Graph Features in Spark 3.0: Integrating Graph Querying and Algorithms in Spark Graph

WIFI SSID:Spark+AISummit | Password: UnifiedDataAnalytics

#UniﬁedDataAnalytics #SparkAISummit

#UnifiedDataAnalytics #SparkAISummit
Graphs are everywhere
3

… and growing
4

Graphs at Spark Summit
5

Property Graphs & Big Data
The Property Graph data model is becoming increasingly mainstream
Cloud graph data services like Azure CosmosDB or Amazon Neptune
Simple graph features in SQLServer 2017, multiple new graph DB products
New graph query language to be standardized by ISO
Neo4j becoming common operational store in retail, ﬁnance, telcos … and more
Increasing interest in graph algorithms over graph data as a basis for AI
Apache® Spark is the leading scale-out clustered memory solution for Big Data
Spark 2: Data viewed as tables (DataFrames), processed by SQL,
in function chains, using queries and user functions,
transforming immutable tabular data sets
6

Graphs are coming to Spark
7
[SPARK-25994]
SPIP: Property Graphs, Cypher Queries, and Algorithms
Goal
● bring Property Graphs and the Cypher Query language to
Spark
● the SparkSQL for graphs
Status
● Accepted by the community
● Implementation still Work in Progress

Demonstration

The Property Graph

The Whiteboard Model Is the Physical Model
Eliminates
Graph-to-Relational
Mapping
In your data
Bridge the gap
between logical
model and DB
models

REVIEW
S
name: “Dan”
born: May 29, 1970
twitter: “@dan”
name: “Ann”
born: Dec 5, 1975
date:
Jan 10, 2011
name: “Cars, Inc”
sector:
“automotive”
Property Graph Model Components
Nodes
• The objects in the graph
• Can have name-value
properties
• Can be labeled
KNOWS
KNOWS
FOLLOWS
REVIEW
S
User User
Relationships
• Relate nodes by type and
direction
• Can have name-value
properties Business

Relational Versus Graph Models
Relational Model Graph Model
REVIEWS
REVIEWS
REVIEWS
Alice
Burgers, Inc
Pizza, Inc
Pretzels
User BusinessUser-Business
Alice
Burgers, Inc
Pretzels
Pizza, Inc

Graphs in Spark 3.0

Tables for Labels
• In Spark Graph, PropertyGraphs are represented by
– Node Tables and Relationship Tables
• Tables are represented by DataFrames
– Require a ﬁxed schema
• Property Graphs have a Graph Type
– Node and relationship types that occur in the graph
– Node and relationship properties and their data type
Property Graph
Node Tables
Rel. Tables
Graph Type

Tables for Labels
:User:ProAccount
name: Alice
:Business
name: Burgers, Inc
:REVIEWS
id name
0 Alice
id name
1 Burgers, Inc
id source target
0 0 1
:User:ProAccount
:Business
:REVIEWS
Graph Type {
:User:ProAccount (
name: STRING
),
:Business (
name: STRING
),
:REVIEWS
}

Creating a graph
Property Graphs are created from a set of DataFrames.
There are two possible options:
- Using Wide Tables
- one DF for nodes and one for relationships
- column name convention identiﬁes label and property
columns
- Using NodeFrames and RelationshipFrames
- requires a single DataFrame per node label combination and
relationship type
- allows mapping DF columns to properties
16

Storing and Loading
business.json
user.json
review.json
Create Node and
Relationship Tables
Create Property Graph Store Property Graph
as Parquet
17

Graph Querying with
Cypher

What is Cypher?
• Declarative query language for graphs
– "SQL for graphs"
• Based on pattern matching
• Supports basic data types for properties
• Functions, aggregations, etc
20

Pattern matching
a b
1
4
3
2
5
1 2
32
4 5
Query graph: Data graph: Result:

Basic Pattern: Alice's reviews?
(:User {name:'Alice'} ) -[:REVIEWS]-> (business:Business)
REVIEWS
User
Forrest
Gump
VAR LABEL
NODE NODE
?
LABEL PROPERTY
RELATIONSHIP
Type

Cypher query structure
• Cypher operates over a graph and returns a table
• Basic structure:
MATCH pattern
WHERE predicate
RETURN/WITH expression AS alias, ...
ORDER BY expression
SKIP ... LIMIT ...

Basic Query:
Businesses Alice has reviewed?
MATCH (user:User)-[r:REVIEWS]->(b:Business)
WHERE user.name = 'Alice'
RETURN b.name, r.rating

Query Comparison: Colleagues of Tom Hanks?
SELECT co.name AS coReviewer, count(co) AS nbrOfCoReviews
FROM User AS user
JOIN UserBusiness AS ub1 ON (user.id = ub1.user_id)
JOIN UserBusiness AS ub2 ON (ub1.b_id = ub2.b_id)
JOIN User AS co ON (co.id = ub2.user_id)
WHERE user.name = "Alice"
GROUP BY co.name
MATCH
(user:User)-[:REVIEWS]->(:Business)<-[:REVIEWS]-(co:User)
WHERE user.name = 'Alice'
RETURN co.name AS coReviewer, count(*) AS nbrOfCoReviews

Variable-length patterns
MATCH (a:User)-[r:KNOWS*2..6]->(other:User)
RETURN a, other, length(r) AS length
Allows the traversal of paths of variable length
Returns all results between the minimum and maximum number of
hops
26

Aggregations
• Cypher supports a number of aggregators
– min(), max(), sum(), avg(), count(), collect(), ...
• When aggregating, non-aggregation projections form a grouping
key:
MATCH (u:User)
RETURN u.name, count(*) AS count
The above query will return the count per unique name
27

Projections
• UNWIND
UNWIND [‘a’, ‘b’, ‘c’] AS list
• WITH
– Behaves similar to RETURN
– Allows projection of values into new variables
– Controls scoping of variables
MATCH (n1)-[r1]->(m1)
WITH n1, collect(r1) AS r // r1, m1 not visible after this
RETURN n1, r
list
‘a’
‘b’
‘c’

Expressions
• Arithmetic (+, -, *, /, %)
• Logical (AND, OR, NOT)
• Comparison (<, <=, =, <>, >=, >)
• Functions
– Math functions (sin(), cos(), asin(), ceil(), ﬂoor())
– Conversion (toInteger(), toFloat(), toString())
– String functions
– Date and Time functions
– Containers (Nodes, Relationships, Lists, Maps)
– …

Cypher in Spark 3.0

In the previous session...
• Property Graph is a growing data model
• Spark Graph will bring Property Graphs
and Cypher to Spark 3.0
• Cypher is "SQL for graphs"
– Based on pattern matching
31

Now that you know Cypher...
val graph: PropertyGraph = …
graph.cypher(
"""
|MATCH (u:User)-[:REVIEWS]->(b:Business)
|WHERE u.name = 'Alice'
|RETURN u.name, b.name
""".stripMargin
).df.show
32

So, what happens in that call?
33

● Distributed executionSpark Core
Spark SQL ● Rule-based query optimization
Query Processing
MATCH (u:User)-[:REVIEWS]->(b:Business)
WHERE u.name = 'Alice'
RETURN u.name, b.name
34
openCypher Frontend
● Shared with Neo4j database system
● Parsing, Rewriting, Normalization
● Semantic Analysis (Scoping, Typing, etc.)
Okapi + Spark Cypher ● Schema and Type handling
● Query translation to DataFrame operations

Query Translation
Logical view
Physical view (DataFrame operations)
NodeTable(Business)
RelTable(REVIEWS)
NodeTable(User)
Result
35

Spark Cypher Architecture
36
● Conversion of expressions
● Typing of expressions
● Translation into Logical Operators
● Basic Logical Optimization
● Column layout computation for intermediate results
● Translation into Relational Operations
openCypher Frontend
Okapi + Spark Cypher
Spark SQL
Spark Core
Intermediate Language
Relational Planning
Logical Planning
● Translation of Relational Operations into DataFrame
transformations
● Expression Conversion to Spark SQL Columns
Spark Cypher

Query(None,
SingleQuery(List(
Match(false,
Pattern(List(
EveryPath(
RelationshipChain(
NodePattern(Some(Variable(u)),List(),None,None),
RelationshipPattern(Some(Variable(UNNAMED18)),List(RelTypeName(REVIEWS)),None,None,OUTGOING,None,false),
NodePattern(Some(Variable(b)),List(),None,None))))),List(),
Some(Where(
Ands(
Set(HasLabels(Variable(u),List(LabelName(User))),
HasLabels(Variable(b),List(LabelName(Business))),
Equals(Property(Variable(u),PropertyKeyName(name)),Parameter( AUTOSTRING0,String))))))),
With(false,ReturnItems(false,List(
AliasedReturnItem(Property(Variable(u),PropertyKeyName(name)),Variable(n.name)),
AliasedReturnItem(Property(Variable(b),PropertyKeyName(name)),Variable(b.name)))),None,None,None,None),
Return(false,ReturnItems(false,List(
AliasedReturnItem(Variable(u.name),Variable(u.name)),
AliasedReturnItem(Variable(b.name),Variable(b.name)))),None,None,None,Set()))))
37
openCypher Frontend
Okapi + Spark
Cypher
Spark SQL
Spark Core
Intermediate
Language
Relational Planning
Logical Planning
Spark Cypher

38
╙──TableResultBlock(OrderedFields(List(u.name :: STRING, b.name :: STRING)), ...)
╙──ProjectBlock(Fields(Map(u.name :: STRING -> u.name :: STRING, b.name :: STRING -> b.name :: STRING)), Set(), ...)
╙──ProjectBlock(Fields(Map(u.name :: STRING -> u.name :: STRING, b.name :: STRING -> b.name :: STRING)), Set(), ...)
╙──MatchBlock(Pattern(
Set(u :: NODE, b :: NODE, UNNAMED18 :: RELATIONSHIP(:REVIEWS)),
Map( UNNAMED18 :: RELATIONSHIP(:REVIEWS) -> DirectedRelationship(Endpoints(u :: NODE, b :: NODE))),
Set(u:User :: BOOLEAN, b:Business :: BOOLEAN, u.name :: STRING = "Alice" :: STRING)
))
╙──SourceBlock(IRCatalogGraph(session.tmp#1)
openCypher Frontend
Okapi + Spark
Cypher
Spark SQL
Spark Core
Intermediate
Language
Relational Planning
Logical Planning
Spark Cypher
• Converting AST patterns into okapi patterns
• Converting AST expressions into okapi expressions
• Typing expressions

WHERE n.name = 'Alice'
Select(List(u.name :: STRING, b.name :: STRING), ...)
╙─Project((b.name :: STRING,Some(b.name :: STRING)), ...)
╙─Project((u.name :: STRING,Some(u.name :: STRING)), ...)
╙─Filter(u.name :: STRING = $ AUTOSTRING0 :: STRING, ...)
╙─Project((u.name :: STRING,None), ...)
╙─Filter(b:Business :: BOOLEAN, ...)
╙─Filter(u:User :: BOOLEAN, ...)
╙─Expand(u :: NODE, UNNAMED18 :: RELATIONSHIP(:REVIEWS), b :: NODE, Directed, ...)
╟─NodeScan(u :: NODE, ...)
║ ╙─Start(LogicalCatalogGraph(session.tmp#1), ...)
╙─NodeScan(b :: NODE , ...)
╙─Start(LogicalCatalogGraph(session.tmp#1), ...)
Convert Intermediate Language Blocks into Logical Query Operators
39
openCypher Frontend
Okapi + Spark
Cypher
Spark SQL
Spark Core
Intermediate
Language
Relational Planning
Logical Planning
Spark Cypher

Select(List(u.name :: STRING, b.name :: STRING), ...)
╙─Project((b.name :: STRING,Some(b.name :: STRING)), ...)
╙─Project((u.name :: STRING,Some(u.name :: STRING)), ...)
╙─Filter(u.name :: STRING = $ AUTOSTRING0 :: STRING, ...)
╙─Project((u.name :: STRING,None), ...)
╙─Expand(u :: NODE, UNNAMED18 :: RELATIONSHIP(:REVIEWS), b :: NODE, Directed, ...)
╟─NodeScan(u :: NODE(:User), ...)
║ ╙─Start(LogicalCatalogGraph(session.tmp#1), ...)
╙─NodeScan(b :: NODE(:Business), ...)
╙─Start(LogicalCatalogGraph(session.tmp#1), ...)
40
openCypher Frontend
Okapi + Spark
Cypher
Spark SQL
Spark Core
Intermediate
Language
Relational Planning
Logical Planning
Spark Cypher
Apply basic optimizations to a Logical Query plan (e.g. label pushdown)

Select(u.name :: STRING, b.name :: STRING), RecordHeader with 2 entries)
╙─Alias(b.name :: STRING AS b.name :: STRING RecordHeader with 15 entries)
╙─Alias(u.name :: STRING AS u.name :: STRING, RecordHeader with 14 entries)
╙─Filter(u.name :: STRING = "Alice", RecordHeader with 13 entries)
╙─Join((target(UNNAMED18 :: RELATIONSHIP(:REVIEWS)) -> b :: NODE)), RecordHeader with 13 entries, InnerJoin)
╟─Join((u :: NODE -> source(UNNAMED18 :: RELATIONSHIP(:REVIEWS))), RecordHeader with 9 entries, InnerJoin)
║ ╟─NodeScan(u :: NODE(:User), RecordHeader 4 entries)
║ ║ ╙─Start(Some(CAPSRecords.unit), session.tmp#1)
║ ╙─RelationshipScan(UNNAMED18 :: RELATIONSHIP(:REVIEWS), RecordHeader with 5 entries)
║ ╙─Start(None)
╙─NodeScan(b :: NODE(:Business), RecordHeader with 4 entries)
╙─Start(Some(CAPSRecords.unit))
Translation of graph operations into relational operations
41
openCypher Frontend
Okapi + Spark
Cypher
Spark SQL
Spark Core
Intermediate
Language
Logical Planning
Relational Planning
Spark Cypher

● Describes the output table of a relational operator
● Maps query expressions (e.g. ‘n.name’ or ‘n:User’) to DataFrame / Table columns
● Used to access columns when evaluating expression during physical execution
● Supports relational operations to reﬂect data changes (e.g.
header.join(otherHeader))
42

Expression Column Name
NodeVar(n) :: CTNODE(:User) n
HasLabel(n, :User) :: CTBOOLEAN ____n:User
Property(n.name) :: CTSTRING ____n_dot_nameSTRING
43
Select(b, name)
╙─Project(b.name as name)
╙─Project(n as b)
╙─NodeScan(n:User)

NodeVar(b) :: CTNODE(:User) n
HasLabel(b, :User) :: CTBOOLEAN ____n:User
Property(b.name) :: CTSTRING ____n_dot_nameSTRING
44
Select(b, name)

SimpleVar(name) :: CTSTRING ____n_dot_nameSTRING
45
Select(b, name)

SimpleVar(name) :: CTSTRING ____n_dot_nameSTRING
46
Select(b, name)

Abstracts relational operators over the relational backends table
Converts OKAPI expressions into backend specific expressions
trait Table[E] {
def header: RecordHeader
def select(expr: String, exprs: String*): Table[E]
def filter(expr: Expr): Table[E]
def distinct: Table[E]
def order(by: SortItem[Expr]*): Table[E]
def group(by: Set[Expr], aggregations: Set[Aggregation]): Table[E]
def join(other: Table[E], joinExprs: Set[(String, String)], joinType: JoinType): Table[E]
def unionAll(other: Table[E]): Table[E]
def add(expr: Expr): Table[E]
def addInto(expr: Expr, into: String): Table[E]
def drop(columns: String*): Table[E]
def rename(oldColumn: Expr, newColumn: String): Table[E]
}
47
openCypher Frontend
Okapi + Spark
Cypher
Spark SQL
Spark Core
Intermediate
Language
Relational Planning
Spark Cypher
Logical Planning

class DataFrameTable(df: DataFrame) extends RelationalTable[DataFrameTable] {
// ...
override def filter(expr: Expr): DataFrameTable = {
new DataFrameTable(df.filter(convertExpression(expr, header)))
}
override def join(other: DataFrameTable, joinExprs: Set[(String, String)], joinType: JoinType): DataFrameTable = {
val joinExpr = joinExprs.map { case (l,r) => df.col(l) === other.df(r) }.reduce(_ && _)
new DataFrameTable(df.join(other.df, joinExpr, joinType))
}
// ...
}
openCypher Frontend
Okapi + Spark
Cypher
Spark SQL
Spark Core
Intermediate
Language
Relational Planning
Logical Planning
Spark Cypher
48

Future improvement ideas
• Common table expressions
• Worst-case optimal joins
• Graph-aware optimisations in Catalyst
• Graph-aware data partitioning
49

Spark Cypher in Action

Extending Spark
Graph with Neo4j
Morpheus

Neo4j Morpheus
• Incubator for SparkCypher
• Extends Cypher language with
multiple-graph features
• Graph catalog
• Property graph data sources for
integration with Neo4j, SQL DBMS, etc.
https://github.com/opencypher/morpheus
52

Get Involved!
• SPIP was accepted in February
• Current status:
– Core development is poc-complete
– PRs in review
• We are not Spark committers
– Help us review / merge
– Contribute to documentation, Python API
53

DON’T FORGET TO RATE
AND REVIEW THE SESSIONS
SEARCH SPARK + AI SUMMIT

Graph Features in Spark 3.0: Integrating Graph Querying and Algorithms in Spark Graph

More Related Content

What's hot (20)

Similar to Graph Features in Spark 3.0: Integrating Graph Querying and Algorithms in Spark Graph (20)

More from Databricks (20)

Recently uploaded (20)

Graph Features in Spark 3.0: Integrating Graph Querying and Algorithms in Spark Graph