Processing Hierarchical Data using Spark Graphx Pregel API

August 3, 2017 by Suraj Bang and Qubole Updated March 2nd, 2020

Today distributed compute engines are backbone of many analytic, batch & streaming applications. Spark provides many advanced features (pivot, analytic window functions, etc.) out of box to transform data. Sometimes there is need to process hierarchical data or perform hierarchical calculations. Many database vendors provide features like “Recursive CTE’s (Common Table Expressions)”[1] or “connect by”[2] SQL clause to query\transform hierarchical data. CTE’s are also known by recursive queries or parent child queries. In this post, we will look at how we can address it with Spark.

Hierarchical Data Overview –

Hierarchical relationships exist where one item of data is the parent of another item. Hierarchical data can be represented using a graph property object model where every row is a vertex(node) and connection is the edge(relationship) that connects the vertices and columns are the properties of the vertex.

Some Use Cases –

Financial Calculations – Subaccounts rolling up into parent account all the way to topmost account
Creating Organizational hierarchy – Manager employee relationship with path
Generating Graph of links between web pages with path
Any kind of Iterative computations involving linked data

Challenges –

Querying hierarchical data in a distributed system has some challenges

The data is connected but it is distributed across partitions and nodes. The implementation to solve this should be optimized for performing iterations and moving the data (shuffle) as needed.
Depth of graph can vary over time – solution should take care of varying depth and should not enforce users to define it before processing.

Solution –

One of the ways of implementing CTE’s in spark is using Graphx Pregel API.

What is Graphx Pregel API?

Graphx[3] is a spark API for graph and graph-parallel computation. Graph algorithms are iterative in nature and properties of vertices depends upon the properties of its directly or indirectly (connected via other vertices) connected vertices. Pregel is a vertex centric graph processing model developed by Google and spark graphx provides an optimized variant of pregel api[4].

How does Pregel API work?

Pregel API processing consists of executing super steps –

Superstep 0:

Pass initial message to all the vertices
Send the value as message to its directly connected vertices

Superstep 1:

Receive messages from the previous steps
Mutate the value

Send the value as message to its directly connected vertices

Repeat Superstep 1 until there are messages to pass and stop when there are no more messages to be passed.

Hierarchical Data for use case –

Table below shows sample employee data that we will be using for generating top down hierarchy. Here the manager for the employee is represented by mgr_id field which has emp_id values.

We will add the following columns as part of processing

Level (Depth)	The level at which vertex stands in the hierarchy
Path	The path from the top most vertex to current vertex in the hierarchy
Root	The top most vertex in the hierarchy, useful when multiple hierarchies exists in the dataset
Iscyclic	Incase if there is bad data, a cyclic relationship exists, then flag it
Isleaf	If vertex has no parent, flag it

Code –

// The code below demonstrates use of Graphx Pregel API - Scala 2.11+ 

// functions to build the top down hierarchy

//setup & call the pregel api
def calcTopLevelHierarcy(vertexDF: DataFrame, edgeDF: DataFrame): RDD[(Any,(Int,Any,String,Int,Int))] = {

// create the vertex RDD
// primary key, root, path
val verticesRDD = vertexDF
  .rdd
  .map{x=> (x.get(0),x.get(1) , x.get(2))}
  .map{ x => (MurmurHash3.stringHash(x._1.toString).toLong, ( x._1.asInstanceOf[Any], x._2.asInstanceOf[Any] , x._3.asInstanceOf[String]) ) }

// create the edge RDD
// top down relationship
val EdgesRDD = edgeDF.rdd.map{x=> (x.get(0),x.get(1))}
  .map{ x => Edge(MurmurHash3.stringHash(x._1.toString).toLong,MurmurHash3.stringHash(x._2.toString).toLong,"topdown" )}

// create graph
val graph = Graph(verticesRDD, EdgesRDD).cache()

val pathSeperator = """/"""

// initialize id,level,root,path,iscyclic, isleaf
val initialMsg = (0L,0,0.asInstanceOf[Any],List("dummy"),0,1)

// add more dummy attributes to the vertices - id, level, root, path, isCyclic, existing value of current vertex to build path, isleaf, pk
val initialGraph = graph.mapVertices((id, v) => (id,0,v._2,List(v._3),0,v._3,1,v._1) )

val hrchyRDD = initialGraph.pregel(initialMsg,
  Int.MaxValue,
  EdgeDirection.Out)(
  setMsg,
  sendMsg,
  mergeMsg)


// build the path from the list
val hrchyOutRDD = hrchyRDD.vertices.map{case(id,v) => (v._8,(v._2,v._3,pathSeperator + v._4.reverse.mkString(pathSeperator),v._5, v._7 )) }

  hrchyOutRDD

}




//mutate the value of the vertices
def setMsg(vertexId: VertexId, value: (Long,Int,Any,List[String], Int,String,Int,Any), message: (Long,Int, Any,List[String],Int,Int)): (Long,Int, Any,List[String],Int,String,Int,Any) = {
  if (message._2 < 1) { //superstep 0 - initialize
    (value._1,value._2+1,value._3,value._4,value._5,value._6,value._7,value._8)
  } else if ( message._5 == 1) { // set isCyclic   
    (value._1, value._2, value._3, value._4, message._5, value._6, value._7,value._8)
  } else if ( message._6 == 0 ) { // set isleaf
    (value._1, value._2, value._3, value._4, value._5, value._6, message._6,value._8)
  } else { // set new values
    ( message._1,value._2+1, message._3, value._6 :: message._4 , value._5,value._6,value._7,value._8)
  }
}



// send the value to vertices
def sendMsg(triplet: EdgeTriplet[(Long,Int,Any,List[String],Int,String,Int,Any), _]): Iterator[(VertexId, (Long,Int,Any,List[String],Int,Int))] = {
  val sourceVertex = triplet.srcAttr
  val destinationVertex = triplet.dstAttr
// check for icyclic
 if (sourceVertex._1 == triplet.dstId || sourceVertex._1 == destinationVertex._1)
    if (destinationVertex._5==0) { //set iscyclic
      Iterator((triplet.dstId, (sourceVertex._1, sourceVertex._2, sourceVertex._3,sourceVertex._4, 1,sourceVertex._7)))
    } else {
     Iterator.empty
    }
  else {
    if (sourceVertex._7==1) //is NOT leaf
    {
      Iterator((triplet.srcId, (sourceVertex._1,sourceVertex._2,sourceVertex._3, sourceVertex._4 ,0, 0 )))
    }
    else { // set new values
      Iterator((triplet.dstId, (sourceVertex._1, sourceVertex._2, sourceVertex._3, sourceVertex._4, 0, 1)))
    }
  }
}



// receive the values from all connected vertices
def mergeMsg(msg1: (Long,Int,Any,List[String],Int,Int), msg2: (Long,Int, Any,List[String],Int,Int)): (Long,Int,Any,List[String],Int,Int) = {
// dummy logic not applicable to the data in this usecase
msg2
}


// Test with some sample data 


val empData = Array(
  ("EMP001", "Bob", "Baker", "CEO", null.asInstanceOf[String])
  , ("EMP002", "Jim", "Lake", "CIO", "EMP001")
  , ("EMP003", "Tim", "Gorab", "MGR", "EMP002")
  , ("EMP004", "Rick", "Summer", "MGR", "EMP002")
  , ("EMP005", "Sam", "Cap", "Lead", "EMP004")
  , ("EMP006", "Ron", "Hubb", "Sr.Dev", "EMP005")
  , ("EMP007", "Cathy", "Watson", "Dev", "EMP006")
  , ("EMP008", "Samantha", "Lion", "Dev", "EMP007")
  , ("EMP009", "Jimmy", "Copper", "Dev", "EMP007")
  , ("EMP010", "Shon", "Taylor", "Intern", "EMP009")
)

// create dataframe with some partitions
val empDF = sc.parallelize(empData, 3).toDF("emp_id","first_name","last_name","title","mgr_id").cache()

// primary key , root, path - dataframe to graphx for vertices
val empVertexDF = empDF.selectExpr("emp_id","concat(first_name,' ',last_name)","concat(last_name,' ',first_name)")

// parent to child - dataframe to graphx for edges
val empEdgeDF = empDF.selectExpr("mgr_id","emp_id").filter("mgr_id is not null")

// call the function
val empHirearchyExtDF = calcTopLevelHierarcy(empVertexDF,empEdgeDF)
  .map{ case(pk,(level,root,path,iscyclic,isleaf)) => (pk.asInstanceOf[String],level,root.asInstanceOf[String],path,iscyclic,isleaf)}
  .toDF("emp_id_pk","level","root","path","iscyclic","isleaf").cache()



// extend original table with new columns
val empHirearchyDF = empHirearchyExtDF.join(empDF , empDF.col("emp_id") === empHirearchyExtDF.col("emp_id_pk")).selectExpr("emp_id","first_name","last_name","title","mgr_id","level","root","path","iscyclic","isleaf")

// print
empHirearchyDF.show()

Output –

Under the hood –

A spark job breaks up into job(s), stage(s) and task(s). Pregel API internally generates multiple jobs due to its iterative nature. A job is generated every time messages are passed to the vertexes. Each job may end up with multiple shuffles as data can be on different nodes.

Things to watch out for are long RDD lineages that are created when working with huge datasets.