ML Feature store

FEAST

• Open source feature store
• Serves features in production
• Prevents feature leakage by building training datasets from batch data
• Automates process of loading and serving features in an online feature store
• Supports feature store management without kubernetes, spark or self-managed infra.
• Feature stores require access to compute layers, offline and online databases and need to directly interface with production systems.

Vertex AI Feature Store:

Challenges in ML Workflows:

• Difficulty in sharing and reusing features across different ML workflows and projects.
• Need to serve features reliably in production with low latency.
• Addressing feature value skew between training and serving.

Main Capabilities:

• Sharing features across the organization.
• Reducing duplicate efforts and eliminating the need for feature reengineering.
• Minimizing redundant work.
• Providing a central repository for storing and serving features.
• Offering search and filter capabilities based on feature name and entity name.
• Managing services to reduce latency and management overhead.
• Mitigating training-serving skew through various measures, including computing feature values only once and point-in-time lookups to avoid data leakage.
• Continuous monitoring of data distribution to detect drift.

Data Model:

• Utilizes a time series data model to store features.
• Organizes resources hierarchically: Feature Store > Entity Type > Feature.
• Feature Store: A container for storing and managing features, shared across teams.
• Entity Type: Groups related features (e.g., movies and users) and organizes them.
• Specifies entity type name, description, and features.
• Entity IDs are used to identify entities uniquely within a feature store.
• Entity IDs must be of type string.
• Feature Value: Captured at specific time points, allowing multiple values for a feature for a single entity.
• Associated with an identifier (Entity ID, Feature ID, Timestamp) for retrieval during serving.

Storage Types:

• Two storage methods: Online storage and Offline storage.
• Online storage retains the latest timestamp values for online serving.
• Offline storage stores data until the retention limit or deletion.
• You can control offline storage costs and retention limits.
• Online serving nodes are virtual machines for serving feature values online.
• They offer low latency, scalability, and reliability.
• The number of online serving nodes depends on online serving requests and ingestion jobs.
• Options for configuring online serving nodes: Auto-scaling or allocating a fixed node count.
• Considerations for auto-scaling, including re-balancing data and error handling.
• Option to set the number of online serving nodes to zero to prevent charges and data deletion.