Fennel

Fennel is a realtime feature platform designed to author, compute, store, serve, monitor, and govern both realtime and batch machine learning features. It uses plain Python and Pandas, avoiding DSLs, Spark, or Flink jobs, and provides automatic backfills for pipelines. Fennel offers fully-managed infrastructure with zero dependencies on production infrastructure and includes a feature repository for reuse across different use cases. It ensures immutability and versioning of features to eliminate offline-online skew due to definition changes and supports unit testing across batch and realtime pipelines. Fennel performs compile-time validation for strict end-to-end lineage validation to prevent runtime errors and allows specifying expected data distributions with alerts for anomalies. It supports a single definition of features across both offline and online scenarios, providing sub-second feature freshness and single-digit millisecond response times. Fennel is SOC2-compliant and deploys inside the customer’s VPC, aiming to lower cloud costs for the same workload and performance. It supports various use cases including fraud detection, risk engineering, credit underwriting, insurance underwriting, recommendations/personalization, search ranking, and marketing personalization. Fennel’s backend is primarily written in Rust, relying heavily on Tokio’s async runtime, and uses Kafka for handling all inflow data and RocksDB for at-rest data storage. Kubernetes is used for maintaining the lifecycle of all running services, and Pulumi is used for provisioning infrastructure as code. PostgreSQL serves as a central metadata store, excluding customer data. Fennel’s architecture includes a replicated fleet of engines for heavy compute tasks, with query servers converting queries to physical plans and executing them using Python executors. Its incremental ingestion and processing approach reduces cloud costs by avoiding redundant computations. Fennel’s in-house Rust-based streaming engine supports CDC, event time, and out-of-order handling, using RocksDB instances on local SSDs for K/V serving to reduce memory overhead and cloud vendor margins. Various cost optimizations include using spot instances, efficient Rust services, AWS Graviton processors, and data tiering.