Skip to content

Traversal Optimization

jbmusso edited this page Jul 11, 2016 · 14 revisions

Attention: this Wiki hosts an outdated version of the TinkerPop framework and Gremlin language documentation.

Please visit the Apache TinkerPop website and latest documentation.

In Gremlin, the same path expression can be written in numerous ways. Gremlin is an imperative language in that the developer explicitly instructs Gremlin which path to take through the graph (as defined by an abstract path description). However, Gremlin doesn’t always do as he is told. There are various optimizations that Gremlin will automatically take advantage of. This section describes such optimizations.

Query Optimization

In Blueprints there are two types of graph indices that are respected: graph and vertex indices. A graph index is global to the entire graph whereby vertices/edges can be efficiently looked-up by their key/value properties. A vertex index is local to a vertex whereby incident edges can be efficiently looked-up by their direction, label, and key/value properties.

Graph Query Optimization

Blueprints provides a method called Graph.query(). This method returns a GraphQuery object that can be configured to filter the vertices and edges of a graph according to their properties. While the implementation depends on the backend graph engine, typically global indicies are utilized (i.e. log(n)-structures, where n is the number of vertices/edges in the graph). In Gremlin, whenever an expression starts with g.V (or g.E) the next series of statements provides predicates to the GraphQuery.

gremlin> g.V.has('name','marko').interval('age',20,40).toString()
==>[GremlinStartPipe, GraphQueryPipe(has,interval,vertex), IdentityPipe, IdentityPipe]

Vertex Query Optimization

Blueprints provides a method called Vertex.query(). This method returns a VertexQuery object that can be configured to filter the edges (and thus vertices) associated with that Vertex according to edge labels, edge properties, edge property intervals, etc. While the implementation depends on the backend graph engine, typically vertex-centric indicies are utilized (i.e. log(m)-structures, where m is the number of incident edges to the vertex). Whenever a pattern such as outE....inV is seen by Gremlin, Gremlin will automatically compile that statement into a VertexQuery object.

gremlin> g.v(1).outE('knows','created').interval('weight',0.5,1.0).has('date',2012)
==>[GremlinStartPipe, VertexQueryPipe(out,[knows, created],has,interval,edge), IdentityPipe, IdentityPipe, InVertexPipe, PropertyPipe(name)]

Also, a range filter on a vertex-to-vertex traversal will compile to a VertexQueryPipe.

gremlin> g.v(1).out('knows')[0..10].toString()
==>[GremlinStartPipe, VertexQueryPipe(out,[knows],range:[0,10],vertex), IdentityPipe]

Note that Gremlin ensures that the pipeline length stays the same by introducing IdentityPipe steps. This ensures that numbered step constructs (e.g. loop(2), back(1)) remain faithful. Though it is always a good idea to toString() an expression when using numbered steps to ensure proper indexing.

If the query optimization is not desired, there is the method GremlinPipeline.optimize(boolean). This method makes it possible to turn off query optimizations.

gremlin> g.v(1).optimize(false).outE('knows','created').interval('weight',0.5,1.0).has('date',2012)
==>[GremlinStartPipe, OutEdgesPipe(knows,created), IntervalFilterPipe, PropertyFilterPipe(date,EQUAL,2012), InVertexPipe, PropertyPipe(name)]

Automatic Path Enabling

Pipes natively supports the recording of the history of a particular traversal. By default, Gremlin does not assume that path information will be required of the traversal unless a path-specific step is called — e.g. path, simplePath. If the path information is required internal to a closure, Gremlin doesn’t know that as it can not interpret what is in a closure. As such, be sure to GremlinPipeline.enablePath() if path information will be required by the expression.

gremlin> g.v(1).as('x').out.loop('x'){it.loops < 3}{it.path.contains(g.v(4))}
Cannot invoke method contains() on null object
Display stack trace? [yN]
gremlin> g.v(1).as('x').out.loop('x'){it.loops < 3}{it.path.contains(g.v(4))}.enablePath()
You can’t perform that action at this time.