The Road to U-SQL: Experiences in Language Design (SQL Konferenz 2017 Keynote)

First Language: APL
 Example (arbitrary dim regression): Const⌹Coeff∘.*⌽0,Dim
 High-Dimensional Arrays/Nested Arrays as Data Model
 Expressions: Mathematics (declarative), parallel
 Control flow: Recursion, GoTo 
 Syntax: Greek and Special Characters (easier to write than read)
Next Languages: Pascal, Modula-2, Oberon, C/C++
 Procedural, imperative (structured control flow)
 One Item: Structured types and Object Models
 Single-node, Parallelism/Distributed via Libraries
Other experiences: Lisp, Prolog
 Functional and Logic
 List as Data Model
 Recursion instead of control flow
Data Processing Languages:
 SQL: Declarative Expressions, Procedural Control Flow
 DataLog: Recursion
 XQuery: Tree Data Model, Declarative/Functional
My Language
History
Imperative vs Declarative
Procedural vs Functional vs Logical
One item vs Sets
Single-node vs Parallel vs Distributed
Programming vs Data Processing

Imperative
Tell the system how to get it
Declarative
Tell the system what you want
Let the system find a way to get it
Declarative leaves options that an optimizer can reason about
Imperative vs
Declarative

Procedural
operates by changing “persistent state” (variables) with each
expression
Allows side-effects in control flow
i = 0;
FOR j FROM 1 TO 100 DO i = i + 1;
i is now 100.
Functional
Transforms input into output, no “persistent state”
No side-effects in control flow
i = 0;
FOR j FROM 1 TO 100 DO i = i + 1;
i is now 1.
Procedural vs
Functional

Single objects
 Requires control flow
 Explicit Parallelism
Sets of objects
 Allows higher-level abstraction expressions
 Implicit Parallelism
 Objects can be
 Object/value
 Tuples
 Trees
 graphs
Meta Data Service provides sharing of data model
objects
One Item vs
Sets
Data Models

Language Parallelism vs Libraries
Scale-up vs Scale-out
Synchronization/Transactions
 Explicit imperative vs Implicit declarative
 ACID support
Single-node vs
Parallel vs
Distributed

Programming Languages
 Long-term data is in a store but not part of the language
model
 Designed for tight coupling of data and application logic
 Often imperative, procedural, one-item object-oriented,
explicit/library based parallelism
Data Processing Languages
 Long-term data is part of the language model
 Data can evolve independently of application
 Declarative and Functional
 Set-based
 Built-in parallelism and implicit/declarative synchronization
Programming
vs
Data
Processing

Writeability vs Readability
Consistency
Familiarity
Context independent
Composable
Mathematical vs natural language
Reserved Keywords?
Syntax Matters

No surprises!
Avoid complexities!
Composable
Optimizable
Implementable
Semantics
Matters

Consortiums/Standard Bodies:
 Slow (it took XQuery 6 years!)
 “Political” interests of participants can
negatively impact design
Individual/Small team:
 More focused
 Risk: different for difference sake
Evolve vs New
Create Language that address demand
How to create
languages

Some sample use cases
Digital Crime Unit – Analyze complex attack patterns
to understand BotNets and to predict and mitigate
future attacks by analyzing log records with
complex custom algorithms
Image Processing – Large-scale image feature
extraction and classification using custom code
Shopping Recommendation – Complex pattern
analysis and prediction over shopping records
using proprietary algorithms

 Declarativity does scaling and
parallelization for you
 Extensibility is bolted on and
not “native”
 hard to work with anything other than
structured data
 difficult to extend with custom code

 Extensibility through custom code
is “native”
 Declarativity is bolted on and
not “native”
 User often has to
care about scale and performance
 SQL is 2nd class within string
 Often no code reuse/
sharing across queries

 Declarativity and Extensibility are
equally native to the language!
Get benefits of both!
Makes it easy for you by unifying:
• Unstructured and structured data processing
• Declarative SQL and custom imperative Code
(C#, Python, R, …)
Scales-up and Scales-out custom code within
declarative framework

The origins
of U-SQL
SCOPE – Microsoft’s internal
Big Data language
• SQL and C# integration model
• Optimization and Scaling model
• Runs 100’000s of jobs daily
Hive
• Complex data types (Maps, Arrays)
• Data format alignment for text files
T-SQL/ANSI SQL
• Many of the SQL capabilities (windowing functions, meta
data model etc.)

U-SQL Language Philosophy
Declarative Query and Transformation Language:
• Uses SQL’s SELECT FROM WHERE with GROUP
BY/Aggregation, Joins, SQL Analytics functions
• Optimizable, Scalable
Expression-flow programming style:
• Easy to use functional lambda composition
• Composable, globally optimizable
Operates on Unstructured & Structured Data
• Schema on read over files
• Relational metadata objects (e.g. database, table)
Extensible from ground up:
• Type system is based on C#
• Expression language IS C#
• User-defined functions (U-SQL and C#)
• User-defined Aggregators (C#)
• User-defined Operators (UDO) (C#)
U-SQL provides the Parallelization and Scale-out
Framework for Usercode
• EXTRACTOR, OUTPUTTER, PROCESSOR, REDUCER,
COMBINER, APPLIER
REFERENCE MyDB.MyAssembly;
CREATE TABLE T( cid int, first_order DateTime
, last_order DateTime, order_count int
, order_amount float, ... );
@o = EXTRACT oid int, cid int, odate DateTime, amount float
FROM "/input/orders.txt"
USING Extractors.Csv();
@c = EXTRACT cid int, name string, city string
FROM "/input/customers.txt"
USING Extractors.Csv();
@j = SELECT c.cid, MIN(o.odate) AS firstorder
, MAX(o.date) AS lastorder, COUNT(o.oid) AS ordercnt
, AGG<MyAgg.MySum>(c.amount) AS totalamount
FROM @c AS c LEFT OUTER JOIN @o AS o ON c.cid == o.cid
WHERE c.city.StartsWith("New")
&& MyNamespace.MyFunction(o.odate) > 10
GROUP BY c.cid;
OUTPUT @j TO "/output/result.txt"
USING new MyData.Write();
INSERT INTO T SELECT * FROM @j;

U-SQL Data Model
Files and Tables
Set-based

Unstructured Data
@s = EXTRACT a string, b int, date DateTime, file string
FROM "filepath/{date:yyyy}/{date:MM}/{date:dd}/{file}.csv"
USING Extractors.Csv(encoding: Encoding.Unicode);
• Pro: Flexible, scaling with file sets and over parts of partitionable files
• Cons: System doesn’t know data distribution, statistics; no indexing
Structured Data
CREATE TABLE T (col1 int
, col2 string
, col3 SQL.MAP<string,string>
, INDEX idx CLUSTERED (col2 ASC)
PARTITIONED BY (col1)
DISTRIBUTED BY HASH (driver_id) );
• Pro: Provides system guarantees about data distributions, statistics, indices to
help performance and scale; object discoverability
• Cons: Needs Schema a priori, Cost of additional storage and generation

U-SQL Familiarity
SQL
C#
Python, R

Familiar Operations
• ORDER BY FETCH n ROWS
• GROUP BY HAVING
• UNION/INTERSECT/EXCEPT
• OVER Expression: Windowing, Analytics, Ranking Functions
• JOINS: INNER, FULL/LEFT/RIGHT OUTER, CROSS, SEMI, ANTI-SEMI-JOIN
• CROSS APPLY
• PIVOT/UNPIVOT (new!)
New Operations
• SET OPERATION BY NAME
SELECT * FROM @left
INTERSECT BY NAME ON (id, *)
SELECT * FROM @right;
• OUTER UNION BY NAME
SELECT * FROM @left
OUTER UNION BY NAME ON (A, K)
SELECT * FROM @right;
• Flexible Column Sets for parameter polymorphism

“Top 5”s
Surprises for
SQL Users
• AS is not as
• C# keywords and SQL keywords overlap
• Future Proofing against new reserved keywords in
both languages: Reserve all upper-case words as
U-SQL keywords
• " vs ' vs []
• = != ==
• Remember: C# expression language
• null IS NOT NULL
• C# nulls are two-valued
• PROCEDURES but no WHILE
• No UPDATE, DELETE, nor
MERGE (yet)

U-SQL Object Model
Reusability and
Discoverability

ADLA Account/Catalog
Database
Schema
[1,n]
[1,n]
[0,n]
tables views TVFs
C# Fns C# UDAgg
Clustered
Index
partitions
C#
Assemblies
C# Extractors
Data
Source
C# Reducers
C# Processors
C# Combiners
C# Outputters
Ext. tables
User
objects
Refers toContains Implemented
and named by
Procedures
Creden-
tials
MD
Name
C# Name
C# Applier
Table Types
Legend
Statistics
C# UDTs

U-SQL Extensibility
 Start Time - End Time - User Name
 5:00 AM - 6:00 AM - ABC
 5:00 AM - 6:00 AM - XYZ
 8:00 AM - 9:00 AM - ABC
 8:00 AM - 10:00 AM - ABC
 10:00 AM - 2:00 PM - ABC
 7:00 AM - 11:00 AM - ABC
 9:00 AM - 11:00 AM - ABC
 11:00 AM - 11:30 AM - ABC
 11:40 PM - 11:59 PM - FOO
Start Time - End Time - User Name
5:00 AM - 6:00 AM - ABC
5:00 AM - 6:00 AM - XYZ
7:00 AM - 2:00 PM - ABC
11:40 PM - 0:40 AM - FOO

U-SQL extensibility
Extend U-SQL with C#/.NET
Built-in operators,
function, aggregates
C# expressions (in SELECT expressions)
User-defined aggregates (UDAGGs)
User-defined functions (UDFs)
User-defined operators (UDOs)

 User-Defined Extractors
 User-Defined Outputters
 User-Defined Processors
 Take one row and produce one row
 Pass-through versus transforming
 User-Defined Appliers
 Take one row and produce 0 to n rows
 Used with OUTER/CROSS APPLY
 User-Defined Combiners
 Combines rowsets (like a user-defined join)
 User-Defined Reducers
 Take n rows and produce m rows (normally m<n)
 Scaled out with explicit U-SQL Syntax that takes a UDO
instance (created as part of the execution):
 EXTRACT
 OUTPUT
What are
UDOs?
Custom Operator Extensions
Scaled out by U-SQL
• PROCESS
• COMBINE
• REDUCE
• CROSS APPLY

 .Net API provided to build UDOs
 Any .Net language usable
 however only C# is first-class in tooling
 Use U-SQL specific .Net DLLs
 Deploying UDOs
 Compile DLL
 Upload DLL to ADLS
 register with U-SQL script
 VisualStudio provides tool support
 UDOs can
 Invoke managed code
 Invoke native code deployed with UDO assemblies
 Invoke other language runtimes (e.g., Python, R)
 be scaled out by U-SQL execution framework
 UDOs cannot
 Communicate between different UDO invocations
 Call Webservices/Reach outside the vertex boundary
How to specify
UDOs?
Provide integration into U-SQL Data
Models (Files and Rowsets)
Integrates into Processing model and
optimization model

U-SQL Scalability and
Performance
Script level optimization
Scales as Data scales

 Automatic "in-lining"
optimized out-of-
the-box
 Per job
parallelization
visibility into execution
 Heatmap to identify
bottlenecks

U-SQL’s designed for Big Data
Analytics
Functional, Declarative, Set-based =>
Scalability and Optimizable
Provides Extensibility with known
Programming Languages
Familiarity: Evolution and Re-use
Summary

The Road to U-SQL: Experiences in Language Design (SQL Konferenz 2017 Keynote)

More Related Content

What's hot (20)

Similar to The Road to U-SQL: Experiences in Language Design (SQL Konferenz 2017 Keynote) (20)

More from Michael Rys (11)

Recently uploaded (20)

The Road to U-SQL: Experiences in Language Design (SQL Konferenz 2017 Keynote)

Editor's Notes