TDB Optimizer

Query execution in TDB involves both static and dynamic optimizations. Static optimizations are transformations of the SPARQL algebra performed before query execution begins; dynamic optimizations involve deciding the best execution approach during the execution phase and can take into account the actual data so far retrieved.

The optimizer has a number of strategies: a statistics based strategy, a fixed strategy and a strategy of no reordering.

For the preferred statistics strategy, the TDB optimizer uses information captured in a per-database statistics file. The file takes the form of a number of rules for approximate matching counts for triple patterns. The statistic file can be automatically generated. The user can add and modify rules to tune the database based on higher level knowledge, such as inverse function properties.

Contents

Quickstart

This section provides a practical how-to.

  1. Load data.
  2. Generate the statistics file. Run tdbstats.
  3. Place the file generated in the database directory with the name stats.opt.

Running tdbstats

Usage:

 tdbstats --loc=DIR|--desc=assemblerFile [--graph=URI]

Choosing the optimizer strategy

TDB chooses the basic graph pattern optimizer by the presence of a file in the database directory.

Optimizer control files

File name Effect
none.opt No reordering - execute triple patterns in the order in the query
fixed.opt Use a built-in reordering based on the number of variables in a triple pattern.
stats.opt The contents of this file are the weighing rules (see below).

The contents of the files none.opt and fixed.opt are not read and don't matter. They can be zero-length files.

If more then one file is found, the choice is made: stats.opt over fixed.opt over none.opt.

The "no reorder" strategy can be useful in investigating the effects. Filter placement still takes place.

Filter placement

One of the key optimization is of filtered basic graph patterns. This optimization decides the best order of triple patterns in a basic graph pattern and also the best point at which to apply the filters within the triple patterns.

Any filter expression of a basic graph pattern is placed immediately after all it's variables will be bound. Conjunctions at the top level in filter expressions are broken into their constituent pieces and placed separately.

Investigating what is going on

TDB can optionally log query execution details. This is controlled by two setting: the logging level and a context setting. Having two setting means it is possible to log some queries and not others.

The logger used is called com.hp.hpl.jena.arq.exec. Message are sent at level "info". So for log4j, the following can be set in the log4j.properties file:

# Execution logging
log4j.logger.com.hp.hpl.jena.arq.info=WARN
log4j.logger.com.hp.hpl.jena.arq.exec=WARN

In versions of TDB before 0.8.7, this is:

log4j.logger.com.hp.hpl.jena.tdb.exec=INFO

The context setting is for key (Java constant) ARQ.symLogExec. To set globally:

TDB.getContext().set(ARQ.symLogExec,true) ;

and it may also be set on an individual query execution using it's local context.

 try(QueryExecutiuon qExec = QueryExecutionFactory.create(...)) {
    qExec.getContext().set(ARQ.symLogExec,true) ;
    ResultSet rs = qExec.execSelect() ;
 }

Use TDB.symLogExec for versions prior to 0.8.8.

On the command line:

 tdbquery --set tdb:logExec=true --file queryfile

TDB version 0.8.3 provides more fine-grained logging controls. Instead of "true", which sets all levels, the following can be used:

Explanation Levels

Level Effect
INFO Log each query
FINE Log each query and it's algebra form after optimization
ALL Log query, algebra and every database access (can be expensive)
NONE No information logged

These can be specified as string, to the command line tools, or using the constants in Explain.InfoLevel.

 qExec.getContext().set(ARQ.symLogExec,Explain.InfoLevel.FINE) ;

Statistics Rule File

The syntax is SSE, a simple format that uses Turtle-syntax for RDF terms, keywords for other terms (for example, the stats marks a statistics data structure), and forms a tree data structure.

The structure of a statistics file takes the form:

(prefix ...
  (stats
    (meta ...)
    rule
    rule
   ))

that is, a meta block and a number of pattern rules.

A simple example:

(prefix ((: <http://example/))
  (stats
    (meta
      (timestamp "2008-10-23T10:35:19.122+01:00"^^<http://www.w3.org/2001/XMLSchema#dateTime>)
      (run@ "2008/10/23 10:35:19")
      (count 11))
    (:p 7)
    (<http://example/q> 7)
  ))

This example statistics file contains some metadata about statistics (time and date the file was generated, size of graph), the frequence count for two predicates http://example/p (written using a prefixed name) and http://example/q (written in full).

The numbers are the estimated counts. They do not have to be exact - they guide the optimizer in choosing one execution plan over another. They do not have to exactly up-to-date providing the relative counts are representative of the data.

Statistics Rule Language

A rule is made up of a triple pattern and a count estimation for the approximate number of matches that the pattern will yeild. This does have to be exact, only an indication.

In addition, the optimizer considers which variables will be bound to RDF terms by the time a triplepatetrn is reached in the execution plan being considered. For example, in the basic graph pattern:

{ ?x  :identifier  1234 .
  ?x  :name        ?name .
}

then ?x will be bound in pattern ?x :name ?name to an RDF term if executed after the pattern ?x :identifier 1234.

A rule is of the form:

( (subj pred obj) count)

where subj, pred, obj are either RDF terms or one of the tokens in the following table:

Statistic rule tokens

Token | Description TERM | Matches any RDF term (URI, Literal, Blank node) VAR | Matches a named variable (e.g. ?x) URI | Matches a URI LITERAL | Matches an RDF literal BNODE | Matches an RDF blank node (in the data) ANY |Matches anything - a term or variable

From the example above, (VAR :identifier TERM) will match ?x :identifier 1234.

(TERM :name VAR) will match ?x :name ?name when in a potential plan where the :identifier triple pattern is first because ?x will be a bound term at that point but not if this triple pattern is considered first.

When searching for a weighting of a triple pattern, the first rule to match is taken.

The rule which says an RDF graph is a set of triples:

((TERM TERM TERM) 1)

is always implicitly present.

BNODE does not match a blank node in the query (which is a variable and matches VAR) but in the data, if it is known that slot of a triple pattern is a blank node.

Abbreviated Rule Form

While a complete rule is of the form:

( (subj pred obj) count)

there is an abbreviated form:

(prediate count)

The abbreviated form is equivalent to writing:

((TERM predicate ANY) X)
((ANY predicate TERM) Y)
((ANY predicate ANY) count)

where for small graphs (less that 100 triples) X=2, Y=4 but Y=40 if the predicate is rdf:type and 2, 10, 1000 for large graphs. Use of "VAR rdf:type Class" can be a quite unselective triple pattern and so there is a preference to move it later in the order of execution to allow more selective patterns reduce the set of possibilities first. The astute reader may notice that ontological information may render it unnecessary (the domain or range of another property implies the class of some resource). TDB does not currently perform this optimization.

These number are merely convenient guesses and the application can use the full rules language for detailed control of pattern weightings.

Defaults

A rule of the form:

(other number)

is used when no matches from other rules (abbreviated or full) when matching a triple pattern that has a URI in the predicate position. If a rule of this form is absent, the default is to place the triple pattern after all known triple patterns; this is the same as specifying -1 as the number. To declare that the rules are complete and no other predicates occur in the data, set this to 0 (zero) because the triple pattern can not match the data (the predicate does not occur).

Generating a statistics file

The command line tdbstats will scan the data and produce a rules file based on the frequency of properties. The output should first go to a temporary file, then that file moved into the database location.

Practical tip: Don't feed the output of this command directly to location/stats.opt because when the command starts it will find an empty statistics file at that location.

Generating statistics for Union Graphs

By default tdbstats only processes the default graph of a dataset. However in some circumstances it is desirable to have the statistics generated over Named Graphs in the dataset.

The tdb:unionDefaultGraph option will cause TDB to synthesize a default graph for SPARQL queries, from the union of all Named Graphs in the dataset.

Ideally the statistics file should be generated against this union graph. This can be achieved using the --graph option as follows:

 tdbstats --graph urn:x-arq:UnionGraph --loc /path/to/indexes

The graph parameter uses a built-in TDB special graph name

Writing Rules

Rule for an inverse functional property:

((VAR :ifp TERM) 1 )

and even if a property is only approximately identifying for resources (e.g. date of birth in a small dataset of people), it is useful to indicate this. Because the counts needed are only approximations so the optimizer can choose one order over another, and does not need to predicate exact counts, rules that are usually right but may be slightly wrong are still useful overall.

Rules involving rdf:type can be useful where they indicate whether a partciualr class is common or not. In some datasets

((VAR rdf:type class) ...)

may help little because a property whose domain is that class, or a subclass, may be more elective. SO a rule like:

((VAR :property VAR) ...)

is more selective.

In other datasets, there may be many classes, each with a small number of instances, in which case

((VAR rdf:type class) ...)

is a useful selective rule.