"Full Stack" Data Science with R

Talk at EARL 2017, San Francisco 2017-06-06

- https://earlconf.com/2017/sanfrancisco/

Transcript

1. "Full Stack" Data Science with R
   Production data science & engineering
   with open source tools
   Gabriela de Queiroz
   Ajay Gopal
   

   View Slide

2. Agenda
   1) Data (Science) Stack Needs
   2) Using R - how, why & where?
   3) Cloud R Stack
   - Example Cloud Infra
   - Choice of Tools
   4) Production Mindset
   - BetteR habits
   - Workflow
   5) Buy or Build?
   This talk
   Born on Twitter
   2
   

   View Slide

3. Gabriela de Queiroz
   Lead Data Scientist
   / : gdequeiroz
   First met R: 2007
   Ajay Gopal, PhD
   Chief Data Scientist
   : ajzz : @aj2z
   First met R: 2011
   About us
   3
   

   View Slide

4. What’s SelfScore?
   4
   Industry: FinTech Startup, Menlo Park, CA
   What we do: use ML models with alternative financial signals to help deserving but underserved
   populations gain access to fair credit, starting with international students (2 products in market).
   Differentiator: Measure borrower’s credit potential instead of their credit history (SSN / FICO etc)
   Team: ~25 & growing (hiring local Sr R Data Engineer, Summer Intern)
   Funding: Series B
   Founded: 2013
   

   View Slide

5. The “Full Stack” Analogy
   5
   Front End
   Back End
   Data Store
   Devops
   APIs
   UX
   Technology
   Puppet, Chef, Ansible, AWS EC2,
   Docker, ECS/GCE, Heroku
   MySQL, PostGres, MongoDB, Redis,
   MemCached etc.
   PHP, JS, Python, Ruby, ORMs, CI, Git
   Restify, Django, Rails, ASP.net, Lambda
   HTML/CSS, JS (Node, React), Bootstrap,
   iOS, Android, Ionic, Cordoba
   Email (SendGrid), SMS (Twilio), Push
   (SNS, Firebase), Msg Frmwks
   Function
   Multi-Channel Engagement
   Optimal Service Delivery
   Platform-agnostic function &
   information availability
   Business Logic
   Identities, Attribs, Relations
   Scaleable Services &
   Contingencies
   Goal: Scalable, Engaging, Valuable Service
   

   View Slide

6. 6
   Front End
   Back End
   Data Store
   Devops
   APIs
   UX
   Technology
   rocker, EMIs, ECS, GCE, other cloud
   tools
   DBI, RMySQL, RPostGreSQL, Redis,
   Hadoop, Kinesis (AWR), Spark etc.
   Your internal pkgs, RServer, CI, Git,
   Chron, (most R packages), sparkR
   shiny, HTML, CSV, rook, googlesheets,
   HtmlWidgets, shinyapps.io, Dropbox
   httr, curl - API interactions for Email,
   SMS, Push, Slack, OR via CI tool
   Function
   Multi-Channel Engagement
   Optimal Service Delivery
   Platform-agnostic function &
   information availability
   Business Logic
   Identities, Attribs, Relations
   Scaleable Services &
   Contingencies
   “Full Stack” Data Science with R
   Generic: rapache, opencpu, plumber
   ML: h2o, Domino Data Lab
   Goal: Scalable, Timely, Economic Decisions
   

   View Slide

7. Detractors
   - Very few hard-core devs
   - Only two major dev
   shops, but no serious
   bandwidth for hire
   - Memory mgmt (still?)
   Why R? 2011 vs Now
   Drivers
   1. Large ecosystem of
   packages
   2. Most other things are
   available as HTTP APIs
   3. Fantastic IDE (RStudio)
   1-pt access to stack
   4. Great recruiting tool
   7
   Good news: it works!
   

   View Slide

8. Minimum Functions Infra Should Support
   1) Retrieve Data
   - Ad / Marketing
   - Sales
   - Transaction
   - 3rd Party / Behavioral
   2) Process (ETL)
   - Fetch, clean up, store
   3) Analyze
   - Cross-Connectivity
   - Aggregation
   - Algorithms
   4) Predict
   - Models in batch
   - REST APIs
   5) Inform
   - Customers (Services & API)
   - Partners
   Eg: Marketing, fulfillment
   - Internal Stakeholders
   Eg: Reporting / Dashboards
   8
   

   View Slide

9. Infra Resource Roles
   1) Bastion (to connect to external world)
   (small, low memory, public IP)
   2) Scheduler (do things triggered by time & events)
   (medium, run CI tools, invoke compute slaves)
   3) Workers (heavy computations)
   (highmem, multi core, stateless)
   4) Storage (data source & sink)
   (external services or internal clusters)
   5) Modeler (H2O Cluster, or similar)
   (cluster, available on demand)
   6) Reporting (scalable web / Shiny server)
   (medium, autoscaled containers)
   9
   

   View Slide

10. Sample AWS Infra
    10
    

    View Slide

11. Choice of Tools
    11
    

    View Slide

12. “Staging” Shiny App
    1. Git Commit to “Dev” branch
    2. Jenkins Sync Repo on Commit
    3. Sync triggers next Jenkins job
    creates Docker container
    4. Next job: AWS cli tools deploy
    Docker container to ECS
    5. “Dev” Shiny app live on staging
    6. API call to notify Slack channel
    Sample Workflows
    Marketing Cost Monitor
    1. Jenkins Rscript fetches
    today’s Adwords spend &
    internal sales data every 5
    minutes.
    2. Rscript runs anomaly
    detection & threshold checks
    3. When check fails, API calls
    from R to alert via Slack and
    Email (eg: SendGrid).
    12
    

    View Slide

13. Full-Stack Data Science
    People
    - Data / Backend Engineer
    - Data Scientist
    - Modeller / Statistician
    - Product Manager
    - Devops Engineer
    Output
    - EDA / ad-hoc
    - Scheduled Reporting
    - Batch Predictions
    - Stream Processing
    - Real-Time Prediction APIs
    Our “product” is scalable, actionable intelligence
    13
    … let’s adopt good software development practices
    

    View Slide

14. BetteR habits:
    1. Write inline and offline tests for your code
    2. Generate informational logs so you can debug later (futile.logger)
    3. Add versioning (github)
    4. Save business logic as functions in package (selfscoRe)
    5. Add examples (Rmd)
    6. Write documentation (Rmd)
    7. Create a web service (Shiny apps)
    8. Put the service in a docker container
    The Production Mindset
    14
    

    View Slide

15. Create and Prioritize Issues/Tickets
    Write Code
    Test Code in Staging
    Review the Code
    Continuous Integration
    Workflow
    15
    

    View Slide

16. Should we buy or should we build?
    VS
    Should my company buy the infra? Should my team build it?
    16
    

    View Slide

17. Buy vs Build Considerations
    BUY
    - If no dev/tech in-house
    - If time-to-market is key
    requires:
    - Custom Development
    - Higher Cost Tolerance
    - Niche engagements
    BUILD
    - If compliance is major factor
    (HIPAA, PCI)
    - If cost control is key
    - Full Control of Features Reqd
    requires:
    - In-house devops competence
    - Longer time-to-market
    - Ongoing maintenance
    17
    

    View Slide

18. “Full Stack” Data Engineer
    More info: bit.ly/LeadDataEngineer
    WE ARE HIRING!
    18
    Img Credits: http://daemon.co.za/2014/04/what-does-full-stack-mean
    

    View Slide