Patrick Deziel & Prema Roman - Reconceptualizing Machine Learning for the Real-time World

87% of data science projects never make it to production due to challenges with MLOps and deployment

Traditional batch machine learning faces limitations:

• Models become outdated quickly
• Long iteration cycles for updates
• High latency between data ingestion and model updates
• Memory issues with large datasets

Real-time/online learning advantages:

• Continuous model updates as new data arrives
• Faster training cycles (learning one record at a time)
• More responsive to changing user behaviors
• Lower latency between data receipt and model updates

Ensign platform features:

• Managed streaming service for data scientists
• Publish/subscribe model for data streams
• Historical data access and querying
• Python and Go SDK support
• Built specifically for ML/data science workflows

Asynchronous data science approach:

• Replaces batch processing with continuous data streams
• Enables concurrent processing of different components
• Separates data ingestion from model training
• Makes deployment and updates more efficient

Best use cases for real-time ML:

• Recommendation systems
• Anomaly detection
• Personalized models
• IoT applications
• Systems requiring immediate feedback loops