Tutorial "Linked Data Query Processing" Part 5 "Query Planning and Optimization" (WWW 2013 Ed.)

Linked Data Query Processing
Tutorial at the 22nd International World Wide Web Conference (WWW 2013)
May 14, 2013
http://db.uwaterloo.ca/LDQTut2013/
5. Query Planning
and Optimization
Olaf Hartig
University of Waterloo

WWW 2013 Tutorial on Linked Data Query Processing [ Query Planning and Optimization ] 2
Query Plan Selection
● Possible assessment criteria:
● Benefit (size of computed query result)
● Cost (overall query execution time)
● Response time (time for returning k solutions)
● To select from candidate plans, criteria must be estimated
● For index-based source selection: estimation may be
based on information recorded in the index [HHK+10]
● For (pure) live exploration: estimation impossible
● No a-priori information available
● Use heuristics instead

Outline
 Heuristics-Based Planning
 Optimizing Link Traversing Iterators
➢ Prefetching
➢ Postponing
 Source Ranking
➢ Harth et al. [HHK+10, UHK+11]
➢ Ladwig and Tran [LT10]

Heuristics-Based Plan Selection [Har11a]
● Four rules:
● DEPENDENCY RULE
● SEED RULE
● INSTANCE SEED RULE
● FILTER RULE
● Tailored to LTBQE implemented by link traversing iterators
● Assumptions about queries:
● Query pattern refers to instance data
● URIs mentioned in the query pattern are the seed URIs

?p ex:affiliated_with <http://.../orgaX>
?p ex:interested_in ?b
?b rdf:type <http://.../Book>
Query
DEPENDENCY RULE
● Dependency: a variable from each triple pattern already
occurs in one of the preceding triple patterns
tp1
= ( ?p , ex:affiliated_with , <http://.../orgaX>) I1
tp2
= ( ?p , ex:interested_in , ?b ) I2
tp3
= ( ?b , rdf:type , <http://.../Book> ) I3
Use a dependency respecting query plan
√

Query
DEPENDENCY RULE
tp1
tp2
tp3

Query
DEPENDENCY RULE
● Rationale:
Avoid
cartesian
products
tp1
tp2
tp3

Recall assumption:
seed URIs = URIs in the query
SEED RULE
● Seed triple pattern of a plan
… is the first triple pattern in the plan, and
… contains at least one HTTP URI
● Rationale:
Good starting point
Use a plan with a seed triple pattern
Query
√
√
√

INSTANCE SEED RULE
● Patterns to avoid:
✗ ?s ex:any_property ?o
✗ ?s rdf:type ex:any_class
● Rationale: URIs for vocabulary terms usually resolve to
vocabulary definitions with little instance data
Avoid a seed triple pattern with vocabulary terms
Query
√

FILTER RULE
● Filtering triple pattern: each variable already occurs in one
of the preceding triple patterns
● For each valuation
consumed as input
a filtering TP can
only report 1 or 0
valuations as
output
● Rationale: Reduce
cost
tp2
tp3
Use a plan where all filtering triple patterns are
as close to the first triple pattern as possible
{ ?p = <http://.../alice> }
{ ?p = <http://.../alice> , ?b = <http://.../b1> }
tp2
' = ( <http://.../alice> , ex:interested_in , ?b )
tp3
' = ( <http://.../b1> , rdf:type , <http://.../Book> )
tp1

Outline
➢ Prefetching
➢ Postponing
 Source Ranking
√

Next?
Next?
tp3
tp1
= ( ?p , ex:affiliated_with , <http://.../orgaX> ) I1
tp2
= ( ?p , ex:interested_in , ?b )
tp2
I2
query-local
dataset
{ ?p = <http://.../alice> }
Link Traversing Iterators May Block!

Next?
Next?
tp3
tp1
tp2
tp2
I2
query-local
dataset
{ ?p = <http://.../alice> }
Initiate look-up(s)
and wait

Next?
Next?
tp3
tp1
tp2
tp2
I2
query-local
dataset
{ ?p = <http://.../alice> }
Prefetching of URIs [HBF09]
Ensure look-up
is finished
Initiate
look-up
in the
background
Initiate look-up(s)
and wait

Next?
Next?
tp3
tp1
tp2
tp2
I2
query-local
dataset
{ ?p = <http://.../alice> }
Prefetching of URIs [HBF09]
Wait until look-up
is finished
Initiate
look-up
in the
background
Initiate look-up(s)
and wait

Postponing Iterator [HBF09]
● Idea: temporarily reject an input solution
if processing it would cause blocking
● Enabled by an extension of the iterator paradigm:
● New function POSTPONE: treat the element most recently
reported by GETNEXT as if it
has not yet been reported
(i.e., “take back” this element)
● Adjusted GETNEXT: either return a (new) next element or
return a formerly postponed element

Outline
➢ Prefetching
➢ Postponing
 Source Ranking
√
√

General Idea of Source Ranking
Rank the URIs resulting from source selection
such that
the ranking represents a priority for lookup
● Possible objectives:
● Report first solutions as early as possible
● Minimize time for computing the first k solutions
● Maximize the number of solutions computed in a
given amount of time

Harth et al. [HHK+10, UHK+11]
● For triple patterns this number is directly available:
● Recall, each QTree bucket stores a set of (URI,count)-pairs
● All query-relevant buckets are known after source selection
For any URI u (selected by the QTree-based approach), let:
rank(u) :═ estimated number of solutions that u contributes to
Root
B
C
AA1
A2
B2
B1
Root
A B
A1 A2
C
B1 B2

Harth et al. [HHK+10, UHK+11]
● For triple patterns this number is directly available:
● Recall, each QTree bucket stores a set of (URI,count)-pairs
● All query-relevant buckets are known after source selection
● For BGPs, estimate the number recursively:
● Recursively determine regions of join-able data
(based on overlapping QTree buckets for each triple pattern)
● For each of these regions, recursively estimate number of
triples the URI contributes to the region
● Factor in the estimated join result cardinality of these regions
(estimated based on overlap between contributing buckets)
For any URI u (selected by the QTree-based approach), let:
rank(u) :═ estimated number of solutions that u contributes to

Ladwig and Tran [LT10]
● Multiple scores
● Triple pattern cardinality
● Triple frequency – inverse source frequency (TF–ISF)
● (URI-specific) join pattern cardinality
● Incoming links
● Assumption: pre-populated index that stores triple pattern
cardinalities and join pattern cardinalities for each URI
● Aggregation of the scores to obtain ranks
● For indexed URIs: weighted summation of all scores
● For non-indexed URIs: weighting of (currently known) in-links
● Ranking is refined at run-time

Metric: Triple Pattern Cardinality [LT10]
● Rationale: data that contains many matching triples
is likely to contribute to many solutions
● Requirement: pre-populated index that stores the cardinalities
● Caveat: some triple patterns have a high
cardinality for almost all URIs
● Example: (?x, rdf:type, ?y)
● These patterns do not discriminate URIs
For a selected URI u, and a triple pattern tp (from the query), let:
card(u, tp) :═ number of triples in the data of u that match tp

Metric: TF–ISF [LT10]
● Idea: adopt TF-IDF concept to weight triple patterns
● Triple Frequency – Inverse Source Frequency (TF–ISF)
● Rationale:
● Importance positively correlates to the number of matching
triples that occur in the data for a URI
● Importance negatively correlates to how often matching
triples occur for all known URIs (i.e., all indexed URIs)
For a selected URI u, a triple pattern tp, and a set of all known
URIs Uknown , let:
tf.isf (u ,tp):=card (u ,tp) ∗ log
( ∣U known∣
{r∈U known ∣ card (r ,tp)>0})

Metric: Join Pattern Cardinality [LT10]
● Rationale: data that matches pairs of (joined) triple patterns
is highly relevant, because it matches a larger
part of the query
● Requirement: these join cardinalities are also pre-computed
and stored in a pre-populated index
For a selected URI u, two triple pattern tpi and tpj , and
query variable v, let:
card(u, tpi , tpj , v) :═ number of solutions produced
by joining tpi and tpj on variable v
using only the data from u

Ladwig and Tran [LT10]
● Multiple scores
● Triple pattern cardinality
● Triple frequency – inverse source frequency (TF–ISF)
● (URI-specific) join pattern cardinality
● Incoming links
● Assumption: pre-populated index that stores triple pattern
cardinalities and join pattern cardinalities for each URI
● Aggregation of the scores to obtain ranks
● For indexed URIs: weighted summation of all scores
● For non-indexed URIs: weighting of (currently known) in-links
● Ranking is refined at run-time

Refinement at Run-Time [LT10]
● During query execution information becomes available
(1) intermediate join results (2) more incoming links
● Use it to adjust scores & ranking (for integrated execution)
● Re-estimate join pattern cardinalities based on samples of
intermediate results (available from hash tables in SHJ)
● Parameters for influencing behavior of ranking process:
● Invalid score threshold: re-rank when the number of URIs
with invalid scores passes this threshold
● Sample size: larger samples give better estimates, but make
the process more costly
● Re-sampling threshold: reuse cached estimates unless the
hash table of join operators grows past this threshold

Outline
➢ Prefetching
➢ Postponing
 Source Ranking
√
√
√

WWW 2013 Tutorial on Linked Data Query Processing [ Introduction ] 30
Tutorial Outline
(1) Introduction
(2) Theoretical Foundations
(3) Source Selection Strategies
(4) Execution Process
(5) Query Planning and Optimization
… Thanks!

These slides have been created by
Olaf Hartig
for the
WWW 2013 tutorial on
Link Data Query Processing
Tutorial Website: http://db.uwaterloo.ca/LDQTut2013/
This work is licensed under a
Creative Commons Attribution-Share Alike 3.0 License
(http://creativecommons.org/licenses/by-sa/3.0/)
(Some of the slides in this slide set have been inspired by
slides from Günter Ladwig [LT10] – Thanks!)

These slides have been created by
Olaf Hartig
for the
WWW 2013 tutorial on
Link Data Query Processing
Tutorial Website: http://db.uwaterloo.ca/LDQTut2013/
This work is licensed under a
Creative Commons Attribution-Share Alike 3.0 License
(http://creativecommons.org/licenses/by-sa/3.0/)
(Slides 24 - 26, 33, and 34 are inspired by slides
from Günter Ladwig [LT10] – Thanks!)

Backup Slides

Metric: Links to Results [LT10]
● Rationale: a URI is more relevant if data from
many relevant URIs mention it
● Links are only discovered at run-time
The “links to results” of a selected URI u is defined by:
where Uprocessed is the set of URIs whose data has already been
processed and links( u1 , u2 ) are the links to URI u1 mentioned
in the data from URI u2.
links(u):={l ∈links(u ,uprocessed )∣u processed ∈U processed }

Metric: Retrieval Cost [LT10]
● Rationale: URIs are more relevant the faster their data can
be retrieved
● Size is available in the pre-populated index
● Bandwidth for any particular host can be approximated
based on past experience or average performance
recorded during the query execution process
The retrieval cost of a selected URI u is defined by:
cost( u) :═ Agg( size(u) , bandwidth(u) )
where size(u) is the of the data from u, and bandwidth(u) is the
bandwidth of the Web server that hosts u.

Tutorial "Linked Data Query Processing" Part 5 "Query Planning and Optimization" (WWW 2013 Ed.)

More Related Content

What's hot (20)

Similar to Tutorial "Linked Data Query Processing" Part 5 "Query Planning and Optimization" (WWW 2013 Ed.) (20)

More from Olaf Hartig (20)

Recently uploaded (20)

Tutorial "Linked Data Query Processing" Part 5 "Query Planning and Optimization" (WWW 2013 Ed.)