Efficient Evaluation of Regular Path Expressions on Streaming XML Data

1
Efficient Evaluation of Regular Path Expressions
on Streaming XML Data

• By - Zachary G. Ives, Alon Y. Levy and Daniel S.
  Weld

2
Table of Contents

• A bit about XML (yes, again)
• Our goal, problem and solution
• Our XML data model
• How to ask questions ?

3
Table of Contents

• X-scan operation and structure
• Digging deep into x-scan
• How good is it ? Performance Evaluation
• Conclusion

4
A Bit About XML (yes, again)

• XML the eXtensible Markup Language
• Become a standard
• Useful for the dissemination and exchange of
  information

5
A Bit About XML (yes, again)

• Advantages
• Simple
• Self-describing nature
• Flexible
• Represents both structured and semi-structured
  data

6
XML Structure

• Consists of
• Elements pairs of matching open and close tags.
• Elements may enclose additional elements or data
  values.
• Attributes included in element tags.
• Attributes are single-valued and describe the
  element.

7
XML Structure (Cont.)

• ID is special attribute which uniquely identify
  the element.
• IDREF form links the other elements in the
  document.
• Combining ID and IDREF forms a graph structure
  rather than just a tree structure.

8
XML Example
We will use this example throughout the rest of
the lecture
9
Our Goal

• Our goal is to perform queries and search
  operations on the XML document.
• Several query languages have been proposed.
• Represents the XML document as a graph.

10
Our Goal (Cont.)

• Represents the query as a regular path expression
  that should be matched against XML source.
• These regular path expressions describe
  traversals along edges in the XML graph.
• The variables in the query are mapped to XML
  elements along these paths.

11
Our Problem

• Most XML query processors
• Loading the data into a local repository
• Building indexes on the repository
• Processing the query
• The repository is either
• Relational database
• An object oriented database
• A repository of semi structured data

12
Our Problem (Cont.)

• The local storing and indexing is expensive.
• Especially when the query is made over streams of
  incoming XML.
• The streams can come from many sources, some fast
  and some slow.
• Sometimes we want some partial answer but as soon
  as possible.

13
Our Solution

• The query can be performed while the data streams
  in.
• The XML-Scan (x-scan) operator does exactly that.
• Used at the lowest level of the query plan and
  supplies data to other operators.

14
The X-Scan Operator

• Input
• An XML data stream.
• Set of regular path expressions.
• Output
• Stream of binding for the variables occurring in
  the expressions.
• The bindings are produced incrementally, as the
  XML data is streaming in.

15
The X-Scan Operator (Cont.)

• The entire graph can be constructed in a single
  pass.
• X-Scan simultaneously.
• Parse the XML data.
• Indexing nodes by their IDs.
• Resolving IDREFs.
• Return the nodes that match the path expressions
  of the query.

16
The X-Scan Operator (Cont.)

• Some issues in the X-Scan operation are
• Deal with possibly cyclic data
• Preserve order of elements
• Remove duplicate bindings that are generated due
  to multiple paths to the same elements

17
Data Model for XML

• Naturally, the XML data model is a graph.
• Each XML tag is an edge labeled with the tag
  name.
• It is directed to a node which label is the tags
  ID. (if it has no ID it gets a number).
• A given element node will have labeled edges
  directed to its attribute values, sub-elements,
  and any other elements referenced via IDREF.

18
Data Model for XML (Cont.)

• Example is always the best way

19
How to ask questions ?

• A variety of query languages have been proposed.
• The key feature in all of these languages is the
  use of regular path expressions over the data.
• Most of them also give the answer to the query as
  XML document.
• X-Scan uses XML-QL.

20
The XML-QL Syntax

• The syntax of XML-QL is
• patterni template is matched against the XML data
  graph from sourcei and the resulted tuples are
  formatted as described in result.

WHERE pattern1 IN source1, pattern2 IN
source2, CONSTRUCT result
21
The XML-QL Syntax (Cont.)

• An XML-QL pattern is a set of nested tags with
  embedded variable names (prefixed by ) that
  specify bindings of graph nodes to variables.
• The CONSTRUCT clause specifies a tree-structured
  set of edges and nodes to add to the output graph
  for each tuple of variable bindings.

22
The XML-QL Syntax (Cont.)

• Again, example is the best way
• Lets look at

WHERE ltdbgt ltlabgt ltnamegtnlt/gt lt_gtltcitygt
clt/gtlt/gt lt/gt ELEMENT_AS l lt/gt IN
fig1.xml CONSTRUCT ltresultgt ltcentergt ltname
gtnlt/gt ltlocationgtclt/gt lt/gt lt/gt
23
The XML-QL Syntax (Cont.)

• As we can see, the result will be

ltresultgt ltcentergt ltnamegtSeattle Bio
Lablt/namegt ltlocationgtSeattlelt/locationgt lt/cente
rgt ltcentergt ltnamegtPMBLlt/namegt ltlocationgtPhila
delphialt/locationgt lt/centergt lt/resultgt
24
The XML-QL Syntax (Cont.)

• If the variable is bound to a node with
  sub-elements, all the sub-graph will be inserted
  to the resulted graph.
• We will use dot-notation to describe the X-Scan
  operation.
• The previous example will rewritten as.
• El root.db.lab
• En El.name
• Ec El._.city

25
The X-Scan Place

• The goal of the X-Scan operator is therefore to
  produce a set of bindings for each pattern in the
  WHERE clause.

26
So, What X-Scan do ?

• Given the XML Stream and a set of regular path
  expressions, outputs a stream of tuples assigning
  binding values to each variable in the set of
  regular path expression.
• The central mechanism is a set of state machines
  that traverse the XML graph, trying to satisfy
  the path expressions.

27
What is it made of ?

• The data components of X-Scan are

28
Where the data flows?

• As the data streams into the system, several
  structures are created
• The data get parsed and stored locally
• A structural index of the XML graph is created
• An ID index records the IDs of all elements and
  their location in the structural index
• A list of references to not-yet-seen element IDs
  is maintained

29
Where the data flows?

• In parallel to the creation of those structures,
  a set of finite state machines perform a DFS over
  the partial structural index.
• When a machine reaches an accepting state, a new
  value is added to the binding-value table of that
  machine.
• Those values are later combine to form the
  complete image.

30
Example problems

• It sounds easy, but yet there some problems to
  meet, for example
• The handling of cycles
• How to prune duplicate bindings as they are
  created ? Remember X-Scan is online operator

31
The State Machines

• As described earlier, we create one regular
  expression for every variable in the query in
  the dot-notation.
• So, we build a finite-state machine for each
  expression.
• State transition is correspond to edge traversals
  in the XML data graph

32
The State Machines (Cont.)

• The end of the path expression yield an accepting
  state, which outputs instances of the
  corresponding variables.
• When one variable is dependent upon other
  variable, the other variable machine accepting
  state is pointing to the state machine of the
  first one.

33
The State Machines (Cont.)

• And back to our example

34
Indexing the XML Graph

• The structural index should allow x-scan to
  quickly traverse the XML data graph.
• Each node in the index contains
• The ID of the element and its offset in the
  document
• Pointers to all the sub-elements, attributes and
  IDREFs of the element.
• Essentially it looks like the graph except for
  the leafs.

35
The Algorithm Step by Step

• X-Scan proceeds by building the structural index
  and running a set of active state machines in
  parallel.
• The core algorithm is in fact the way those state
  machines run, lets focus on that by running our
  example.

36
The Algorithm Step by Step

• Initially, only the top level machine is active.
• When a machine M reaches an accepting state, it
  produces a binding b for its variable, writes it
  and the parent value to its table and activates
  all of its dependent state machines.

37
The Algorithm Step by Step

• Those machines remain active while x-scan is
  scanning b or any element accessible by a path
  from b.
• The final output of x-scan is the equi-join of
  all the appropriate tables.

38
The Algorithm By Example

• Ml is initialized on state 1 as the only active
  machine.

39
The Algorithm By Example

• The root got a db edge, so the machine is
  pushed to its stack and moving to state 2 with
  value node 1

40
The Algorithm By Example

• Next, following the first outgoing edge, pushing
  the old state value, and setting Ml to state 3
  with value baselab

41
The Algorithm By Example

• Since it now in accepting state
• the baselab value is written to the Ml table
• Ml is suspended
• Mn and Mc are activated

42
The Algorithm By Example

• The next edge takes Mn from state 4 to 5
• And Mc run on the loop back to state 6
• Both machines have 2 as binding value

43
The Algorithm By Example

• Since Mn is now in an accept state x-scan writes
  lt2,baselabgt into Mns table.
• Since no edges remain for exploration, x-scan
  pops the stack and backs up the state machines,
  resetting Mn to state 4 and Mc to state 6

44
The Algorithm By Example

• The next edge is labeled location so
• Mn stay in state 4
• Mc also stay in state 6 but advanced to node 3
• Then Mc is advanced to state 7 on the city edge
  to node 4

45
The Algorithm By Example

• At this point x-scan writes lt4,baselabgt into
  Mcs table.
• It can also produce the first tuple of bindings
  ltl/baselab,n/2,c/4gt

46
The Algorithm By Example

• X-Scan keeps running Mc but no more cities are
  found
• It pops back up to baselab
• Running Mc along the IDREF to smith1 gives no
  more cities

47
The Algorithm By Example

• Now, Mn and Mc are deactivated and the control
  return to Ml
• X-scan pops up to node 1 to state 2
• The other lab edge yield another tuple
  ltl/lab2,n/6,c/7gt

48
Where should we go ?

• On occasion x-scan will encounter an IDREF to a
  node that has not yet been parsed.
• Unknown node simply will not be in the ID index.

49
Where should we go ?

• When X-Scan hits such unseen reference
• It pauses all the relevant state machines
• Adds an entry to the list of unresolved IDREFs
• ltdesired ID value, referrers addressgt
• Continue to parse and build the structural index

50
Where should we go ?

• Once the target element is parsed x-scan
• fills its address into each referring IDREF in
  the structural index
• Removes the entry from the list of unresolved
  IDREFs
• Awakens the state machines and proceeds

51
Are We Going in Circles ?

• Sometimes the input XML graph contains a cycles.
• X-Scan must not get trap in an infinite loop.

52
Are We Going in Circles ?

• Considering the following XML graph

a
1
a
2
b
a
a
4
3
53
Are We Going in Circles ?

• If we refuse to move in circles, we will miss the
  answer to the query
• Exroot._.b.a
• But if we allow moving in circles we going to get
  in trouble with this one
• Eyroot._.z

54
Are We Going in Circles ?

• What can we do ?
• Now we going to use the stack
• The stack contains pairs of the form
• (binding, state)
• Describing which bindings have been associated
  with states of the machine along the current path.

55
Are We Going in Circles ?

• Since x-scan uses deterministic finite state
  machines, returning to a previous state with the
  same binding will not add any new possible
  actions.
• So, when a machine enters a state, it should
  checks to see that this state has not been bound
  to the same binding along the current path.

56
Are We Going in Circles ?

• Is it working for our example ?
• Look at those state machines.

57
Some Enhancements

• It is important to prevent the operator from
  spending time evaluating paths that are not
  useful.
• There are two enhancements.
• Selection Push-Down.
• Duplicate elimination.

58
Selection Push-Down

• The query optimizer creates and push selection
  operators down into the x-scan operation.
• Works only on attributes since they are single
  valued.
• So X-Scan evaluates all node attribute edges
  before sub-elements edges.

59
Selection Push-Down

• Here also, the best way to explain is by example.

WHERE ltdbgt ltlab managersmith1gt ltnamegtnlt
/gt lt_gtltcitygtclt/gtlt/gt lt/gt ELEMENT_AS
l lt/gt IN fig1.xml CONSTRUCT ltresultgtltcentergt
ltnamegtnlt/gt ltlocationgtclt/gt lt/gtlt/gt
60
Selection Push-Down

• The query plan generator must create an
  additional temporary variable temp1 and a regular
  path expression
• Etemp1El._at_manager
• It also adds a selection predicate
• Etemp1smith1

61
Selection Push-Down

• Now, for the second lab, since it got only ID
  attribute, as X-Scan iterates through all lab2
  attributes it finds no manager attribute.
• So it can short-circuit on this sub-graph.
• Discarding the value of l and ignoring its
  children.

62
Duplicate Elimination

• Sometimes we can visit an element multiple times
  through different paths.
• This can produce duplicate binding tuples.
• The naive way is to do some post-processing
  stage.
• It can be done smartly so there is no need to
  save the entire history.

63
There is a big one on the way

• Main memory may not be large enough to handle all
  of the index structures.
• The way to handle is by
• Paging the XML source document
• Paging the structural index
• Conventional buffer manager using LRU or some
  other similar policy is sufficient.

64
There is a big one on the way

• But what about the ID lookup index, list of
  unresolved IDREFs and the state machine stack?
• They use either a B-Tree or a multilevel
  hashtable
• The size of stack is bounded by the product of
  number of variables and the longest non-repeating
  path.
• Inactive state machine stack can be naturally
  paged

65
How good is it ?

• They used the IBM XML4C parser version 3.0.1 with
  the SAX parser API to implement the X-Scan.
• The SAX API provides callbacks to the code as
  elements are read, and so allowing X-Scan to
  evaluate streaming XML data

66
How good is it ?

• They compare the X-Scan against
• Stanfords Lore semi-structured/XML database
  system.
• A commercial OO-based XML repository
• The experiments were performed with locally
  stored XML files
• X-Scan lose some of its advantages

67
How good is it ?

• Also this X-Scan implementation didnt include
  selection predicates.
• All the queries were performed on a single
  processor 450MHz Pentium II with 256 MB of
  memory.
• X-Scan and the OO-based system run on Windows-NT
  and Lore run on Linux.

68
How good is it ?

• The queries they had performed included the
  following documents.
• Mondial and VLDB contains many references whereas
  the rest are mostly tree structures.

69
How good is it ?
70
How good is it ?

• Conclusions.
• Neither Lore nor the commercial system scale up
  well to queries across multi-megabyte data files.
• They failed particularly on files that contain
  graph structure.
• X-Scan scale better in all cases.

71
How good is it ?

• Another experiments on synthetic XML data files
  were conducted.
• Those XMLs are random generated.
• They was to check the scalability of X-Scan.
• They averaged three different runs across each of
  the three different random graphs of the same
  generation parameters.

72
How good is it ?

• The first experiments were conducted on a
  tree-structured data.
• Therefore they didnt have to build the
  structural index, ID and IDREFs tables.
• This was to check how good the state machines
  work.

73
How good is it ?

• The results were

74
How good is it ?

• Next, they wanted to check for the cost of the
  graph indexing and resolving references.
• Without traversing the IDREFs.
• They took the same graphs from the previous tests
  and change back the DTD so it will considered as
  graph.
• The results were.

75
How good is it ?
76
How good is it ?
77
How good is it ?

• Next, they wanted to check the effectiveness of
  the structural index when called to evaluate such
  reference edges.
• The results were.

78
How good is it ?
79
How good is it
80
Conclusion

• X-Scan differs in three key ways from previous
  works
• The structural index allow more efficient
  traversing without splitting the data to table or
  objects that should be re combined later
• X-Scan state machines are based on the query
  rather then on the data source not reusable,
  but we always reread the data
• X-Scan is pipelined and produces bindings as data
  is being streamed into the system

81
Conclusion (Cont.)

• Another points regarding to X-Scan are
• It handles cycles well
• It preserves the document order and structure
• Eliminate duplicate tuples
• The state machines are independent and so can run
  in parallel
• X-Scan is very efficient, typically imposing 8
  overhead on top of the time required to parse the
  XML document

82
THE END