Thanks Jack for your suggestions. Regards, Modassar
On Fri, Dec 26, 2014 at 6:04 PM, Jack Krupansky <jack.krupan...@gmail.com> wrote: > Either you have too little RAM on each node or too much data on each node. > > You may need to shard the data much more heavily so that the total work on > a single query is distributed in parallel to more nodes, each node having a > much smaller amount of data to work on. > > First, always make sure that the entire Lucene index for each node fits > entirely in the system memory available for file system caching. Otherwise > the queries will be I/O bound. Check your current queries to see if that is > the case - are the nodes compute bound or I/O bound? If I/O bound, add more > system memory until the queries are no longer I/O bound. If compute bound, > shard more heavily until the query latency becomes acceptable. > > > > -- Jack Krupansky > > On Fri, Dec 26, 2014 at 1:02 AM, Modassar Ather <modather1...@gmail.com> > wrote: > > > Thanks for your suggestions Erick. > > > > This may be one of those situations where you really have to > > push back at the users and understand why they insist on these > > kinds of queries. They must be very patient since it won't be > > very performant. That said, I've seen this pattern; there are > > certainly valid conditions under which response times can be > > many seconds if there are few users and they are doing very > > complex/expert-level things. > > > > We have tried educating the users but it did not work because they are > used > > to the old way. They feel that wildcard gives more control over the > results > > and may not fully understand stemming. > > > > Regards, > > Modassar > > > > On Thu, Dec 25, 2014 at 3:17 AM, Erick Erickson <erickerick...@gmail.com > > > > wrote: > > > > > There's no magic bullet here that I know of. If your requirements > > > are to support these huge, many-wildcard queries then you only > > > have a few choices: > > > > > > 1> redo the index. I was surprised at how little it bloated the > > > index as far as memory required is concerned to add ngrams. > > > The key here is that there really aren't very many unique terms. > > > If you use bigrams, then there are only maybe 36^2 distinct > > > combinations. (assuming English and including numbers). > > > > > > 2> Increase the number of shards, putting many fewer docs > > > on each shard. > > > > > > 3> give each shard a lot more memory. This isn't actually one > > > of my preferred solutions since GC issues may raise their ugly > > > heads here. > > > > > > 4> insert creative solution here. > > > > > > This may be one of those situations where you really have to > > > push back at the users and understand why they insist on these > > > kinds of queries. They must be very patient since it won't be > > > very performant. That said, I've seen this pattern; there are > > > certainly valid conditions under which response times can be > > > many seconds if there are few users and they are doing very > > > complex/expert-level things. > > > > > > Now, all that said, wildcards are often examples of poor habits > > > or habits learned in DB systems where the only hammer was > > > %whatever%. I've seen situations where users didn't > > > understand that Solr broke the input stream up into words. And > > > stemmed. And WordDelimiterFilterFactory did all the magic > > > for finding, say D.C. and DC. So it's worth looking at the actual > > > queries that are sent, perhaps talking to users and understanding > > > what they _want_ out of the system, then perhaps educating them > > > as to better ways to get what they want. > > > > > > Literally I've seen people insist on entering queries that > > > wildcarded _everything_ both pre and post wildcards because > > > they didn't realize that Solr tokenizes... > > > > > > Once you hit an OOM, all bets are off as Shawn outlined. > > > > > > Best, > > > Erick > > > > > > On Wed, Dec 24, 2014 at 1:57 AM, Modassar Ather < > modather1...@gmail.com> > > > wrote: > > > > Thanks for your response. > > > > > > > > How many items in the collection ? > > > > There are about 100 millions documents. > > > > > > > > How are configured cache in solrconfig.xml ? > > > > Each cache has size attribute as 128. > > > > > > > > Can you provide a sample of the query ? > > > > Does it fail immediately after solrcloud startup or after several > > hours ? > > > > It is a query with many terms(more than a thousand) and phrase where > > > > phrases have many wildcards in it. > > > > Once such query is executed there are many zookeeper related > exceptions > > > and > > > > with a couple of such queries executed it goes for OutOfMemory. > > > > > > > > Thanks, > > > > Modassar > > > > > > > > > > > > On Wed, Dec 24, 2014 at 1:37 PM, Dominique Bejean < > > > dominique.bej...@eolya.fr > > > >> wrote: > > > > > > > >> And you didn’t give how many RAM on each servers ? > > > >> > > > >> 2014-12-24 8:17 GMT+01:00 Dominique Bejean < > dominique.bej...@eolya.fr > > >: > > > >> > > > >> > Modassar, > > > >> > > > > >> > How many items in the collection ? > > > >> > I mean how many documents per collection ? 1 million, 10 millions, > > …? > > > >> > > > > >> > How are configured cache in solrconfig.xml ? > > > >> > What are the size attribute value for each cache ? > > > >> > > > > >> > Can you provide a sample of the query ? > > > >> > Does it fail immediately after solrcloud startup or after several > > > hours ? > > > >> > > > > >> > Dominique > > > >> > > > > >> > 2014-12-24 6:20 GMT+01:00 Modassar Ather <modather1...@gmail.com > >: > > > >> > > > > >> >> Thanks for your suggestions. > > > >> >> > > > >> >> I will look into the link provided. > > > >> >> http://wiki.apache.org/solr/SolrPerformanceProblems#Java_Heap > > > >> >> > > > >> >> This is usually an anti-pattern. The very first thing > > > >> >> I'd be doing is trying to not do this. See ngrams for infix > > > >> >> queries, or shingles or ReverseWildcardFilterFactory or..... > > > >> >> > > > >> >> We cannot avoid multiple wildcards since that's is our user's > > > >> requirement. > > > >> >> We try to discourage it but the users insist on firing such > > queries. > > > >> Also, > > > >> >> ngrams etc. can be tried but our index is already huge and ngrams > > may > > > >> >> further add lot to it. We are OK with such queries failing as > long > > as > > > >> >> other > > > >> >> queries are not affected. > > > >> >> > > > >> >> > > > >> >> Please find the details below. > > > >> >> > > > >> >> So, how many nodes in the cluster ? > > > >> >> There are total 4 nodes on the cluster. > > > >> >> > > > >> >> How many shards and replicas for the collection ? > > > >> >> There are 4 shards and no replica for any of them. > > > >> >> > > > >> >> How many items in the collection ? > > > >> >> If I understand the question correctly there are two collection > on > > > each > > > >> >> node and there size on each node is approximately 190GB and > 130GB. > > > >> >> > > > >> >> What is the size of the index ? > > > >> >> There are two collection on each node and there size on each node > > is > > > >> >> approximately 190GB and 130GB. > > > >> >> > > > >> >> How is updated the collection (frequency, how many items per > days, > > > what > > > >> is > > > >> >> your hard commit strategy) ? > > > >> >> It is an optimized index and read-only. There are no > inter-mediate > > > >> update. > > > >> >> > > > >> >> How are configured cache in solrconfig.xml ? > > > >> >> Filter cache, query result cache and document cache are enabled. > > > >> >> Auto-warming is also done. > > > >> >> > > > >> >> Can you provide all other JVM parameters ? > > > >> >> -Xms20g -Xmx24g -XX:+UseConcMarkSweepGC > > > >> >> > > > >> >> Thanks again, > > > >> >> Modassar > > > >> >> > > > >> >> On Wed, Dec 24, 2014 at 2:29 AM, Dominique Bejean < > > > >> >> dominique.bej...@eolya.fr > > > >> >> > wrote: > > > >> >> > > > >> >> > Hi, > > > >> >> > > > > >> >> > I agree Erick it could be a good think to have more details > about > > > your > > > >> >> > configuration and collection. > > > >> >> > > > > >> >> > Your heap size is 32Gb. How many RAM on each servers ? > > > >> >> > > > > >> >> > By « 4 shard Solr cluster », you mean a 4 nodes Solr servers > or a > > > >> >> > collection with 4 shards ? > > > >> >> > > > > >> >> > So, how many nodes in the cluster ? > > > >> >> > How many shards and replicas for the collection ? > > > >> >> > How many items in the collection ? > > > >> >> > What is the size of the index ? > > > >> >> > How is updated the collection (frequency, how many items per > > days, > > > >> what > > > >> >> is > > > >> >> > your hard commit strategy) ? > > > >> >> > How are configured cache in solrconfig.xml ? > > > >> >> > Can you provide all other JVM parameters ? > > > >> >> > > > > >> >> > Regards > > > >> >> > > > > >> >> > Dominique > > > >> >> > > > > >> >> > 2014-12-23 17:50 GMT+01:00 Erick Erickson < > > erickerick...@gmail.com > > > >: > > > >> >> > > > > >> >> > > Second most important part of your message: > > > >> >> > > "When executing a huge query with many wildcards inside it > the > > > >> server" > > > >> >> > > > > > >> >> > > This is usually an anti-pattern. The very first thing > > > >> >> > > I'd be doing is trying to not do this. See ngrams for infix > > > >> >> > > queries, or shingles or ReverseWildcardFilterFactory or..... > > > >> >> > > > > > >> >> > > And if your corpus is very large with many unique terms it's > > even > > > >> >> > > worse, but you haven't really told us about that yet. > > > >> >> > > > > > >> >> > > Best, > > > >> >> > > Erick > > > >> >> > > > > > >> >> > > On Tue, Dec 23, 2014 at 8:30 AM, Shawn Heisey < > > > apa...@elyograg.org> > > > >> >> > wrote: > > > >> >> > > > On 12/23/2014 4:34 AM, Modassar Ather wrote: > > > >> >> > > >> Hi, > > > >> >> > > >> > > > >> >> > > >> I have a setup of 4 shard Solr cluster with embedded > > > zookeeper on > > > >> >> one > > > >> >> > of > > > >> >> > > >> them. The zkClient time out is set to 30 seconds, -Xms is > > 20g > > > and > > > >> >> -Xms > > > >> >> > > is > > > >> >> > > >> 24g. > > > >> >> > > >> When executing a huge query with many wildcards inside it > > the > > > >> >> server > > > >> >> > > >> crashes and becomes non-responsive. Even the dashboard > does > > > not > > > >> >> > responds > > > >> >> > > >> and shows connection lost error. This requires me to > restart > > > the > > > >> >> > > servers. > > > >> >> > > > > > > >> >> > > > Here's the important part of your message: > > > >> >> > > > > > > >> >> > > > *Caused by: java.lang.OutOfMemoryError: Java heap space* > > > >> >> > > > > > > >> >> > > > > > > >> >> > > > Your heap is not big enough for what Solr has been asked to > > do. > > > >> You > > > >> >> > > > need to either increase your heap size or change your > > > >> configuration > > > >> >> so > > > >> >> > > > that it uses less memory. > > > >> >> > > > > > > >> >> > > > > > http://wiki.apache.org/solr/SolrPerformanceProblems#Java_Heap > > > >> >> > > > > > > >> >> > > > Most programs have pretty much undefined behavior when an > > OOME > > > >> >> occurs. > > > >> >> > > > Lucene's IndexWriter has been hardened so that it tries > > > extremely > > > >> >> hard > > > >> >> > > > to avoid index corruption when OOME strikes, and I believe > > that > > > >> >> works > > > >> >> > > > well enough that we can call it nearly bulletproof ... but > > the > > > >> rest > > > >> >> of > > > >> >> > > > Lucene and Solr will make no guarantees. > > > >> >> > > > > > > >> >> > > > It's very difficult to have definable program behavior when > > an > > > >> OOME > > > >> >> > > > happens, because you simply cannot know the precise point > > > during > > > >> >> > program > > > >> >> > > > execution where it will happen, or what isn't going to work > > > >> because > > > >> >> > Java > > > >> >> > > > did not have memory space to create an object. Going > > > unresponsive > > > >> >> is > > > >> >> > > > not surprising. > > > >> >> > > > > > > >> >> > > > If you can solve your heap problem, note that you may run > > into > > > >> other > > > >> >> > > > performance issues discussed on the wiki page that I > linked. > > > >> >> > > > > > > >> >> > > > Thanks, > > > >> >> > > > Shawn > > > >> >> > > > > > > >> >> > > > > > >> >> > > > > >> >> > > > >> >> > > > >> >> On Wed, Dec 24, 2014 at 2:29 AM, Dominique Bejean < > > > >> >> dominique.bej...@eolya.fr > > > >> >> > wrote: > > > >> >> > > > >> >> > Hi, > > > >> >> > > > > >> >> > I agree Erick it could be a good think to have more details > about > > > your > > > >> >> > configuration and collection. > > > >> >> > > > > >> >> > Your heap size is 32Gb. How many RAM on each servers ? > > > >> >> > > > > >> >> > By « 4 shard Solr cluster », you mean a 4 nodes Solr servers > or a > > > >> >> > collection with 4 shards ? > > > >> >> > > > > >> >> > So, how many nodes in the cluster ? > > > >> >> > How many shards and replicas for the collection ? > > > >> >> > How many items in the collection ? > > > >> >> > What is the size of the index ? > > > >> >> > How is updated the collection (frequency, how many items per > > days, > > > >> what > > > >> >> is > > > >> >> > your hard commit strategy) ? > > > >> >> > How are configured cache in solrconfig.xml ? > > > >> >> > Can you provide all other JVM parameters ? > > > >> >> > > > > >> >> > Regards > > > >> >> > > > > >> >> > Dominique > > > >> >> > > > > >> >> > 2014-12-23 17:50 GMT+01:00 Erick Erickson < > > erickerick...@gmail.com > > > >: > > > >> >> > > > > >> >> > > Second most important part of your message: > > > >> >> > > "When executing a huge query with many wildcards inside it > the > > > >> server" > > > >> >> > > > > > >> >> > > This is usually an anti-pattern. The very first thing > > > >> >> > > I'd be doing is trying to not do this. See ngrams for infix > > > >> >> > > queries, or shingles or ReverseWildcardFilterFactory or..... > > > >> >> > > > > > >> >> > > And if your corpus is very large with many unique terms it's > > even > > > >> >> > > worse, but you haven't really told us about that yet. > > > >> >> > > > > > >> >> > > Best, > > > >> >> > > Erick > > > >> >> > > > > > >> >> > > On Tue, Dec 23, 2014 at 8:30 AM, Shawn Heisey < > > > apa...@elyograg.org> > > > >> >> > wrote: > > > >> >> > > > On 12/23/2014 4:34 AM, Modassar Ather wrote: > > > >> >> > > >> Hi, > > > >> >> > > >> > > > >> >> > > >> I have a setup of 4 shard Solr cluster with embedded > > > zookeeper on > > > >> >> one > > > >> >> > of > > > >> >> > > >> them. The zkClient time out is set to 30 seconds, -Xms is > > 20g > > > and > > > >> >> -Xms > > > >> >> > > is > > > >> >> > > >> 24g. > > > >> >> > > >> When executing a huge query with many wildcards inside it > > the > > > >> >> server > > > >> >> > > >> crashes and becomes non-responsive. Even the dashboard > does > > > not > > > >> >> > responds > > > >> >> > > >> and shows connection lost error. This requires me to > restart > > > the > > > >> >> > > servers. > > > >> >> > > > > > > >> >> > > > Here's the important part of your message: > > > >> >> > > > > > > >> >> > > > *Caused by: java.lang.OutOfMemoryError: Java heap space* > > > >> >> > > > > > > >> >> > > > > > > >> >> > > > Your heap is not big enough for what Solr has been asked to > > do. > > > >> You > > > >> >> > > > need to either increase your heap size or change your > > > >> configuration > > > >> >> so > > > >> >> > > > that it uses less memory. > > > >> >> > > > > > > >> >> > > > > > http://wiki.apache.org/solr/SolrPerformanceProblems#Java_Heap > > > >> >> > > > > > > >> >> > > > Most programs have pretty much undefined behavior when an > > OOME > > > >> >> occurs. > > > >> >> > > > Lucene's IndexWriter has been hardened so that it tries > > > extremely > > > >> >> hard > > > >> >> > > > to avoid index corruption when OOME strikes, and I believe > > that > > > >> >> works > > > >> >> > > > well enough that we can call it nearly bulletproof ... but > > the > > > >> rest > > > >> >> of > > > >> >> > > > Lucene and Solr will make no guarantees. > > > >> >> > > > > > > >> >> > > > It's very difficult to have definable program behavior when > > an > > > >> OOME > > > >> >> > > > happens, because you simply cannot know the precise point > > > during > > > >> >> > program > > > >> >> > > > execution where it will happen, or what isn't going to work > > > >> because > > > >> >> > Java > > > >> >> > > > did not have memory space to create an object. Going > > > unresponsive > > > >> >> is > > > >> >> > > > not surprising. > > > >> >> > > > > > > >> >> > > > If you can solve your heap problem, note that you may run > > into > > > >> other > > > >> >> > > > performance issues discussed on the wiki page that I > linked. > > > >> >> > > > > > > >> >> > > > Thanks, > > > >> >> > > > Shawn > > > >> >> > > > > > > >> >> > > > > > >> >> > > > > >> >> > > > >> > > > > >> > > > > >> > > > > > >