eRum 2020: Reflections on R in Production

R in Production - Reflections on two years of solutions engineering at RStudio

Transcript

1. R in Production Reflections on two years of solutions engineering at RStudio
2. What is Solutions Engineering? Industrial Research Business Management Human Resources Government Work Regulated Environments Big Data Applications Cloud Infrastructure R in Production What is there to learn? What are the needs? What are the problems? Solutions Engineers!
3. Stages of R in Production Build a Sandbox (Proof of Concept) Invest in Learning Develop Best Practices Extend your Domain Integrate and Interoperate
4. “R Admin” - Analytic Administrator Role A data scientist who: Onboards new tools, deploys solutions, supports existing standards Works closely with IT to maintain, upgrade and scale analytic environments Influences others in the organization to be more effective Passionate about making R a legitimate analytic standard within the organization Check out the RViews Blog: Analytics Administration for R by Nathan Stephens
5. Times when I’ve felt entirely alone... A data scientist who: Onboards new tools, deploys solutions, supports existing standards Works closely with IT to maintain, upgrade and scale analytic environments Influences others in the organization to be more effective Passionate about making R a legitimate analytic standard within the organization Works at a giant tech company and Is very concerned with the improvement of daily work and analytic infrastructure
6. Times when I’ve felt entirely alone... A data scientist who: Onboards new tools, deploys solutions, supports existing standards Works closely with IT to maintain, upgrade and scale analytic environments Influences others in the organization to be more effective Passionate about making R a legitimate analytic standard within the organization Works at a tiny startup and Is very concerned with the improvement of daily work and analytic infrastructure
7. RStudio Solutions Engineering is where I found my community
8. Phoenix is the most important project in the company. They’ve spent $20M over three years. And yet, here she is, trying to help, and they won’t spend $5k on more disk space. And now she won’t get a Dev environment for five months! She buries her head in her hands and silently screams down at her keyboard. ...None of the meetings on her calendar seem interesting anymore. It’s just people complaining about waiting. Waiting for something. Waiting for someone. Everyone is just waiting. And she wants no part of it right now.
9. Trauma noun 1. an interpretation of an experience tied to a severely painful emotion You need a team of empathetic witnesses. You need people to encourage you to keep going - to encourage your work when others don’t understand. - Benjamin Hardy, PhD
10. The Five Ideals from The Unicorn Project by Gene Kim 1 - Locality & Simplicity 2 - Focus, Flow & Joy 3 - Improvement of Daily Work 4 - Psychological Safety 5 - Customer Focus
11. Ideal 1 - Locality & Simplicity
12. Classic DevOps Silo Diagram Dev Silo IT/Ops Silo Focus on THE FEAR “Hey - could you just put this thing in production real quick?” “Uh.. I just deployed this little change, and something might be broken”
13. Can Data Scientists... Independently develop, test, and deploy value to customers? - Should data scientists be trusted with this responsibility? - Who are the customers in this situation? - What does deploying value entail? The First Ideal: To what degree do teams have the capabilities and the authority to get what they need done? - Gene Kim
14. Challenges for the R User Organizational • Legitimizing R • Working with IT Technical • Experience • Education • Exposure Credibility Crisis Management Plan
15. Independently develop, test, and deploy value to customers The “Hour-Long-Talk” of Data Products - Rambling, Cluttered - Parts that work well - Parts that work not-so well Local Development EDA, Prototyping, Iteration The “Lightning-Talk” of Data Products - Targeted - Elegant - Streamlined - Optimized Production Development
16. Joe Cheng “Shiny in Production” RStudio Conf 2019 Keynote
17. Turn a Prototype into a Production Application Performance Workflow 1. Use shinyloadtest to see if app is fast enough 2. If not, use profvis to see what’s making it slow 3. Optimize a. Move work out of shiny (very often) b. Make code faster (very often) c. Use caching (sometimes) d. Use async (occasionally) 4. Repeat!
18. Start by answering some questions… - What is a Shiny Application? - Who is the audience? - What is your service level agreement definition? (SLA) - What does your analytic architecture look like today? - What are your goals for evolving this architecture? - How will monitoring be handled? - Who is responsible for maintenance? Make work visible, Define shared goals, Build a checklist, Iterate Developing Trust is Challenging What does ‘Production’ mean? Keep it up: unplanned outages are rare or nonexistent Keep it safe: data, functionality, and code are all kept safe from unauthorized users Keep it correct: works as intended, provides the right answers Keep it snappy: fast response times, ability to predict needed capacity for expected traffic
19. R (& Shiny) in Production Journey Code Profiling Version Control Testing Deployment/Release Access/Security Performance Tuning Shared Goal: Shorten the distance between development and production Shared Goal: The improvement of daily work Shared Goal: Reduce the risk of deploying a breaking change Testing! Automated Testing! Getting a Sandbox!
20. Shared Goal: Shorten the distance between development and production ADVOCATE FOR A SANDBOX PUBLISHING ENVIRONMENT B. User Acceptance Testing A. Automated Snapshot Testing
21. When developers begin to think of infrastructure as part of their application, stability and performance become normative. - Jeff Geerling “Ansible for DevOps”
22. Learning Environments for Zero Dollars
23. • Deployment is any push of code to an environment (test, prod) • Release is when that code (feature) is made available to users Application-based release patterns vs. Environment-based release patterns DevOps Learning: Decouple deployment from release Shared Goal: Reduce the risk of deploying a breaking change!
24. The DevOps Handbook 1. Accelerate Flow - Make work visible - Limit Work in Progress (WIP) - Reduce Batch Sizes - Reduce the number of handoffs - Continually identify and elevate constraints - Eliminate hardships and waste 2. Utilize Feedback - See problems as they occur - Swarm to solve problems and build new knowledge - Keep pushing quality closer to the source - Enable optimizing for downstream work centers 3. Learn and Experiment - Enable organizational learning and a safety culture - Institutionalize the improvement of daily work - Transform local discoveries into global improvements - Inject resilience patterns into daily work Three principles form the underpinnings of DevOps:
25. More Books! bit.ly/devops-bookshelf
26. Join the Community How to be an empathetic witness: - Listen - Ask good questions - Never judge - Never advise* but offer your own experiences (citation: Benjamin Hardy, PhD) solutions.rstudio.com community.rstudio.com #radmins