GraphX Programming Guide
- Getting Started
- The Property Graph
- Graph Operators
- Pregel API
- Graph Builders
- Vertex and Edge RDDs
- Optimized Representation
- Graph Algorithms
GraphX is a new component in Spark for graphs and graph-parallel computation. At a high level, GraphX extends the Spark RDD by introducing a new Graph abstraction: a directed multigraph with properties attached to each vertex and edge. To support graph computation, GraphX exposes a set of fundamental operators (e.g., subgraph, joinVertices, and aggregateMessages) as well as an optimized variant of the Pregel API. In addition, GraphX includes a growing collection of graph algorithms and builders to simplify graph analytics tasks.
Migrating from Spark 1.1
GraphX in Spark 1.2 contains a few user facing API changes:
To improve performance we have introduced a new version of
aggregateMessageswhich takes the messages previously returned from
mapReduceTripletsthrough a callback (
EdgeContext) rather than by return value. We are deprecating
mapReduceTripletsand encourage users to consult the transition guide.
In Spark 1.0 and 1.1, the type signature of
EdgeRDD[ED, VD]to enable some caching optimizations. We have since discovered a more elegant solution and have restored the type signature to the more natural
To get started you first need to import Spark and GraphX into your project, as follows:
If you are not using the Spark shell you will also need a
SparkContext. To learn more about
getting started with Spark refer to the Spark Quick Start Guide.
The Property Graph
The property graph is a directed multigraph
with user defined objects attached to each vertex and edge. A directed multigraph is a directed
graph with potentially multiple parallel edges sharing the same source and destination vertex. The
ability to support parallel edges simplifies modeling scenarios where there can be multiple
relationships (e.g., co-worker and friend) between the same vertices. Each vertex is keyed by a
unique 64-bit long identifier (
VertexID). GraphX does not impose any ordering constraints on
the vertex identifiers. Similarly, edges have corresponding source and destination vertex
The property graph is parameterized over the vertex (
VD) and edge (
ED) types. These
are the types of the objects associated with each vertex and edge respectively.
GraphX optimizes the representation of vertex and edge types when they are primitive data types (e.g., int, double, etc…) reducing the in memory footprint by storing them in specialized arrays.
In some cases it may be desirable to have vertices with different property types in the same graph. This can be accomplished through inheritance. For example to model users and products as a bipartite graph we might do the following:
Like RDDs, property graphs are immutable, distributed, and fault-tolerant. Changes to the values or structure of the graph are accomplished by producing a new graph with the desired changes. Note that substantial parts of the original graph (i.e., unaffected structure, attributes, and indicies) are reused in the new graph reducing the cost of this inherently functional data structure. The graph is partitioned across the executors using a range of vertex partitioning heuristics. As with RDDs, each partition of the graph can be recreated on a different machine in the event of a failure.
Logically the property graph corresponds to a pair of typed collections (RDDs) encoding the properties for each vertex and edge. As a consequence, the graph class contains members to access the vertices and edges of the graph:
EdgeRDD[ED] extend and are optimized versions of
RDD[Edge[ED]] respectively. Both
EdgeRDD[ED] provide additional
functionality built around graph computation and leverage internal optimizations. We discuss the
EdgeRDD API in greater detail in the section on vertex and edge
RDDs but for now they can be thought of as simply RDDs of the form:
RDD[(VertexID, VD)] and
Example Property Graph
Suppose we want to construct a property graph consisting of the various collaborators on the GraphX project. The vertex property might contain the username and occupation. We could annotate edges with a string describing the relationships between collaborators:
The resulting graph would have the type signature:
There are numerous ways to construct a property graph from raw files, RDDs, and even synthetic generators and these are discussed in more detail in the section on graph builders. Probably the most general method is to use the Graph object. For example the following code constructs a graph from a collection of RDDs:
In the above example we make use of the
Edge case class. Edges have a
srcId and a
dstId corresponding to the source and destination vertex identifiers. In addition, the
class has an
attr member which stores the edge property.
We can deconstruct a graph into the respective vertex and edge views by using the
graph.edges members respectively.
VertexRDD[(String, String)]which extends
RDD[(VertexID, (String, String))]and so we use the scala
caseexpression to deconstruct the tuple. On the other hand,
Edge[String]objects. We could have also used the case class type constructor as in the following:
In addition to the vertex and edge views of the property graph, GraphX also exposes a triplet view.
The triplet view logically joins the vertex and edge properties yielding an
RDD[EdgeTriplet[VD, ED]] containing instances of the
EdgeTriplet class. This
join can be expressed in the following SQL expression:
or graphically as:
EdgeTriplet class extends the
Edge class by adding the
dstAttr members which contain the source and destination properties respectively. We can use the
triplet view of a graph to render a collection of strings describing relationships between users.
Just as RDDs have basic operations like
reduceByKey, property graphs also
have a collection of basic operators that take user defined functions and produce new graphs with
transformed properties and structure. The core operators that have optimized implementations are
Graph and convenient operators that are expressed as a compositions of the
core operators are defined in
GraphOps. However, thanks to Scala implicits the
GraphOps are automatically available as members of
Graph. For example, we can
compute the in-degree of each vertex (defined in
GraphOps) by the following:
The reason for differentiating between core graph operations and
GraphOps is to be
able to support different graph representations in the future. Each graph representation must
provide implementations of the core operations and reuse many of the useful operations defined in
Summary List of Operators
The following is a quick summary of the functionality defined in both
GraphOps but presented as members of Graph for simplicity. Note that some function
signatures have been simplified (e.g., default arguments and type constraints removed) and some more
advanced functionality has been removed so please consult the API docs for the official list of
Like the RDD
map operator, the property graph contains the following:
Each of these operators yields a new graph with the vertex or edge properties modified by the user
Note that in each case the graph structure is unaffected. This is a key feature of these operators which allows the resulting graph to reuse the structural indices of the original graph. The following snippets are logically equivalent, but the first one does not preserve the structural indices and would not benefit from the GraphX system optimizations:
mapVerticesto preserve the indices:
These operators are often used to initialize the graph for a particular computation or project away unnecessary properties. For example, given a graph with the out degrees as the vertex properties (we describe how to construct such a graph later), we initialize it for PageRank:
Currently GraphX supports only a simple set of commonly used structural operators and we expect to add more in the future. The following is a list of the basic structural operators.
reverse operator returns a new graph with all the edge directions reversed.
This can be useful when, for example, trying to compute the inverse PageRank. Because the reverse
operation does not modify vertex or edge properties or change the number of edges, it can be
implemented efficiently without data movement or duplication.
subgraph operator takes vertex and edge predicates and returns the graph
containing only the vertices that satisfy the vertex predicate (evaluate to true) and edges that
satisfy the edge predicate and connect vertices that satisfy the vertex predicate. The
operator can be used in number of situations to restrict the graph to the vertices and edges of
interest or eliminate broken links. For example in the following code we remove broken links:
Note in the above example only the vertex predicate is provided. The
subgraphoperator defaults to
trueif the vertex or edge predicates are not provided.
mask operator constructs a subgraph by returning a graph that contains the
vertices and edges that are also found in the input graph. This can be used in conjunction with the
subgraph operator to restrict a graph based on the properties in another related graph. For
example, we might run connected components using the graph with missing vertices and then restrict
the answer to the valid subgraph.
groupEdges operator merges parallel edges (i.e., duplicate edges between
pairs of vertices) in the multigraph. In many numerical applications, parallel edges can be added
(their weights combined) into a single edge thereby reducing the size of the graph.
In many cases it is necessary to join data from external collections (RDDs) with graphs. For example, we might have extra user properties that we want to merge with an existing graph or we might want to pull vertex properties from one graph into another. These tasks can be accomplished using the join operators. Below we list the key join operators:
joinVertices operator joins the vertices with the input RDD and
returns a new graph with the vertex properties obtained by applying the user defined
to the result of the joined vertices. Vertices without a matching value in the RDD retain their
Note that if the RDD contains more than one value for a given vertex only one will be used. It is therefore recommended that the input RDD be made unique using the following which will also pre-index the resulting values to substantially accelerate the subsequent join.
The more general
outerJoinVertices behaves similarly to
except that the user defined
map function is applied to all vertices and can change the vertex
property type. Because not all vertices may have a matching value in the input RDD the
function takes an
Option type. For example, we can setup a graph for PageRank by initializing
vertex properties with their
You may have noticed the multiple parameter lists (e.g.,
f(a)(b)) curried function pattern used in the above examples. While we could have equally written
f(a,b)this would mean that type inference on
bwould not depend on
a. As a consequence, the user would need to provide type annotation for the user defined function:
A key step in many graph analytics tasks is aggregating information about the neighborhood of each vertex. For example, we might want to know the number of followers each user has or the average age of the the followers of each user. Many iterative graph algorithms (e.g., PageRank, Shortest Path, and connected components) repeatedly aggregate properties of neighboring vertices (e.g., current PageRank Value, shortest path to the source, and smallest reachable vertex id).
To improve performance the primary aggregation operator changed from
graph.mapReduceTripletsto the new
graph.AggregateMessages. While the changes in the API are relatively small, we provide a transition guide below.
Aggregate Messages (aggregateMessages)
The core aggregation operation in GraphX is
This operator applies a user defined
sendMsg function to each edge triplet in the graph
and then uses the
mergeMsg function to aggregate those messages at their destination vertex.
The user defined
sendMsg function takes an
EdgeContext, which exposes the
source and destination attributes along with the edge attribute and functions
sendToDst) to send
messages to the source and destination attributes. Think of
sendMsg as the map
function in map-reduce.
The user defined
mergeMsg function takes two messages destined to the same vertex and
yields a single message. Think of
mergeMsg as the reduce function in map-reduce.
aggregateMessages operator returns a
containing the aggregate message (of type
Msg) destined to each vertex. Vertices that did not
receive a message are not included in the returned
aggregateMessages takes an optional
tripletsFields which indicates what data is accessed in the
(i.e., the source vertex attribute but not the destination vertex attribute).
The possible options for the
tripletsFields are defined in
the default value is
TripletFields.All which indicates that the user
sendMsg function may access any of the fields in the
tripletFields argument can be used to notify GraphX that only part of the
EdgeContext will be needed allowing GraphX to select an optimized join strategy.
For example if we are computing the average age of the followers of each user we would only require
the source field and so we would use
TripletFields.Src to indicate that we
only require the source field
In earlier versions of GraphX we used byte code inspection to infer the
TripletFieldshowever we have found that bytecode inspection to be slightly unreliable and instead opted for more explicit user control.
In the following example we use the
aggregateMessages operator to
compute the average age of the more senior followers of each user.
aggregateMessagesoperation performs optimally when the messages (and the sums of messages) are constant sized (e.g., floats and addition instead of lists and concatenation).
Map Reduce Triplets Transition Guide (Legacy)
In earlier versions of GraphX neighborhood aggregation was accomplished using the
mapReduceTriplets operator takes a user defined map function which
is applied to each triplet and can yield messages which are aggregated using the user defined
However, we found the user of the returned iterator to be expensive and it inhibited our ability to
apply additional optimizations (e.g., local vertex renumbering).
aggregateMessages we introduced the EdgeContext which exposes the
triplet fields and also functions to explicitly send messages to the source and destination vertex.
Furthermore we removed bytecode inspection and instead require the user to indicate what fields
in the triplet are actually required.
The following code block using
can be rewritten using
Computing Degree Information
A common aggregation task is computing the degree of each vertex: the number of edges adjacent to
each vertex. In the context of directed graphs it is often necessary to know the in-degree,
out-degree, and the total degree of each vertex. The
GraphOps class contains a
collection of operators to compute the degrees of each vertex. For example in the following we
compute the max in, out, and total degrees:
In some cases it may be easier to express computation by collecting neighboring vertices and their
attributes at each vertex. This can be easily accomplished using the
collectNeighborIds and the
These operators can be quite costly as they duplicate information and require substantial communication. If possible try expressing the same computation using the
Caching and Uncaching
In Spark, RDDs are not persisted in memory by default. To avoid recomputation, they must be explicitly cached when using them multiple times (see the Spark Programming Guide). Graphs in GraphX behave the same way. When using a graph multiple times, make sure to call
Graph.cache() on it first.
In iterative computations, uncaching may also be necessary for best performance. By default, cached RDDs and graphs will remain in memory until memory pressure forces them to be evicted in LRU order. For iterative computation, intermediate results from previous iterations will fill up the cache. Though they will eventually be evicted, the unnecessary data stored in memory will slow down garbage collection. It would be more efficient to uncache intermediate results as soon as they are no longer necessary. This involves materializing (caching and forcing) a graph or RDD every iteration, uncaching all other datasets, and only using the materialized dataset in future iterations. However, because graphs are composed of multiple RDDs, it can be difficult to unpersist them correctly. For iterative computation we recommend using the Pregel API, which correctly unpersists intermediate results.
Graphs are inherently recursive data structures as properties of vertices depend on properties of their neighbors which in turn depend on properties of their neighbors. As a consequence many important graph algorithms iteratively recompute the properties of each vertex until a fixed-point condition is reached. A range of graph-parallel abstractions have been proposed to express these iterative algorithms. GraphX exposes a variant of the Pregel API.
At a high level the Pregel operator in GraphX is a bulk-synchronous parallel messaging abstraction constrained to the topology of the graph. The Pregel operator executes in a series of super steps in which vertices receive the sum of their inbound messages from the previous super step, compute a new value for the vertex property, and then send messages to neighboring vertices in the next super step. Unlike Pregel, messages are computed in parallel as a function of the edge triplet and the message computation has access to both the source and destination vertex attributes. Vertices that do not receive a message are skipped within a super step. The Pregel operators terminates iteration and returns the final graph when there are no messages remaining.
Note, unlike more standard Pregel implementations, vertices in GraphX can only send messages to neighboring vertices and the message construction is done in parallel using a user defined messaging function. These constraints allow additional optimization within GraphX.
The following is the type signature of the Pregel operator as well as a sketch of its implementation (note calls to graph.cache have been removed):
Notice that Pregel takes two argument lists (i.e.,
graph.pregel(list1)(list2)). The first
argument list contains configuration parameters including the initial message, the maximum number of
iterations, and the edge direction in which to send messages (by default along out edges). The
second argument list contains the user defined functions for receiving messages (the vertex program
vprog), computing messages (
sendMsg), and combining messages
We can use the Pregel operator to express computation such as single source shortest path in the following example.
GraphX provides several ways of building a graph from a collection of vertices and edges in an RDD or on disk. None of the graph builders repartitions the graph’s edges by default; instead, edges are left in their default partitions (such as their original blocks in HDFS).
Graph.groupEdges requires the graph to be repartitioned because it assumes identical edges will be colocated on the same partition, so you must call
Graph.partitionBy before calling
GraphLoader.edgeListFile provides a way to load a graph from a list of edges on disk. It parses an adjacency list of (source vertex ID, destination vertex ID) pairs of the following form, skipping comment lines that begin with
# This is a comment 2 1 4 1 1 2
It creates a
Graph from the specified edges, automatically creating any vertices mentioned by edges. All vertex and edge attributes default to 1. The
canonicalOrientation argument allows reorienting edges in the positive direction (
srcId < dstId), which is required by the connected components algorithm. The
minEdgePartitions argument specifies the minimum number of edge partitions to generate; there may be more edge partitions than specified if, for example, the HDFS file has more blocks.
Graph.apply allows creating a graph from RDDs of vertices and edges. Duplicate vertices are picked arbitrarily and vertices found in the edge RDD but not the vertex RDD are assigned the default attribute.
Graph.fromEdges allows creating a graph from only an RDD of edges, automatically creating any vertices mentioned by edges and assigning them the default value.
Graph.fromEdgeTuples allows creating a graph from only an RDD of edge tuples, assigning the edges the value 1, and automatically creating any vertices mentioned by edges and assigning them the default value. It also supports deduplicating the edges; to deduplicate, pass
Some of a
PartitionStrategy as the
uniqueEdges parameter (for example,
uniqueEdges = Some(PartitionStrategy.RandomVertexCut)). A partition strategy is necessary to colocate identical edges on the same partition so they can be deduplicated.
Vertex and Edge RDDs
RDD views of the vertices and edges stored within the graph. However, because
GraphX maintains the vertices and edges in optimized data structures and these data structures
provide additional functionality, the vertices and edges are returned as
respectively. In this section we review some of the additional useful functionality in these types.
RDD[(VertexID, A)] and adds the additional constraint that each
VertexID occurs only once. Moreover,
VertexRDD[A] represents a set of vertices each with an
attribute of type
A. Internally, this is achieved by storing the vertex attributes in a reusable
hash-map data-structure. As a consequence if two
VertexRDDs are derived from the same base
VertexRDD (e.g., by
mapValues) they can be joined in constant time without hash
evaluations. To leverage this indexed data structure, the
VertexRDD exposes the following
Notice, for example, how the
filter operator returns an
VertexRDD. Filter is actually
implemented using a
BitSet thereby reusing the index and preserving the ability to do fast joins
VertexRDDs. Likewise, the
mapValues operators do not allow the
map function to
VertexID thereby enabling the same
HashMap data structures to be reused. Both the
innerJoin are able to identify when joining two
VertexRDDs derived from the same
HashMap and implement the join by linear scan rather than costly point lookups.
aggregateUsingIndex operator is useful for efficient construction of a new
VertexRDD from an
RDD[(VertexID, A)]. Conceptually, if I have constructed a
VertexRDD[B] over a set of vertices,
which is a super-set of the vertices in some
RDD[(VertexID, A)] then I can reuse the index to
both aggregate and then subsequently index the
RDD[(VertexID, A)]. For example:
EdgeRDD[ED], which extends
RDD[Edge[ED]] organizes the edges in blocks partitioned using one
of the various partitioning strategies defined in
each partition, edge attributes and adjacency structure, are stored separately enabling maximum
reuse when changing attribute values.
The three additional functions exposed by the
In most applications we have found that operations on the
EdgeRDD are accomplished through the
graph operators or rely on operations defined in the base
While a detailed description of the optimizations used in the GraphX representation of distributed graphs is beyond the scope of this guide, some high-level understanding may aid in the design of scalable algorithms as well as optimal use of the API. GraphX adopts a vertex-cut approach to distributed graph partitioning:
Rather than splitting graphs along edges, GraphX partitions the graph along vertices which can
reduce both the communication and storage overhead. Logically, this corresponds to assigning edges
to machines and allowing vertices to span multiple machines. The exact method of assigning edges
depends on the
PartitionStrategy and there are several tradeoffs to the
various heuristics. Users can choose between different strategies by repartitioning the graph with
Graph.partitionBy operator. The default partitioning strategy is to use
the initial partitioning of the edges as provided on graph construction. However, users can easily
switch to 2D-partitioning or other heuristics included in GraphX.
Once the edges have be partitioned the key challenge to efficient graph-parallel computation is
efficiently joining vertex attributes with the edges. Because real-world graphs typically have more
edges than vertices, we move vertex attributes to the edges. Because not all partitions will
contain edges adjacent to all vertices we internally maintain a routing table which identifies where
to broadcast vertices when implementing the join required for operations like
GraphX includes a set of graph algorithms to simplify analytics tasks. The algorithms are contained in the
org.apache.spark.graphx.lib package and can be accessed directly as methods on
GraphOps. This section describes the algorithms and how they are used.
PageRank measures the importance of each vertex in a graph, assuming an edge from u to v represents an endorsement of v’s importance by u. For example, if a Twitter user is followed by many others, the user will be ranked highly.
GraphX comes with static and dynamic implementations of PageRank as methods on the
PageRank object. Static PageRank runs for a fixed number of iterations, while dynamic PageRank runs until the ranks converge (i.e., stop changing by more than a specified tolerance).
GraphOps allows calling these algorithms directly as methods on
GraphX also includes an example social network dataset that we can run PageRank on. A set of users is given in
graphx/data/users.txt, and a set of relationships between users is given in
graphx/data/followers.txt. We compute the PageRank of each user as follows:
The connected components algorithm labels each connected component of the graph with the ID of its lowest-numbered vertex. For example, in a social network, connected components can approximate clusters. GraphX contains an implementation of the algorithm in the
ConnectedComponents object, and we compute the connected components of the example social network dataset from the PageRank section as follows:
A vertex is part of a triangle when it has two adjacent vertices with an edge between them. GraphX implements a triangle counting algorithm in the
TriangleCount object that determines the number of triangles passing through each vertex, providing a measure of clustering. We compute the triangle count of the social network dataset from the PageRank section. Note that
TriangleCount requires the edges to be in canonical orientation (
srcId < dstId) and the graph to be partitioned using
Suppose I want to build a graph from some text files, restrict the graph to important relationships and users, run page-rank on the sub-graph, and then finally return attributes associated with the top users. I can do all of this in just a few lines with GraphX: