Skip to content

Latest commit

 

History

History
689 lines (520 loc) · 35.6 KB

README.md

File metadata and controls

689 lines (520 loc) · 35.6 KB

Spark Dgraph Connector

This project provides an Apache Spark connector for Dgraph databases in Scala and Python. It comes with a Spark Data Source to read graphs from a Dgraph cluster directly into DataFrames, GraphX or GraphFrames. The connector supports filter pushdown, projection pushdown and partitioning by orthogonal dimensions predicates and nodes.

Example Scala code:

import org.apache.spark.sql.DataFrame

val target = "localhost:9080"

import org.apache.spark.graphx._
import uk.co.gresearch.spark.dgraph.graphx._
val graph: Graph[VertexProperty, EdgeProperty] = spark.read.dgraph.graphx(target)
val edges: RDD[Edge[EdgeProperty]] = spark.read.dgraph.edges(target)
val vertices: RDD[(VertexId, VertexProperty)] = spark.read.dgraph.vertices(target)

import org.graphframes.GraphFrame
import uk.co.gresearch.spark.dgraph.graphframes._
val graph: GraphFrame = spark.read.dgraph.graphframes(target)
val edges: DataFrame = spark.read.dgraph.edges(target)
val vertices: DataFrame = spark.read.dgraph.vertices(target)

import org.apache.spark.sql.DataFrame
import uk.co.gresearch.spark.dgraph.connector._
val triples: DataFrame = spark.read.dgraph.triples(target)
val edges: DataFrame = spark.read.dgraph.edges(target)
val nodes: DataFrame = spark.read.dgraph.nodes(target)

Example Python code (pyspark ≥3.0, see PySpark Shell and Python script):

from pyspark.sql import DataFrame
from gresearch.spark.dgraph.connector import *

triples: DataFrame = spark.read.dgraph.triples("localhost:9080")
edges: DataFrame = spark.read.dgraph.edges("localhost:9080")
nodes: DataFrame = spark.read.dgraph.nodes("localhost:9080")

Features

The connector provides the following features:

  • Scala, Java, and Python API: Supports Spark's Scala, Java, and Python Dataset API.
  • Spark Graph API support: Supports Spark Graph APIs GraphX and GraphFrames.
  • Various graph representations: Graph data are available as Spark Datasets in various formats: Triples, Nodes and Edge Datasets, fully typed, wide or long format.
  • Filter Pushdown: Many filters applied to the Graph Datasets will be pushed down to Dgraph so that only the relevant sub-graph is read from the cluster.
  • Projection Pushdown: Only actually used columns of the Graph Datasets will be read from the Dgraph cluster.
  • Multi-language strings: Strings values of the same predicate can be stored in multiple languages.
  • Spark metrics: The connector collects Spark metrics per partition providing insights in throughout and timing of the communication to the Dgraph cluster.
  • Graph Partitioning and Streaming: Graph data are partitioned and streamed in small chunks from Dgraph into Spark. This guarantees that graphs of any size can be read into Spark.

Limitations

The connector has the following known limitations:

  • Read-only: The connector does not support mutating the graph (issue #8).
  • Namespaces: The connector can only read the default namespace (issue #148).
  • Authorization: No authorization against Dgraph supported (issue #149).
  • Limited Lifetime of Transactions: The connector optionally reads all partitions within the same transaction, but concurrent mutations reduce the lifetime of that transaction.
  • Language tags: The node source in wide mode cannot read string values with language tags. All other sources and modes can read language strings.
  • Filtering on language string: The connector does not support filtering predicates with Dgraph @lang directives.
  • Facets: The connector cannot read facets.

Using Spark Dgraph Connector

The Spark Dgraph Connector is available for Spark 3.0, 3.1, 3.2, 3.3, 3.4 and 3.5, with Scala 2.12 and 2.13. Use Maven artifact ID spark-dgraph-connector_2.12 or spark-dgraph-connector_2.12. The Spark version is part of the package version, i.e. 0.11.0-3.0, 0.11.0-3.1, 0.11.0-3.2, 0.11.0-3.3, 0.11.0-3.4 and 0.11.0-3.5, respectively.

SBT

Add this line to your build.sbt file to use the latest version for Spark 3.5:

libraryDependencies += "uk.co.gresearch.spark" %% "spark-dgraph-connector" % "0.11.0-3.5"

Maven

Add this dependency to your pom.xml file to use the latest version:

<dependency>
  <groupId>uk.co.gresearch.spark</groupId>
  <artifactId>spark-dgraph-connector_2.13</artifactId>
  <version>0.11.0-3.5</version>
</dependency>

Spark Shell

Launch the Scala Spark REPL (Spark ≥3.0.0) with the Spark Dgraph Connector dependency (version ≥0.5.0) as follows:

spark-shell --packages uk.co.gresearch.spark:spark-dgraph-connector_2.12:0.11.0-3.5

PySpark Shell and Python script

Launch the Python Spark REPL (pyspark ≥3.0.0) with the Spark Dgraph Connector dependency (version ≥0.5.0) as follows:

pyspark --packages uk.co.gresearch.spark:spark-dgraph-connector_2.12:0.11.0-3.5

Run your Python scripts that use PySpark (pyspark ≥3.0.0) and the Spark Dgraph Connector (version ≥0.5.0) via spark-submit:

spark-submit --packages uk.co.gresearch.spark:spark-dgraph-connector_2.12:0.11.0-3.5 [script.py]

Examples

The following examples use a local Dgraph (≥20.03.3) instance setup as described in the Dgraph Quickstart Guide. Run Step 1 to start an instance, a DROP_ALL for Dgraph ≥20.07.0 only, Step 2 to load example graph data, and Step 3 to add a schema. These steps are provided in the following scripts:

./dgraph-instance.start.sh
./dgraph-instance.drop-all.sh  # for Dgraph ≥20.07.0 only
./dgraph-instance.schema.sh
./dgraph-instance.insert.sh

The Dgraph version can optionally be set via DGRAPH_TEST_CLUSTER_VERSION environment variable.

The connection to Dgraph can be established via a target, which is the hostname and gRPC port of a Dgraph Alpha node in the form <hostname>:<port>. With our example instance started above, we can use localhost:9080 as the target.

Dgraph UI

The Dgraph UI Ratel can be used to query your local Dgraph instance. Open a browser and got to https://play.dgraph.io/?latest. Connect to your local Dgraph via http://localhost:8080.

GraphX

You can load the entire Dgraph database into an Apache Spark GraphX graph. For example:

import uk.co.gresearch.spark.dgraph.graphx._

val graph = spark.read.dgraph.graphx("localhost:9080")

Example code to perform a PageRank computation on this graph to test that the connector is working:

val pageRank = graph.pageRank(0.0001)
pageRank.vertices.foreach(println)

GraphFrames

You can load the entire Dgraph database into a GraphFrames graph. For example:

import uk.co.gresearch.spark.dgraph.graphframes._

val graph: GraphFrame = spark.read.dgraph.graphframes("localhost:9080")

Example code to perform a PageRank computation on this graph to test that the connector is working:

val pageRank = graph.pageRank.maxIter(10)
pageRank.run().triplets.show(false)

Note: Predicates get renamed when they are loaded from the Dgraph database. Any . (dot) in the name is replaced by a _ (underscore). To guarantee uniqueness of names, underscores in the original predicate names are replaced by two underscores. For instance, predicates dgraph.type and release_date become dgraph_type and release__date, respectively.

DataFrame

Dgraph data can be loaded into Spark DataFrames in various forms:

  • Triples
    • fully typed values
    • string values
  • Nodes
    • fully typed properties
    • wide schema
  • Edges

Typed Triples

You can load the entire Dgraph database as triples into an Apache Spark DataFrame. For example:

import uk.co.gresearch.spark.dgraph.connector._

val triples = spark.read.dgraph.triples("localhost:9080")

The returned DataFrame has the following schema:

root
 |-- subject: long (nullable = false)
 |-- predicate: string (nullable = true)
 |-- objectUid: long (nullable = true)
 |-- objectString: string (nullable = true)
 |-- objectLong: long (nullable = true)
 |-- objectDouble: double (nullable = true)
 |-- objectTimestamp: timestamp (nullable = true)
 |-- objectBoolean: boolean (nullable = true)
 |-- objectGeo: string (nullable = true)
 |-- objectPassword: string (nullable = true)
 |-- objectType: string (nullable = true)

The object value gets stored in exactly one of the object* (except objectType) columns, depending on the type of the value. The objectType column provides the type of the object. Here is an example:

subject predicate objectString objectLong objectDouble objectTimestamp objectBoolean objectGeo objectPassword objectType
1 dgraph.type Person null null null null null null string
1 name Luke Skywalker null null null null null null string
2 dgraph.type Person null null null null null null string
2 name Princess Leia null null null null null null string
3 dgraph.type Film null null null null null null string
3 name Star Wars: Episode IV - A New Hope null null null null null null string
3 release_date null null null 1977-05-25 00:00:00 null null null timestamp
3 revenue null null 7.75E8 null null null null double
3 running_time null 121 null null null null null long

This model allows you to store the fully-typed triples in a DataFrame.

String Triples

The triples can also be loaded in an un-typed, narrow form:

import uk.co.gresearch.spark.dgraph.connector._

spark
  .read
  .option(TriplesModeOption, TriplesModeStringOption)
  .dgraph.triples("localhost:9080")
  .show

The resulting DataFrame has the following schema:

root
 |-- subject: long (nullable = false)
 |-- predicate: string (nullable = true)
 |-- objectString: string (nullable = true)
 |-- objectType: string (nullable = true)

The object value gets stored as a string in objectString, and objectType provides you with the actual type of the object. Here is an example:

subject predicate objectString objectType
1 dgraph.type Person string
1 name Luke Skywalker string
2 dgraph.type Person string
2 name Princess Leia string
3 dgraph.type Film string
3 revenue 7.75E8 double
3 running_time 121 long
3 starring 1 uid
3 starring 2 uid
3 starring 6 uid
3 director 7 uid
3 name Star Wars: Episode IV - A New Hope string
3 release_date 1977-05-25 00:00:00.0 timestamp

Typed Nodes

You can load all nodes into a DataFrame in a fully-typed form. This contains all the nodes' properties but no edges to other nodes:

import uk.co.gresearch.spark.dgraph.connector._

spark.read.dgraph.nodes("localhost:9080")

The returned DataFrame has the following schema:

root
 |-- subject: long (nullable = false)
 |-- predicate: string (nullable = true)
 |-- objectString: string (nullable = true)
 |-- objectLong: long (nullable = true)
 |-- objectDouble: double (nullable = true)
 |-- objectTimestamp: timestamp (nullable = true)
 |-- objectBoolean: boolean (nullable = true)
 |-- objectGeo: string (nullable = true)
 |-- objectPassword: string (nullable = true)
 |-- objectType: string (nullable = true)

The schema of the returned DataFrame is very similar to the typed triples schema, except that there is no objectUid column linking to other nodes. Here is an example:

subject predicate objectString objectLong objectDouble objectTimestamp objectBoolean objectGeo objectPassword objectType
1 dgraph.type Person null null null null null null string
1 name Luke Skywalker null null null null null null string
2 dgraph.type Person null null null null null null string
2 name Princess Leia null null null null null null string
3 dgraph.type Film null null null null null null string
3 revenue null null 7.75E8 null null null null double
3 running_time null 121 null null null null null long
3 name Star Wars: Episode IV - A New Hope null null null null null null string
3 release_date null null null 1977-05-25 00:00:00 null null null timestamp

Wide Nodes

Nodes can also be loaded in a wide, fully-typed format:

import uk.co.gresearch.spark.dgraph.connector._

spark
  .read
  .option(NodesModeOption, NodesModeWideOption)
  .dgraph.nodes("localhost:9080")

The returned DataFrame has the following schema format, which is dependent on the schema of the underlying Dgraph database. Node properties are stored in typed columns and are ordered alphabetically (property columns start after the subject column):

root
 |-- subject: long (nullable = false)
 |-- dgraph.graphql.schema: string (nullable = true)
 |-- dgraph.type: string (nullable = true)
 |-- name: string (nullable = true)
 |-- release_date: timestamp (nullable = true)
 |-- revenue: double (nullable = true)
 |-- running_time: long (nullable = true)

Note: The graph schema could become very large and therefore the DataFrame could become prohibitively wide.

subject dgraph.graphql.schema dgraph.type name release_date revenue running_time
1 null Person Luke Skywalker null null null
2 null Person Princess Leia null null null
3 null Film Star Wars: Episode IV - A New Hope 1977-05-25 00:00:00 7.75E8 121
4 null Film Star Wars: Episode VI - Return of the Jedi 1983-05-25 00:00:00 5.72E8 131
5 null Film Star Trek: The Motion Picture 1979-12-07 00:00:00 1.39E8 132
6 null Person Han Solo null null null
7 null Person George Lucas null null null
8 null Person Irvin Kernshner null null null
9 null Person Richard Marquand null null null
10 null Film Star Wars: Episode V - The Empire Strikes Back 1980-05-21 00:00:00 5.34E8 124

Note: The Wide Nodes source enforces the predicate partitioner to produce a single partition.

Edges

Edges can be loaded as follows:

import uk.co.gresearch.spark.dgraph.connector._

spark.read.dgraph.edges("localhost:9080")

The returned DataFrame has the following simple schema:

root
 |-- subject: long (nullable = false)
 |-- predicate: string (nullable = true)
 |-- objectUid: long (nullable = false)

Though there is only a single object column for the destination node, it is called objectUid to align with the DataFrame schemata above.

subject predicate objectUid
3 starring 1
3 starring 2
3 starring 6
3 director 7
4 starring 1
4 starring 2
4 starring 6
4 director 9
10 starring 1
10 starring 2
10 starring 6
10 director 8

Language Strings

Predicates marked in the Dgraph schema with the @lang directive can store string values in multiple languages at a time:

{
  set {
   _:sw3 <title> "Star Wars: Episode VI - Return of the Jedi" .
   _:sw3 <title@en> "Star Wars: Episode VI - Return of the Jedi" .
   _:sw3 <title@zh> "星際大戰六部曲:絕地大反攻" .
   _:sw3 <title@th> "สตาร์ วอร์ส เอพพิโซด 6: การกลับมาของเจได" .
   _:sw3 <title@de> "Die Rückkehr der Jedi-Ritter" .
  }
}

The connector reads all these languages. Each of the predicate name contains the language in the form predicate@language:

subject predicate objectString objectType
6 dgraph.type Film string
6 title Star Wars: Episode VI - Return of the Jedi string
6 title@en Star Wars: Episode VI - Return of the Jedi string
6 title@zh 星際大戰六部曲:絕地大反攻 string
6 title@th สตาร์ วอร์ส เอพพิโซด 6: การกลับมาของเจได string
6 title@de Die Rückkehr der Jedi-Ritter string

Transactions

Dgraph isolates reads from writes through transactions. Since the connector initiates multiple reads while fetching the entire graph (partitioning), writes called mutations should be isolated in order to get a consistent snapshot of the graph.

Setting the dgraph.transaction.mode option to "read" will cause the connector to read all partitions within the same transaction. However, this will cause an exception on the Dgraph cluster when too many mutations occur while reading partitions. With that option set to "none", no such exception will occur but reads are not isolated from writes.

Filter Pushdown

The connector supports filter pushdown to improve efficiency when reading only sub-graphs. This is supported only in conjunction with the predicate partitioner. Spark filters can only be pushed for some column and data source because columns may have different meaning. Columns can be of the following types:

Column Type Description Type Columns Sources
subject column the subject of the row long subject all DataFrame sources
predicate column the predicate of the row string predicate all but Wide Nodes source
predicate value column the value of a specific predicate, column name is predicate name any one column for each predicate in the schema, e.g. dgraph.type Wide Nodes source
object value columns object value of the row long
string
long
double
timestamp
boolean
geo
password
objectUid
objectString
objectLong
objectDouble
objectTimestamp
objectBoolean
objectGeo
objectPassword
all but Wide Nodes source
the String Triples source has only objectString
the Typed Nodes source lacks the objectUid
the Edges source has only objectUid
object type column the type of the object string objectType String Triples source, Typed Triples source and Typed Nodes source

The following table lists all supported Spark filters:

Spark Filter Supported Columns Example
EqualTo
  • .where($"subject" === 1L)
  • .where($"predicate" === "dgraph.type")
  • .where($"dgraph.type" === "Person")
  • .where($"objectType" === "string")
  • .where($"objectLong" === 123)
In
  • .where($"subject".isin(1L,2L))
  • .where($"predicate".isin("release_date", "revenue"))
  • .where($"dgraph.type".isin("Person","Film"))
  • .where($"objectType".isin("string","long"))
  • .where($"objectLong".isin(123,456))
IsNotNull
  • .where($"dgraph.type".isNotNull)
  • .where($"objectLong".isNotNull)

Projection Pushdown

The connector supports projection pushdown to improve efficiency when reading only sub-graphs. A projection in Spark terms is a select operation that selects only a subset of a DataFrame's columns. The Wide Nodes source supports projection pushdown on all predicate value columns.

Filter and Projection Pushdown Example

The following query uses filter and projection pushdown. First we define a wide node DataFrame:

val df =
  spark.read
    .options(Map(
      NodesModeOption -> NodesModeWideOption,
      PartitionerOption -> PredicatePartitionerOption
    ))
    .dgraph.nodes("localhost:9080")

Then we select some columns (projection) and rows (filter):

df
  .select($"subject", $"`dgraph.type`", $"revenue")  // projection
  .where($"revenue".isNotNull)                       // filter
  .show()

This selects the columns subject, dgraph.type and revenue for only those rows that actually have a value for revenue. The underlying query to Dgraph simplifies from (the full graph):

{
  pred1 as var(func: has(<dgraph.graphql.schema>))
  pred2 as var(func: has(<dgraph.graphql.xid>))
  pred3 as var(func: has(<dgraph.type>))
  pred4 as var(func: has(<name>))
  pred5 as var(func: has(<release_date>))
  pred6 as var(func: has(<revenue>))
  pred7 as var(func: has(<running_time>))

  result (func: uid(pred1,pred2,pred3,pred4,pred5,pred6,pred7)) {
    uid
    <dgraph.graphql.schema>
    <dgraph.graphql.xid>
    <dgraph.type>
    <name>
    <release_date>
    <revenue>
    <running_time>
  }
}

to (selected predicates and nodes only):

{
  pred1 as var(func: has(<revenue>))

  result (func: uid(pred1)) {
    uid
    <dgraph.type>
    <revenue>
  }
}

The response is faster as only relevant data are transferred between Dgraph and Spark.

subject dgraph.type revenue
4 Film 7.75E8
5 Film 5.34E8
6 Film 5.72E8
9 Film 1.39E8

Metrics

The connector (Spark ≥3.0 only) collects metrics per partition that provide insights in throughout and timing of the communication to the Dgraph cluster. For each request to Dgraph (a chunk), the number of received bytes, uids and retrieval time are recorded and summed per partition. The values can be seen on the Spark UI for the respective stages that performs the read:

Dgraph metrics as shown on Spark UI Stages page

The connector uses Spark Accumulators to collect these metrics. They can be accessed by the Spark driver via a SparkListener:

val handler = new SparkListener {
  override def onStageCompleted(stageCompleted: SparkListenerStageCompleted): Unit =
    stageCompleted.stageInfo.accumulables.values.foreach(println)
}

spark.sparkContext.addSparkListener(handler)
spark.read.dgraph.triples("localhost:9080").count()

The following metrics are available:

Metric Description
Dgraph Bytes Size of JSON responses from the Dgraph cluster in Byte.
Dgraph Chunks Number of requests sent to the Dgraph cluster.
Dgraph Time Time waited for Dgraph to respond in Seconds.
Dgraph Uids Number of Uids read.

Partitioning

Partitioning your Dgraph graph is essential to be able to load large quantities of graph data into Spark. Spark splits data into partitions, where ideally all partitions have the same size and are of decent size. Partitions that are too large will kill your Spark executor as they won't fit into memory. When partitions are too small your Spark job becomes inefficient and slow, but will not fail.

Each partition connects to the Dgraph cluster and reads a specific sub-graph. Partitions are non-overlapping.

This connector provides various ways to partition your graph. When the default partitioning does not work for your specific use case, try a more appropriate partitioning scheme.

Partitioner

The following Partitioner implementations are available:

Partitioner partition by Description Use Case
Singleton nothing Provides a single partition for the entire graph. Unit Tests and small graphs that fit into a single partition. Can be used for large graphs if combined with a "by uid" partitioner.
Predicate predicate Provides multiple partitions with at most P predicates per partition where P defaults to 1000. Picks multiple predicates from the same Dgraph group. Large graphs where each predicate fits into a partition, otherwise combine with Uid Range partitioner. Skewness of predicates reflects skewness of partitions.
Uid Range uids Each partition has at most N uids where N defaults to 1000000. Large graphs where single uids fit into a partition. Can be combined with any predicate partitioner, otherwise induces internal Dgraph cluster communication across groups.
Predicate + Uid Range (default) predicates + uids Partitions by predicate first (see Predicate Partitioner), then each partition gets partitioned by uid (see Uid Partitioner) Graphs of any size.

Partitioning by Predicates

The Dgraph data can be partitioned by predicates. Each partition then contains a distinct set of predicates. The number of predicates per partition can be configured via dgraph.partitioner.predicate.predicatesPerPartition, which defaults to 1000.

Predicate partitions connect only to alpha nodes that contain those predicates. Hence, these reads are all locally to the alpha nodes and induce no Dgraph cluster internal communication.

Partitioning by Uids

A uid represents a node or vertex in Dgraph terminology. An "Uid Range" partitioning splits the graph by the subject of the graph triples. This can be combined with predicate partitioning, which serves as an orthogonal partitioning. Without predicate partitioning, uid partitioning induces internal Dgraph cluster communication across the groups.

The uid partitioning always works on top of a predicate partitioner. If none is defined a singleton partitioner is used. The number of uids of each underlying partition has to be estimated. Once the number of uids is estimated, the partition is further split into ranges of that uid space.

The space of existing uids is split into ranges of N uids per partition. The N defaults to 1000000 and can be configured via dgraph.partitioner.uidRange.uidsPerPartition. The uids are allocated to partitions in ascending order. Such a split will not be done if more than dgraph.partitioner.uidRange.maxPartitions partitions would be created. This defaults to 10000. If vertex size is skewed and a function of uid, then partitions will be skewed as well.

Note: With uid partitioning, the chunk size configured via dgraph.chunkSize should be at least a 10th of the number of uids per partition configured via dgraph.partitioner.uidRange.uidsPerPartition to avoid inefficiency due to chunks overlapping with partition borders. When your result is sparse w.r.t. the uid space set the chunk size to 100th or less.

Streamed Partitions

The connector reads each partition from Dgraph in a streamed fashion. It splits up a partition into smaller chunks, where each chunk contains 100000 uids. This chunk size can be configured via dgraph.chunkSize. Each chunk sends a single query to Dgraph. The chunk size limits the size of the result. Due to the low memory footprint of the connector, Spark could read your entire graph via a single partition (you would have to repartition the read DataFrame to make Spark shuffle the data properly). However, this would be slow, but it proves the connector can handle any size of graph with fixed executor memory requirement.

Shaded Dependencies

This connector comes packaged with a Java Dgraph client and all its dependencies. This is necessary because the Dgraph client requires Guava ≥20.0, where ≥24.1.1-jre is recommended, as well as Protobuf Java ≥3.4.0. Running this in a Spark deployment would cause these dependencies to conflict with older versions deployed with Spark. Therefore, these newer versions are provided with the connector as shaded dependencies (renamed package names).

Please refer to the NOTICE file for licences of that third-party software.

Logging

The connector uses Spark's Log4j standard logging framework. Add the following line to your log4j.properties to set the log level of the connector specifically:

log4j.logger.uk.co.gresearch.spark.dgraph.connector=DEBUG

See SPARK_HOME/conf/log4j.properties.template for a template file.

Testing

Some unit tests require a Dgraph (≥20.03.3) cluster running at localhost:9080. It has to be set up as described in the Examples section. If that cluster is not running, the unit tests will launch and set up such a cluster for you. This requires docker to be installed on your machine and will make the tests take longer. If you run those tests frequently it is recommended you run the cluster setup yourself.

You can set the Dgraph version that is started automatically by setting environment variable DGRAPH_TEST_CLUSTER_VERSION. The default version is defined in uk.co.gresearch.spark.dgraph.DgraphTestCluster.DgraphDefaultVersion.

The Python code can be tested with pytest:

PYTHONPATH="python:python/test" python -m pytest python/test