Feature Store: Training/Serving Parity, Feature Reuse

A feature store is a system that computes, stores, and serves ML features consistently for both model training (batch) and online inference (low-latency).

Feature Store: When to Build One, When to Buy, When to Skip Entirely

A feature store centralizes the computation and serving of ML features so that the same feature definition produces the same values whether a model is being trained on last year's data or scored in production milliseconds after a user event. The two-store architecture — an offline store (usually a data warehouse or object storage, optimized for large historical reads) and an online store (usually a low-latency KV store, optimized for point lookups) — is designed to eliminate the class of bugs where a model trains on one definition of a feature and scores against another. Feature stores also enable reuse: a feature computed once (say, 'user's last-30-day purchase count') can serve many models rather than being reimplemented per team.

The problems it actually solves

(1) Training/serving skew — the leading cause of models that look great in offline eval and mysteriously underperform in production. A feature store makes the transformation logic single-sourced, so batch and online use identical code. (2) Feature reuse across teams — instead of every model re-implementing 'is this user a paying customer,' the definition lives once and multiple models consume it. (3) Backfills and point-in-time correctness — computing features 'as of' a historical timestamp for training without leakage of future information. (4) Feature freshness observability — knowing when a feature pipeline broke and stopped updating, before a model consumes stale data.

When you don't need one

Fewer than 3-5 production models: the coordination overhead outweighs benefits. Models that all consume batch predictions (no online serving): the offline store is just your data warehouse; a full feature store is overkill. Small feature surface (dozens, not hundreds of features): a well-organized dbt project plus a Redis cache achieves 80% of the benefit with 10% of the operational load. Prototyping and R&D: adds friction to experimentation without paying back until models are production-bound.

Build vs. buy vs. open source

Buy (Tecton, Databricks Feature Store, SageMaker Feature Store, Vertex AI Feature Store): fastest to production, tight cloud integration, ongoing spend that scales with usage. Sensible for teams whose bottleneck is not infrastructure but shipping models. Open source (Feast, Hopsworks): more control, less lock-in, more operational burden. Feast in particular is deliberately lightweight and pairs well with your existing warehouse + KV store. Build: reserved for orgs whose scale, latency, or compliance requirements exceed vendor capabilities. Almost always the wrong answer for the first two years.

Getting adoption from data science

A feature store only works if data scientists actually use it for training, not just as a serving layer glued on late. The adoption arc: (1) migrate an existing model's feature pipeline into the store, prove training/serving parity, ship. (2) new models default to the store; opt-out requires justification. (3) audit teams that keep bypassing and understand why (usually: painful DX, missing feature types, review latency for new feature definitions). Fix the DX rather than mandating usage. A feature store nobody uses because it's slower than a notebook is worse than no feature store.

Operational realities

Feature pipeline SLAs are now user-facing SLAs — if a feature is stale, an online model may score using yesterday's value. Monitor freshness per feature, alert on breaches. Cost management: online stores (DynamoDB, Redis, Cassandra) can become surprisingly expensive at scale; audit unused features quarterly and retire them (feature stores accumulate dead features aggressively). Governance: PII-classified features need access controls at the store layer; document lineage from source data to feature to model to prediction for audit trails. Backup and disaster recovery: treat the online store like production infra — restore drills, replication, capacity planning.

Frequently asked questions

Can a data warehouse serve as a feature store?
For batch inference and training, yes — a well-organized warehouse with dbt models is functionally a feature store minus the online serving. For online inference at low latency, you need an online layer (Redis, DynamoDB, Cassandra) synchronized from the warehouse.
Feast vs. Tecton — how to choose?
Feast if you have engineering capacity to operate infrastructure and want to avoid vendor lock-in. Tecton if you'd rather spend money than time and want features like managed pipelines and materialization out of the box.
How does a feature store interact with a lakehouse?
Well. Lakehouse (Delta, Iceberg) is a natural offline store; feature stores like Databricks Feature Store and Tecton integrate directly. The online layer stays separate (KV store) and is populated by materialization jobs reading from the lakehouse.

Related fundraising guides (40)

Investor directory · Fundraising library · Articles A–Z · Company funding database