Gelly is a Java Graph API for Flink. It contains a set of methods and utilities which aim to simplify the development of graph analysis applications in Flink. In Gelly, graphs can be transformed and modified using high-level functions similar to the ones provided by the batch processing API. Gelly provides methods to create, transform and modify graphs, as well as a library of graph algorithms.
Gelly is currently part of the staging Maven project. All relevant classes are located in the org.apache.flink.graph package.
Add the following dependency to your pom.xml to use Gelly.
<dependency> <groupId>org.apache.flink</groupId> <artifactId>flink-gelly</artifactId> <version>{{site.version}}</version> </dependency>
Note that Gelly is currently not part of the binary distribution. See linking with it for cluster execution here.
The remaining sections provide a description of available methods and present several examples of how to use Gelly and how to mix it with the Flink Java API. After reading this guide, you might also want to check the {% gh_link /flink-staging/flink-gelly/src/main/java/org/apache/flink/graph/example/ “Gelly examples” %}.
In Gelly, a Graph is represented by a DataSet of vertices and a DataSet of edges.
The Graph nodes are represented by the Vertex type. A Vertex is defined by a unique ID and a value. Vertex IDs should implement the Comparable interface. Vertices without value can be represented by setting the value type to NullValue.
{% highlight java %} // create a new vertex with a Long ID and a String value Vertex<Long, String> v = new Vertex<Long, String>(1L, “foo”);
// create a new vertex with a Long ID and no value Vertex<Long, NullValue> v = new Vertex<Long, NullValue>(1L, NullValue.getInstance()); {% endhighlight %}
The graph edges are represented by the Edge type. An Edge is defined by a source ID (the ID of the source Vertex), a target ID (the ID of the target Vertex) and an optional value. The source and target IDs should be of the same type as the Vertex IDs. Edges with no value have a NullValue value type.
{% highlight java %} Edge<Long, Double> e = new Edge<Long, Double>(1L, 2L, 0.5);
// reverse the source and target of this edge Edge<Long, Double> reversed = e.reverse();
Double weight = e.getValue(); // weight = 0.5 {% endhighlight %}
You can create a Graph in the following ways:
DataSet of edges and an optional DataSet of vertices:{% highlight java %} ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
DataSet<Vertex<String, Long>> vertices = ...
DataSet<Edge<String, Double>> edges = ...
Graph<String, Long, Double> graph = Graph.fromDataSet(vertices, edges, env); {% endhighlight %}
DataSet of Tuple3 and an optional DataSet of Tuple2. In this case, Gelly will convert each Tuple3 to an Edge, where the first field will be the source ID, the second field will be the target ID and the third field will be the edge value. Equivalently, each Tuple2 will be converted to a Vertex, where the first field will be the vertex ID and the second field will be the vertex value:{% highlight java %} ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
DataSet<Tuple2<String, Long>> vertexTuples = env.readCsvFile(“path/to/vertex/input”);
DataSet<Tuple3<String, String, Double>> edgeTuples = env.readCsvFile(“path/to/edge/input”);
Graph<String, Long, Double> graph = Graph.fromTupleDataSet(vertexTuples, edgeTuples, env); {% endhighlight %}
Collection of edges and an optional Collection of vertices:{% highlight java %} ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
List<Vertex<Long, Long>> vertexList = new ArrayList...
List<Edge<Long, String>> edgeList = new ArrayList...
Graph<Long, Long, String> graph = Graph.fromCollection(vertexList, edgeList, env); {% endhighlight %}
If no vertex input is provided during Graph creation, Gelly will automatically produce the Vertex DataSet from the edge input. In this case, the created vertices will have no values. Alternatively, you can provide a MapFunction as an argument to the creation method, in order to initialize the Vertex values:
{% highlight java %} ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
// initialize the vertex value to be equal to the vertex ID Graph<Long, Long, String> graph = Graph.fromCollection(edges, new MapFunction<Long, Long>() { public Long map(Long value) { return value; } }, env); {% endhighlight %}
Gelly includes the following methods for retrieving various Graph properties and metrics:
{% highlight java %} // get the Vertex DataSet DataSet<Vertex<K, VV>> getVertices()
// get the Edge DataSet DataSet<Edge<K, EV>> getEdges()
// get the IDs of the vertices as a DataSet DataSet getVertexIds()
// get the source-target pairs of the edge IDs as a DataSet DataSet<Tuple2<K, K>> getEdgeIds()
// get a DataSet of <vertex ID, in-degree> pairs for all vertices DataSet<Tuple2<K, Long>> inDegrees()
// get a DataSet of <vertex ID, out-degree> pairs for all vertices DataSet<Tuple2<K, Long>> outDegrees()
// get a DataSet of <vertex ID, degree> pairs for all vertices, where degree is the sum of in- and out- degrees DataSet<Tuple2<K, Long>> getDegrees()
// get the number of vertices long numberOfVertices()
// get the number of edges long numberOfEdges()
// get a DataSet of Triplets<srcVertex, trgVertex, edge> DataSet<Triplet<K, VV, EV>> getTriplets()
{% endhighlight %}
mapVertices and mapEdges return a new Graph, where the IDs of the vertices (or edges) remain unchanged, while the values are transformed according to the provided user-defined map function. The map functions also allow changing the type of the vertex or edge values.{% highlight java %} ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment(); Graph<Long, Long, Long> graph = Graph.fromDataSet(vertices, edges, env);
// increment each vertex value by one Graph<Long, Long, Long> updatedGraph = graph.mapVertices( new MapFunction<Vertex<Long, Long>, Long>() { public Long map(Vertex<Long, Long> value) { return value.getValue() + 1; } }); {% endhighlight %}
Graph. filterOnEdges will create a sub-graph of the original graph, keeping only the edges that satisfy the provided predicate. Note that the vertex dataset will not be modified. Respectively, filterOnVertices applies a filter on the vertices of the graph. Edges whose source and/or target do not satisfy the vertex predicate are removed from the resulting edge dataset. The subgraph method can be used to apply a filter function to the vertices and the edges at the same time.{% highlight java %} Graph<Long, Long, Long> graph = ...
graph.subgraph( new FilterFunction<Vertex<Long, Long>>() { public boolean filter(Vertex<Long, Long> vertex) { // keep only vertices with positive values return (vertex.getValue() > 0); } }, new FilterFunction<Edge<Long, Long>>() { public boolean filter(Edge<Long, Long> edge) { // keep only edges with negative values return (edge.getValue() < 0); } }) {% endhighlight %}
joinWithVertices joins the vertices with a Tuple2 input data set. The join is performed using the vertex ID and the first field of the Tuple2 input as the join keys. The method returns a new Graph where the vertex values have been updated according to a provided user-defined map function. Similarly, an input dataset can be joined with the edges, using one of three methods. joinWithEdges expects an input DataSet of Tuple3 and joins on the composite key of both source and target vertex IDs. joinWithEdgesOnSource expects a DataSet of Tuple2 and joins on the source key of the edges and the first attribute of the input dataset and joinWithEdgesOnTarget expects a DataSet of Tuple2 and joins on the target key of the edges and the first attribute of the input dataset. All three methods apply a map function on the edge and the input data set values. Note that if the input dataset contains a key multiple times, all Gelly join methods will only consider the first value encountered.{% highlight java %} Graph<Long, Double, Double> network = ...
DataSet<Tuple2<Long, Long>> vertexOutDegrees = network.outDegrees();
// assign the transition probabilities as the edge weights Graph<Long, Double, Double> networkWithWeights = network.joinWithEdgesOnSource(vertexOutDegrees, new MapFunction<Tuple2<Double, Long>, Double>() { public Double map(Tuple2<Double, Long> value) { return value.f0 / value.f1; } }); {% endhighlight %}
Reverse: the reverse() method returns a new Graph where the direction of all edges has been reversed.
Undirected: In Gelly, a Graph is always directed. Undirected graphs can be represented by adding all opposite-direction edges to a graph. For this purpose, Gelly provides the getUndirected() method.
Union: Gelly's union() method performs a union operation on the vertex and edge sets of the specified graph and current graph. Duplicate vertices are removed from the resulting Graph, while if duplicate edges exists, these will be maintained.
difference() method performs a difference on the vertex and edge sets of the current graph and specified graph.Gelly includes the following methods for adding and removing vertices and edges from an input Graph:
{% highlight java %} // adds a Vertex to the Graph. If the Vertex already exists, it will not be added again. Graph<K, VV, EV> addVertex(final Vertex<K, VV> vertex)
// adds a list of vertices to the Graph. If the vertices already exist in the graph, they will not be added once more. Graph<K, VV, EV> addVertices(List<Vertex<K, VV>> verticesToAdd)
// adds an Edge to the Graph. If the source and target vertices do not exist in the graph, they will also be added. Graph<K, VV, EV> addEdge(Vertex<K, VV> source, Vertex<K, VV> target, EV edgeValue)
// adds a list of edges to the Graph. When adding an edge for a non-existing set of vertices, the edge is considered invalid and ignored. Graph<K, VV, EV> addEdges(List<Edge<K, EV>> newEdges)
// removes the given Vertex and its edges from the Graph. Graph<K, VV, EV> removeVertex(Vertex<K, VV> vertex)
// removes the given list of vertices and their edges from the Graph Graph<K, VV, EV> removeVertices(List<Vertex<K, VV>> verticesToBeRemoved)
// removes all edges that match the given Edge from the Graph. Graph<K, VV, EV> removeEdge(Edge<K, EV> edge)
// removes all edges that match the edges in the given list Graph<K, VV, EV> removeEdges(List<Edge<K, EV>> edgesToBeRemoved)
{% endhighlight %}
Neighborhood methods allow vertices to perform an aggregation on their first-hop neighborhood. reduceOnEdges() can be used to compute an aggregation on the values of the neighboring edges of a vertex and reduceOnNeighbors() can be used to compute an aggregation on the values of the neighboring vertices. These methods assume associative and commutative aggregations and exploit combiners internally, significantly improving performance. The neighborhood scope is defined by the EdgeDirection parameter, which takes the values IN, OUT or ALL. IN will gather all in-coming edges (neighbors) of a vertex, OUT will gather all out-going edges (neighbors), while ALL will gather all edges (neighbors).
For example, assume that you want to select the minimum weight of all out-edges for each vertex in the following graph:
The following code will collect the out-edges for each vertex and apply the SelectMinWeight() user-defined function on each of the resulting neighborhoods:
{% highlight java %} Graph<Long, Long, Double> graph = ...
DataSet<Tuple2<Long, Double>> minWeights = graph.reduceOnEdges(new SelectMinWeight(), EdgeDirection.OUT);
// user-defined function to select the minimum weight static final class SelectMinWeight implements ReduceEdgesFunction {
@Override
public Double reduceEdges(Double firstEdgeValue, Double secondEdgeValue) {
return Math.min(firstEdgeValue, secondEdgeValue);
}
} {% endhighlight %}
Similarly, assume that you would like to compute the sum of the values of all in-coming neighbors, for every vertex. The following code will collect the in-coming neighbors for each vertex and apply the SumValues() user-defined function on each neighborhood:
{% highlight java %} Graph<Long, Long, Double> graph = ...
DataSet<Tuple2<Long, Long>> verticesWithSum = graph.reduceOnNeighbors(new SumValues(), EdgeDirection.IN);
// user-defined function to sum the neighbor values static final class SumValues implements ReduceNeighborsFunction {
@Override
public Long reduceNeighbors(Long firstNeighbor, Long secondNeighbor) {
return firstNeighbor + secondNeighbor;
}
} {% endhighlight %}
When the aggregation function is not associative and commutative or when it is desirable to return more than one values per vertex, one can use the more general groupReduceOnEdges() and groupReduceOnNeighbors() methods. These methods return zero, one or more values per vertex and provide access to the whole neighborhood.
For example, the following code will output all the vertex pairs which are connected with an edge having a weight of 0.5 or more:
{% highlight java %} Graph<Long, Long, Double> graph = ...
DataSet<Tuple2<Vertex<Long, Long>, Vertex<Long, Long>>> vertexPairs = graph.groupReduceOnNeighbors(new SelectLargeWeightNeighbors(), EdgeDirection.OUT);
// user-defined function to select the neighbors which have edges with weight > 0.5 static final class SelectLargeWeightNeighbors implements NeighborsFunctionWithVertexValue<Long, Long, Double, Tuple2<Vertex<Long, Long>, Vertex<Long, Long>>> {
@Override
public void iterateNeighbors(Vertex<Long, Long> vertex,
Iterable<Tuple2<Edge<Long, Double>, Vertex<Long, Long>>> neighbors,
Collector<Tuple2<Vertex<Long, Long>, Vertex<Long, Long>>> out) {
for (Tuple2<Edge<Long, Double>, Vertex<Long, Long>> neighbor : neighbors) {
if (neighbor.f0.f2 > 0.5) {
out.collect(new Tuple2<Vertex<Long, Long>, Vertex<Long, Long>>(vertex, neighbor.f1));
}
}
}
} {% endhighlight %}
When the aggregation computation does not require access to the vertex value (for which the aggregation is performed), it is advised to use the more efficient EdgesFunction and NeighborsFunction for the user-defined functions. When access to the vertex value is required, one should use EdgesFunctionWithVertexValue and NeighborsFunctionWithVertexValue instead.
Gelly exploits Flink's efficient iteration operators to support large-scale iterative graph processing. Currently, we provide implementations of the popular vertex-centric iterative model and a variation of Gather-Sum-Apply. In the following sections, we describe these models and show how you can use them in Gelly.
The vertex-centric model, also known as “think like a vertex” model, expresses computation from the perspective of a vertex in the graph. The computation proceeds in synchronized iteration steps, called supersteps. In each superstep, a vertex produces messages for other vertices and updates its value based on the messages it receives. To use vertex-centric iterations in Gelly, the user only needs to define how a vertex behaves in each superstep:
Gelly wraps Flink‘s Spargel API to provide methods for vertex-centric iterations. The user only needs to implement two functions, corresponding to the phases above: a VertexUpdateFunction, which defines how a vertex will update its value based on the received messages and a MessagingFunction, which allows a vertex to send out messages for the next superstep. These functions and the maximum number of iterations to run are given as parameters to Gelly’s runVertexCentricIteration. This method will execute the vertex-centric iteration on the input Graph and return a new Graph, with updated vertex values.
A vertex-centric iteration can be extended with information such as the total number of vertices, the in degree and out degree. Additionally, the neighborhood type (in/out/all) over which to run the vertex-centric iteration can be specified. By default, the updates from the in-neighbors are used to modify the current vertex's state and messages are sent to out-neighbors.
Let us consider computing Single-Source-Shortest-Paths with vertex-centric iterations on the following graph and let vertex 1 be the source. In each superstep, each vertex sends a candidate distance message to all its neighbors. The message value is the sum of the current value of the vertex and the edge weight connecting this vertex with its neighbor. Upon receiving candidate distance messages, each vertex calculates the minimum distance and, if a shorter path has been discovered, it updates its value. If a vertex does not change its value during a superstep, then it does not produce messages for its neighbors for the next superstep. The algorithm converges when there are no value updates.
{% highlight java %} // read the input graph Graph<Long, Double, Double> graph = ...
// define the maximum number of iterations int maxIterations = 10;
// Execute the vertex-centric iteration Graph<Long, Double, Double> result = graph.runVertexCentricIteration( new VertexDistanceUpdater(), new MinDistanceMessenger(), maxIterations);
// Extract the vertices as the result DataSet<Vertex<Long, Double>> singleSourceShortestPaths = result.getVertices();
// - - - UDFs - - - //
// messaging public static final class MinDistanceMessenger extends MessagingFunction<K, Double, Double, Double> {
public void sendMessages(Vertex<K, Double> vertex) {
for (Edge<K, Double> edge : getEdges()) {
sendMessageTo(edge.getTarget(), vertex.getValue() + edge.getValue());
}
}
}
// vertex update public static final class VertexDistanceUpdater extends VertexUpdateFunction<K, Double, Double> {
public void updateVertex(Vertex<K, Double> vertex, MessageIterator<Double> inMessages) {
Double minDistance = Double.MAX_VALUE;
for (double msg : inMessages) {
if (msg < minDistance) {
minDistance = msg;
}
}
if (vertex.getValue() > minDistance) {
setNewVertexValue(minDistance);
}
}
}
{% endhighlight %}
A vertex-centric iteration can be configured using a VertexCentricConfiguration object. Currently, the following parameters can be specified:
Name: The name for the vertex-centric iteration. The name is displayed in logs and messages and can be specified using the setName() method.
Parallelism: The parallelism for the iteration. It can be set using the setParallelism() method.
Solution set in unmanaged memory: Defines whether the solution set is kept in managed memory (Flink's internal way of keeping objects in serialized form) or as a simple object map. By default, the solution set runs in managed memory. This property can be set using the setSolutionSetUnmanagedMemory() method.
Aggregators: Iteration aggregators can be registered using the registerAggregator() method. An iteration aggregator combines all aggregates globally once per superstep and makes them available in the next superstep. Registered aggregators can be accessed inside the user-defined VertexUpdateFunction and MessagingFunction.
Broadcast Variables: DataSets can be added as Broadcast Variables to the VertexUpdateFunction and MessagingFunction, using the addBroadcastSetForUpdateFunction() and addBroadcastSetForMessagingFunction() methods, respectively.
Number of Vertices: Accessing the total number of vertices within the iteration. This property can be set using the setOptNumVertices() method. The number of vertices can then be accessed in the vertex update function and in the messaging function using the getNumberOfVertices() method. If the option is not set in the configuration, this method will return -1.
Degrees: Accessing the in/out degree for a vertex within an iteration. This property can be set using the setOptDegrees() method. The in/out degrees can then be accessed in the vertex update function and in the messaging function, per vertex using the getInDegree() and getOutDegree() methods. If the degrees option is not set in the configuration, these methods will return -1.
Messaging Direction: By default, a vertex sends messages to its out-neighbors and updates its value based on messages received from its in-neighbors. This configuration option allows users to change the messaging direction to either EdgeDirection.IN, EdgeDirection.OUT, EdgeDirection.ALL. The messaging direction also dictates the update direction which would be EdgeDirection.OUT, EdgeDirection.IN and EdgeDirection.ALL, respectively. This property can be set using the setDirection() method.
{% highlight java %}
Graph<Long, Double, Double> graph = ...
// configure the iteration VertexCentricConfiguration parameters = new VertexCentricConfiguration();
// set the iteration name parameters.setName(“Gelly Iteration”);
// set the parallelism parameters.setParallelism(16);
// register an aggregator parameters.registerAggregator(“sumAggregator”, new LongSumAggregator());
// run the vertex-centric iteration, also passing the configuration parameters Graph<Long, Double, Double> result = graph.runVertexCentricIteration( new VertexUpdater(), new Messenger(), maxIterations, parameters);
// user-defined functions public static final class VertexUpdater extends VertexUpdateFunction {
LongSumAggregator aggregator = new LongSumAggregator();
public void preSuperstep() {
// retrieve the Aggregator
aggregator = getIterationAggregator("sumAggregator");
}
public void updateVertex(Long vertexKey, Long vertexValue, MessageIterator inMessages) {
//do some computation
Long partialValue = ...
// aggregate the partial value
aggregator.aggregate(partialValue);
// update the vertex value
setNewVertexValue(...);
}
}
public static final class Messenger extends MessagingFunction {...}
{% endhighlight %}
The following example illustrates the usage of the degree as well as the number of vertices options.
{% highlight java %}
Graph<Long, Double, Double> graph = ...
// configure the iteration VertexCentricConfiguration parameters = new VertexCentricConfiguration();
// set the number of vertices option to true parameters.setOptNumVertices(true);
// set the degree option to true parameters.setOptDegrees(true);
// run the vertex-centric iteration, also passing the configuration parameters Graph<Long, Double, Double> result = graph.runVertexCentricIteration( new VertexUpdater(), new Messenger(), maxIterations, parameters);
// user-defined functions public static final class VertexUpdater { ... // get the number of vertices long numVertices = getNumberOfVertices(); ... }
public static final class Messenger { ... // retrieve the vertex out-degree outDegree = getOutDegree(); ... }
{% endhighlight %}
The following example illustrates the usage of the edge direction option. Vertices update their values to contain a list of all their in-neighbors.
{% highlight java %}
Graph<Long, HashSet, Double> graph = ...
// configure the iteration VertexCentricConfiguration parameters = new VertexCentricConfiguration();
// set the messaging direction parameters.setDirection(EdgeDirection.IN);
// run the vertex-centric iteration, also passing the configuration parameters DataSet<Vertex<Long, HashSet>> result = graph.runVertexCentricIteration( new VertexUpdater(), new Messenger(), maxIterations, parameters) .getVertices();
// user-defined functions public static final class VertexUpdater { @Override public void updateVertex(Vertex<Long, HashSet> vertex, MessageIterator messages) throws Exception { vertex.getValue().clear();
for(long msg : messages) {
vertex.getValue().add(msg);
}
setNewVertexValue(vertex.getValue());
}
}
public static final class Messenger { @Override public void sendMessages(Vertex<Long, HashSet> vertex) throws Exception { for (Edge<Long, Long> edge : getEdges()) { sendMessageTo(edge.getSource(), vertex.getId()); } } }
{% endhighlight %}
Like in the vertex-centric model, Gather-Sum-Apply also proceeds in synchronized iterative steps, called supersteps. Each superstep consists of the following three phases:
Let us consider computing Single-Source-Shortest-Paths with GSA on the following graph and let vertex 1 be the source. During the Gather phase, we calculate the new candidate distances, by adding each vertex value with the edge weight. In Sum, the candidate distances are grouped by vertex ID and the minimum distance is chosen. In Apply, the newly calculated distance is compared to the current vertex value and the minimum of the two is assigned as the new value of the vertex.
Notice that, if a vertex does not change its value during a superstep, it will not calculate candidate distance during the next superstep. The algorithm converges when no vertex changes value. The resulting graph after the algorithm converges is shown below.
To implement this example in Gelly GSA, the user only needs to call the runGatherSumApplyIteration method on the input graph and provide the GatherFunction, SumFunction and ApplyFunction UDFs. Iteration synchronization, grouping, value updates and convergence are handled by the system:
{% highlight java %} // read the input graph Graph<Long, Double, Double> graph = ...
// define the maximum number of iterations int maxIterations = 10;
// Execute the GSA iteration Graph<Long, Double, Double> result = graph.runGatherSumApplyIteration( new CalculateDistances(), new ChooseMinDistance(), new UpdateDistance(), maxIterations);
// Extract the vertices as the result DataSet<Vertex<Long, Double>> singleSourceShortestPaths = result.getVertices();
// - - - UDFs - - - //
// Gather private static final class CalculateDistances extends GatherFunction<Double, Double, Double> {
public Double gather(Neighbor<Double, Double> neighbor) {
return neighbor.getNeighborValue() + neighbor.getEdgeValue();
}
}
// Sum private static final class ChooseMinDistance extends SumFunction<Double, Double, Double> {
public Double sum(Double newValue, Double currentValue) {
return Math.min(newValue, currentValue);
}
}
// Apply private static final class UpdateDistance extends ApplyFunction<Long, Double, Double> {
public void apply(Double newDistance, Double oldDistance) {
if (newDistance < oldDistance) {
setResult(newDistance);
}
}
}
{% endhighlight %}
Note that gather takes a Neighbor type as an argument. This is a convenience type which simply wraps a vertex with its neighboring edge.
For more examples of how to implement algorithms with the Gather-Sum-Apply model, check the {% gh_link /flink-staging/flink-gelly/src/main/java/org/apache/flink/graph/example/GSAPageRank.java “GSAPageRank” %} and {% gh_link /flink-staging/flink-gelly/src/main/java/org/apache/flink/graph/example/GSAConnectedComponents.java “GSAConnectedComponents” %} examples of Gelly.
A GSA iteration can be configured using a GSAConfiguration object. Currently, the following parameters can be specified:
Name: The name for the GSA iteration. The name is displayed in logs and messages and can be specified using the setName() method.
Parallelism: The parallelism for the iteration. It can be set using the setParallelism() method.
Solution set in unmanaged memory: Defines whether the solution set is kept in managed memory (Flink's internal way of keeping objects in serialized form) or as a simple object map. By default, the solution set runs in managed memory. This property can be set using the setSolutionSetUnmanagedMemory() method.
Aggregators: Iteration aggregators can be registered using the registerAggregator() method. An iteration aggregator combines all aggregates globally once per superstep and makes them available in the next superstep. Registered aggregators can be accessed inside the user-defined GatherFunction, SumFunction and ApplyFunction.
Broadcast Variables: DataSets can be added as Broadcast Variables to the GatherFunction, SumFunction and ApplyFunction, using the methods addBroadcastSetForGatherFunction(), addBroadcastSetForSumFunction() and addBroadcastSetForApplyFunction methods, respectively.
Number of Vertices: Accessing the total number of vertices within the iteration. This property can be set using the setOptNumVertices() method. The number of vertices can then be accessed in the gather, sum and/or apply functions by using the getNumberOfVertices() method. If the option is not set in the configuration, this method will return -1.
Neighbor Direction: By default values are gathered from the out neighbors of the Vertex. This can be modified using the setDirection() method.
The following example illustrates the usage of the number of vertices option.
{% highlight java %}
Graph<Long, Double, Double> graph = ...
// configure the iteration GSAConfiguration parameters = new GSAConfiguration();
// set the number of vertices option to true parameters.setOptNumVertices(true);
// run the gather-sum-apply iteration, also passing the configuration parameters Graph<Long, Long, Long> result = graph.runGatherSumApplyIteration( new Gather(), new Sum(), new Apply(), maxIterations, parameters);
// user-defined functions public static final class Gather { ... // get the number of vertices long numVertices = getNumberOfVertices(); ... }
public static final class Sum { ... // get the number of vertices long numVertices = getNumberOfVertices(); ... }
public static final class Apply { ... // get the number of vertices long numVertices = getNumberOfVertices(); ... }
{% endhighlight %}
The following example illustrates the usage of the edge direction option. {% highlight java %}
Graph<Long, HashSet, Double> graph = ...
// configure the iteration GSAConfiguration parameters = new GSAConfiguration();
// set the messaging direction parameters.setDirection(EdgeDirection.IN);
// run the gather-sum-apply iteration, also passing the configuration parameters DataSet<Vertex<Long, HashSet>> result = graph.runGatherSumApplyIteration( new Gather(), new Sum(), new Apply(), maxIterations, parameters) .getVertices(); {% endhighlight %} Back to top
As seen in the examples above, Gather-Sum-Apply iterations are quite similar to vertex-centric iterations. In fact, any algorithm which can be expressed as a GSA iteration can also be written in the vertex-centric model. The messaging phase of the vertex-centric model is equivalent to the Gather and Sum steps of GSA: Gather can be seen as the phase where the messages are produced and Sum as the phase where they are routed to the target vertex. Similarly, the value update phase corresponds to the Apply step.
The main difference between the two implementations is that the Gather phase of GSA parallelizes the computation over the edges, while the messaging phase distributes the computation over the vertices. Using the SSSP examples above, we see that in the first superstep of the vertex-centric case, vertices 1, 2 and 3 produce messages in parallel. Vertex 1 produces 3 messages, while vertices 2 and 3 produce one message each. In the GSA case on the other hand, the computation is parallelized over the edges: the three candidate distance values of vertex 1 are produced in parallel. Thus, if the Gather step contains “heavy” computation, it might be a better idea to use GSA and spread out the computation, instead of burdening a single vertex. Another case when parallelizing over the edges might prove to be more efficient is when the input graph is skewed (some vertices have a lot more neighbors than others).
Another difference between the two implementations is that the vertex-centric implementation uses a coGroup operator internally, while GSA uses a reduce. Therefore, if the function that combines neighbor values (messages) requires the whole group of values for the computation, vertex-centric should be used. If the update function is associative and commutative, then the GSA's reducer is expected to give a more efficient implementation, as it can make use of a combiner.
Another thing to note is that GSA works strictly on neighborhoods, while in the vertex-centric model, a vertex can send a message to any vertex, given that it knows its vertex ID, regardless of whether it is a neighbor. Finally, in Gelly's vertex-centric implementation, one can choose the messaging direction, i.e. the direction in which updates propagate. GSA does not support this yet, so each vertex will be updated based on the values of its in-neighbors only.
Gelly provides a simple utility for performing validation checks on input graphs. Depending on the application context, a graph may or may not be valid according to certain criteria. For example, a user might need to validate whether their graph contains duplicate edges or whether its structure is bipartite. In order to validate a graph, one can define a custom GraphValidator and implement its validate() method. InvalidVertexIdsValidator is Gelly's pre-defined validator. It checks that the edge set contains valid vertex IDs, i.e. that all edge IDs also exist in the vertex IDs set.
{% highlight java %} ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
// create a list of vertices with IDs = {1, 2, 3, 4, 5} List<Vertex<Long, Long>> vertices = ...
// create a list of edges with IDs = {(1, 2) (1, 3), (2, 4), (5, 6)} List<Edge<Long, Long>> edges = ...
Graph<Long, Long, Long> graph = Graph.fromCollection(vertices, edges, env);
// will return false: 6 is an invalid ID graph.validate(new InvalidVertexIdsValidator<Long, Long, Long>());
{% endhighlight %}
Gelly has a growing collection of graph algorithms for easily analyzing large-scale Graphs. So far, the following library methods are implemented:
Gelly's library methods can be used by simply calling the run() method on the input graph:
{% highlight java %} ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
Graph<Long, Long, NullValue> graph = ...
// run Label Propagation for 30 iterations to detect communities on the input graph DataSet<Vertex<Long, Long>> verticesWithCommunity = graph.run( new LabelPropagation(30)).getVertices();
// print the result verticesWithCommunity.print();
env.execute(); {% endhighlight %}
Gelly provides the old Spargel API functionality through its vertex-centric iteration methods. Applications can be easily migrated from one API to the other, using the following general guidelines:
Vertex and Edge Types: In Spargel, vertices and edges are defined using tuples (Tuple2 for vertices, Tuple2 for edges without values, Tuple3 for edges with values). Gelly has a more intuitive graph representation by introducing the Vertex and Edge types.
Methods for Plain Edges and for Valued Edges: In Spargel, there are separate methods for edges with values and edges without values when running the vertex centric iteration (i.e. withValuedEdges(), withPlainEdges()). In Gelly, this distinction is no longer needede because an edge with no value will simply have a NullValue type.
OutgoingEdge: Spargel's OutgoingEdge is replaced by Edge in Gelly.
Running a Vertex Centric Iteration: In Spargel, an iteration is run by calling the runOperation() method on a VertexCentricIteration. The edge type (plain or valued) dictates the method to be called. The arguments are: a data set of edges, the vertex update function, the messaging function and the maximum number of iterations. The result is a DataSet<Tuple2<vertexId, vertexValue>> representing the updated vertices. In Gelly, an iteration is run by calling runVertexCentricIteration() on a Graph. The parameters given to this method are the vertex update function, the messaging function and the maximum number of iterations. The result is a new Graph with updated vertex values.
Configuring a Vertex Centric Iteration: In Spargel, an iteration is configured by directly setting a set of parameters on the VertexCentricIteration instance (e.g. iteration.setName("Spargel Iteration")). In Gelly, a vertex-centric iteration is configured using the IterationConfiguration object (e.g. iterationConfiguration.setName("Gelly Iteration”)). An instance of this object is then passed as a final parameter to the runVertexCentricIteration() method.
Record API: Spargel's Record API was completely removed from Gelly.
In the following section, we present a step-by-step example for porting the Connected Components algorithm from Spargel to Gelly.
In Spargel, the edges and vertices are defined by a DataSet<Tuple2<IdType, EdgeValue>> and a DataSet<Tuple2<IdType, VertexValue>> respectively:
{% highlight java %} // Spargel API ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment(); DataSet<Tuple2<Long, Long>> edges = ... DataSet<Tuple2<Long, Long>> initialVertices = vertexIds.map(new IdAssigner());
DataSet<Tuple2<Long, Long>> result = initialVertices.runOperation( VertexCentricIteration.withPlainEdges(edges, new CCUpdater(), new CCMessenger(), maxIterations));
result.print(); env.execute(“Spargel Connected Components”); {% endhighlight %}
In this algorithm, initially, each vertex has its own ID as a value (is in its own component). Hence, the need for IdAssigner(), which is used to initialize the vertex values.
In Gelly, the edges and vertices have a more intuitive definition: they are represented by separate types Edge, Vertex. After defining the edge data set, we can create a Graph from it.
{% highlight java %} // Gelly API ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment(); DataSet<Edge<Long, NullValue>> edges = ...
Graph<Long, Long, NullValue> graph = Graph.fromDataSet(edges, new MapFunction<Long, Long>() { @Override public Long map(Long value) throws Exception { return value; } }, env);
DataSet<Vertex<Long, Long>> result = graph.runVertexCentricIteration(new CCUpdater(), new CCMessenger(), maxIterations) .getVertices();
result.print(); env.execute(“Gelly Connected Components”); {% endhighlight %}
Notice that when assigning the initial vertex IDs, there is no need to perform a separate map operation. The value is specified directly in the fromDataSet() method. Instead of calling runOperation() on the set of vertices, runVertexCentricIteration() is called on the Graph instance. As previously stated, runVertexCentricIteration returns a new Graph with the updated vertex values. In order to retrieve the result (since for this algorithm we are only interested in the vertex ids and their corresponding values), we will call the getVertices() method.
The user-defined VertexUpdateFunction and MessagingFunction remain unchanged in Gelly, so you can reuse them without any changes.
In the Connected Components algorithm, the vertices propagate their current component ID in iterations, each time adopting a new value from the received neighbor IDs, provided that the value is smaller than the current minimum. To this end, we iterate over all received messages and update the vertex value, if necessary:
{% highlight java %} public static final class CCUpdater extends VertexUpdateFunction<Long, Long, Long> { @Override public void updateVertex(Long vertexKey, Long vertexValue, MessageIterator inMessages) { long min = Long.MAX_VALUE; for (long msg : inMessages) { min = Math.min(min, msg); } if (min < vertexValue) { setNewVertexValue(min); } } } {% endhighlight %}
The messages in each superstep consist of the current component ID seen by the vertex: {% highlight java %} public static final class CCMessenger extends MessagingFunction<Long, Long, Long, NullValue> { @Override public void sendMessages(Long vertexId, Long componentId) { sendMessageToAllNeighbors(componentId); } } {% endhighlight %}
Similarly to Spargel, if the value of a vertex does not change during a superstep, it will not send any messages in the superstep. This allows to do incremental updates to the hot (changing) parts of the graph, while leaving cold (steady) parts untouched.
The computation terminates after a specified maximum number of supersteps -OR- when the vertex states stop changing.