Mohamed Sabri: Operationalize machine learning with Kubeflow
0 likes351 views
This document summarizes a hands-on workshop on Kubeflow Pipeline. The workshop will cover requirements, an introduction to the presenter Mohamed Sabri, and their approach of strategizing, shaping, and spreading knowledge. It then discusses operationalizing machine learning (MLOps) and provides an analysis, design, coaching, and implementation framework. Deliverables include an implemented MLOps environment, training sessions, design documents, and a recommendations roadmap. The rest of the document discusses MLOps architectures, challenges, example technologies and tools, a use case, and deployment workflows from notebooks to production.
Mohamed Sabri: Operationalize machine learning with Kubeflow
- 3. MOHAMED SABRI Mentor and data science leader
- 7. This is no traditional consulting where we burn your cash Our philosophy We deliver concrete value with full transparancy
- 9. Delivrables The implementation of a viable environment in MLOps MLOps training sessions High-Level and Low- Level Design Document A report with recommendations and roadmap The following deliverables to be expected during and after our mandate:
- 10. How to define an MLOps efficient architecture? What is the level of expertise in MLOps do we have or willing to hire? What type of inference in ML are we looking for? How many machine learning projects do we have inline in short/mid/long term? Automation vs Resource scalability? Which type of vendors is the company working with? Or open source vs Enterprise Is the data science team following the state of the art when it comes to source coding and versioning?
- 12. Microservices design environment Machine learning experimentation Performance monitoring Model serving Retraining pipelines Model registry Development environment Dashboard for monitoring Data analysis Trigger retraining Pipelines deployment Experiment tracking Source code & versioning Automatic detection of new models Framework for microservices Push new model for deployment Automated pipelines Challenges • Data scientists need more education about code submission and production-ready code. • The customer is looking to scale the environment for all the organization’s machine learning projects. • No clear performance metrics have been defined by the customer to evaluate model performance in production. • Helping the customer identify the right resources internally to maintain the environment
- 13. Microservices design environment Challenges • Reaching a low latency (maximum 15 ms) to allow a fast reactivity after model inference. • Handling a large volume of data points per second (between 1 million to 10 million per second) • Scaling the streams, environment based on data volume with no buffer. • Automatically updating the machine learning model if required. Model registry Domain events Update docker image connectors Machine learning experimentation Push new model for deployment Logs & KPI storage Real time monitoring Steam processing Data analysis Development environment
- 14. Some technologies and tools End to end platform Continuous delivery platform Commercial Microservice deployment Automation and pipelines Experimentation tracking and versioning Open source
- 15. MLOps 101
- 16. For you, what is MLOps ? Why is it necessary ?
- 17. MLOps is not just about deployment
- 18. MLOps is like DevOps but for ML • Continuous integration (CI) CI is about testing and validating code and components, but also testing and validating data, data schemas, and models. • Continuous delivery (CD) CD is about a system (an ML training pipeline) that should automatically deploy another service (model prediction service). • Continuous training (CT) CT is concerned with automatically retraining and serving the models.
- 19. Kubernetes
- 21. Kubeflow
- 23. Our use case
- 24. From the notebook to production
- 25. Our architecture Data extraction Data pre- processing Building classifier Trigger deployment Model registry (persistent volume) Data pre- processing Inference model Automated Training pipeline ML Engine Integration with app
- 26. Q&A