[!todo] Seed note. A starting point, not a finished note yet.
A feature store is the data layer of classic ML infrastructure: it computes features once, stores them, and serves the same values to both the training pipeline (offline, batch) and the live model (online, low-latency). Its whole reason to exist is preventing training/serving skew, where a feature is calculated one way when the model learns and a different way when it predicts, a bug no unit test catches. Feast, Tecton, and the Databricks Feature Store are the common implementations; the layer matured enough by 2025 that Tecton folded into Databricks. In the LLM-era stack its structural role is taken over by the vector store, which is why the two are the same row in the comparison table. Seeded from ML Infrastructure.