Feature Store  ·  View 01 of 21  ·  Context and scope

System Context

Who reads a feature vector, where the signals come from, and what the platform deliberately does not own.

Editable source SVG draw.io All views
Consumers ETA model service p99 25 ms Payment risk model Store ranking 500 keys/call Surge pricing Training jobs SageMaker People Data scientist 60, 9 teams ML engineer Feature group owner Data governance Platform SRE Sources — read, never written Order & payment events Kinesis Courier telemetry 1.4M ev/s peak Clickstream Analytics warehouse Redshift Weather & traffic feed Feature Store 1,400 features · 45 models Out of scope Model registry BI reporting defines onboards owns classifies operates events telemetry sessions dimensions conditions reads vector read batch keys vector read datasets not owned not owned Feature Store — System Context External / third party Person or role synchronous event / async batch The platform reads every source and writes to none of them. Model training and serving sit outside the boundary. v 1.0 · owner Data Platform Architecture · date 2026-09

Decisions

  • The platform reads every source and writes to none of them. A feature store that writes back becomes a second system of record for data it does not own.
  • Model training, model serving and the model registry sit outside the boundary. The platform serves them; making it host them would couple a data contract to a deployment cadence.
  • Five consumer surfaces, one of which is a batch consumer with entirely different latency needs — which is why there are two access paths rather than one.

Assumptions

  • 40M monthly active users, 1.1M partner stores, 180k couriers, 2.2M orders/day peaking at 4,500/minute.
  • 45 models in production owned by 9 teams; 1,400 features in 120 groups over 6 entity types.
  • Courier telemetry peaks at 1.4M events/second into materialisation.

Risks

  • A source team changes an event schema without notice; the platform's only defence is the contract guard, which stops a bad batch but cannot restore a missed window.
  • The weather and traffic feed is a third party with no SLA to the platform; features derived from it are the ones most likely to be stale at peak.
All views High-Level Architecture  ▶
Solution Architecture v1.0 · Amazon Web Services · Data Platform Architecture · 2026-09 · 21 views · 14 architecture decision records01-system-context