You are viewing a plain text version of this content. The canonical link for it is here.
Posted to solr-user@lucene.apache.org by Stephen Lewis Bianamara <st...@gmail.com> on 2021/03/02 22:47:21 UTC
Caffeine Cache Metrics Broken?
Hi SOLR Community,
I'm investigating a weird behavior I've observed in the admin page for
caffeine cache metrics. It looks to me like on the older caches, warm-up
queries were not counted toward hit/miss ratios, which of course makes
sense, but on Caffeine cache it looks like they are. I'm using solr 8.3.
Obviously this makes measuring its true impact a little tough. Is this by
any chance a known issue and already fixed in later versions?
Thanks!
Stephen
Re: Caffeine Cache Metrics Broken?
Posted by Shawn Heisey <ap...@elyograg.org>.
On 3/8/2021 11:06 PM, Stephen Lewis Bianamara wrote:
> So if you change cache types from a cache you've had for a long time to a
> new one, the cumulative numbers may seem oddly low. In that case it would
> be best to perform a reset of the process in your solr cluster for accurate
> metrics.
If you change cache types, you have to restart Solr or reload the core
in order to actually make the change active. Restart or reload will
reset all counters to zero.
Thanks,
Shawn
Re: Caffeine Cache Metrics Broken?
Posted by Stephen Lewis Bianamara <st...@gmail.com>.
Okay, I have a hypothesis after some testing. It seems like what might be
happening is that the cumulative stats are computed per process and per
cache type.
So if you change cache types from a cache you've had for a long time to a
new one, the cumulative numbers may seem oddly low. In that case it would
be best to perform a reset of the process in your solr cluster for accurate
metrics.
Does that seem right to you?
Thanks,
Stephen
On Fri, Mar 5, 2021, 10:32 AM Stephen Lewis Bianamara <
stephen.bianamara@gmail.com> wrote:
> Thanks Shawn. Something seems different between the two because Caffeine
> Cache is having much higher volume per hour than our previous
> implementation was. So I guess it is then more likely that it is something
> actually expected due to a change in what is getting kept/warmed, so I'll
> look into this more and get back to you if that doesn't end up making sense
> based on what I observe.
>
> Thanks again,
> Stephen
>
> On Tue, Mar 2, 2021 at 6:35 PM Shawn Heisey <ap...@elyograg.org> wrote:
>
>> On 3/2/2021 3:47 PM, Stephen Lewis Bianamara wrote:
>> > I'm investigating a weird behavior I've observed in the admin page for
>> > caffeine cache metrics. It looks to me like on the older caches, warm-up
>> > queries were not counted toward hit/miss ratios, which of course makes
>> > sense, but on Caffeine cache it looks like they are. I'm using solr 8.3.
>> >
>> > Obviously this makes measuring its true impact a little tough. Is this
>> by
>> > any chance a known issue and already fixed in later versions?
>>
>> The earlier cache implementations are entirely native to Solr -- all the
>> source code is include in the Solr codebase.
>>
>> Caffeine is a third-party cache implementation that has been integrated
>> into Solr. Some of the metrics might come directly from Caffeine, not
>> Solr code.
>>
>> I would expect warming queries to be counted on any of the cache
>> implementations. One of the reasons that the warming capability exists
>> is to pre-populate the caches before actual queries begin. If warming
>> queries are somehow excluded, then the cache metrics would not be correct.
>>
>> I looked into the code and did not find anything that would keep warming
>> queries from affecting stats. But it is always possible that I just
>> didn't know what to look for.
>>
>> In the master branch (Solr 9.0), CaffeineCache is currently the only
>> implementation available.
>>
>> Thanks,
>> Shawn
>>
>
Re: Caffeine Cache Metrics Broken?
Posted by Stephen Lewis Bianamara <st...@gmail.com>.
Thanks Shawn. Something seems different between the two because Caffeine
Cache is having much higher volume per hour than our previous
implementation was. So I guess it is then more likely that it is something
actually expected due to a change in what is getting kept/warmed, so I'll
look into this more and get back to you if that doesn't end up making sense
based on what I observe.
Thanks again,
Stephen
On Tue, Mar 2, 2021 at 6:35 PM Shawn Heisey <ap...@elyograg.org> wrote:
> On 3/2/2021 3:47 PM, Stephen Lewis Bianamara wrote:
> > I'm investigating a weird behavior I've observed in the admin page for
> > caffeine cache metrics. It looks to me like on the older caches, warm-up
> > queries were not counted toward hit/miss ratios, which of course makes
> > sense, but on Caffeine cache it looks like they are. I'm using solr 8.3.
> >
> > Obviously this makes measuring its true impact a little tough. Is this by
> > any chance a known issue and already fixed in later versions?
>
> The earlier cache implementations are entirely native to Solr -- all the
> source code is include in the Solr codebase.
>
> Caffeine is a third-party cache implementation that has been integrated
> into Solr. Some of the metrics might come directly from Caffeine, not
> Solr code.
>
> I would expect warming queries to be counted on any of the cache
> implementations. One of the reasons that the warming capability exists
> is to pre-populate the caches before actual queries begin. If warming
> queries are somehow excluded, then the cache metrics would not be correct.
>
> I looked into the code and did not find anything that would keep warming
> queries from affecting stats. But it is always possible that I just
> didn't know what to look for.
>
> In the master branch (Solr 9.0), CaffeineCache is currently the only
> implementation available.
>
> Thanks,
> Shawn
>
Re: Caffeine Cache Metrics Broken?
Posted by Shawn Heisey <ap...@elyograg.org>.
On 3/2/2021 3:47 PM, Stephen Lewis Bianamara wrote:
> I'm investigating a weird behavior I've observed in the admin page for
> caffeine cache metrics. It looks to me like on the older caches, warm-up
> queries were not counted toward hit/miss ratios, which of course makes
> sense, but on Caffeine cache it looks like they are. I'm using solr 8.3.
>
> Obviously this makes measuring its true impact a little tough. Is this by
> any chance a known issue and already fixed in later versions?
The earlier cache implementations are entirely native to Solr -- all the
source code is include in the Solr codebase.
Caffeine is a third-party cache implementation that has been integrated
into Solr. Some of the metrics might come directly from Caffeine, not
Solr code.
I would expect warming queries to be counted on any of the cache
implementations. One of the reasons that the warming capability exists
is to pre-populate the caches before actual queries begin. If warming
queries are somehow excluded, then the cache metrics would not be correct.
I looked into the code and did not find anything that would keep warming
queries from affecting stats. But it is always possible that I just
didn't know what to look for.
In the master branch (Solr 9.0), CaffeineCache is currently the only
implementation available.
Thanks,
Shawn