Hi Erick,
first of all thanks a lot for your response!
I suppose that in my case is happening exactly what you describe as "GC Hell"
because I see continuous GC cycles and solr is not showing OOM errors.
I absolutely agree with you that this is a bad query but I was wondering if
there is any setting (both solr side or jvm side) to avoid that a single query
can tear down an entire cluster without the possibility of an automatic
recovery.
I will implement in our APIs a check on luke request handler to prevent the
execution of expensive queries.
Anyway, I my opinion it could be useful a parameter like a timeAllowed working
on the whole query process or even better something like a memoryAllowed
parameter to kill a query if it's too expensive in terms of timing or memory
consumption.
I'll study a little bit better streaming capabilities, it seems really powerful.
Best regards,
Matteo Diarena
Direttore Innovazione
Volo.com S.r.l. (www.volocom.it - volo...@pec.it)
Via Luigi Rizzo, 8/1 - 20151 MILANO
Via Leone XIII, 95 - 00165 ROMA
Tel +39 02 89453024 / +39 02 89453023
Fax +39 02 89453500
Mobile +39 345 2129244
m.diar...@volocom.it
-----Messaggio originale-----
Da: Erick Erickson <erickerick...@gmail.com>
Inviato: 30 January 2019 17:44
A: solr-user <solr-user@lucene.apache.org>
Oggetto: Re: SolrCloud become unresponsive after huge pivot facet query
My suggestion is "don't do that" ;).
Ok, seriously. Conceptually what you have is an N-dimnensional matrix.
Each "dimension" is
one of your pivot fields with one cell for each unique value in the field. So
the size is (cardinality of field 1) x (cardinality of field 2) * (cardinality
of field 3) .....
To make matters worse, the results from each shard need to be aggregated, so
you're correct that you're shoving potentially a _very_ large set of data
across your network that then has to be sorted into the final packet.
You don't indicate that you have OOM errors so what I suspect is happening is
you're in "GC hell". Each GC cycle recovers just enough memory to continue for
a very short time, then stops for another GC cycle. Rinse, repeat. Timeout.
For more concrete suggestions.
1> You can use the "Luke request handler" to find the cardinality of
the fields and then
have a blacklist of fields so you wind up rejecting these queries up front.
2> Consider the streaming capabilities. "Rollup" can be used for
high cardinality fields.
see: https://lucene.apache.org/solr/guide/6_6/stream-decorators.html.
NOTE:
"Facet" streams push the faceting down to the replicas, which you don't
want to use in this case as it'll be the same problem. The facet streams
are faster when they can be used, but I doubt you can in your case.
BTW, as chance would have it, Joel B. just explained this to me ;).
Best,
Erick
On Wed, Jan 30, 2019 at 3:41 AM Matteo Diarena <m.diar...@volocom.it> wrote:
>
> Dear all,
> we have a solrcloud cluster with the following features:
> - 3 zookeeper nodes
> - 4 solr nodes with:
> - 4 CPU
> - 16GB RAM
>
> Each solr instance is configured as follow:
> SOLR_JAVA_MEM="-Xms2g -Xmx8g"
> SOLR_OPTS="$SOLR_OPTS -Dlucene.cms.override_spins=false
> -Dlucene.cms.override_core_count=4"
>
> On the cluster we created a collection with 5 shards each with 2 replicas for
> a total of 10 replicas.
>
> The full index size is less than 2 GB and under normal usage the used heap
> space is between 200MB and 500MB.
>
> Unfortunately if we try to perform a query like the this:
>
> .../select?q=*:*&fq=ActionType:MAILOPENING&facet=true&rows=0&facet.piv
> ot=FIELD_ObjectId,FIELD_MailId&f.FIELD_ObjectId.facet.pivot.mincount=0
> &f.FIELD_ObjectId.facet.limit=-1&f.FIELD_ObjectId.facet.pivot.mincount
> =0&f.FIELD_ObjectId.facet.limit=-1
>
> where FIELD_ObjectId and FIELD_MailId are high cardinality fields all the
> heap space is used and the entire solr cluster becomes really slow and
> unresponsive.
> The solr instance is not killed and the heap space is never released so the
> only way is to get the cluster up again is to restart all the solr instances.
>
> I know that the problem is the wrong query but I'd like to know how I can
> avoid this kind of problems.
> Is there a way to limit the memory usage during query execution to avoid a
> single query to hang a cluster?
>
> I tried to disable all caches and to investigate the heap dump but I didn't
> manage to find any good solution.
> I also thought that an issue could be the really big search response exchange
> between shards. Is it possible?
>
> Actually the cluster is not in production so I can easily perform tests or
> get all needed data.
>
> Any suggestion is welcome.
>
> Thanks a lot,
> Matteo Diarena
> Direttore Innovazione
>
> Volo.com S.r.l. (www.volocom.it<http://www.volocom.it/> -
> volo...@pec.it<mailto:volo...@pec.it>)
> Via Luigi Rizzo, 8/1 - 20151 MILANO
> Via Leone XIII, 95 - 00165 ROMA
>
> Tel +39 02 89453024 / +39 02 89453023
> Fax +39 02 89453500
> Mobile +39 345 2129244
> m.diar...@volocom.it<mailto:m.diar...@volocom.it>
>
