Industrializing Machine Learning on an Enterprise Azure Platform

Sodexo as the world leader of services, we believe that quality of life is created when we integrate our Food services, Facilities management, employee benefits and more… Our ambition is to positively impact one billion consumers worldwide. Sodexo has launched a POC for food services that is satisfying our business, but how do we come from a local pure Python Jupyter notebook to a product that can be scalable to cover all our consumers? To do so we have not only used DataBricks but we have made it interacting with many Azure services: Data Factory, Log Analytics, DevOps, CosmosDB, Kubernetes, Data Lake and Key Vault. This talk deal about what are those services, how we use Databricks, how it’s interacting with those services, and what are our feedbacks and experiences.

展开查看详情

1.WIFI SSID:Spark+AISummit | Password: UnifiedDataAnalytics

2.Industrializing ML on Azure with DataBricks RADJI Yannick, SODEXO #UnifiedDataAnalytics #SparkAISummit

3.Who am I? Using: • Spark since 2015 • DataBricks since 2018 Work for: • Industry • Consulting • Services Data Data Data Scientist Engineer Architect 3

4.Agenda • Sodexo • Food service use case • What was there? • What are we using know? • Our development practices 4

5.Sodexo Key figures €20.4bn World leader in Quality of Life Services 460,000 employees in consolidated revenues 72 100 #1 million France-based private employer countries consumers worldwide* served daily *2018 Fortune 500 ranking 5

6.Sodexo We are the global leader in quality of life services > Corporate Services ▪ BUSINESS > Energy & Resources & ADMINISTRATIONS > Government & Agencies > Sports & Leisure On-site Services > Health Care ▪ HEALTH CARE & SENIORS > Seniors > Universities ▪ EDUCATION > Schools Benefits & Rewards ▪ EMPLOYEE EXPERIENCE Services ▪ MOBILITY AND EXPENSE ▪ CONCIERGE SERVICES Personal & Home Services ▪ HOME CARE ▪ CHILD CARE 6

7.Sodexo Business & Administrations REVENUES BY CLIENT SUB-SEGMENT KEY FIGURES 56% 10,938 276,573 of Group million euro employees revenues in revenues 7

8.Food service use case 8

9.What was there? • Dozens of Jupyter’s notebooks running on Azure HDInsight Apache Spark on Hive Query on HDInsight HDInsight Azure Blob Azure Data Storage Lake Store 9

10.What was there? • Unversionned, no CI/CD • Hard to maintain & operate with high costs • No orchestration • Not versatile • Not secure • Poor performances • No live Data • No re-usable design • Library conflict 10

11. Current (macro) Architecture Azure Container Registry Batch data Azure Data Azure Azure Lake DataBricks Kubernetes Site Store Prep & Train Serve managers Azure Azure Azure Azure Data DevOps Monitor KeyVault Factory 11

12.Azure DataBricks Why migrate? • Same tools used by data scientist exploration on notebooks & industrialisation by data engineer • Managed Spark cluster • Large compatibility of Spark’s versions • Centralize all the ETL in Spark whereas split it on several technologies 12

13.Azure DataBricks Feedbacks: • Cluster’s autoScaling gives fine results • Pools helps to speed up jobs 13

14.Azure DataFactory Why we use it? • Managed service • Native integrated to all azure services and on premise DB • Easy orchestrate & scheduling • No limitations on data volume or on the number of files 14

15.Azure DataFactory Lesson learned? • Call to DataBricks via web activity or DataBricks activity • Cost saving by using a DataBricks cluster job and argparse compatible • You can use 1 template for all jobs 15

16.Azure DataFactory 16

17.Azure Monitor Why we use it? • Monitor, analyze daily run and send mails alerts • Same tool to get metrics and logs from all the packages of our software and services of Azure platform (including Databricks) 17

18.Azure Monitor source: docs.azuredatabricks.net 18

19.Azure Monitor source: docs.azuredatabricks.net 19

20.Azure Monitor Lesson learned? • It uses KQL language with specificities so there is a little learning curve is to plan • You can do your own library but rather use SDK for python, java, JS, C# & .NET to send app logs 20

21.Azure Data Lake Why we use it? • Used to store the sources CSV files and output parquet files of DataBricks jobs Lesson learned? • To easily use it, mount the Data Lake with dbutils • Data Lake Storage Gen2 is converging of Azure Blob storage and Azure Data Lake Storage Gen1 – not available in all regions • Large number of files, propagating the permissions can take long 21

22.Azure Kubernetes Why we use it? • To host our dashboard, REST API & DataBase Lesson learned? • The managed service save you some DevOps time to set up the cluster • Configure a secure network can be challenging 22

23.Azure DevOps • Azure Repos • Azure Pipelines source: docs.microsoft.com 23

24.Azure DevOps • Azure Boards • Azure Test Plans • Azure Artifacts source: docs.microsoft.com 24

25.Azure DevOps Retro, Sprint Refinment planning Commits Review, link to UAT tasks Unit & Integration DEV tests are passing Pull Request 25

26.• Our development practices Use naming conventions • Validate inputs config or parameters • New code should come with new tests • Documents • Avoid repetitive code • Use a logger do not print • Take the time to think about the right data structure to use (a dictionary? Tuple? Dataframe? Array?) • Break down responsibilities between the classes • Code should be easy to explain to a third person (avoid complex code) • Developers should used an unify development environment • Use dockerized environments to make simulation of real interactions • Keep development, UAT, and production as similar as possible • One codebase protected in revision control, many releases • Strictly separate build and release stages 26

27.Summary • To industrialiaze you need to design an architecture that is : – Versatile – Easy to maintain – Operate with low costs • Azure provides managed services to: – Orchestrate – Store – Set up all your DevOps – Host • Finally you need a good team with development practices 27

28.DON’T FORGET TO RATE AND REVIEW THE SESSIONS INTERNAL POSITIONS OPEN! Please send me any suggestion or your resume to: yannick.radji@sodexo.com SEARCH SPARK + AI SUMMIT

29. Q&A #UnifiedDataAnalytics #SparkAISummit