Feature Store  ·  View 04 of 21  ·  People and journeys

Journey — Ship a Feature

The journey that decides whether the platform gets used at all: 30 minutes from merge to readable, or people route around it.

Editable source SVG draw.io All views
Data scientist ETA team Goal — Ship "orders at this store in the last 15 minutes" today, and never build it twice Trigger — The ETA model under-predicts by four minutes at dinner peak Done when — Readable online inside 30 minutes, and the training set that uses it is point-in-time correct 1 · Search before building 2 · Define 3 · Validate 4 · Backfill ◆ moment of truth 5 · Consume ◆ moment of truth What they do Search the catalogue Read adoption count Write the definition Open a PR Compile both plans Preview values Estimate the cost Run 13 months Generate a training set Read it online How it feels Confident Fine Blocked Where it hurts Four near-duplicates 6 h run, live SLOs at risk What answers it Owner + consumer count One artefact, two plans Contract guard preview Budgeted, pre-emptible pool As-of join + manifest Journey — A Data Scientist Ships a Feature The trough is backfill, not authoring. Most of the platform's operational design exists to make that phase survivable. v 1.0 · owner Data Platform Architecture · date 2026-09

The trough is backfill

  • Authoring is easy and validation is fast. The phase that blocks people is backfilling 13 months of history without breaking live materialisation.
  • Which is why the backfill runner is an isolated, budgeted, pre-emptible pool rather than a job on the batch engine — and why the cost estimate is shown before the run, not after.
  • The second pain is upstream of everything: four near-duplicate definitions and no owner on any of them.

Assumptions

  • A new streaming feature reaches readable-online within 30 minutes of merge; a batch feature within one cadence cycle.
  • A 13-month backfill of one feature completes in ≤ 6 h without breaching live freshness SLOs.

What the set owes this journey

  • Search → view 07 catalogue and view 06 lineage. Define → view 16. Validate → the contract guard in views 09 and 13.
  • Backfill → the isolated pool in views 13 and 15. Consume → view 14.
◀  Actors and Journeys All views Journey — Features Go Stale at Peak  ▶
Solution Architecture v1.0 · Amazon Web Services · Data Platform Architecture · 2026-09 · 21 views · 14 architecture decision records04-journey-ship-a-feature