Build a feature store using Feast with offline/online serving, feature pipelines, point-in-time joins, and monitoring.
You are an ML platform engineer. Build a production feature store using Feast. ## Configuration - **Offline Store**: [OFFLINE e.g. BigQuery / Redshift / Snowflake / file-based] - **Online Store**: [ONLINE e.g. Redis / DynamoDB / Bigtable / PostgreSQL] - **Registry**: [REGISTRY e.g. local file / S3 / GCS] - **ML Tasks**: [TASKS e.g. fraud detection, recommendation, churn prediction] - **Data Sources**: [SOURCES e.g. event streams, batch ETL, real-time APIs] ## Feature Store Components ### 1. Feature Definitions (`features/`) For each task in [TASKS], define: - Entity definitions (user, item, transaction, session) - Feature views with: - Source: [SOURCES] configuration - Schema: feature names, types, descriptions - TTL (Time-to-Live) for online features - Batch vs streaming materialization - On-demand feature views for real-time transformations - Feature services grouping features per model ### 2. Data Pipeline (`pipelines/`) - Batch materialization: schedule offline-to-online sync - Streaming ingestion from [SOURCES] via Kafka/Kinesis - Feature transformation logic (aggregations, embeddings, encodings) - Data quality checks during materialization - Backfill pipeline for historical features ### 3. Point-in-Time Joins (`training/`) - Historical feature retrieval for training data - Entity DataFrame with timestamps for PIT correctness - Prevent data leakage with proper timestamp handling - Efficient retrieval with partition pruning - Training dataset export to Parquet/CSV ### 4. Online Serving (`serving/`) - Low-latency feature retrieval for inference - Batch feature retrieval for batch predictions - Feature freshness monitoring - Fallback values for missing features - Caching layer configuration ### 5. Feature Discovery (`discovery/`) - Feature catalog with search and browse - Feature lineage tracking - Feature usage analytics (which models use which features) - Data quality metrics per feature - Feature documentation auto-generation ### 6. Monitoring (`monitoring/`) - Feature drift detection (distribution changes) - Feature freshness alerts - Serving latency tracking - Data quality scorecards - Grafana dashboards for feature store health Provide complete Feast project with all configs, pipelines, and infrastructure code.
Free to copy and use. Compatible with Claude 4 Opus, GPT-5, Gemini 2.0 Flash.
Configure your offline and online stores. Define feature views for your ML tasks. Run feast apply to register features. Schedule materialization jobs for online store sync.
Initial release
Sign in and download this prompt to leave a review.