Compared with self-hosting Feast, what do Databricks Feature Store, SageMaker Feature Store and Agent Platform Feature Store each run for you?
answer
- who runs the online database
- offline history stays where it is
- sync from offline to online
- each service's own read call
- Feast is not a database
basics
~20 sFeast layers over storage you run: your online database, your materialization schedule, your feature server. Each managed service provisions its online store, keeps it in sync with its offline store, and serves reads through its own API.
solid answer
~50 sFeast's docs say it is not a database or an orchestrator, so a self-hosted Feast team runs the online store, the `feast materialize-incremental` schedule and `feast serve`. **Databricks Feature Store** keeps feature tables as Unity Catalog Delta tables, provisions a Lakebase online store with `create_online_store`, syncs it with `publish_table`, and Model Serving looks the features up. **SageMaker Feature Store** takes `PutRecord` writes into a feature group whose online store keeps the latest record and whose offline store appends Parquet to your S3 bucket for Athena; `GetRecord` reads. **Agent Platform Feature Store** treats your BigQuery tables as the offline store and syncs `FeatureView`s into a Bigtable-served `FeatureOnlineStore`, read with `fetchFeatureValues`. **Hopsworks** runs both stores, as a hosted app or on your own Kubernetes. Computing the feature values stays your job, except with Databricks' Feature Views (Public Preview).
go deeper
Recall the four pieces Feast leaves to you: offline store, online database, materialization schedule and feature server. Then name what each service provisions instead.
Explain how values reach each online store: publish_table on Databricks, PutRecord on SageMaker, feature view sync on Agent Platform. Know which call reads them back.
Point out what stays yours after buying: computing the features, designing keys and event times, and deciding freshness. A managed store moves operations, not modelling.
Frame the choice as who operates the online tier and sync, and whose API the applications depend on. Price that against the platform team you actually have.
## The build option: what Feast leaves to you **Feast** (release 0.66) is a Python SDK, a CLI and a registry. Its own documentation is explicit about what it is *not*: not a database, not an ETL system and not a data orchestration tool. It is a layer over storage that someone else provides. A team that self-hosts Feast therefore runs, itself: - the **offline store** Feast reads history from — a warehouse or lake named under `offline_store` in `feature_store.yaml`; - the **online store** database that holds the latest values — named under `online_store`; - the **schedule** that runs `feast materialize-incremental`, because Feast does not schedule its own jobs; - the **feature server** (`feast serve`, port 6566 by default) when applications read over HTTP; - capacity, upgrades, credentials and on-call for every one of those pieces. A managed feature store is best understood as a subset of that list taken off your hands. Which subset differs per service. ## What each managed service runs | Service | Offline store | Online store | How values get online | Online read | |---|---|---|---|---| | **Databricks Feature Store** | Delta tables in Unity Catalog with a primary key | Online Feature Store on Lakebase | `fe.publish_table` sets up a sync pipeline | Model Serving automatic feature lookup, Feature Serving endpoints | | **Amazon SageMaker Feature Store** | Parquet in your S3 bucket, queried with Athena | `Standard` (default), `Standard_V2` or `InMemory` | `PutRecord` writes the record to the group's stores | `GetRecord`, `BatchGetRecord` | | **Agent Platform Feature Store** | Your own BigQuery tables or views | `FeatureOnlineStore` with Bigtable online serving | A `FeatureView` syncs from BigQuery on a cron schedule or continuously | `fetchFeatureValues` | | **Hopsworks** | Offline feature groups | Online-enabled feature groups, stored in RonDB | `insert` on the feature group | `get_feature_vector` on a feature view | ## Databricks Feature Store Any **Delta table in Unity Catalog with a primary key** can be a feature table. You still own the pipeline that writes it; only the declarative **Feature Views**, in Public Preview, let Databricks compute and manage the feature pipelines. `fe.create_online_store(name=..., capacity="CU_2")` provisions a Lakebase Autoscaling instance, `fe.publish_table` creates the online table and keeps it in sync, and a model logged with `fe.log_model` gets its features looked up by Model Serving at request time. Apps outside Databricks read through a Feature Serving endpoint. The older Workspace Feature Store is deprecated. ## Amazon SageMaker Feature Store A **feature group** must name a **record identifier** feature and an **event time** feature. The synchronous `PutRecord` call ingests a record. The online store keeps only the record with the latest event time per identifier. The offline store appends every record to an S3 bucket you name, as Parquet registered in the AWS Glue Data Catalog, so Athena can query it. Two details catch newcomers: 1. The SDK does not create an online store unless you enable it. 2. The `InMemory` online tier has no offline store at all. ## Agent Platform Feature Store Google's documentation now calls the service **Feature Store on Gemini Enterprise Agent Platform**, or Agent Platform Feature Store; Vertex AI Feature Store is the former name. Your **BigQuery tables or views are the offline store** — the service provisions no separate copy. You create a `FeatureOnlineStore` and one or more `FeatureView` resources that sync from BigQuery, optionally registering `FeatureGroup` and `Feature` resources first. Optimized online serving is deprecated, so Bigtable online serving is the path for new stores. ## Hopsworks Hopsworks keeps **feature groups** (a primary key and an optional `event_time`) and reads through **feature views**. Online-enabled groups live in RonDB. Of the four, it is the one you can also install yourself, with Helm on Kubernetes in a cloud or on-premises, as well as use as a hosted serverless app. ## What none of them takes away - **Feature computation**, apart from Databricks' declarative Feature Views: you still write the pipeline that produces the values. - **Key and event-time design**: every service asks you for a primary key or record identifier, and for a timestamp when history matters. - **Freshness decisions**: someone still chooses between a scheduled sync and a continuous one, and pays for that choice. The build-versus-buy question is therefore narrower than it sounds. You are deciding who runs the online database, the sync and the read endpoint, and whose API your applications are written against.
- Which of these services still leaves feature computation to you?All of them, for the common path. Databricks feature tables are populated by your own pipeline; SageMaker ingests what you `PutRecord`; Agent Platform serves what is already in BigQuery; Hopsworks stores what you `insert`. The exception is Databricks' declarative Feature Views, where Databricks creates and manages the pipelines, and those are in Public Preview.
- Why is Hopsworks a different kind of option from the other three?Its docs offer it as a hosted serverless app and as a platform you install with Helm on Kubernetes in AWS, Azure, GCP or on-premises. That lets a team keep one feature-store API across environments while still choosing who operates it. The other three each run only inside their own cloud or platform.
- A managed feature store computes every feature for you from raw data.
- Feast stores feature values itself, so it needs no separate online database.
- Agent Platform Feature Store copies BigQuery data into its own offline store.
- SageMaker's online store keeps every historical record for training.
- Buying a managed store removes any need to design keys and event times.