Proprietary Feature Stores Hold
Overview
Proprietary feature stores without open export paths create lock-in for training and serving teams. Feast and warehouse-native patterns reduce duplicate storage and speed agent feature access (Feast).
Hold new contracts that cannot export features to open tables or APIs. Negotiate exit clauses and batch export before production dependence.
Adoption Signals
- Google deprecated Vertex AI Feature Store (Legacy) on 17 February 2026, stopped adding features on 17 May 2026 and will sunset its APIs on 17 February 2027 (Vertex AI deprecations).
- Google's successor, now documented as Agent Platform Feature Store, keeps feature data in BigQuery tables or views that form the offline store, so the vendor itself moved to a warehouse-native design (Agent Platform Feature Store).
- Tecton joined Databricks in August 2025 (Databricks blog), and Redis acquired Featureform in October 2025 to fold it into its own offering (SD Times).
- Lakehouse vendors now ship feature serving on their own storage: Databricks runs its Online Feature Store on Lakebase, synchronised from offline feature tables (Databricks Online Feature Store).
- The open alternative is maintained and broad: Feast is Apache-2.0, released v0.66.0 on 21 August 2026, and supports offline stores including Snowflake, BigQuery, Redshift, Spark and PostgreSQL and online stores including Redis, DynamoDB and PostgreSQL (Feast GitHub).
Risks
- Forced migrations. The Vertex AI Feature Store (Legacy) shutdown moves customers onto a successor whose offline store is a set of BigQuery tables or views (Vertex AI deprecations, Agent Platform Feature Store); comparable work applies to any proprietary store that is renamed, merged or retired, on the vendor's timeline.
- Acquisition roadmap risk. After Tecton joined Databricks, the announcement contained no dated commitments for existing customers (Databricks blog); contracts without exit terms leave customers dependent on the acquirer's product plans.
- Logic lock-in beyond data lock-in. Transformations, aggregation windows and backfill semantics defined in a vendor SDK have no portable format; teams that only negotiated data export still face a rewrite.
- Duplicate storage and governance. A separate offline store beside the lakehouse or warehouse duplicates data, access policies and lineage, which complicates GDPR deletion and audit for regulated teams.
Pros & Cons
Advantages
- They promise real-time serving without building it: Databricks describes Tecton's managed platform as delivering sub-10 ms latency and 99.99% uptime, which is hard to reach with in-house streaming jobs and online caches.
- A single SDK handles point-in-time correct training sets, backfills and online serving, removing a class of training/serving skew bugs that small ML teams struggle to prevent on their own.
- Cloud ML suites bundle a feature store with notebooks, training and model serving, so teams already on that cloud get one contract, one IAM model and one support channel.
Disadvantages
- Vendors change or retire these products on their own schedule: Google deprecated Vertex AI Feature Store (Legacy) in February 2026 and will shut its APIs down in February 2027, moving customers onto a successor with a different, BigQuery-based design.
- Standalone feature platforms are being absorbed by larger vendors: Tecton joined Databricks in 2025 and its website now redirects there, and Redis acquired Featureform, leaving customers with roadmaps they do not control.
- Exporting feature values is not an exit: feature logic, backfill jobs and online serving contracts live in the vendor's SDK and must be rewritten when moving to Feast or a warehouse-native approach.
Recommendation
Hold new commitments to feature stores that keep feature definitions and offline data in a proprietary format. For new ML and agent feature work, start from feature tables in your existing warehouse or lakehouse (Delta, Iceberg, BigQuery or Snowflake tables) and add Feast or the platform's native feature serving only for the online path. Keep transformations in SQL, dbt or plain Python so they survive a change of serving layer.
If a managed feature platform is still required for millisecond real-time features, treat it as a replaceable serving component: keep the offline source of truth in open tables you own, write exit terms and data export into the contract, and document every feature's logic outside the vendor SDK. Existing users of Vertex AI Feature Store (Legacy) or Tecton should schedule their migration or renegotiation now rather than at the vendor's deadline.
Sources
Consulting
From radar to implementation
You have assessed the technologies — we help you bring them into production systems safely and measurably.
