This document provides an overview of the Java API of VIATRA Query, describing the features that will help you to integrate it into any Java application. This page is a detailed technical documentation, for basic usage information consult the the getting started tutorial instead.

The most typical way of using the VIATRA Query API is to make use of the generated code that is found in the "src-gen" folder of your VIATRA Query project. This generated code provides easy and type-safe access to most of VIATRA Query’s features from Java code.

VIATRA Query also supports a "dynamic", generic API that allows to make use of patterns without relying on the generated code. The generic API shares functionality with the base classes of the "generated" API, and for most scenarios there is no performance difference between the two. A notable exception for this rule of thumb are check() expressions, where the generated code that is invoked through the generated API will execute the Java code instead of interpreting Xbase.

1. Most important classes and their relationships

For every pattern definition, the VIATRA Query tooling generates a few classes:

QuerySpecification

represents a pattern from a VQL file for the API specification. Not required to use for basic querying, only for generic APIs and fine-tuning settings.

Match

binds the parameters of a query specification to elements from a model; represents a single result of a query or filters the results

Matcher

provides functionality to retrieve the results of the query

MatchProcessor

used to handle query results in a functional style (similar to Java Stream handling).

1.1. Match

A Match object represents a single match of the pattern, i.e. a tuple of objects whose members point to corresponding elements of the instance model (or scalar values) that the pattern is matched against. It is essentially a Data Transfer Object that is used to extract query result information from VIATRA Query, with an SQL analogy you could think of it as one "row" of the result set of a query. The generated fields correspond to the pattern header parameters.

You can also use Match objects to specify fixed input parameters to a query (while other fields can be left unspecified) - analogously to a "prepared" SQL statement that accepts input parameter bindings. In this case, the input Match will act as a filter (mask) and the results of you queries will also be instances of this class (where parameters already have the values given in the input). See below for further details.

The code example below shows the ApplicationInstancesMatch class generated for the applicationInstances pattern (with a single parameter AI). The generated class implements the interface IPatternMatch through the BasePatternMatch internal implementation class.

Immutable matches returned by pattern matchers never include null as a parameter value and can never change after initialized. Mutable matches, e.g. ones created for filtering uses null to represent unset values; but such matches are never returned from the matcher.
public abstract class ApplicationInstancesMatch extends BasePatternMatch {
  /** getters and setters for each parameter */
  public ApplicationInstance getAI();
  public void setAI(final ApplicationInstance pAI);

  public String prettyPrint();
  public int hashCode();
  public boolean equals(final Object obj);

  /** "reflective" calls **/
  public ApplicationInstancesQuerySpecification specification();
  public String patternName();
  public List<String> parameterNames();
  public Object get(final String parameterName);
  public boolean set(final String parameterName, final Object newValue);
  public Object[] toArray();
  public ApplicationInstancesMatch toImmutable();

  /* Mutable and immutable match instantiation */
  public static ApplicationInstancesMatch newEmptyMatch();
  public static ApplicationInstancesMatch newMutableMatch(final ApplicationInstance pAI);
  public static ApplicationInstancesMatch newMatch(final ApplicationInstance pAI);
}

1.2. Matcher

The Matcher is the main entry point of the VIATRA Query API, with pattern-specific query methods. It provides access to the three key features of VIATRA Query:

  • First of all it provides means to initialize a pattern matcher for a given Query engine.

  • After the initialization of the engine, the Matcher provides getter methods to retrieve the contents of the match set. For easy iteration over the match set it provides a convenience method (forEachMatch) as well, as this is the most frequent use case in our observation. Of course it contains other handy features (e.g.: countMatches, hasMatch) to help integration.

  • Finally, it provides means to efficiently track the changes in the match set in an event-driven fashion.

The example generated source code below demonstrates the ApplicationInstancesMatcher class generated for the eClassNames pattern from the running example. The matcher class implements the ViatraQueryMatcher generic interface, and its implementation code extends the BaseGeneratedMatcher internal class, inheriting several useful methods. In the listing below, we show some methods that are not actually part of generated code, but conform to the interface ViatraQueryMatcher and are accessible through inheritance from BaseGeneratedMatcher.

Each function of the pattern matcher API has an overridden version that accepts a (partial) match as input parameters. These input matches may be both mutable and immutable, and can contain null values. However, VIATRA matchers always return immutable matches without null values. This means, there is no need to check for null when processing matches.
public class EClassNamesMatcher implements ViatraQueryMatcher<EClassNamesMatch> {
  /** factory method **/
  public static ApplicationInstancesMatcher on(final ViatraQueryEngine engine);

  /** access to match set **/
  public Collection<ApplicationInstancesMatch> getAllMatches(); // inherited
  public Collection<ApplicationInstancesMatch> getAllMatches(final ApplicationInstance pAI);
  public Stream<ApplicationInstancesMatch> streamAllMatches(); // inherited
  public Stream<ApplicationInstancesMatch> streamAllMatches(final ApplicationInstance pAI);
  public Optional<ApplicationInstancesMatch> getOneArbitraryMatch(); // inherited
  public Optional<ApplicationInstancesMatch> getOneArbitraryMatch(final ApplicationInstance pAI);
  public boolean hasMatch(); // inherited
  public boolean hasMatch(final ApplicationInstance pAI);
  public int countMatches(); // inherited
  public int countMatches(final ApplicationInstance pAI);

  /** Retrieve the set of values that occur in matches.**/
  public Set<ApplicationInstance> getAllValuesOfAI() {}

  /** iterate over matches using a lambda **/
  public void forEachMatch(Consumer<? super EClassNamesMatch> processor); // inherited
  public void forEachMatch(final ApplicationInstance pAI, final Consumer<? super ApplicationInstancesMatch> processor);
  public void forOneArbitraryMatch(Consumer<? super EClassNamesMatch> processor); // inherited
  public boolean forOneArbitraryMatch(final ApplicationInstance pAI, final Consumer<? super ApplicationInstancesMatch> processor) {}

  /** Returns a new (partial) Match object for the matcher.
   *  This can be used e.g. to call the matcher with a partial match. **/
  public ApplicationInstancesMatch newMatch(final ApplicationInstance pAI);

  /** Access query specification */
  public static IQuerySpecification<ApplicationInstancesMatcher> querySpecification();
}

Starting with VIATRA 2.0 the matcher API also returns stream of matches. These streams can be used for processing the streams functionally, greatly extending the similar capabilities provided by the forEachMatch calls available since earlier version. Furthermore, relying on these streams might provide better performance: (1) the use of these streams does not necessitate the copying of the match set, and (2) the pattern matcher is allowed to postpone the match set calculation until the next match is necessary.

The new Stream-based APIs cannot handle if the underlying model is changed during match processing. If a snapshot of the match set is required, either rely on the similar getAllMatches() call or collect the results of the stream in end-user code.

1.3. Query Specification

A pattern-specific specification that can instantiate a Matcher class in a type-safe way. You can get an instance of it via the Matcher class’s specification() method. The recommended way to instantiate a Matcher is with an ViatraQueryEngine. In both cases if the pattern is already registered (with the same root in the case of the Notifier method) then only a lightweight reference is created which points to the existing engine.

The code sample extends the BaseGeneratedQuerySpecification class.

/**
 * A pattern-specific query specification that can instantiate EClassNamesMatcher in a type-safe way.
 */
public final class ApplicationInstancesQuerySpecification extends BaseGeneratedEMFQuerySpecification<ApplicationInstancesMatcher> {

  /** Singleton instance access */
  public static ApplicationInstancesQuerySpecification instance();

  /** Instantiate matches and matchers */
  public ApplicationInstancesMatcher instantiate();
  public ApplicationInstancesMatch newEmptyMatch();
  public ApplicationInstancesMatch newMatch(final Object... parameters);

}

2. Lifecycle management

In VIATRA Query, all pattern matching (query evaluation) is carried out in ViatraQueryEngine instances that are accessed through the user-friendly generated classes of the public API. The ViatraQueryEngine associated to your patterns can be accessed and managed through the EngineManager singleton class, to track and manipulate their lifecycles.

A ViatraQueryEngine is instantiated with a Scope implementation that describes the model the query should work with. By default, in case of EMF it is recommended to initialize an EMFScope instance with the ResourceSet containing the EMF model. For more details about scopes see Query Scopes.

By default, for each scope a single, managed ViatraQueryEngine is created, which is shared by all objects that access VIATRA Query’s features through the generated API. The ViatraQueryEngine is attached to the scope and it is retained on the heap as long as the model itself is there. It will listen on update notifications stemming from the given model in order to maintain live results. If you release all references to the model (e.g. unload the resource), the ViatraQueryEngine can also be garbage collected (as long as there are no other inbound references on it).

In all, for most (basic) scenarios, the following workflow should be followed:

  • initialize/load the model

  • initialize your ViatraQueryEngine instance

  • initialize pattern matchers, or groups of pattern matchers and use them

  • if you release the model and your ViatraQueryEngine instance, all resources will be freed by the garbage collector.

For advanced scenarios (if you wish to manage lifecycles at a more finegrained level), you have the option of creating unmanaged ViatraQueryEngines and dispose of them independently of your instance model. For most use-cases though, we recommend the use of managed engines, this is the default and optimized behavior, as these engines can share common indices and caches to save memory and CPU time. The EngineManager ensures that there will be no duplicated engine for the same model root (Notifier) object. Creating an unmanaged engine will give you certain additional benefits, however additional considerations should be applied.

If you want to remove the matchers from the engine you can call the wipe() method on it. It discards any pattern matcher caches and forgets the known patterns. The base index built directly on the underlying EMF model, however, is kept in memory to allow reuse when new pattern matchers are built. If you don’t want to use it anymore call the dispose() instead, to completely disconnect and dismantle the engine.

Never call wipe or dispose on any engine that were not explicitly created by you; any created matcher over the engine becomes unusable.

3. Typical programming patterns

In the followings, we provide short source code samples (with some explanations) that cover the most important use-cases supported by the VIATRA Query API.

3.1. Loading an instance model and executing a query

// get all matches of the pattern
// initialization
// phase 1: (managed) ViatraQueryEngine
ViatraQueryEngine engine = ViatraQueryEngine.on(new EMFScope(resource /* or resourceSet */));
// phase 2: the matcher itself
EObjectMatcher matcher = EObjectMatcher.on(engine);
// get all matches of the pattern
Collection<EObjectMatch> matches = matcher.getAllMatches();
// process matches, produce some output
StringBuilder results = new StringBuilder();
prettyPrintMatches(results, matches);

3.2. Using the MatchProcessor

With the MatchProcessor you can iterate over the matches of a pattern quite easily:

matcher2.forEachMatch(new EClassNamesProcessor() {
 @Override
 public void process(EClass c, String n) {
  results.append("\tEClass: " + c.toString() + "\n");
 }
});

3.3. Matching with partially bound input parameters

An important aspect of VIATRA Query queries is that they are bidirectional in the sense that they accept input bindings, to filter/project the result set with a given input constraint. The following example illustrates the usage of the match processor with an input binding that restricts the result set to the cases where the second parameter (the name of the EClass) takes the value "A":

matcher2.forEachMatch( matcher2.newMatch(null, "A") , new EClassNamesProcessor() {
 @Override
 public void process(EClass c, String n) {
  results.append("\tEClass with name A: " + c.toString() + "\n");
 }
});

// alternatively:
matcher2.forEachMatch(null, "A" , new EClassNamesProcessor() {
 @Override
 public void process(EClass c, String n) {
  results.append("\tEClass with name A: " + c.toString() + "\n");
 }
});

The input bindings may be used for all match result set methods.

Additionally, the getAllValuesOf…​ methods allow you to perform projections of the result set to one of the parameters:

// projections
for (EClass ec: matcher2.getAllValuesOfc(matcher2.newMatch(null,"A")))
{
 results.append("\tEClass with name A: " + ec.toString() + "\n");
}

3.4. Initialization of pattern groups

Using pattern groups is important for performance. By default, VIATRA Query performs a traversal of the instance model when a matcher is accessed through the ViatraQueryEngine for the first time. If you wish to use several pattern matchers, it is a good idea to make use of the generated pattern group class and prepare the ViatraQueryEngine to perform a combined traversal (with minimal additional overhead) so that any additional Matcher initializations avoid re-traversals.

// phase 1: (managed) ViatraQueryEngine
ViatraQueryEngine engine = ViatraQueryEngine.on(new EMFScope(resource));
// phase 2: the group of pattern matchers
HeadlessQueries patternGroup = HeadlessQueries.instance();
patternGroup.prepare(engine);
// from here on everything is the same
EObjectMatcher matcher = EObjectMatcher.on(engine);
// get all matches of the pattern
Collection<EObjectMatch> matches = matcher.getAllMatches();
prettyPrintMatches(results, matches);
// ... //
// matching with partially bound input parameters
// because EClassNamesMatcher is included in the patterngroup, *no new traversal* will be done here
EClassNamesMatcher matcher2 = EClassNamesMatcher.on(engine);

4. Parsing Patterns

VIATRA provides an API to parse patterns from a text and creates query specifications from them that can be used similar to generated query specifications. This is based on the generic pattern matcher API that differs from the generated one in two key aspects:

  • it can be used to apply queries and use other VIATRA Query features without generating code and loading the resulting bundles into the running configuration. In other words, you just need to supply the EMF-based in-memory representation (an instance of the Pattern class)

  • the generic API is not "type safe" in the sense that the Java types of your pattern variables is not known and needs to be handled dynamically (typically by type casting).

To use this API, the code from the org.eclipse.viatra.query.patternlanguage.emf plug-in has to be added to the classpath (in standalone applications, rely on the Maven dependency org.eclipse.viatra:viatra-query-language). This will add further transitive dependencies, most notable on Xtext and Google Guice to your application.
Using the Pattern Parser API
final StringBuilder results = new StringBuilder();
Resource resource = loadModel(modelURI);

// Initializing Xtext-based resource parser (once per Java application)
new EMFPatternLanguageStandaloneSetup().createInjectorAndDoEMFRegistration();

// Parse pattern definition
PatternParsingResults parseResults = PatternParserBuilder.instance()
        .parse("import \"http://org.eclipse.viatra/model/cps\" \n"
                + "\n"
                + "pattern hostIpAddress(host: HostInstance, ip : java String) {\n"
                + "    HostInstance.nodeIp(host,ip);\n"
                + "}");
ViatraQueryEngine engine = ViatraQueryEngine.on(new EMFScope(resource));

parseResults.getQuerySpecification("hostIpAddress").ifPresent(specification -> {
    ViatraQueryMatcher<?> matcher = engine.getMatcher(specification);
    prettyPrintMatches(results, matcher.getAllMatches());
});

return results.toString();
In VIATRA 2.1 the pattern parser API was updated to support more advanced cases, like updating previously loaded patterns (see below for details). In previous versions, the pattern parser could be initialized by calling PatternParser.parser().parse(…​), but this call is deprecated in VIATRA version 2.1.

The pattern parser can be initialized in two modes: a basic mode (initialized with calling either the parse or build methods of the PatternParserBuilder class does not supports updating query definitions after being loaded; while in advanced mode (initialized with the buildAdvanced method of PatterParserBuilder) previously loaded specifications can be updated, and the resulting query specifications are updated (and revalidated) as necessary. Given the more complex infrastructure required for advanced mode, it is recommended to rely on the basic mode unless reparsing pattern is truly necessary, such as when integrating VIATRA in an environment where the user may specify custom queries with complex dependencies between them.

5. VIATRA Query Base

VIATRA Query provides a light-weight indexer library called Base that aims to provide several useful (some would even argue critical) features for querying EMF models:

  • inverse navigation along EReferences

  • finding and incrementally tracking all model elements by attribute value/type (i.e. inverse navigation along EAttributes)

  • incrementally computing transitive reachability along given reference types (i.e. transitive closure of an EMF model)

  • getting and tracking all the (direct) instances of a given EClass

The point of VIATRA Query Base is to provide all of these in an incremental way, which means that once the query evaluator is attached to an EMF model, as long as it stays attached, the query results can be retrieved instantly (as the query result cache is automatically updated). VIATRA Query Base is a lightweight, small Java library that can be integrated easily to any EMF-based tool as it can be used in a stand-alone way, without the rest of VIATRA Query.

We are aware that some of the functionality can be found in some Ecore utility classes (for example ECrossReferenceAdapter). These standard implementations are non-incremental, and are thus do not scale well in scenarios where high query evaluation performance is necessary (such as e.g. on-the-fly well-formedness validation or live view maintenance). VIATRA Query Base has an additional important feature that is not present elsewhere: it contains very efficient implementations of transitive closure that can be used e.g. to maintain reachability regions incrementally, in very large EMF instance models.

5.1. Extracting reachability paths from transitive closure

Beyond the support for querying reachability information between nodes in the model, the TransitiveClosureHelper class also provides the functionality to retrieve paths between pairs of nodes. The getPathFinder method returns an IGraphPathFinder object, which exposes the following operations:

Deque<V> getPath(V sourceNode, V targetNode)

Returns an arbitrary path from the source node to the target node (if such exists).

Iterable<Deque<V>> getShortestPaths(V sourceNode, V targetNode)

Returns the collection of shortest paths from the source node to the target node (if such exists).

Iterable<Deque<V>> getAllPaths(V sourceNode, V targetNode)

Returns the collection of paths from the source node to the target node (if such exists).

Iterable<Deque<V>> getAllPathsToTargets(V sourceNode, Set<V> targetNodes)

Returns the collection of paths from the source node to any of the target nodes (if such exists).

Internally these operations use a depth-first-search traversal and rely on the information which is incrementally maintained by the transitive closure component.

6. Query Scopes

VIATRA Query uses the concept of Scopes to define the entire model to search for results. For queries over EMF models, the EMFScope class defines such scopes. When initializing a ViatraQueryEngine, it is required to specify this scope by creating a new instance of EMFScope.

This instance might be created from one or more Notifier instances (ResourceSet: includes all model elements stored in the ResourceSet; Resource: includes all elements inside the corresponding Resource; EObject: includes all elements in the containment subtree of the object itself).

In most cases, it is recommended to include the entire ResourceSet as the query scope; however, if required, it is possible to

6.1. Using Filtered Input Models During Pattern Matching

In several cases it is beneficial to not include all Resources from a ResourceSet during pattern matching, but consider more than one. Such cases might include Xtext/Xbase languages or JaMoPP-based instances that include resources representing the classes of the Java library.

In case of EMF models, the EMFScope instance may also set some base index options to filter out containment subtrees from being indexed both by the Base Indexer and the Rete networks, by providing a filter implementation to the VIATRA Query Engine. These options include the IBaseIndexResourceFilter and IBaseIndexObjectFilter instances that can be used to filter out entire resources or containment subtrees, respectively.

Sample usage (by filtering out Java classes referred by JaMoPP):

ResourceSet resourceSet = ...; //Use a Resource Set as the root of the engine
BaseIndexOptions options = new BaseIndexOptions().withResourceFilterConfiguration(new IBaseIndexResourceFilter() {

  @Override
  public boolean isResourceFiltered(Resource resource) {
    // PathMap URI scheme is used to refer to JDK classes
    return "pathmap".equals(resource.getURI().scheme());
  }
});
//Initializing scope with custom options
EMFScope scope = new EMFScope(resourceSet, options);
ViatraQueryEngine engine = ViatraQueryEngine.on(scope);
there are some issues to be considered while using this API:
  • If a Resource or containment subtree is filtered out, it is filtered out entirely. It is not possible to re-add some lower-level contents.

  • In case of the query scope is set to a subset of the entire model (e.g only one EMF resource within the resource set), model elements within the scope of the engine may have references pointing to elements outside the scope; these are called dangling edges. Previous versions of VIATRA made the assumption that the model is self-contained and free of dangling edges; the behavior of the query engine was ''unspecified'' (potentially incorrect match sets) if the model did not have this property. In VIATRA 1.6, this behavior was cleaned up by adding a new indexer mode that drops this assumption, and (with a minor cost to performance) always checks both ends of all indexed edges to be in-scope. For backward compatibility, the old behavior is used by default, but you can manually change this using the corresponding base index option as below. For new code we suggest to use this option to drop the dangling-free assumption, as it provides more consistent and intuitive results in a lot of cases; in a future VIATRA release this will be the new default.

BaseIndexOptions options = new BaseIndexOptions().withDanglingFreeAssumption(false);
ResourceSet rSet = new ResourceSetImpl();
EMFScope scope = new EMFScope(rSet, options);
ViatraQueryEngine engine = ViatraQueryEngine.on(scope);

Since version 0.9, there is a possibility to refer to alternative search engines in addition to Rete-based incremental engines; version 1.0 includes a local search based search algorithm usable with the VIATRA Query matcher API.

Since version 1.4, the Local Search engine is considered stable, and users are encuraged to use it in applications where incrementality is not crucial. The Local Search engine reuses the same matcher API used in VIATRA Query.

  • When is local search the most beneficial?

    • A single, batch evaluation of models

    • Memory limit is severe and the Rete network does not fit into the memory

    • When all calls have one or more parameters bound, resulting in simple traversal

  • Harder cases

    • Repeated model execution

    • Query evaluation requires expensive model traversal (think about iterating over all instances in a model)

7.1. Using Local Search

The most important steps to perform:

  • Add a dependency to the optional plug-in org.eclipse.viatra.query.runtime.localsearch

  • Explicitly ask for a local search-based matcher when initializing the matcher instance:

IQuerySpecification<?> specification = ...;
QueryEvaluationHint hint = LocalSearchHints.getDefault().build();
AdvancedViatraQueryEngine.from(queryEngine).getMatcher(specification, hint);
  • Or alternatively, set the local search as default for a query engine:

// Access the default local search hint
QueryEvaluationHint localSearchHint = LocalSearchHints.getDefault().build();

// Build an engine options with the local search hint
ViatraQueryEngineOptions options = ViatraQueryEngineOptions.
		defineOptions().
		withDefaultHint(localSearchHint).
                withDefaultBackend(localSearchHint.getQueryBackendFactory()). // this line is needed in 1.4 due to bug 507777
		build();

//Access the query engine
ViatraQueryEngine queryEngine = ViatraQueryEngine.on(scope, options);
  • After initialization, the existing pattern matcher API constructs can be used over the local search engine.

It is also possible to declare specific patterns to be executed by Local Search in the VQL file using the search, although this setting may be overridden by the hints given at the matcher creation.

search pattern minCPUs(n : java Integer) {
	n == min find cpus(_hi1, #_);
}

Parameterization of the planner algorithm is possible via the hint mechanism. Currently (version 1.7) the following hints are available by using the LocalSearchHints builder class:

Use Base

allow/disallow the usage of the index at runtime. Its value may be true or false. The default value is true.

Row Count

An internal parameter, bigger values often mean longer plan generation times, and potentially search plans with lower cost. Its value may be a positive int, the default value is 4.

Cost Function

The cost function to be used by the planner. Must implement org.eclipse.viatra.query.runtime.localsearch.planner.cost.ICostFunction

Flatten call predicate

The predicate to control which pattern composition calls shall be flattened before planning. By deafult all called patterns are flattened.

For example, to disable the use of base index:

IQuerySpecification<?> specification = ...;
QueryEvaluationHint hint = LocalSearchHints.getDefault().setUseBase(false).build();
AdvancedViatraQueryEngine.from(queryEngine).getMatcher(specification, hint);

7.3. Cost function

The default cost function estimates operation costs based on the statistical structure of the model, which is obtained using the base index. This is true even if USE_BASE_INDEX is set to false, in which case a plan is created which does not rely on the base index at execution time. Since 1.4.0 the base index is capable of providing only statistical information with much less overhead compared to instance indexing. To avoid using base index even in the planning phase, the cost function can be replaced to another implementation. For this purpose, two alternative implementations are provided:

  • VariableBindingBasedCostFunction estimates the operation costs using the number of variables it binds. This cost function usually results in lower performance executions.

  • The abstract class StatisticsBasedConstraintCostFunction can be used to provide model statistics from different sources, e.g. a previously populated map:

final Map<IInputKey, Long> statistics = ..
QueryEvaluationHint hint = LocalSearchHints.getDefault().setCostFunction(new StatisticsBasedConstraintCostFunction(){
  public long countTuples(IConstraintEvaluationContext input, IInputKey supplierKey){
    return statistics.get(supplierKey);
  }
}).build();

The latter is advised to be used if the model is expected to be changed after the planning phase to ensure that the planing is based on a realistic model statistics which resembles the actual structure which the pattern is executed on.

We plan on providing a simpler way of setting up model statistics in later versions; this kind of setup might be changed.

7.4. Known limitations

  • A local search matcher cannot provide change notifications on pattern matches. If asked, an UnsupportedOperationException is thrown.

  • As of version 1.4, it is not possible to combine different pattern matching algorithms for the evaluation of a single pattern. Either the entire search must use Rete or Local search based algorithms.

  • The Local Search engine currently is not able to execute recursive queries. See http://bugs.eclipse.org/458278 for more details.

8. Providing Query Evaluation Hints

It is possible to pass extra information to the runtime of VIATRA Query using evaluation hints, such as information about the structure of the model or requirements for the evaluation. In version 1.4, the handling of such hints were greatly enhanced, allowing the following ways to pass hints:

  1. The Query engine might be initialized with default hints using the static method AdvancedQueryEngine#createUnmanagedEngine(QueryScope, ViatraQueryEngineOptions). The hints provided inside the query engine options are the default hints used by all matchers, but can be overridden using the following options.

  2. A pattern definition can be extended with hints, e.g. for backend selection in the pattern language. Such hints will be generated into the generated query specification code.

  3. When accessing a new pattern matcher through the Query Engine, further override hints might be presented using AdvancedQueryEngine#getMatcher(IQuerySpecification, QueryEvaluationHint). Such hints override both the engine default and the pattern default hints.

In version 1.4 the hints are mostly used to fine tune the local search based pattern matcher, but their usage is gradually being extended. See classes ReteHintOptions and LocalSearchHints for hint options provided by the query backends. As of 2.0, the delete and rederive (DRED) hint option is available on the UI as well.

9. Query specification registry

The query specification registry, available since ''VIATRA 1.3'' is used to manage query specifications provided by multiple connectors which can dynamically add and remove specifications. Users can read the contents of the registry through views that are also dynamically updated when the registry is changed by the connectors.

9.1. Basic usage

The most common usage of the registry will be to get a registered query specification based on its fully qualified name. You can access the registry through a singleton instance:

IQuerySpecificationRegistry registry = org.eclipse.viatra.query.runtime.registry.QuerySpecificationRegistry.getInstance();
IQuerySpecification<?> specification = registry.getDefaultView().getEntry("my.registered.query.fqn").get();

The default view lets you access the contents of the registry, the entry returned is a provider for the query specification that returns it when requested through the get() method.

9.2. Advanced usage

9.2.1. Views

To get an always up to date view of the registry, you can either: * request a default view that will contain on specification marked to be included in this view (e.g. queries registered through the queryspecification extension point) * create a new view that may use either a filter or a factory for defining which specifications should be included in the view

IQuerySpecificationRegistry registry = QuerySpecificationRegistry.getInstance();
// access default view
IDefaultRegistryView defaultView = registry.getDefaultView();

// create new view
IRegistryView simpleView = registry.createView();

// create filtered view
IRegistryView filteredView = registry.createView(new IRegistryViewFilter() {
  @Override
  public boolean isEntryRelevant(IQuerySpecificationRegistryEntry entry) {
    // return true to include in view
  }
});

// create specific view instance
boolean allowDuplicateFQNs = false;
IRegistryView ownView = registry.createView(new IRegistryViewFactory() {
  return new AbstractRegistryView(registry, allowDuplicateFQNs) {
    @Override
    protected boolean isEntryRelevant(IQuerySpecificationRegistryEntry entry) {
      // return true to include in view
    }
  }
);

Once you have a view instance, you can access the contents of the registry by requesting the entries from the view or adding a listener that will be notified when the view changes.

Default views add a few additional utilities that are made possible by also restricting what is included in them. Default views will only contain entries that are marked explicitly to be included and will not allow different specifications with the same fully qualified name. In return, you can request a single entry by its FQN (since at most one can exist) and also request a query group that contains all entries.

9.2.2. Listening to view changes

The contents of the registry may change after a view is created. When you access the view to get its entries, it will always return the current state of the registry. If you want to get notified when the contents of your view change, you can add a listener to the view:

IQuerySpecificationRegistry registry = QuerySpecificationRegistry.getInstance();
IRegistryView myView = registry.createView();
IQuerySpecificationRegistryChangeListener listener = new IQuerySpecificationRegistryChangeListener() {
  @Override
  public void entryAdded(IQuerySpecificationRegistryEntry entry) {
    // process addition
  }

  @Override
  public void entryRemoved(IQuerySpecificationRegistryEntry entry) {
    // process removal
  }
});
myView.addViewListener(listener);

// when you don't need to get notifications any more
myView.removeViewListener(listener);

Important note: your code has to keep a reference to your view otherwise it will be garbage collected. The registry uses weak references to created views in order to free users from having to manually dispose views.

9.2.3. Adding specifications to the registry

The registry is supplied with specifications through sources. You can add your own source connector as a source and dynamically add and remove your own specifications.

IQuerySpecificationRegistry registry = QuerySpecificationRegistry.getInstance();
// initialize your connector
IRegistrySourceConnector connector;

// add connector
boolean sourceAdded = registry.addSource(connector);

// [...]

// remove your source when needed
boolean sourceRemoved = registry.removeSource(connector);

We already have some connector implementations for the most common use cases. For example, you can create a connector with a simple add and remove method for query specifications:

IRegistrySourceConnector connector = new SpecificationMapSourceConnector("my.source.identifier", true /* include these in default view; fqn clashes are errors */);

IQuerySpecification<?> specification = /* available from somewhere */

IQuerySpecificationProvider provider = new SingletonQuerySpecificationProvider(specification);

// add specification to source
connector.addQuerySpecificationProvider(provider);

// remove specification by FQN
connector.removeQuerySpecificationProvider(specification.getFullyQualifiedName());
The default view assumes all queries loaded there have a single qualified name. If this cannot ensured, the source should not be added to the default views and specific views are to be created accordingly.

10. Performance tuning and special engine modes

10.1. Query groups and coalescing model traversals

If you initialize a new query that requires the indexing of some EMF types for which the current engine instance has not yet built an index, then the base index of the VIATRA engine will traverse the entire scope to build the index. It can make a great difference if such expensive re-traversals are avoided, and the engine traverses the model only once to build indexes for all queries.

The easiest wax to do this would be to use <code>IQueryGroup.prepare(engine)</code> for a group of queries. Such a group is generated for every query file, and any other custom group can be manually assembled with <code>GenericQueryGroup</code>.

IQueryGroup queries = ...
ViatraQueryEngine engine = ...
queries.prepare(engine);

For advanced use cases, it is possible to directly control indexing traversals in an arbitrary code block, such that any index constructions are coalesced into a single traversal:

ViatraQueryEngine engine = ...
engine.getBaseIndex().coalesceTraversals(new Callable<Void>() {
    @Override
    public Void call() throws Exception {
        // obtain matchers, etc.
        return null;
    }
});

10.2. Delaying query result updates

As of version 1.6, the advanced query API now includes a feature that lets users temporarily "turn off" query result maintenance in the incremental query backend. During such a code block, only the base model indexer is updated, query results remain stale until the end of the block. The advantage is that it is possible to save significant execution time when changing the model in a way that partially undoes itself, e.g. a large part of the model is removed and then re-added.

AdvancedViatraQueryEngine engine = ...
engine.delayUpdatePropagation(new Callable<Void>() {
    @Override
    public Void call() throws Exception {
        // perform extensive changes in model that largely cancel each other out
        return null;
    }
});

10.3. Run-once Query Engine

This page describes how VIATRA Query can be used to carry out one-time query evaluation which is useful in the following cases:

  • You want less (steady-state) memory consumption instead of incremental evaluation.

  • You have derived features that are not well-behaving, more precisely do not always send correct notifications when its value changes, but you want to include them in queries.

  • You like the query language of VIATRA Query, but you don’t need incremental evaluation and the batch performance is better than the sum of model modification overheads between query usages.

These scenarios are now supported by a "run-once" query engine that will perform the evaluation on a given query and return the match set then dispose of the Rete network and base index to free up memory.

The local search engine provided by VIATRA should perform better for these cases and it is recommended to use that instead. This functionality predates local search support and is kept for backward compatibility.

10.3.1. Example

The most up-to-date sample source code to this page is found in Git here: http://git.eclipse.org/c/viatra/org.eclipse.viatra.git/tree/examples/minilibrary Most notably,

10.3.2. Usage

10.3.2.1. Run-once then dispose

The API of the run-once query engine is very simple, just instantiate the engine with the constructor using the proper scope (EObject, Resource or ResourceSet) and call the getAllMatches with a query specfication:

RunOnceQueryEngine engine = new RunOnceQueryEngine(notifier);
// using generated query specification
Collection<SumOfPagesInLibraryMatch> allMatches = engine.getAllMatches(SumOfPagesInLibraryMatcher.querySpecification());
// if you only have Pattern object
IQuerySpecification<ViatraQueryMatcher<IPatternMatch>> specification = (IQuerySpecification<ViatraQueryMatcher<IPatternMatch>>) QuerySpecificationRegistry.getOrCreateQuerySpecification(BooksWithMultipleAuthorsMatcher.querySpecification().getPattern());
Collection<IPatternMatch> matches = engine.getAllMatches(specification);

Note that each invocation of getAllMatches will traverse the model completely, index the classes, features and data types that are required for the query, collect the match set than dispose the indexes.

10.3.3. Automatic re-sampling

In many cases, the derived features are only a small part of the queries and it would be better to keep the indices once they are built. However, in this case, we need a way to update the values of all derived features that are indexed.

The run-once query engine supports automatic re-sampling by listening to model modifications and updating values before returning match results.The following example shows how you can enable this mode:

RunOnceQueryEngine engine = new RunOnceQueryEngine(notifier);
engine.setAutomaticResampling(true); // enable re-sampling mode
Collection<SumOfPagesInLibraryMatch> allMatches = engine.getAllMatches(SumOfPagesInLibraryMatcher.querySpecification());
// some model modification
// only re-sampling of derived features, not complete traversal
allMatches = engine.getAllMatches(SumOfPagesInLibraryMatcher.querySpecification());

If you no longer need automatic re-sampling, you can turn it off. In this case the engine that was kept incrementally updated is removed from memory.

engine.setAutomaticResampling(false); // disable re-sampling mode, indices removed

Finally, if the value of derived features change without any model modifications (not recommended), you can tell the engine to run the re-sampling next time:

engine.resampleOnNextCall();

11. Logging in VIATRA Query

VIATRA Query logs error messages and some trace information using log4j. If you need to debug your application and would like to see these messages, you can set the log level in different hierarchy levels. Since we use standard log4j, you can configure logging both with configuration files or through API calls.

  • All loggers are children of a top-level default logger, that can be accessed from ViatraQueryLoggingUtil.getDefaultLogger(), just call setLevel(Level.DEBUG) on the returned logger to see all messages (of course you can use other levels as well).

  • Each engine has it’s own logger that is shared with the Base Index and the matchers as well. If you want to see all messages related to all engines, call ViatraQueryLoggingUtil.getLogger(ViatraQueryEngine.class) and set the level.

  • Some other classes also use their own loggers and the same approach is used, they get the loggers based on their class, so retrieving that logger and setting the level will work as well.

11.1. Configuration problems

log4j uses a properties file as a configuration for its root logger. However, since this configuration is usually supplied by developers of applications, we do not package it in VIATRA Query. This means you may encounter the following on your console if no configuration was supplied:

 log4j:WARN No appenders could be found for logger (org.eclipse.viatra.query.runtime.util.ViatraQueryLoggingUtil).
 log4j:WARN Please initialize the log4j system properly.

There are several cases where this can occur:

  • You have Xtext SDK installed, which has a plugin fragment called org.eclipse.xtext.logging that supplies a log4j configuration. Make sure that the fragment is selected in your Runtime Configuration.

  • You are using the tooling of VIATRA Query without the Xtext SDK, you will see the above warning, but since the patternlanguage.emf plugins also inject appenders to the loggers of VIATRA Query, log messages will be correctly displayed.

  • You are using only the runtime part of VIATRA Query that has no Xtext dependency. You have to provide your own properties file (standalone execution) or fragment (OSGi execution), see http://www.eclipsezone.com/eclipse/forums/t99588.html

  • Alternatively, if you just want to make sure that log messages appear in the console no matter what other configuration happens, you can call ViatraQueryLoggingUtil.setupConsoleAppenderForDefaultLogger() which will do exactly what its name says. Since appenders and log levels are separate, you will still have to set the log level on the loggers you want to see messages from.

  • If you wish to completely turn the logger of, call ViatraQueryLoggingUtil.getDefaultLogger().setLevel(Level.OFF);.