Unified client for Confluent and Apicurio Schema Registries with automatic schema caching, retry logic, and Avro serialization/deserialization. Demonstrates robust integration patterns with fallback mechanisms and can be extracted as a standalone library for any project requiring schema registry integration.

Core feature store abstraction providing a unified interface for online (Redis) and offline (MinIO/S3, Iceberg) feature storage. Implements pluggable backend design, entity/feature versioning, and time-windowed aggregations. Can be decoupled from the platform and used independently in ML workflows requiring feature serving.

Implements Population Stability Index (PSI) for detecting drift in categorical and numerical features. Includes configurable windowing, baseline comparison, and alerting mechanisms. The detection logic is isolated and can be applied to any data stream monitoring scenario without dependencies on the broader platform.

Prefetch engine that analyzes historical access patterns to predict and preload hot data into cache before ML training execution. Uses frequency-based scoring and LRU eviction to optimize cache hit ratios. The algorithm is self-contained and adaptable to various caching systems beyond the demonstrated Redis implementation.