Uploaded byStepan Pushkarev
DevOps for DataScience
The document discusses the integration of DevOps practices into data science, focusing on deploying analytics into production and the responsibilities of data scientists and data engineers. It outlines the importance of feedback loops, testing, and monitoring in the deployment process, as well as the challenges faced in ensuring analytics reliability. The presentation also introduces Hydrosphere's tools, like Mist and Swirl, for managing machine learning models and analytics services efficiently.
More Related Content
Similar to DevOps for DataScience
DevOps for DataScience
- 1. DevOps for DataScience by Stepan Pushkarev CTO of Hydrosphere.io
- 2. DevOps is acatchy buzzword to optimise things
- 3.
- 4. Is there lifeafter marriage data science? Dating, Flowers, Dreams Marriage Happily lived forever? Collect & prepare data Build ML Model
- 5. This talk isfor people who are married aware of “other 99% of data science” Dating, Flowers, Dreams Marriage Happily lived forever? Collect & prepare data Build ML Model
- 6. This talk isNOT about - Setting up Apache Spark/Hadoop cluster - Configuring CI/CD tools like Jenkins - Configuring monitoring tools & dashboards - Agile/DevOps brainwashing & consulting story
- 7. Agenda - Challenges indeploying analytics into production - Deploying analytics as a service - Feedback loops: testing, monitoring, analytics of analytics
- 8. Why do companieshire data scientists?
- 9. Why do companieshire data scientists? To make products smarter.
- 10. What is adeliverable of data scientist and data engineer?
- 11. What is adeliverable of data scientist? Academic paper? ML Model? R/Python script? Jupiter Notebook? BI Dashboard?
- 12. What has tobe a deliverable of data scientist? Data pipelines and machine learning models that deployed as pluggable, testable, supportable, monitorable analytics services.
- 13. Option 1: Engineerto implement academic paper
- 14. Option 2: Engineerto re-implement R/Python script
- 15. Option 3: Runnotebook as it is using cron
- 16. Option 3: Runnotebook as it is using cron
- 17. Option 4: Buildsoftware to eat the world Data Science
- 18. Eating data science ©Daniel Tunkelang - Where should you put your data scientists? - www.slideshare.net/dtunkelang/where-should-you-put-your-data-scientists Step 1 (management): Integrate data scientists into cross-functional teams
- 19. Eating data science Step2 (operations): Make environments scalable and elastic. Finally.
- 20. Eating data science Step3: Make data scientists to write less code
- 21. Eating data science Step4: Deploy analytics as services
- 22. Step 5: Usefeedback loops: testing, monitoring, analytics for analytics Build ML Model Test Monitor, maintain, analyze Deploy as a service Collect & prepare data
- 23. Agenda - Challenges indeploying analytics into production - Deploying analytics as a service - Feedback loops: testing, monitoring, analytics of analytics
- 24. Deploying analytics asa service - Defines deliverable for Data Scientist / Data Engineer. - Plugs analytics into end-to-end products through API. - With the right tooling allows Data Scientist to deploy it in self serve
- 25. Look around -proprietary ML based APIs - Alchemy API - Google Prediction API - Cloud Vision API - Azure ML Can we do our own on top of Apache Spark?
- 26. Bad Practice #1.Business logic in Spark? WTF?
- 27. Bad Practice #2.Database as API Execute reporting job Mark Job as complete & save result Poll for new tasks Poll for resultSet a flag to build a report
- 28. Bad Practice #3.Low level HTTP API When Data Scientists design an API...
- 29. Hydrosphere Mist -a service for exposing analytics jobs and machine learning models as web services
- 30. Types of analyticsservices - Enterprise Analytics services - Reactive or Streaming services - Realtime ML services
- 31. Enterprise analytics services -Could not be pre-calculated - On-demand parametrized jobs - Requires large scale processing - Reporting - Simulation (pricing, bank stress testing, taxi rides) - Forecasting (ad campaign, energy savings, others) - Ad-hoc analytics tools for business users
- 32.
- 33. Reactive or Streamingservices
- 34. Reactive or StreamingReporting services
- 35.
- 36. Realtime Machine LearningServices Train models in Apache Spark and deploy it for realtime low latency serving/scoring with high throughput
- 37. PMML is notan option Spark ML, TensorFlow, H2O, Vowpal Wabbit, and every new ML library invents uses own serialisation format
- 38. Format is notan issue if we re-define a deliverable for ML model xml, json, parquet, pojo, other Single row Serving / Scoring layer Large Scale, Batch processing engine Monitoring, testing integration Deliverable artifact for Machine Learning Model
- 39. Repository Zooming out MLLib modelTensorFlow model Other model Unified Serving/Scoring API
- 40.
- 41. Agenda - Challenges indeploying analytics into production - Deploying analytics as a service - Feedback loops: testing, monitoring, analytics of analytics
- 42. Testing, monitoring, analyticsof analytics - Poorly discussed in community. - We are in production, baby! - Regression. - State matters. Model lifetime is limited. - Data drifts, pipelines and model fail silently. ● Saves time ● Saves money ● Saves lifes
- 43.
- 44. TDD world doesnot work here Pff… easy: - Unit tests - by platform developers - Integration tests - often impossible Not clear who and not clear how: - Regression - Data Validation - Production testing - Data and ML pipelines quality monitoring
- 45. Need either “DataQA” & “Data Ops” people or … AI (formula for the next 10 000 startups - take something and add AI)
- 46. Smart data structuresand dumb code works a lot better than the other way around
- 47. Can we developDSL and Data Structure which is smart enough to learn from data patterns, trends and anomalies to be self-QAed?
- 48. QA view: auniverse vs. big data analytics system People observe and monitor signals from stars to check that universe is not broken today
- 49.
- 50. And Marijuana tomake sense out of it
- 51. Metrics processing, monitoring,correlation insights... ...Isn’t it a big data analytics task on its own?
- 53. ML pipeline Kafka Analyticsjobs for metrics Emit Metrics Stream it back into Spark Context Use insights to make our data structures smart Solution: loop of analytics for analytics
- 54. Benefits ● Don’t needto talk to Ops! :) ● Already have Apache Spark and Kafka in place ● Data Scientist in the loop! ● Unlimited flexibility in analytics, correlation and using ML for ML ● Models could feeded back into Smart self QA-ed data structures.
- 55. Hydrosphere Swirl -a system that creates a swirl of analytics for analytics
- 56. Original ML pipeline Kafka Streaming or BatchSwirl jobs Hydrosphere Swirl Plug, modify, deploy, run jobs & consume results Metrics definition, Notebook integration Hydrosphere Mist (1) Emit metrics Hydrosphere Swirl: Vision
- 57.
- 58. Classify by sentiment Twitter Prepare Data Serve adsto user Hydrosphere Swirl Invalid records 10/sec 2k/sec0.8Ratio Clicks Swirl Demo: Serve Ads to users with positive Tweets
- 59. Classify by sentiment Twitter Ingest & transform Serveads to user Hydrosphere Swirl Invalid records 20k/sec 10/sec0.2 Clicks Data pipeline is broken Ratio Swirl Demo: Serve Ads to users with positive Tweets
- 60. Twitter Ingest & transform Serve adsto user Hydrosphere Swirl Invalid records 10/sec 10/sec0.2 Clicks New ML model deployment Deployed bug in ML code Ratio Swirl Demo: Serve Ads to users with positive Tweets
- 61. Thank you Looking for -Feedback - Advisors, mentors & partners - Pilots and early adopters Stay in touch - @hydrospheredata - https://github.com/Hydrospheredata - http://hydrosphere.io/ - spushkarev@hydrosphere.io