Optimizationin XSLT and XQuery
Michael Kay
2
Challenges
• XSLT/XQuery are high-level declarative languages: performance depends on good optimization
• Performance also depends on good programming!
• How can users write good programs if they don’t know what the optimizer is doing?
3
What is optimization?
• Widest sense:– Everything that’s done to make your
query go fast
• Narrower sense:– Expression rewriting: replacing the
code that you write with equivalent, faster code that has the same effect
4
Main performance contributors
• Efficient internal coding
• Tree model for documents
• Streamed execution (pipelining)+ lazy evaluation
• Rewrite optimizations– Including join optimization
• Tail recursion
• XSLT template rule matching
5
Databases vs. in-memory processors
• Databases– 90% of optimization is about finding and using indexes
– You can spend more time building the data to reduce query costs
– Indexes are long-lived
– Queries may be repeatable or one-off
• In-memory processors– Loading the data is a significant part of the overall cost
– Memory utilization needs to be minimized
– Indexes, if used, are transient
– Queries/Stylesheets may be repeatable or one-off
6
The Saxon TinyTree Model
• Requirements:– Low memory footprint– Fast construction– Fast access paths– Support for document order
• Non-requirement:– In-situ update
7
TinyTree example
<root> <a>12</a> <b>Prague</b></root>
assume whitespace is stripped
nr depth kind name next/owner
type data
0 0 DOC - - - -
1 1 ELEM 3028(root)
0 9735 -
2 2 ELEM 3164(a)
4 INT
3 3 TEXT - 2 - ►12
4 2 ELEM 3165(b)
1 STR
5 3 TEXT - 4 - ►”Prague”
6 0 STOP - - - -
8
TinyTree: key points
• No “object-per-node” overhead• Names held as integer codes• Fast child navigation• Fast document order comparison• Extra information added dynamically
if needed:– preceding-sibling pointers– Base-uri, line numbers etc– Indexes
9
Streaming (Pipelining)
• Common practice in set-based languages– Functional programming languages– SQL
• Each node in the expression tree can deliver its results incrementally to the parent node
• Can be implemented as pull or push (Saxon uses both)
10
Example: filter expressions
Class FilterExpressionIterator {
public Item next() {
while (true) {
Item item = base.next(); if (item == EOS) return EOS;
if (matches(item, predicate)) return item;
}
}
$nodes[x=1]
filter
$nodes =
x 1
11
Example: Many-to-One Comparisons
Class ManyToOneComparisonEvaluator {
public boolean evaluate () { while (true) {
Item item = lhs.next();
if (item = rhs) return true;
}
return false;
}
x=1=
x 1
12
Benefits of Streaming
• Saves memory– No memory for intermediate results– Allocating and de-allocating memory
takes time
• Early exit, for example in– (a/b/c/d)[1]– book[author = ‘Smith’]– exists(//@xml:space)
13
Lazy Evaluation
• Closely associated with streaming
• Variables and function arguments are not evaluated until the value is needed
• Benefits:– The value might never be needed– Only part of the value might be needed
(early exit)– Memory is used for the minimum time
14
Compile-time Expression Rewrites
General approach:1. Parse the source code into an expression tree
2. Resolve references (variables, functions)
3. Decorate the tree with attributes– Type of an expression– Dependencies of an expression– Other properties, e.g. whether a node-set is sorted
4. Scan the tree repeatedly to identify expressions that can be replaced by faster equivalents
15
Two kinds of rewrites
• Rewrites that could have been done by the programmer– count(A) > 3 ►exists(A[4])
• Rewrites that use constructs not available to the programmer– A[position()=last()] ► A[isLast()]
16
Some important rewrites
• Sort removal– Not sorting path expressions where the
result is already sorted
• Constant subexpressions– Evaluated at compile time where possible
• Extracting subexpressions from loops
• Distributing WHERE conditions
• + many ad-hoc rewrites
17
Some rewrites that Saxon doesn’t yet do
• Inline expansion of variable references
• Inline expansion of function calls
• Detecting common subexpressions
• Creating new global variables
18
Type Checkingand its effect on performance
• XQuery and XSLT 2.0 allow you to declare types of variables and functions– But it’s not mandatory
• Main benefit is better error detection
• Type information can also be used by the optimizer– With Saxon, this rarely makes a big
difference
19
“Optimistic static type checking”
• The static type of an expression S is compared with the required type R
• Possible outcomes:– S is a subtype of R: no action needed– S overlaps with R: run-time type-checking
code is generated– S and R are disjoint: static error reported
• Special case:– integer* and string* overlap
(both allow the empty sequence)
20
Join Optimization
• Less important in XQuery than in SQL– Except that some people write XQuery
as if it were SQL
• General strategy in Saxon-SA:– Distribute the join predicates (turn
WHERE clauses into filter expressions)
– Use indexed lookup for predicates where appropriate
21
Indexes in Saxon-SA
• Explicit user-defined indexes– xsl:key
• Implicit document-level indexes– //a/b/c[@id=$param]
• Implicit sequence-level indexes– $abc[@id = $param]
• Hash tables for many-to-many “=“– book[keyword = $keywords]
22
Some tips for effective indexing
• Declare your types
• Avoid untypedAtomic – use a schema
• Use “eq” rather than “=“
23
Tail Recursion
• See the printed paper
24
Conclusions
• Optimization techniques are similar for XSLT and XQuery– But vary between database products
and in-memory processors
• Compile-time techniques– Type analysis– Expression rewriting
• Run-time techniques– Streaming/pipelining– etc