Re: [Owlim-discussion] Poor performance for Sparql queries with property path optional elements
Hello Ruslan, Thanks for such an exhaustive answer! For now, we can just avoid using those optional property path elements in our Sparql queries and convert them to use unions instead - it seems to work properly for us and shouldn't be very time-consuming. This is still just a workaround though and it would be really nice if Sesame developers fixed this. -- Best regards, Krzysztof Sielski Poznan Supercomputing and Networking Center W dniu 2012-06-18 15:53, Ruslan Velkov pisze: Hi Krzysztof, Many thanks for reporting this and for providing a test class! What we introduced in 5.1 is the Sesame's QueryJoinOptimizer which rearranges joins so that if there are sub-select clauses in the query they will be evaluated first and in the best possible order (in terms of number of variables shared by their respective projections). This optimizer allows for fast and efficient evaluation of nested SELECT clauses. What I saw as a by-product of this optimizer on q2 was this: [Original query plan without applying the QueryJoinOptimizer] Projection ProjectionElemList ProjectionElem name Join Join StatementPattern Var (name=-const-1, value=person://1, anonymous) Var (name=-const-2, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-2-0, anonymous) Union ZeroLengthPath Var (name=-const-2-0, anonymous) Var (name=-const-3-1, anonymous) StatementPattern Var (name=-const-2-0, anonymous) Var (name=-const-3, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-3-1, anonymous) StatementPattern Var (name=-const-3-1, anonymous) Var (name=-const-4, value=http://xmlns.com/foaf/0.1/name, anonymous) Var (name=name) [Query plan after applying the QueryJoinOptimizer] Projection ProjectionElemList ProjectionElem name Join StatementPattern Var (name=-const-1, value=person://1, anonymous) Var (name=-const-2, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-2-0, anonymous) Join StatementPattern Var (name=-const-3-1, anonymous) Var (name=-const-4, value=http://xmlns.com/foaf/0.1/name, anonymous) Var (name=name) Union ZeroLengthPath Var (name=-const-2-0, anonymous) Var (name=-const-3-1, anonymous) StatementPattern Var (name=-const-2-0, anonymous) Var (name=-const-3, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-3-1, anonymous) As you can see, using the first query model we'll evaluate person://1 http://xmlns.com/foaf/0.1/knows -const-2-0 and then -const-2-0 http://xmlns.com/foaf/0.1/knows -const-3-1 with the already bound -const-2-0 and finally we'll evaluate the last statement using the binding for -const-3-1. This is the correct ordering. What we can see from the second query model is evaluating person://1 http://xmlns.com/foaf/0.1/knows -const-2-0 and then the last statement -const-3-1 http://xmlns.com/foaf/0.1/name name and then the optional second statement, but as far as the the first two patterns in this order don't share any variables a Cartesian product will be formed (the first pattern has 999 results and the second one has 23334 results, hence 23,310,666 iterations, very few of which will succeed). Unfortunately, there is no way to turn this optimizer off and even there was one, other queries would become much slower (namely the ones with sub-selects). There can be introduced a parameter that switches the optimizer off as a workaround, but you may encounter problems when using queries with sub-selects, so in that case you should arrange the sub-selects manually and they should be the first thing to a appear in a query (in case you use such queries along the ones with problematic property path evaluation). Another workaround could be using an ASK query to switch that optimizer on/off at runtime, but this will be a rather clumsy approach (queries can be evaluated from multiple threads asynchronously, so you won't have guarantee when exactly you use the optimizer and when not). The real solution is a fix in Sesame to be provided (we'll communicate the issue with the Sesame guys). So the fastest solution will be to provide a parameter which will statically forbid the optimizer (i.e. at initialization time). Will that be ok in your case? Hth, Ruslan On 06/18/2012 02:55 PM, Krzysztof Sielski wrote: Hello, We noticed that after migrating to Owlim SE 5.1.5183 our queries which use property paths with optional elements are evaluated very slowly (in contrast to previous releases). Their direct equivalents using UNION would return the same results much faster. This is an example: [query for particular
Re: [Owlim-discussion] Poor performance for Sparql queries with property path optional elements
Hi Marek, The new behaviour after we introduced the QueryJoinOptimizer is a feature we were looking for and it took much time to be implemented and stabilized in Sesame because in the beginning it was implemented like a different query model node (SPARQLIntersection), then this node disappeared from the Sesame code base and the implementation was introduced in EvaluationStrategyImpl.evaluate(Join), see http://www.openrdf.org/issues/browse/SES-953. This issue was fixed in Sesame 2.6.5 and the fix was available for Owlim 5.0, but our release notes were rather large because they covered all bug fixes and new features, so we didn't mention the concrete Sesame bug fixes and improvements, just mentioned that we use Sesame version 2.6.5 (that is why you can't find this issue in our release notes). Some of the sub-select handling code went into the QueryJoinOptimizer, but this happened after releasing Owlim 5.0 and was available in Sesame 2.6.6, from where we took that optimizer and introduced it into our code (version 5.1). The thing which was really improved was namely the sub-selects, we experimented with the BSBM Business Intelligence benchmark in particular, and that optimizer gave wonderful results (some of the queries contained sub-selects, sub-select nested into sub-selects, sub-selects interconnected just with a single filter, etc.). I hope this answers your question what actually was improved by using this optimizer. The bad think with this optimizer is that it touches the whole query model, not just the sub-selects, hence the side effects we are experiencing. After communicating the issue with Jeen it became clear that apart from several small fixes in Sesame, we may be able to fix this issue in Owlim by supplying statistics to the QueryJoinOptimizer in order to influence its decision-making when reordering the query. However, the statistics that we use in our query optimization strategy are much different from what Sesame uses in its own one, so it may take some time. If it is achievable, you'll be notified with the fix. Cheers, Ruslan On 06/18/2012 07:18 PM, Marek Šurek wrote: Hi Ruslan, we have the similar issue (see http://www.mail-archive.com/owlim-discussion@ontotext.com/msg01626.html), which is probably based on the same thing. I don't understand from your response whether the new behaviour is bug or feature. I carefully looked at release notes and nothing so serious as total change query optimizer, which dramatically changes behaviour of subselects, was not mentioned! I see(and really appreciate) that you are looking for shorttime hotfix solution introducing some parameter but for now I don't know what to do. Lot of our bussiness code uses subselects as it is powerful feature. We rely on your answer whether this feature/bug is just problem of few days or it is permanent state. The parameter is great as hotfix, but I can't test and optimize all queries and try in which case it runs faster and after some minor bugfix suddenly without any warning I can start from begining. As it is not well documented, I would like to ask, what are positive aspects of using new query optimizer? Where can we see improvements? What kind of queries should ran faster? Thank you for your time and looking forward for your response. Best regards, Marek *From:* Ruslan Velkov rus...@sirma.bg *To:* owlim-discussion@ontotext.com *Sent:* Monday, 18 June 2012, 15:53 *Subject:* Re: [Owlim-discussion] Poor performance for Sparql queries with property path optional elements Hi Krzysztof, Many thanks for reporting this and for providing a test class! What we introduced in 5.1 is the Sesame's QueryJoinOptimizer which rearranges joins so that if there are sub-select clauses in the query they will be evaluated first and in the best possible order (in terms of number of variables shared by their respective projections). This optimizer allows for fast and efficient evaluation of nested SELECT clauses. What I saw as a by-product of this optimizer on q2 was this: [Original query plan without applying the QueryJoinOptimizer] Projection ProjectionElemList ProjectionElem name Join Join StatementPattern Var (name=-const-1, value=person://1, anonymous) Var (name=-const-2, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-2-0, anonymous) Union ZeroLengthPath Var (name=-const-2-0, anonymous) Var (name=-const-3-1, anonymous) StatementPattern Var (name=-const-2-0, anonymous) Var (name=-const-3, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-3-1, anonymous) StatementPattern Var (name=-const-3-1, anonymous) Var (name=-const-4, value=http://xmlns.com/foaf/0.1/name, anonymous
Re: [Owlim-discussion] Poor performance for Sparql queries with property path optional elements
Thank you Ruslan, I will revert back to OWLIM 5.0 b5123 using Sesame 2.6.5 where all things seems fines (except the removeRepository, but I can live with it). I was just confused as I made tons of query testing to realise what are performance strong and weak queries and suddenly it had changed in bad way and I was worried about project destiny. Good luck with bugfixing and looking forward for next release. Best regards, Marek From: Ruslan Velkov rus...@sirma.bg To: Marek Šurek marek_su...@yahoo.co.uk Cc: owlim-discussion@ontotext.com owlim-discussion@ontotext.com; Barry Bishop barry.bis...@ontotext.com Sent: Tuesday, 19 June 2012, 13:51 Subject: Re: [Owlim-discussion] Poor performance for Sparql queries with property path optional elements Hi Marek, The new behaviour after we introduced the QueryJoinOptimizer is a feature we were looking for and it took much time to be implemented and stabilized in Sesame because in the beginning it was implemented like a different query model node (SPARQLIntersection), then this node disappeared from the Sesame code base and the implementation was introduced in EvaluationStrategyImpl.evaluate(Join), see http://www.openrdf.org/issues/browse/SES-953. This issue was fixed in Sesame 2.6.5 and the fix was available for Owlim 5.0, but our release notes were rather large because they covered all bug fixes and new features, so we didn't mention the concrete Sesame bug fixes and improvements, just mentioned that we use Sesame version 2.6.5 (that is why you can't find this issue in our release notes). Some of the sub-select handling code went into the QueryJoinOptimizer, but this happened after releasing Owlim 5.0 and was available in Sesame 2.6.6, from where we took that optimizer and introduced it into our code (version 5.1). The thing which was really improved was namely the sub-selects, we experimented with the BSBM Business Intelligence benchmark in particular, and that optimizer gave wonderful results (some of the queries contained sub-selects, sub-select nested into sub-selects, sub-selects interconnected just with a single filter, etc.). I hope this answers your question what actually was improved by using this optimizer. The bad think with this optimizer is that it touches the whole query model, not just the sub-selects, hence the side effects we are experiencing. After communicating the issue with Jeen it became clear that apart from several small fixes in Sesame, we may be able to fix this issue in Owlim by supplying statistics to the QueryJoinOptimizer in order to influence its decision-making when reordering the query. However, the statistics that we use in our query optimization strategy are much different from what Sesame uses in its own one, so it may take some time. If it is achievable, you'll be notified with the fix. Cheers, Ruslan On 06/18/2012 07:18 PM, Marek Šurek wrote: Hi Ruslan, we have the similar issue (see http://www.mail-archive.com/owlim-discussion@ontotext.com/msg01626.html), which is probably based on the same thing. I don't understand from your response whether the new behaviour is bug or feature. I carefully looked at release notes and nothing so serious as total change query optimizer, which dramatically changes behaviour of subselects, was not mentioned! I see(and really appreciate) that you are looking for shorttime hotfix solution introducing some parameter but for now I don't know what to do. Lot of our bussiness code uses subselects as it is powerful feature. We rely on your answer whether this feature/bug is just problem of few days or it is permanent state. The parameter is great as hotfix, but I can't test and optimize all queries and try in which case it runs faster and after some minor bugfix suddenly without any warning I can start from begining. As it is not well documented, I would like to ask, what are positive aspects of using new query optimizer? Where can we see improvements? What kind of queries should ran faster? Thank you for your time and looking forward for your response. Best regards, Marek From: Ruslan Velkov rus...@sirma.bg To: owlim-discussion@ontotext.com Sent: Monday, 18 June 2012, 15:53 Subject: Re: [Owlim-discussion] Poor performance for Sparql queries with property path optional elements Hi Krzysztof, Many thanks for reporting this and for providing a test class! What we introduced in 5.1 is the Sesame's QueryJoinOptimizer which rearranges joins so that if there are sub-select clauses in the query they will be evaluated first and in the best possible order
[Owlim-discussion] Poor performance for Sparql queries with property path optional elements
Hello, We noticed that after migrating to Owlim SE 5.1.5183 our queries which use property paths with optional elements are evaluated very slowly (in contrast to previous releases). Their direct equivalents using UNION would return the same results much faster. This is an example: [query for particular person's friends' names and their friends' names] (q1) PREFIX foaf:http://xmlns.com/foaf/0.1/ select * WHERE { {person://1 foaf:knows/foaf:name ?name} UNION {person://1 foaf:knows/foaf:knows/foaf:name ?name} } (q2) PREFIX foaf:http://xmlns.com/foaf/0.1/ select * WHERE { person://1 foaf:knows/foaf:knows?/foaf:name ?name } Both queries seem to be equivalent and we used (q2) as it is more concise and elegant but now (q1) is much faster: Executing query q1 Result count: 2 in 0,006000s. Executing query q2 Result count: 2 in 7,034000s. As before, I attached a simple class that creates a local repository, inserts some data and executes the queries to show you the problem. -- Best regards, Krzysztof Sielski Poznan Supercomputing and Networking Center import java.io.File; import java.util.HashMap; import java.util.Map; import java.util.Random; import org.openrdf.model.BNode; import org.openrdf.model.Graph; import org.openrdf.model.Resource; import org.openrdf.model.URI; import org.openrdf.model.impl.GraphImpl; import org.openrdf.model.impl.LiteralImpl; import org.openrdf.model.impl.StatementImpl; import org.openrdf.model.impl.URIImpl; import org.openrdf.model.impl.ValueFactoryImpl; import org.openrdf.model.vocabulary.RDF; import org.openrdf.model.vocabulary.RDFS; import org.openrdf.query.QueryLanguage; import org.openrdf.query.TupleQueryResult; import org.openrdf.repository.Repository; import org.openrdf.repository.RepositoryConnection; import org.openrdf.repository.config.RepositoryConfig; import org.openrdf.repository.config.RepositoryConfigException; import org.openrdf.repository.config.RepositoryConfigSchema; import org.openrdf.repository.manager.LocalRepositoryManager; import org.openrdf.repository.manager.RepositoryManager; import org.openrdf.repository.sail.config.SailRepositorySchema; import org.openrdf.sail.config.SailConfigSchema; /** * * @author Krzysztof Sielski */ public class OwlimTestCaseOptionalPathElement { private static final String REPO_ID = repo; private static URI foaf_knows = new URIImpl(http://xmlns.com/foaf/0.1/knows;); private static URI foaf_name = new URIImpl(http://xmlns.com/foaf/0.1/name;); private static final String q1 = + PREFIX foaf:http://xmlns.com/foaf/0.1/ + select * WHERE { + {person://1 foaf:knows/foaf:name ?name} +UNION + {person://1 foaf:knows/foaf:knows/foaf:name ?name} + } ; private static final String q2 = + PREFIX foaf:http://xmlns.com/foaf/0.1/ + select * WHERE { + person://1 foaf:knows/foaf:knows?/foaf:name ?name + } ; public static void main(String[] args) throws Exception { RepositoryManager manager = new LocalRepositoryManager(new File(.)); manager.initialize(); try { initRepository(manager); insertInitialData(manager); executeQueries(manager); } finally { manager.shutDown(); } } private static void executeQueries(RepositoryManager manager) throws Exception { Repository repo = manager.getRepository(REPO_ID); repo.initialize(); RepositoryConnection con = repo.getConnection(); con.setAutoCommit(false); System.out.println(Executing query q1); executeQuery(q1, con); System.out.println(Executing query q2); executeQuery(q2, con); con.close(); repo.shutDown(); } private static void executeQuery(String query, RepositoryConnection con) { try { TupleQueryResult result = con.prepareTupleQuery(QueryLanguage.SPARQL, query).evaluate(); int resultCount = 0; long time = System.currentTimeMillis(); while (result.hasNext()) { result.next(); resultCount++; } time = System.currentTimeMillis() - time; System.out.printf(Result count: %d in %fs.\n, resultCount, time / 1000.0); } catch (Exception e) { e.printStackTrace();
Re: [Owlim-discussion] Poor performance for Sparql queries with property path optional elements
Hi Ruslan, we have the similar issue (see http://www.mail-archive.com/owlim-discussion@ontotext.com/msg01626.html), which is probably based on the same thing. I don't understand from your response whether the new behaviour is bug or feature. I carefully looked at release notes and nothing so serious as total change query optimizer, which dramatically changes behaviour of subselects, was not mentioned! I see(and really appreciate) that you are looking for shorttime hotfix solution introducing some parameter but for now I don't know what to do. Lot of our bussiness code uses subselects as it is powerful feature. We rely on your answer whether this feature/bug is just problem of few days or it is permanent state. The parameter is great as hotfix, but I can't test and optimize all queries and try in which case it runs faster and after some minor bugfix suddenly without any warning I can start from begining. As it is not well documented, I would like to ask, what are positive aspects of using new query optimizer? Where can we see improvements? What kind of queries should ran faster? Thank you for your time and looking forward for your response. Best regards, Marek From: Ruslan Velkov rus...@sirma.bg To: owlim-discussion@ontotext.com Sent: Monday, 18 June 2012, 15:53 Subject: Re: [Owlim-discussion] Poor performance for Sparql queries with property path optional elements Hi Krzysztof, Many thanks for reporting this and for providing a test class! What we introduced in 5.1 is the Sesame's QueryJoinOptimizer which rearranges joins so that if there are sub-select clauses in the query they will be evaluated first and in the best possible order (in terms of number of variables shared by their respective projections). This optimizer allows for fast and efficient evaluation of nested SELECT clauses. What I saw as a by-product of this optimizer on q2 was this: [Original query plan without applying the QueryJoinOptimizer] Projection ProjectionElemList ProjectionElem name Join Join StatementPattern Var (name=-const-1, value=person://1, anonymous) Var (name=-const-2, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-2-0, anonymous) Union ZeroLengthPath Var (name=-const-2-0, anonymous) Var (name=-const-3-1, anonymous) StatementPattern Var (name=-const-2-0, anonymous) Var (name=-const-3, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-3-1, anonymous) StatementPattern Var (name=-const-3-1, anonymous) Var (name=-const-4, value=http://xmlns.com/foaf/0.1/name, anonymous) Var (name=name) [Query plan after applying the QueryJoinOptimizer] Projection ProjectionElemList ProjectionElem name Join StatementPattern Var (name=-const-1, value=person://1, anonymous) Var (name=-const-2, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-2-0, anonymous) Join StatementPattern Var (name=-const-3-1, anonymous) Var (name=-const-4, value=http://xmlns.com/foaf/0.1/name, anonymous) Var (name=name) Union ZeroLengthPath Var (name=-const-2-0, anonymous) Var (name=-const-3-1, anonymous) StatementPattern Var (name=-const-2-0, anonymous) Var (name=-const-3, value=http://xmlns.com/foaf/0.1/knows, anonymous) Var (name=-const-3-1, anonymous) As you can see, using the first query model we'll evaluate person://1 http://xmlns.com/foaf/0.1/knows -const-2-0 and then -const-2-0 http://xmlns.com/foaf/0.1/knows -const-3-1 with the already bound -const-2-0 and finally we'll evaluate the last statement using the binding for -const-3-1. This is the correct ordering. What we can see from the second query model is evaluating person://1 http://xmlns.com/foaf/0.1/knows -const-2-0 and then the last statement -const-3-1 http://xmlns.com/foaf/0.1/name name and then the optional second statement, but as far as the the first two patterns in this order don't share any variables a Cartesian product will be formed (the first pattern has 999 results and the second one has 23334 results, hence 23,310,666 iterations, very few of which will succeed). Unfortunately, there is no way to turn this optimizer off and even there was one, other queries would become much slower (namely the ones with sub-selects). There can be introduced a parameter that switches the optimizer off as a workaround, but you may encounter problems when using queries with sub-selects, so in that case you should arrange the sub-selects manually and they should be the first thing to a appear in a query (in case you