school_compare/docs/ARCHITECTURE.md

Architecture

This describes the implementation as reviewed on 2026-09-14. It distinguishes current behaviour from improvements still to be implemented.

Request flow

Browser → Next.js public routes
          ├─ /api/* proxy → FastAPI → cached DataFrames / PostgreSQL marts
          │                         ├─ Typesense (search and suggestions)
          │                         └─ postcodes.io (postcode lookup)
          └─ /admin, /cms-api, /blog → Payload → payload schema + media volume

Next.js server rendering → FastAPI directly through FASTAPI_URL

nextjs-app/lib/api.ts contains typed fetch wrappers and revalidation defaults. The proxy is nextjs-app/app/(frontend)/api/[...path]/route.ts. Payload uses /cms-api so its routes do not collide with the FastAPI proxy. The proxy denies /api/flags; server-side rendering reads flags directly from FastAPI.

Data ownership

Layer	Owner and role
Source data	GIAS, DfE EES, Ofsted, finance, deprivation and council admission-distance sources
raw	Singer taps and the PostgreSQL target configured in pipeline/meltano.yml
Staging/intermediate/marts	dbt models in pipeline/transform; marts are materialized tables
marts.dim_school, marts.dim_location	School identity and location, filtered to supported England establishments
marts.fact_*	Performance and supplementary datasets; coverage and years vary
Typesense schools alias	Search documents built by pipeline/scripts/sync_typesense.py
payload	CMS collections and migrations in nextjs-app/; independent of dbt
Media volume	Uploaded blog media; requires backup and cannot be regenerated from school datasets

backend/models.py maps existing marts for reading. It does not create the school schema. There is no startup schema-version migration or CSV reimport. Payload's nextjs-app/migrations/ is active and must not be confused with the removed legacy backend migration code.

Coordinates normally come from GIAS British National Grid coordinates transformed by PostGIS in dim_location.sql. pipeline/scripts/geocode_postcodes.py is a manual fallback utility, not a task wired into the current school-data DAG. Backend postcode searches also use postcodes.io; that lookup does not populate school coordinates in the database.

Backend boundaries

• app.py: routes, middleware, search filtering, sitemap/place publication and response assembly.
• data_loader.py: SQL loading, process-local DataFrame caches, Typesense calls, postcode lookups, supplementary queries and benchmark calculation.
• database.py: synchronous SQLAlchemy engine and sessions.
• schemas.py: metric definitions, column mappings and display metadata; despite its name this is not a collection of Pydantic API response models.
• places.py and localities.py: place registry and curated locality information.
• flags.py: Unleash-backed feature flags, disabled when no server is configured.
• gias_codes.py / ofsted_codes.py: source-code translation and display rules.

Search starts from a cached latest-row-per-school snapshot. Detail pages read history from the full DataFrame and supplementary data from marts. Comparisons batch supplementary queries across selected URNs. Async routes still contain synchronous dependency calls; a fully asynchronous database layer is not present.

Frontend boundaries

app/(frontend) owns the public root layout and pages. app/(payload) owns the CMS root layout. Do not add a shared app/layout.tsx: these groups deliberately have separate root layouts. Root metadata files remain in app/.

Server pages fetch initial data and pass it to client views. Client state uses React hooks, URL search parameters and the comparison context/localStorage. There is no SWR dependency. Leaflet maps are loaded through dynamic wrappers; Chart.js renders performance and comparison charts.

components/school/ contains detail sections, with section decisions and data preparation in lib/schoolSections.ts. The nearby-schools section is selected in backend/nearby_schools.py — hard filters decide eligibility (phase, provision, selectivity, gender) and distance alone decides the order, capped per phase — and served on /api/schools/{urn}. Its rules are presentation logic, deliberately kept out of marts.* so they can be tuned by deploy rather than by pipeline run. lib/types.ts contains manually maintained API types. payload-types.ts and the Payload import map are generated artifacts.

Publication and caching today

1. Airflow DAGs extract and validate source data, then run selected dbt builds.
2. Relevant DAGs rebuild Typesense and swap the schools alias.
3. They call POST /api/admin/reload with X-API-Key. It builds and validates replacement DataFrames, places, reverse membership and sitemaps off the request loop, then publishes them together. Failure returns 503 and preserves live data.
4. A separate weekly sitemap DAG can regenerate the derived publication from the current DataFrame without clearing the live registry first.

GIAS is scheduled daily, Ofsted monthly, and annual datasets are manually triggered. The DAG definitions are authoritative for selectors and dependencies.

Caches exist in several independent layers: backend DataFrames and registries, backend HTTP Cache-Control/ETags, Next.js fetch/page revalidation, and browser or shared HTTP caches where configured. Place fetches request a one-week revalidation interval. HTTP ETags are computed after route execution, not before database work.

Typesense publication validates every import response and the final document count before switching aliases. A session-scoped PostgreSQL advisory lock serialises index reads/publication across DAGs. The previous collection remains available for rollback; old unaliased collections are pruned after success. Failed drafts are retained until a later successful cleanup, because an uncertain alias-update response must never cause deletion of a potentially live index.

The backend snapshot swap is process-local and assumes the current single-worker deployment. It is not an atomic transaction spanning PostgreSQL marts, Typesense and Next.js caches. Next.js caches are not explicitly purged by the pipeline. School search retrieves a relevance-ordered candidate prefix (currently capped at 1,000 URNs) before applying API filters. This keeps scoped searches useful while putting a hard ceiling on Typesense round trips; only a dependency failure invokes substring fallback, not a valid empty match set.

Deployment references

See DEPLOY.md. PR checks include frontend typechecking/tests, backend unit tests, image builds and AI review. Staging journeys run after merging. Staging runs are serialised across builds, deployment and E2E. Build-stamped frontend/backend identities are checked before and after journeys. Only then are the captured image digests marked verified. Promotion resolves and validates the complete verified image set before retagging production. See the runbook for first-rollout requirements and remaining integration checks.