Skip to content

ADR-0005-streamlit-demo-layer.md

Latest commit

 

History

History
44 lines (35 loc) · 2.32 KB

ADR-0005-streamlit-demo-layer.md

File metadata and controls

44 lines (35 loc) · 2.32 KB
Raw

ADR-0005: Streamlit demo as the public consuming layer

  • Status: Accepted
  • Date: 2026-07-08

Context

The pipeline is rigorous but backend-only — there is nothing a reader can click to see it work. A zero-setup hosted demo is the single highest-ROI "showable" artifact for a data project, and Streamlit is a common default. Fusing an ML/AI layer into this repo would read as a shallow bolt-on — so the demo must stay a thin consuming layer, not a new genre. Power BI (the planned consuming layer) can't be shared publicly on the free tier, which blunts its value as the public artifact.

Decision

Add a small Streamlit app (app/) as the public consuming layer, alongside (not instead of) Power BI. Two pages over the existing gold rollups: usage trends and station explorer.

Why this doesn't violate the "no web app" non-goal

The non-goal targets product framing — users, accounts, a thing pretending to be a startup. A read-only analytics dashboard is the same role Power BI already plays in the locked stack: a consuming layer on gold. It adds no auth, no writes, no users. Framed honestly in the README as a demo layer, not a product.

Why DuckDB + committed Parquet, not a live Snowflake connection

  • Trial independence: the Snowflake trial suspends ~2026-08-06, after which queries are blocked. A demo wired to Snowflake would die with it. So gold rollups are exported to app/gold_export/*.parquet (ingestion/export_gold_to_parquet.py) and the app reads them via DuckDB — it runs forever with no warehouse.
  • Free hosting: Streamlit Community Cloud runs a public-repo app for free; committed Parquet (9 MB, well under limits) needs no secrets or external DB.
  • Only rollups ship: daily_journey_stats, station_daily_flows, dim_station, dim_date — never the 41M-row fact_journey (too big for git; the rollups exist for exactly this).

Consequences

  • The ML demand-forecasting extension stays a separate future project (breadth belongs across projects, per the research) — the Streamlit app is a viz layer, not ML.
  • Adds Streamlit + Altair to the dev deps; app/requirements.txt pins them for Cloud.
  • One manual step remains the user's: connecting the repo in Streamlit Community Cloud (a GitHub-authorize click), same hands-on posture as the Power BI dashboard.