
> for Solr metrics responses, and it's also _suggestive_ that all registered
> handlers can have distributed requests when this is quite false, thus
> confusing people.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
