SYSTEM Cited by 2 sources

Feast

Definition

Feast (feast.dev) is an open-source feature store — feature definitions + orchestration layer + serving APIs for ML teams. Feast's design explicitly separates feature definitions (what the feature is) from infrastructure concerns (where it's stored, how it's served) via a pluggable adapter ecosystem.

Why Dropbox chose it

In the Dash feature-store post, Dropbox evaluated the feature-store landscape — Feast, Hopsworks, Featureform, Feathr, Databricks, Tecton — and picked Feast for two reasons:

1. Clean separation of feature definitions from infrastructure. ML engineers can write PySpark transformations for feature engineering without dealing with serving/storage/orchestration plumbing.
2. Modular architecture + extensive adapter ecosystem. Integration with existing infrastructure was straightforward — specifically, the DynamoDB adapter plumbed directly into Dynovault, Dropbox's internal DynamoDB- compatible store.

(Source: sources/2025-12-18-dropbox-feature-store-powering-real-time-ai-dash)

Adapter model

An "adapter" in Feast's terminology is a provided interface that integrates its framework with different backend systems. Adapters cover offline stores, online stores, compute engines, and orchestration. The adapter model is what made it possible to reuse Feast's definitions/orchestration while swapping in Dropbox-specific infrastructure.

Python serving limit

Feast's canonical online-serving path is a Python SDK. Dropbox hit a concurrency ceiling on it (CPU-bound JSON parsing + GIL) and replaced just that layer with a Go service, keeping Feast for definitions + orchestration. This is a broader lesson: a Python control / orchestration layer composed with a performance-critical serving layer in a non-GIL language is a robust architectural split for ML infrastructure.

(Source: sources/2025-12-18-dropbox-feature-store-powering-real-time-ai-dash)

Seen in

• sources/2025-12-18-dropbox-feature-store-powering-real-time-ai-dash — Feast as the orchestration + definitions layer of Dash's feature store, with its Python serving path replaced by a Go service.
• sources/2026-01-06-expedia-powering-vector-embedding-capabilities — non-feature use case: Feast as the metadata / discoverability layer for embedding collections in Expedia's Embedding Store Service. Each collection registered in Feast with its associated service + embedding model / version. Demonstrates that Feast's declarative-definitions + adapter ecosystem properties apply cleanly to vector embeddings and not only to numerical/categorical feature views — a non-obvious platform extension.

Related

• systems/dash-feature-store
• systems/expedia-embedding-store — first wiki instance of Feast used as an embedding-platform metadata layer.
• systems/dynovault — online store behind Feast's DynamoDB adapter at Dropbox.
• concepts/feature-store
• concepts/embedding-collection — the unit Feast registers in the Expedia case.