MLOps Virtual Event: Automating ML at Scale
•
0 likes•729 views
ML is transforming many industries but operating ML systems at scale is complex as it involves many teams, constant data and model updates, and moving from development to production. ML platforms aim to help with this by providing software to manage the entire ML lifecycle from data to experimentation to production deployment through a consistent interface. Desirable features for an ML platform include ease of use, integration with data infrastructure for governance, and collaboration functions to enable sharing of code, data, models and experiments. Databricks provides an open source ML platform that integrates with data lakes and a data science workspace to help organizations perform MLOps at scale.
Report
Share
MLOps Virtual Event: Automating ML at Scale
- 1. MLOps Virtual Event: Automating ML at Scale Matei Zaharia Chief Technologist, Databricks @matei_zaharia
- 2. ML is Transforming All Major Industries Healthcare Logistics Telecom Government Banking High Tech Oil & Gas Agriculture Retail Travel
- 3. But ML is Different from Traditional Software Traditional Software Goal: meet a functional specification Quality depends only on application code Pick one software stack Machine Learning Goal: optimize a metric (e.g. prediction accuracy) Quality depends on training data and tuning parameters Constantly evaluate and combine new libraries for the same task
- 4. So Operating ML is Complex! § Many teams and systems involved § Constantly update data & metrics § Hard to move from development to production environments Data Prep Training Deployment Raw Data ML ENGINEER APPLICATION DEVELOPER DATA ENGINEER
- 5. So Operating ML is Complex! § Many teams and systems involved § Constantly update data & metrics § Hard to move from development to production environments Data Prep Training Deployment Raw Data ML ENGINEER APPLICATION DEVELOPER DATA ENGINEER ML teams often spend >50% of time maintaining existing models
- 6. Response: ML Platforms Software to manage the ML development and operations process, from data to experimentation to production Examples: Google TFX, Facebook FBLearner, Uber Michelangelo, MLflow Typical functionality: ▪ Data management ▪ Experiment management ▪ Model management ▪ Deployment for inference ▪ Reproducibility ▪ Testing & monitoring All through a consistent interface!
- 7. Desirable Features for an ML Platform
- 8. Desirable Features for an ML Platform 1. Ease of adoption by data scientists, engineers, and model users ▪ How much work does it take to use? What ML libraries are supported? Etc.
- 9. Desirable Features for an ML Platform 1. Ease of adoption by data scientists, engineers, and model users ▪ How much work does it take to use? What ML libraries are supported? Etc. 2. Integration with data infrastructure to support data versioning, monitoring, and governance across data pipeline & ML steps
- 10. Desirable Features for an ML Platform 1. Ease of adoption by data scientists, engineers, and model users ▪ How much work does it take to use? What ML libraries are supported? Etc. 2. Integration with data infrastructure to support data versioning, monitoring, and governance across data pipeline & ML steps 3. Collaboration functions to enable sharing code, data, features, experiments and models in a central place (securely!)
- 11. Our MLOps Approach in Databricks § Every org’s requirements will be different, and will change over time § Provide a general platform that is easy to integrate with diverse tools Open source machine learning platform Transactional, versioned data lake storage Data science & ML workspace
- 12. In This Webinar § How we and other organizations perform MLOps at scale § Demos and experience from two customers § Live Q&A with presenters