DevOps for data science: automate the boring stuff and leverage the OSS ecosystem

Transcript

1. DevOps for Data Science? Automate the boring stuff and leverage the OSS ecosystem PyCon Africa – August 6th, 2020 Tania Allard, PhD @ixek
2. @ixek @trallard trallard.dev
3. About Me I Python I am also a GDE for Tensorflow I love mechanical keyboards My dog usually barks while I am giving online talks
4. These slides https://bit.ly/mlops-pyconafrica
5. background ML and Data Science in 2020 Table of Contents 1 What is even MLOps? And why you’d need it… 2 Mlops 101 Getting started with MLOps 3 @ixek
6. background ML and Data Science in 2020 01
7. Where have we been? The Garner hype cycle @ixek
8. Data Scientist It’s never been easier to run ML experiments ML engineer /SRE Machine learning in production is hard y’all! Every team @ixek
9. • Tools like scikit-learn and Keras make it easy to create models in a few lines • Techniques like transfer learning make our lifes easier • More Compute! All the GPUs! From the DS perspective
10. The new unicorn Must have Analytical skills Software eng. Programming Data engineering Data visualization Also must have Containerization End-to-end ML pipeline CI /CD /Versioning Deep learning / NLP / etc. Privacy and security @ixek
11. MLOps What is it? 02
12. Where is my unicorn? A mythical data scientist who can code, write unit tests AND resist the lure of a deep neural network when logistic regression will do.
13. The origin of devops Software developers: Need to move and iterate fast Operation team: Stability and availability of services is priority @ixek
14. DevOps is the union of people, process, and products to enable continuous delivery of value into production - Donovan Brown “ @ixek
15. Automate Automate everything you can (data processing, model training) Feedback Get feedback on new ideas fast (test immediately) No manual handoffs Provide early testing opportunities DevOps principles @ixek
16. Continuous integration – software engineering Based on test results – no waiting time* Quick testing Automated build Project source code in version control Code changes Automate Feedback iterate @ixek
17. Technical considerations • Reliance on metrics (e.g. accuracy, specifity) • Data visualization • Required domain knowledge So what about ML? @ixek
18. More than ML code / model @ixek
19. The origin of mlops Data scientist: • Need to move and iterate fast • Use my loved frameworks • Scalable • Minimal wait: test, stage production SRE/ML Engineers: • Reuse of tooling and platforms • Uptime • Monitoring • Reliability and stability @ixek
20. Continuous integration – software engineering Improve model based on outputs/outcomes Sought metrics Automated training / data processing Project source code in version control. Data lineage. Code& data changes Automate Feedback iterate @ixek
21. Getting started 101 MLOps 03
22. RECYCLE YOUR ECOSYSTEM 1 Collaboration Version control (Git, Mercurial) OSS dev platform / CI /CD (GitHub, GitLab, Travis) 2 automation Leverage your deployment infrastructure (CI / CD, Make) 3 Mix-match Use the OSS libraries you love and leverage cloud computing* @ixek
23. MlOps step by step ENV #1 CI/CD Pipeline Process Train Stage Serve Data Distributed Cloud ENV #2 Data Scientist SRE/ML Engineers @ixek
24. CI/CD Pipeline Process Train Stage Serve Data Distributed Cloud First, I check in my code. ENV #1 ENV #2 Data Scientist Data Scientist SRE/ML Engineers MlOps step by step @ixek
25. Version control @ixek
26. CI/CD Pipeline Process Train Stage Serve Data Distributed Cloud That kicks off a CI/CD Pipeline. ENV #1 ENV #2 Data Scientist SRE/ML Engineers MlOps step by step @ixek
27. Kicking CI /CD Push changes GitHub actions @ixek
28. CI/CD Pipeline Process Train Stage Serve Data Distributed Cloud And now do a training run on the processed data ENV #1 ENV #2 Data Scientist SRE/ML Engineers MlOps step by step @ixek
29. Not only tests Can leverage to do the training or data processing Vision Venus has a beautiful name and is the second planet from the Sun. It’s atmosphere is extremely poisonous @ixek
30. CI/CD Pipeline Process Train Stage Serve Data Distributed Cloud Actually need to update the parameters ENV #1 ENV #2 Data Scientist SRE/ML Engineers MlOps step by step
31. Parameters update? No problem check in to version control @ixek
32. Updated reporting Embed reports and metrics to your Pull request @ixek
33. Updated reporting Embed reports and metrics to your Pull request @ixek
34. CI/CD Pipeline Process Train Stage Serve Data Distributed Cloud Model is optimized and working! Let’s roll out to production. ENV #1 ENV #2 Data Scientist SRE/ML Engineers MlOps step by step @ixek
35. CI/CD Pipeline Process Train Stage Serve Data Distributed Cloud Trigger the CI/CD pipeline one last time. ENV #1 ENV #2 Data Scientist SRE/ML Engineers MlOps step by step @ixek
36. CI/CD Pipeline Process Train Stage Serve Data Distributed Cloud And roll out to the world! ENV #1 ENV #2 Data Scientist SRE/ML Engineers MlOps step by step @ixek
37. But there is more @ixek
38. But there is more @ixek
39. But there is more @ixek
40. In brief MLOps allows you to be more efficient with the tools you use and love @ixek
41. RECYCLE YOUR ECOSYSTEM 1 Collaboration Version control (Git, Mercurial) OSS dev platform / CI /CD (GitHub, GitLab, Travis) 2 automation Leverage your deployment infrastructure (CI / CD, Make) 3 Mix-match Use the OSS libraries you love and leverage cloud computing* @ixek
42. These slides https://bit.ly/mlops-pyconafrica
43. Thanks! @ixek @trallard trallard.dev