1. Lecture 9: XML Compression

2. Semistructured Data / XML
loosely structured (no restrictions on tags & nesting relationships)
no schema required
XML
under the “semistructured” umbrella
self-describing
the standard for information representation & exchange

3. XML data file can be modeled in a tree form
<Staff>
<Name>
<FirstName> Raymond </FirstName>
<LastName> Wong </LastName>
</Name>
<Login> wong </Login>
<Ext> 5932 </Ext>
</Staff>
Staff
Name
Login
Ext
“wong”
“5932”
“Raymond”
“Wong”
FirstName
LastName

4. XPath evaluation
<a><b><c>12</c><d>7</d></b><b><c>7</c></b></a> a b b
/ a / b [c = “12”] c d c 12 7 7

5. Query evaluation
Top-down
Bottom-up
Hybrid

6. XPath evaluation
<a><b><c>12</c><d>7</d></b><b><c>7</c></b></a> a b b
/ a / b [c = “12”] c d c 12 7 7

7. XPath evaluation
<a><b><c>12</c><d>7</d></b><b><c>7</c></b></a> a b b
/ a / b [c = “12”]
<b><c>12</c><d>7</d></b> c d c 12 7 7

8. Path indexing
Traversing graph/tree almost = query processing for semistructured / XML data
Normally, it requires to traverse the data from the root and return all nodes X reachable by a path matching the given regular path expression
Motivation: allows the system to answer regular path expressions without traversing the whole graph/tree

9. Major Criteria for indexing
Speed up the search (by cutting the search space down)
Relatively smaller size than the original data graph/tree
Easy to maintain (during data loading during updates)

10. An Example of DAG Data root o12 o1 o2 o3 o4 o5 o6 o7 o8 o9 o10 o11 o13
member
dept
support
staff
name
phone

11. Index graph based on language-equivalence
a reduced graph that summarizes all paths from the root in the data graph
The paths from root to o12
staff
dept/member
support/member

12. Language-equivalent nodes
Let L(x) := {w |  a path from the root to x labeled w}
The set L(x) may be infinite when there are cycles
Nodes x, y are language-equivalent (x  y) if L(x) = L(y)
We construct index I by taking the nodes to be the equivalent classes for 

13. Language-equivalent The paths from root to o3
staff
dept/member
Paths to o4 happen to be exactly the same 2 sequences
Same for o8 and o12
o3  o4  o8  o12

14. Equivalence classes o3  o4  o8  o12 o1  o2  o7 o12  o13
root
o12 o1 o2 o3 o4 o5 o6 o7 o8 o9
o10
o11
o13
member
dept
support
staff
name
phone
o3  o4  o8  o12
o1  o2  o7
o12  o13
o5  o6  o9
o10
o11

15. The index graph root o1, o2, o7 o3, o4, o8, o12 o12, o13 o5, o6, o9
member
support
staff
dept
name
phone

16. Query processing based on the index graph
root
o1, o2, o7
o3, o4, o8, o12
o12, o13
o5, o6, o9
o10
o11
member
support
staff
dept
name
phone
dept/member/(name | phone)
-> dept/member/name UNION dept/member/phone
-> {o5, o6, o9} UNION {o10}
-> {o5, o6, o9, o10}

17. About this indexing scheme
The index graph is never > the data
In practice, the index graph is small enough to fit in memory
Construct the index is however a problem
check two nodes are language-equivalent is very expensive (are PSPACE)
approximation based on bisimulation exists

18. A Data Guide root dept support staff o11 o1, o2, o7 o3, o4, o8, o12
member
phone
member
name
o12, o13
o3, o4, o8, o12
o5, o6, o9
o10
phone
name
o5, o6, o9
o10

19. About Data Guide unique labels at each node
(hence) extents are no longer disjoint
query processing proceeds as before
size of the index may >= data size
good for data that is regular & has no cycles

20. XML-Specific Compressors
Unqueriable Compression (e.g. XMill):
Full-chunked: data commonalities eliminated
Very good compression ratio
Queriable Compression (e.g. XGrind, XPRESS):
Fine-grained: data commonalities ignored
Inadequate compression ratio and time
Support simple path queries with atomic predicate

21. XMill First specialized compressor for XML data
SAX parser for parsing XML data
Still using gzip as its underlying compressor
Clever grouping of data into containers for compression
Compress XML via three basic techniques
Compress the structure separately from the data
Group the data values according to their types
Apply semantic (specialized) compressors:
Downloadable:

22. XMill Architecture:

23. An Example:Web Server Logs
ASCII File 15.9 MB (gzipped 1.6MB):
|GET / HTTP/1.0|text/html|200|1997/10/01-00:00:02|-|4478|-|-|
XML-ized apache web log inflates to 24.2 MB (gzipped 2.1MB):
<apache:entry>
<apache:host> </apache:host>
<apache:requestLine> GET / HTTP/1.0 </apache:requestLine>
<apache:contentType> text/html </apache:contentType>
<apache:statusCode> 200</apache:statusCode>
<apache:date> 1997/10/01-00:00:02</apache:date>
<apache:byteCount> 4478</apache:byteCount>
<apache:referer> </apache:referer>
<apache:userAgent> Mozilla/3.1$[$ja$]$(I)</apache:userAgent>
</apache:entry>

24. How Xmill Works: Three Ideas
Compress the structure separately from the data:
gzip Structure
gzip Data
<apache:entry>
<apache:host> </apache:host>
. . .
</apache:entry>
GET / HTTP/1.0
text/html
200 … +
=1.75MB

25. How Xmill Works: Three Ideas
Group the data values according to their types:
gzip Structure
gzip Data1
gzip Data2
<apache:entry>
. . .
</apache:entry> …
GET / HTTP/1.0
GET / HTTP/1.1 … + +
=1.33MB

26. How Xmill Works: Three Ideas
Apply semantic (specialized) compressors:
gzip Structure + gzip c1(Data1) + gzip c2(Data2) + ...
=0.82MB
Examples:
8, 16, 32-bit integer encoding (signed/unsigned)
differential compressing (e.g. 1999, 1995, 2001, 2000, 1995, ...)
compress lists, records (e.g  4 bytes)
Need user input to select the semantic compressor

27. Experiments

28. XML Compression

29. Compression Time

30. Transfer Time (& Decode)

31. XGRIND (Tolani & Haritsa, 2002)
Encodes elements and attributes using XMill’s approach
DTD-conscious: enumerated attributes with k possible values are encoded using a log2 k-bit scheme
Data values are encoded using non-adaptive Huffman coding
Requires two passes over the input document
Separate statistical model for each element/attribute
Homomorphic compression: compressed document retains original structure
June 24, 2008
XML Compression Techniques 31

32. XML Compression Techniques
XGRIND
Original Fragment:
Compressed Fragment:
<student name=“Alice“> <a1>78</a1> <a2>86</a2> <midterm>91</midterm> <project>87</project> </student>
T0 A0 nahuff(Alice) T1 nahuff(78) / T2 nahuff(86) / T3 nahuff(91) / T4 nahuff(87) / /
June 24, 2008
XML Compression Techniques 32

33. XML Compression Techniques
XGRIND
Many queries can be carried out entirely in compressed domain
Exact-match, prefix-match
Some others require only decompression of relevant values
Range, substring
Queryability comes at the expense of achievable compression ratio: typically within 65-75% that of XMill
June 24, 2008
XML Compression Techniques 33

34. ISX Requirements Space does matter for many applications
Generally reducing space improves cache locality
Indirection is expensive
Support fast navigations
Support fast insertion and deletion
Support efficient joins
Separate topology, text and schema

35. ISX Goal
To find a space-efficient storage scheme for XML data without compromising both query and update performances

36. Proposed Storage Structure
The ISX Structure

37. Sample DBLP XML Fragment

38. Balanced Parenthesis Encoding

39. Node Navigations

40. Topology Tiers No. of ) No. of ( No. of text nodes
Min, max of forward excess
Min, max of backward excess

41. Primitive operators

42. Topology Tiers No. of ) No. of ( No. of text nodes
Min, max of forward excess
Min, max of backward excess
Excess 2
Where is the close tag?

43. Tier 2 excess

44. Efficient Updates

45. Example
100 MB DBLP document
5 million XML nodes
ISX: 1MB topology

46. Another example 100M DBLP MSXML ISX Runtime (loading) 329MB 67MB
Core Duo 1.83GHz
1GB RAM
5400 RPM Harddrive
MS Vista
100M DBLP
MSXML
ISX
Runtime (loading)
329MB
67MB
Loading time
17.8s
0.67s
Runtime (//www)
333MB
//www
1.814s
0.143s
5M DBLP
MSXML
ISX
Runtime (loading)
15MB
4MB
Loading time
0.54s
0.035s
Runtime (//www)
21MB
//www
0.096s
0.004s

47. ISX Features

48. Experiments Setup Fixed at 64MB memory buffer Up to 16 GB XML document
E.g. 16 GB DBLP contains > 770 million nodes
NO index or query optimization has been employed for ISX (except for ISX Stream where TurboXPath algorithm has been employed)

49. Storage Size (ISX vs NoK)

50. Storage Size (ISX, XMill, XGrind): DBLP

51. Storage Size (ISX, XMill): TreeBank

52. Bulk Loading Performance

53. Queries

54. Q1: //inproceedings

55. Q5: //article[.//month/text() = “July”]//title

56. Other queries

57. XPath 13 axes We can navigate along 13 axes: ancestor ancestor-or-self
attribute
child
descendant
descendant-or-self
following
following-sibling
namespace
parent
preceding
preceding-sibling
self

58. Node Navigation

59. Full document traversal

60. Update (Insertion) Performance

61. ISX Summary Small storage footprint Small runtime footprint
Fast and consistent performance on navigational access
Superior query performance (further indexing / query optimization can be added)
Superior update performance

62. Compressing and Searching XML Data Via Two Zips
Paolo Ferragina et al.
Slides modified from P. Ferragina’s

63. An XML excerpt It is verbose ! ... <dblp> <book>
<author> Donald E. Knuth </author>
<title> The TeXbook </title>
<publisher> Addison-Wesley </publisher>
<year> 1986 </year>
</book>
<article>
<author> Ronald W. Moore </author>
<title> An Analysis of Alpha-Beta Pruning </title>
<pages> </pages>
<year> 1975 </year>
<volume> 6 </volume>
<journal> Artificial Intelligence </journal>
</article>
...
</dblp>
It is verbose !

64. A tree interpretation... XML document exploration  Tree navigation
XML document search  Labeled subpath searches
Subset of XPath [W3C]

65. The Problem XML-native search engines
We wish to devise a compressed representation for a labeled tree T that efficiently supports some operations:
Navigational operations: parent(u), child(u, i), child(u, i, c)
Subpath searches: given a sequence P of k labels
Content searches: subpath + substring search
Visualization operation: given a node, visualize its descending subtree
XML-aware compressors (like XMill, XmlPpm, ScmPpm,...)
need the whole decompression
XML-native search engines
might exploit this tool as a core block for
query optimization and (compressed) storage
XML-queriable compressors (like XPress, XGrind, XQzip,...)
poor compression and scan of the whole (compressed) file
Summary indexes (like Dataguide, 1-index or 2-index)
large space and do not support “content” searches
Theoretically do exist many solutions, starting from [Jacobson, IEEE Focs ’89]
no subpath/content searches, and poor performance on labeled trees

66. A transform for “labeled trees” [Ferragina et al, IEEE Focs ’05]
We proposed the XBW-transform that mimics on trees the nice structural properties of the Burrows-and-Wheeler Trasform on strings
The XBW linearizes the tree T in 2 arrays s.t.:
the compression of T reduces to use any compressor (gzip, bzip,...) over these two arrays
the indexing of T reduces to implement simple rank/select query operations over these two arrays

67. The XBW-Transform Sa Sp Step 1. C B A D c a b C B D c a A b e C B C
D B C
A C
D A C
Step 1.
Visit the tree in pre-order.
For each node, write down its label
and the labels on its upward path
Permutation
of tree nodes
upward labeled paths

68. The XBW-Transform Sa Sp Step 2. C B A D c a b C b a D c B A e A C B C
D A C
D B C
Step 2.
Stably sort according to Sp
upward labeled paths

69. The XBW-Transform Sp Slast Sa Key fact Step 3. C B A D c a b XBW 1 C bC b a D c B A e
A C
B C C
D A C
D B C
Key fact
Nodes correspond to items in <Slast,Sa>
Step 3.
Add a binary array Slast marking the
rows corresponding to last children

70. XBzip – a simple XML compressor
Tags, Attributes and symbol =
XBW is compressible:
Sa and Spcdata are locally homogeneous
Slast has some structure
Pcdata

71. Some structural propertiesB A D c a b C
XBW B
Slast Sa Sp 1 C b a D c B A e
A C
B C C
D A C
D B C B A B D c b a D D a c a c b
Two useful properties:
Children are contiguous and delimited by 1s
Children reflect the order of their parents

72. XBW is navigational C Sp Slast Sa A 2 B 5 C 9 D 12 C B A D c a b C XBWC b a D c B A e
A C
B C C
D A C
D B C A B
Select in Slast
the 2° item 1
from here... D c b a D D a
Get_children c a c b
Rank(B,Sa)=2
XBW is navigational:
Rank-Select data structures on Slast and Sa
The array C of |S| integers

73. XBW is searchable (count subpaths)
D 12 C B A D c a b
P[i+1]
XBW-index
Slast Sa Sp
P = B D 1 C b a D c B A e
A C
B C C
D A C
D B C fr
Rows whose
Sp starts
with ‘B’ lr
Their children
have upward
path = ‘D B’
Inductive step:
Pick the next char in P[i+1], i.e. ‘D’
Search for the first and last ‘D’ in Sa[fr,lr]
 Jump to their children
XBW is searchable:
Rank-Select data structures on Slast and Sa
Array C of |S| integers fr lr
2 occurrences of P
because of two 1s