Type Based XML Projection

Size: px

Start display at page:

Download "Type Based XML Projection"

Olivia Briggs
5 years ago
Views:

1 Type Based XML Projection VLDB 2006, Seoul, Korea Véronique Benzaken Giuseppe Castagna Dario Colazzo Kim Nguy ên : Équipe Bases de Données, LRI, Université Paris-Sud 11, Orsay, France : Équipe Langage, DI, École Normale Supérieure, Paris, France 1 / 30

2 Outline 1 Introduction 2 Notations 3 Algorithm 4 Formal results 5 Experiments 6 Conclusion 2 / 30

3 Outline 1 Introduction 2 Notations 3 Algorithm 4 Formal results 5 Experiments 6 Conclusion 3 / 30

4 Main memory query (XQuery) processing 4 / 30

5 Main memory query (XQuery) processing Pros : Lightweight system 4 / 30

6 Main memory query (XQuery) processing Pros : Lightweight system Easy to program and integrate into an existing architecture 4 / 30

7 Main memory query (XQuery) processing Pros : Lightweight system Easy to program and integrate into an existing architecture Cons : Unlike DBMS solution, whole document in memory (DOM lots of metadata) 4 / 30

8 Main memory query (XQuery) processing Pros : Lightweight system Easy to program and integrate into an existing architecture Cons : Unlike DBMS solution, whole document in memory (DOM lots of metadata) 4 / 30

9 We can use pruning to reduce the amount of data needed in main memory! 5 / 30

10 We can use pruning to reduce the amount of data needed in main memory! Exemple : /descendant-or-self::title/child::text() 5 / 30

11 We can use pruning to reduce the amount of data needed in main memory! Exemple : /descendant-or-self::title/child::text() <bib> <book year="1994"> <title>tcp/ip Illustrated</title> <author><last>stevens</last><first>w.</first></author> <publisher>addison-wesley</publisher> <price>65.95</price> </book> <book year="1992"> <title>advanced Programming in the Unix environment</title> <author><last>stevens</last><first>w.</first></author> <publisher>addison-wesley</publisher> <price>65.95</price> </book>... </bib> 5 / 30

12 Projecting XML Documents, Amélie Marian and Jérome Siméon, VLDB Queries + document based optimizations, it has some drawbacks : Does not take into account backward axis. Performances degrade in the presence of // Indeed, with a query like : /descendant-or-self::title/child::text() one has to iterate through the whole document. 6 / 30

13 Solution : Queries + type based optimisations By using a DTD and a given set of queries, we can statically infer a projector for the set of queries and use it to prune the document. Main advantages : Pruning is efficient Take into account // and backward axis Pruning is sound : executing the query on the projected document gives the same result as on the original. Pruning is precise and even exact for a large class of documents 7 / 30

14 Outline 1 Introduction 2 Notations 3 Algorithm 4 Formal results 5 Experiments 6 Conclusion 8 / 30

15 Definition (DTD) A DTD is a pair (X, E) where X is a distinguished name (the root) and E is a set of productions of the form where each R i is of the form : {X 1 R 1,..., X n R n } a i [Regexp] or String where a i is a unique tag name and Regexp a regular expression of X i. Names(E) is the set of names occuring in E. 9 / 30

16 Definition (Projector) For a given DTD (X, E) a type projector for (X, E) is a set of names P such that : 1 P Names(E) 2 X i P there is at least one derivation from X to X i in E 10 / 30

17 Definition (Pruning) The pruning (or projection) of a document D valid w.r.t a DTD (X, E) with the projector P is a document D where every node not generated by a name in P is erased (replaced by the empty sequence). 11 / 30

18 Definition (Pruning) The pruning (or projection) of a document D valid w.r.t a DTD (X, E) with the projector P is a document D where every node not generated by a name in P is erased (replaced by the empty sequence). <!ELEMENT a <!ELEMENT b <!ELEMENT c <!ELEMENT d <!ELEMENT e (b,c)> (d,e)> (d,e)> (#PCDATA)> (#PCDATA)> b d a e c 11 / 30

19 /descendant-or-self::c 12 / 30

20 X a a[x b, X c ] X b b[x d, X e ] X c c[x d, X e ] X d String X e String <a> <d>lotsofdata</d> <e>superlongstring<e/> <c> <d>bar</d> <e>foo<e/> </c> </a> /descendant-or-self::c 12 / 30

21 /descendant-or-self::c X a a[x b, X c ] X b b[x d, X e ] X c c[x d, X e ] X d String X e String <a> <d>lotsofdata</d> <e>superlongstring<e/> <c> <d>bar</d> <e>foo<e/> </c> </a> P = {X a, X c, X d, X e } <a> <d>lotsofdata</d> <e>superlongstring<e/> <c> <d>foo</d> <e>bar<e/> </c> </a> 12 / 30

22 /descendant-or-self::c X a a[x b, X c ] X b b[x d, X e ] X c c[x d, X e ] X d String X e String <a> <d>lotsofdata</d> <e>superlongstring<e/> <c> <d>bar</d> <e>foo<e/> </c> </a> P = {X a, X c, X d, X e } <a> <d>lotsofdata</d> <e>superlongstring<e/> <c> <d>foo</d> <e>bar<e/> </c> </a> 12 / 30

23 Definition (XPath (1.0)) Q ::= Axis :: Test Axis :: Test[Expr] Q/Q Expr := Q Expr op Expr f(expr,..., Expr) Base Axis ::= self child descendant parent ancestor... Test ::= tag node() text() Definition (Simple path) Q ::= Axis :: Test Axis :: Test[Cond] Q/Q SPath ::= Axis :: Test Axis :: Test/SPath Cond := SPath Cond or Cond 13 / 30

24 Outline 1 Introduction 2 Notations 3 Algorithm 4 Formal results 5 Experiments 6 Conclusion 14 / 30

25 Overall process Static inference : DTD 15 / 30

26 Overall process Static inference : DTD Q2 Q3 Q1 Q4 Q5 15 / 30

27 Overall process Static inference : DTD Q2 Q3 Q1 Q4 Q5 P 15 / 30

28 Overall process Runtime : D1 D3 D2 D1 15 / 30

29 Overall process Runtime : D1 D3 D2 D1 P 15 / 30

30 Overall process Runtime : D1 D3 D2 D1 P D 1 D 1 D 2 D 1 15 / 30

31 Overall process Runtime : D1 D3 D2 D1 P D 1 D 1 D 2 D 1 Q2 Q3 Q1 Q4 Q5 D1 R3 R2 R1 15 / 30

32 Static inference Q2 Q3 Q1 Q4 Q5 16 / 30

33 Static inference Q2 Q3 Q1 Q4 Q5 XPath XPath 1 XPath 2 XPath 7 XPath 3 XPath XPath 1 XPath 6 XPath 4 XPath XPath / 30

34 Static inference Q2 Q3 Q1 Q4 Q5 XPath XPath 1 XPath 2 XPath 7 XPath 3 XPath XPath 1 XPath 6 XPath 4 XPath XPath 1 5 SPath XPath 11 SPath 2 SPath 7 SPath 3 SPath XPath 1 XPath 6 SPath 4 SPath SPath / 30

35 Static inference Q2 Q3 Q1 Q4 Q5 XPath XPath 1 XPath 2 XPath 7 XPath 3 XPath XPath 1 XPath 6 XPath 4 XPath XPath 1 5 SPath XPath 11 SPath 2 SPath 7 SPath 3 SPath XPath 1 XPath 6 SPath 4 SPath SPath 1 5 SPath 1 16 / 30

36 Static inference Q2 Q3 Q1 Q4 Q5 XPath XPath 1 XPath 2 XPath 7 XPath 3 XPath XPath 1 XPath 6 XPath 4 XPath XPath 1 5 SPath XPath 11 SPath 2 SPath 7 SPath 3 SPath XPath 1 XPath 6 SPath 4 SPath SPath 1 5 SPath 1 DTD 16 / 30

37 Static inference Q2 Q3 Q1 Q4 Q5 XPath XPath 1 XPath 2 XPath 7 XPath 3 XPath XPath 1 XPath 6 XPath 4 XPath XPath 1 5 SPath XPath 11 SPath 2 SPath 7 SPath 3 SPath XPath 1 XPath 6 SPath 4 SPath SPath 1 5 SPath 1 Type Projector Inference DTD Type Inference 16 / 30

38 Static inference Q2 Q3 Q1 Q4 Q5 XPath XPath 1 XPath 2 XPath 7 XPath 3 XPath XPath 1 XPath 6 XPath 4 XPath XPath 1 5 SPath XPath 11 SPath 2 SPath 7 SPath 3 SPath XPath 1 XPath 6 SPath 4 SPath SPath 1 5 SPath 1 Type Projector Inference DTD Type Inference P1 16 / 30

39 Static inference Q2 Q3 Q1 Q4 Q5 XPath XPath 1 XPath 2 XPath 7 XPath 3 XPath XPath 1 XPath 6 XPath 4 XPath XPath 1 5 SPath XPath 11 SPath 2 SPath 7 SPath 3 SPath XPath 1 XPath 6 SPath 4 SPath SPath 1 5 SPath SPath 12 3 DTD Type Projector Inference Type Inference P1 P2 P1 16 / 30

40 Static inference Q2 Q3 Q1 Q4 Q5 XPath XPath 1 XPath 2 XPath 7 XPath 3 XPath XPath 1 XPath 6 XPath 4 XPath XPath 1 5 SPath XPath 11 SPath 2 SPath 7 SPath 3 SPath XPath 1 XPath 6 SPath 4 SPath SPath 1 5 SPath SPath 12 3 Type Projector Inference DTD Type Inference P1 P2 P1 P 16 / 30

41 Type inference typeinf(dtd,{x},path) = T Type T is the set of names of types of nodes in the result. 17 / 30

42 Type inference typeinf(dtd,{x},path) = T Type T is the set of names of types of nodes in the result. T = empty query. 17 / 30

43 Type inference typeinf(dtd,{x},path) = T Type T is the set of names of types of nodes in the result. T = empty query. /self::a/child::b/child::d 17 / 30

44 typeinf(dtd,{x},path) = T Type T is the set of names of types of nodes in the result. T = empty query. /self::a/child::b/child::d 1 typeinf(dtd,{x a },self::a)= {X a } a Type inference b c d e 17 / 30

45 typeinf(dtd,{x},path) = T Type T is the set of names of types of nodes in the result. T = empty query. /self::a/child::b/child::d 1 typeinf(dtd,{x a },self::a)= {X a } 2 typeinf(dtd,{x a },child::b)= {X b } Type inference a b c d e 17 / 30

46 Type inference typeinf(dtd,{x},path) = T Type T is the set of names of types of nodes in the result. T = empty query. /self::a/child::b/child::d 1 typeinf(dtd,{x a },self::a)= {X a } 2 typeinf(dtd,{x a },child::b)= {X b } 3 typeinf(dtd,{x b },child::d)= {X d } b a c d e 17 / 30

47 Type inference typeinf(dtd,{x},path) = T Type T is the set of names of types of nodes in the result. T = empty query. /self::a/child::b/child::d 1 typeinf(dtd,{x a },self::a)= {X a } 2 typeinf(dtd,{x a },child::b)= {X b } 3 typeinf(dtd,{x b },child::d)= {X d } P = {X a, X b, X d } b a c d e 17 / 30

48 //*/self::b/child::d 18 / 30

49 //*/self::b/child::d 1 typeinf(dtd,{x a },//*) a b c d e 18 / 30

50 //*/self::b/child::d 1 typeinf(dtd,{x a },//*) a b c d e 18 / 30

51 //*/self::b/child::d 1 typeinf(dtd,{x a },//*) 1 typeinf(dtd,{x b },self::b/child::d)= {X d } a b c d e 18 / 30

52 //*/self::b/child::d 1 typeinf(dtd,{x a },//*) 1 typeinf(dtd,{x b },self::b/child::d)= {X d } 2 typeinf(dtd,{x c },self::b/child::d)= b a c d e 18 / 30

53 //*/self::b/child::d 1 typeinf(dtd,{x a },//*) 1 typeinf(dtd,{x b },self::b/child::d)= {X d } 2 typeinf(dtd,{x c },self::b/child::d)= 3 typeinf(dtd,{x d },self::b/child::d)= b a c d e 18 / 30

54 //*/self::b/child::d 1 typeinf(dtd,{x a },//*) 1 typeinf(dtd,{x b },self::b/child::d)= {X d } 2 typeinf(dtd,{x c },self::b/child::d)= 3 typeinf(dtd,{x d },self::b/child::d)= 4 typeinf(dtd,{x e },self::b/child::d)= b a c d e 18 / 30

55 //*/self::b/child::d 1 typeinf(dtd,{x a },//*) 2 typeinf(dtd,{x b },self::b)= {X b } a b c d e 18 / 30

56 //*/self::b/child::d 1 typeinf(dtd,{x a },//*) 2 typeinf(dtd,{x b },self::b)= {X b } 3 typeinf(dtd,{x b },child::d)= {X d } b a c d e 18 / 30

57 //*/self::b/child::d 1 typeinf(dtd,{x a },//*) 2 typeinf(dtd,{x b },self::b)= {X b } 3 typeinf(dtd,{x b },child::d)= {X d } P = {X a, X b, X d } b a c d e 18 / 30

58 /self::a/child::b/child::d/parent::node()/child::d 19 / 30

59 /self::a/child::b/child::d/parent::node()/child::d 1 typeinf(dtd,{x a },self::a)= {X a } 2 typeinf(dtd,{x a },child::b)= {X b } 3 typeinf(dtd,{x b },child::d)= {X d } b a c d e 19 / 30

60 /self::a/child::b/child::d/parent::node()/child::d 1 typeinf(dtd,{x a },self::a)= {X a } 2 typeinf(dtd,{x a },child::b)= {X b } 3 typeinf(dtd,{x b },child::d)= {X d } 4 typeinf(dtd,{x d },parent::node())= {X b, X c } b a c d e 19 / 30

61 /self::a/child::b/child::d/parent::node()/child::d 1 typeinf(dtd,{x a },{X a },self::a)= {X a } a b c d e 19 / 30

62 /self::a/child::b/child::d/parent::node()/child::d 1 typeinf(dtd,{x a },{X a },self::a)= {X a } 2 typeinf(dtd,{x a },{X a },child::b)= {X b } a b c d e 19 / 30

63 /self::a/child::b/child::d/parent::node()/child::d 1 typeinf(dtd,{x a },{X a },self::a)= {X a } 2 typeinf(dtd,{x a },{X a },child::b)= {X b } 3 typeinf(dtd,{x b },{X a, X b },child::d)= {X d } b a c d e 19 / 30

64 /self::a/child::b/child::d/parent::node()/child::d 1 typeinf(dtd,{x a },{X a },self::a)= {X a } 2 typeinf(dtd,{x a },{X a },child::b)= {X b } 3 typeinf(dtd,{x b },{X a, X b },child::d)= {X d } 4 typeinf(dtd,{x d },{X a, X b, X d },parent:: node())= {X b } b a c d e 19 / 30

65 /self::a/child::b/child::d/parent::node()/child::d 1 typeinf(dtd,{x a },{X a },self::a)= {X a } 2 typeinf(dtd,{x a },{X a },child::b)= {X b } 3 typeinf(dtd,{x b },{X a, X b },child::d)= {X d } 4 typeinf(dtd,{x d },{X a, X b, X d },parent:: node())= {X b } 5 typeinf(dtd,{x b },{X a, X b, X d },child::d)= {X d } b a c d e 19 / 30

66 /self::a/child::b/child::d/parent::node()/child::d 1 typeinf(dtd,{x a },{X a },self::a)= {X a } 2 typeinf(dtd,{x a },{X a },child::b)= {X b } 3 typeinf(dtd,{x b },{X a, X b },child::d)= {X d } 4 typeinf(dtd,{x d },{X a, X b, X d },parent:: node())= {X b } 5 typeinf(dtd,{x b },{X a, X b, X d },child::d)= {X d } b a c P = {X a, X b, X d } d e 19 / 30

67 Outline 1 Introduction 2 Notations 3 Algorithm 4 Formal results 5 Experiments 6 Conclusion 20 / 30

68 Soundness Theorem (Soundness) Let D be a document valid w.r.t a DTD (X, E) and p a path. Let P the type projector deduced from p. Let D be the projection of D with P. eval(p,d) = eval(p,d ) 21 / 30

69 Completeness Pruning is precise / 30

70 Completeness Pruning is precise... Theorem (Completeness) P = {X 1,..., X n } the type projector associated with a path p and a DTD (X, E). Let P = P\{X i }. There exists D a document and it s projection D such that : eval(p,d) eval(p,d ) Completeness holds with : some restrictions on the DTD (star-guarded, non-recursive,... ) 22 / 30

71 Completeness Pruning is precise... Theorem (Completeness) P = {X 1,..., X n } the type projector associated with a path p and a DTD (X, E). Let P = P\{X i }. There exists D a document and it s projection D such that : eval(p,d) eval(p,d ) Completeness holds with : some restrictions on the DTD (star-guarded, non-recursive,... ) some restrictions on the path (no upward axis in predicates,... ) 22 / 30

72 Completeness Pruning is precise... Theorem (Completeness) P = {X 1,..., X n } the type projector associated with a path p and a DTD (X, E). Let P = P\{X i }. There exists D a document and it s projection D such that : eval(p,d) eval(p,d ) Completeness holds with : some restrictions on the DTD (star-guarded, non-recursive,... ) some restrictions on the path (no upward axis in predicates,... ) 22 / 30

73 Completeness Pruning is precise... Theorem (Completeness) P = {X 1,..., X n } the type projector associated with a path p and a DTD (X, E). Let P = P\{X i }. There exists D a document and it s projection D such that : eval(p,d) eval(p,d ) Completeness holds with : some restrictions on the DTD (star-guarded, non-recursive,... ) some restrictions on the path (no upward axis in predicates,... ) 22 / 30

74 Outline 1 Introduction 2 Notations 3 Algorithm 4 Formal results 5 Experiments 6 Conclusion 23 / 30

75 Protocol : Linux desktop, 512MB Ram, 3Ghz x86 CPU (and no swap). Implementation in OCaml. Validity of the result checked with the Galax query engine. Pruning tested on XMark and XPathMark benchmarks. 24 / 30

76 Protocol : Linux desktop, 512MB Ram, 3Ghz x86 CPU (and no swap). Implementation in OCaml. Validity of the result checked with the Galax query engine. Pruning tested on XMark and XPathMark benchmarks. Prunning is one pass bufferless traversal of the document. In practice, computing the type projector is fast. 24 / 30

77 Memory (in MB) QM03 QM06 QM07 QM14 QM15 QM19 QP01 QP02 QP03 QP04 QP05 QP06 QP07 QP08 QP09 QP10 QP11 QP12 QP13 QP21 QP23 Original document Query Axes present in the query Projected document Runtime conditions Memory used to process a 56 MB document with Galax. 25 / 30

78 Gain QM03 QM06 QM07 QM14 QM15 QM19 Original Size (MB) Pruned Size(MB) 25 5, Memory Usage (MB) % of original size Gain in Speed ( faster) / 30

79 Comparison with previous work Qualitative : with one exception, type projectors are always equal or more efficient. 27 / 30

80 Comparison with previous work Qualitative : with one exception, type projectors are always equal or more efficient. Performances : pruning is linear in time (in the size of the document) and constant in memory. It can be done while validating the document. 27 / 30

81 Comparison with previous work Qualitative : with one exception, type projectors are always equal or more efficient. Performances : pruning is linear in time (in the size of the document) and constant in memory. It can be done while validating the document. Features : handles backward as well as following/preceding axes. 27 / 30

82 Outline 1 Introduction 2 Notations 3 Algorithm 4 Formal results 5 Experiments 6 Conclusion 28 / 30

83 Conclusion Formal foundations ensure validity of the pruning and it s efficiency. 29 / 30

84 Conclusion Formal foundations ensure validity of the pruning and it s efficiency. Handle any XQuery query, either directly or by approximating it. 29 / 30

85 Conclusion Formal foundations ensure validity of the pruning and it s efficiency. Handle any XQuery query, either directly or by approximating it. Whereas approximations are made for runtime conditions, the technique still preserve a high degree of precision. 29 / 30

86 Conclusion Formal foundations ensure validity of the pruning and it s efficiency. Handle any XQuery query, either directly or by approximating it. Whereas approximations are made for runtime conditions, the technique still preserve a high degree of precision. No additional cost at runtime. 29 / 30

87 Future work Many directions : adapt our approach to work on untyped documents by using data-guides or path summaries. 30 / 30

88 Future work Many directions : adapt our approach to work on untyped documents by using data-guides or path summaries. extend the formalism to handle XML-Schema rather than DTDs. 30 / 30

89 Future work Many directions : adapt our approach to work on untyped documents by using data-guides or path summaries. extend the formalism to handle XML-Schema rather than DTDs. integration with classical databases techniques. 30 / 30

90 Future work Many directions : adapt our approach to work on untyped documents by using data-guides or path summaries. extend the formalism to handle XML-Schema rather than DTDs. integration with classical databases techniques. integration with a query engine. 30 / 30

Type-Based XML Projection

Type-Based XML Projection Véronique Benzaken 1 Giuseppe Castagna 2 Dario Colazzo 1 Kim Nguyê n 1 1 LRI, Université Paris-Sud 11, Orsay - France 2 École Normale Supérieure de Paris - France ABSTRACT XML