Hi to all, we moved solr with patched lucene's FieldCache in production environment. During tests we noticed random ConcurrentModificationException calling the getCacheEntries method due to this bug
https://issues.apache.org/jira/browse/LUCENE-2273 We applied that patch as well, and added an abstract int getCacheSize() method to FieldCache abstract class and its implementation in abstract Cache inner class in CacheFieldImpl that returns the cache size without instantiating a CacheEntry array. Response time are slower on cache purging but acceptable from the user point of view. Regards, Matteo On 22 June 2010 22:41, Matteo Fiandesio <matteo.fiande...@gmail.com> wrote: > Fields over i'm sorting to are dynamic so one query sorts on > erick_time_1,erick_timeA_1 and other sorts on erick_time_2 and so > on.What we see in the heap are a lot of arrays,most of them,filled > with 0s maybe due to the fact that this timestamps fields are not > present in all the documents. > > By the way, > I have a script that generates the OOM in 10 minutes on our solr > instance and with the temporary patch it runned without any problems. > The side effect is that when the cache is purged next query that > regenerates the cache is a little bit slower. > > I'm aware that the solution is unelegant and we are investigating to > solve the problem in another way. > Regards, > Matteo > > > On 22 June 2010 19:25, Erick Erickson <erickerick...@gmail.com> wrote: >> Hmmm, I'm missing something here then. Sorting over 15 fields of type long >> shouldn't use much memory, even if all the values are unique. When you say >> "12-15 dynamic fields", are you talking about 12-15 fields per query out of >> XXX total fields? And is XXX large? At a guess, how many different fields >> do >> you think you're sorting over cumulative by the time you get your OOM? >> Note if you sort over the field "erick_time" in 10 different queries, I'm >> only counting that as 1 field. I guess another way of asking this is >> "how many dynamic fields are there total?". >> >> If this is really a sorting issue, you should be able to force this to >> happen >> almost immediately by firing off enough sort queries at the server. It'll >> tell you a lot if you can't make this happen, even on a relatively small >> test machine. >> >> Best >> Erick >> >> On Tue, Jun 22, 2010 at 12:59 PM, Matteo Fiandesio < >> matteo.fiande...@gmail.com> wrote: >> >>> Hi Erick, >>> the index is quite small (1691145 docs) but sorting is massive and >>> often on unique timestamp fields. >>> >>> OOM occur after a range of time between three and four hours. >>> Depending as well if users browse a part of the application. >>> >>> We use solrj to make the queries so we did not use Readers objects >>> directly. >>> >>> Without sorting we don't see the problem >>> Regards, >>> Matteo >>> >>> On 22 June 2010 17:01, Erick Erickson <erickerick...@gmail.com> wrote: >>> > Hmmmm.. A couple of details I'm wondering about. How many >>> > documents are we talking about in your index? Do you get >>> > OOMs when you start fresh or does it take a while? >>> > >>> > You've done some good investigations, so it seems like there >>> > could well be something else going on here than just "the usual >>> > suspects" of sorting.... >>> > >>> > I'm wondering if you aren't really closing readers somehow. >>> > Are you updating your index frequently and re-opening readers often? >>> > If so, how? >>> > >>> > I'm assuming that if you do NOT sort on all these fields, you don't have >>> > the problem, is that true? >>> > >>> > Best >>> > Erick >>> > >>> > On Fri, Jun 18, 2010 at 10:52 AM, Matteo Fiandesio < >>> > matteo.fiande...@gmail.com> wrote: >>> > >>> >> Hello, >>> >> we are experiencing OOM exceptions in our single core solr instance >>> >> (on a (huge) amazon EC2 machine). >>> >> We investigated a lot in the mailing list and through jmap/jhat dump >>> >> analyzing and the problem resides in the lucene FieldCache that fills >>> >> the heap and blows up the server. >>> >> >>> >> Our index is quite small but we have a lot of sort queries on fields >>> >> that are dynamic,of type long representing timestamps and are not >>> >> present in all the documents. >>> >> Those queries apply sorting on 12-15 of those fields. >>> >> >>> >> We are using solr 1.4 in production and the dump shows a lot of >>> >> Integer/Character and Byte Array filled up with 0s. >>> >> With solr's trunk code things does not change. >>> >> >>> >> In the mailing list we saw a lot of messages related to this issues: >>> >> we tried truncating the dates to day precision,using missingSortLast = >>> >> true,changing the field type from slong to long,setting autowarming to >>> >> different values,disabling and enabling caches with different values >>> >> but we did not manage to solve the problem. >>> >> >>> >> We were thinking to implement an LRUFieldCache field type to manage >>> >> the FieldCache as an LRU and preventing but, before starting a new >>> >> development, we want to be sure that we are not doing anything wrong >>> >> in the solr configuration or in the index generation. >>> >> >>> >> Any help would be appreciated. >>> >> Regards, >>> >> Matteo >>> >> >>> > >>> >> >
