1. Building 1 million predictions per second using SQL-R
11/30/2017 1:54 AM
Building 1 million predictions per second using SQL-R
Amit Banerjee
Sr. PM Manager, Data Group (TIGER)
© 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

2. Agenda Data Science Process Bringing Analytics to Data
Demo Using Lending Club Data
Optimization Tips

3. Data Science Process – CRISP-DM
Data scientist is able to perform all these steps in SQL Server 2016
CRISP-DM Model– General data science framework

4. In-Database Analytics at Scale
Jack Henry A leading provider for banking solutions for credit unions across Americas
Business User
20M Vehicle Loans
Age, Original Balance, Interest Rate, Loan Remaining Months, Credit Score R
Store
Predictions
Visualize
ColumnStore
In-Database Analytics at Scale
In-memory OLTP
PowerBI Dashboard
Prepare for analytics
Use Case: Predict vehicle loan charge off (default) based on attributes like interest rate, credit scores etc
Input: A subset of 20 million row of vehicle loan data in SQL Server - columns including branch location, customer profiles, interest rate, loan age etc..
Expected Result: Probability score of loans get charged off (Higher the score, higher the probability of loan get charged off)
Build PowerBI report using probability score to show healthiness of vehicle loans across different branches
Build what if scenario in business application

5. Using SQL Server R Services
11/30/2017 1:54 AM
Using SQL Server R Services
Bringing Analytics to the Data
Data already in SQL
Use T-SQL know-hows to do ETL
Use the power of in-memory OLTP and column store indexing to enhance speed of ETL
RevoScaleR package to provide parallelism and scale
Making the data travel
Data sources not in SQL
Data sinks not in SQL
Complex ETL needed
Long running R script
© 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

6. sp_execute_external EXEC sp_execute_external_script
@language = N'language' , @script = N'script',
@input_data_1 = ] 'input_data_1' [ = ] N'input_data_1_name' ] [ = 'output_data_1_name' ] [ = 0 | 1 ] [ = ]
data_type [ OUT | OUTPUT ] [ ,...n ]' [ = ] 'value1' [ OUT | OUTPUT ] [ ,...n ] [ WITH <execute_option> ] [;]
<execute_option>::= { { RESULT SETS UNDEFINED } | { RESULT SETS NONE } | { RESULT SETS ( <result_sets_definition> ) } }
<result_sets_definition> ::= { ( { column_name data_type [ COLLATE collation_name ] [ NULL | NOT NULL ] } [,...n ] ) | AS OBJECT [ db_name . [ schema_name ] . | schema_name . ] {table_name | view_name | table_valued_function_name } | AS TYPE [ schema_name.]table_type_name }
EXEC sp_execute_external_script
@language = N'R' = N'iris_data <- iris;' = N'' = N'iris_data' WITH RESULT SETS (("Sepal.Length" float not null, "Sepal.Width" float not null, "Petal.Length" float not null, "Petal.Width" float not null, "Species" varchar(100))); END; go

7. Building a 1 million predictions per second
Fast Models
Correct attribute selection
Fast Ingestion
Fast Reads
Uniform resource usage
Use R Studio to design your models or you could even use T-SQL to design your models and test

8. Building a 1 million predictions per second
Fast Models
Correct attribute selection
Fast Ingestion
Fast Reads
Uniform resource usage
Important to select the right attributes to ensure that the correct attributes are selected

9. Building a 1 million predictions per second
Fast Models
Correct attribute selection
Fast Ingestion
Fast Reads
Uniform resource usage
In-Memory OLTP can help here. Staging tables can be schema-only.

10. Building a 1 million predictions per second
Fast Models
Correct attribute selection
Fast Ingestion
Fast Reads
Uniform resource usage
Most of these models will read large datasets. The I/O can be reduced which tends to be slowest part using Columnstore indexes

11. Building a 1 million predictions per second
Fast Models
Correct attribute selection
Fast Ingestion
Fast Reads
Uniform resource usage
You want all the CPUs to be used in parallel. This can be achieved with Resource Governor group assignment for the available CPUs/NUMA nodes. Pick the right CPU optimized machines

12. SQL Server as Scoring Engine
Deployment Using:
Triggers
Powershell scripts
SQL agent jobs

13. DEMO Using public dataset of Lending Club
Using G5 instance of Azure Data Science VM (DSVM)
Following Data Science Process using SQL Server 2016 R Services

14. References Loan Classification using SQL Server 2016 R Services
A walkthrough of Loan Classification using SQL Server 2016 R Services
Using MicrosoftML in SQL-Server
GitHub SQL Server Samples

15. Microsoft Data Amp WHERE DATA GETS TO WORK
Put data, analytics and artificial intelligence into the heart of your solutions. Get the latest on big data and machine learning innovations.
Join us online April 19, 2017 at 8AM PT
microsoft.com/data-amp

16. Questions?