Infrastructure

Overview

Layer	What lives there
Supabase	PostgreSQL database, auth, realtime, row-level security
AWS	Data processing pipeline, Lambda artifact builds, S3 data bucket for geodata
Cloudflare	DNS for both domains, and the Worker CDN serving data.burglarysupport.com

The webapp talks directly to Supabase. Processed geodata (district boundaries, crime stats) is written to S3 by the Lambda pipeline and served via a Cloudflare Worker at data.burglarysupport.com.

Terraform config lives in terraform/. The AWS module is at terraform/modules/aws/.

---

Data processing pipeline

Triggered weekly by EventBridge on Sunday at 04:00 UTC. The same workflow can be triggered manually via the BSN CLI or by starting an execution in the Step Functions console.

Step Function: burglary-data-processing-workflow

DataIngestion (geodata-ingestion)
  └─ ParallelProcessing
       ├─ DistrictProcessing (geodata-districts) → NeighborCalculation (geodata-neighbors)
       └─ FetchCrimeData (data-fetcher)
  └─ CrimeDataEnrichment (geodata-crime-enricher)
  └─ MetadataCreation (geodata-metadata)
  └─ ProcessingSucceeded / ProcessingFailed

All tasks retry up to 3 times with exponential backoff on Lambda service errors.

Lambda functions

Function name	Memory	Timeout	Role
burglary-geodata-ingestion	1 024 MB	10 min	Download raw postcode data; publish the monolith and the per-outcode shards
burglary-geodata-districts	1 024 MB	10 min	Compute district bounds
burglary-geodata-neighbors	512 MB	5 min	Compute district neighbour relationships
burglary-data-fetcher	512 MB	5 min	Fetch crime data from UK Police API
burglary-geodata-crime-enricher	2 048 MB	15 min	Join crime records to districts (3 years of data)
burglary-geodata-metadata	256 MB	3 min	Write aggregate metadata
burglary-supabase-keepalive	128 MB	30 s	Ping Supabase daily to prevent free-tier suspension

Published geodata

The ingestion function writes the postcode dataset (~1.73M postcodes) in two shapes, for two different consumers:

Key	Size	Consumer
geodata/postcode_locations.json	~103 MB	geodata-crime-enricher only — it loads the whole thing in memory to join coordinates onto crimes. Never fetch this from a browser.
geodata/postcodes/{outcode}.json	~25 KB each, ~2 900 files	The webapp’s postcode lookup (packages/webapp/src/shared/lib/geocode.ts). Keyed by space-stripped postcode, values are [lat, lng].
geodata/postcode_district_data.json	~744 KB	The map — district centroids and bounds.

The shards exist because the monolith is not addressable per postcode: resolving one postcode from it means downloading all 103 MB. Splitting by outward code is what makes client-side geocoding possible without a third-party lookup; postcodes.io is only a fallback for postcodes newer than the last pipeline run.

Shards are rewritten whenever the source ZIP’s hash changes, at 25 uploads in parallel, and the run fails loudly if any shard fails rather than leaving a partially-published set. Worth watching on the first run against the function’s 10-minute timeout.

All functions run Node.js 20 on arm64. Processed data is read and written to the S3 bucket bsn-production-data-eu-west-2 (in the dedicated BSN account, 582093722364). Lambda artifacts are pulled from s3://bsn-production-data-eu-west-2/lambda-artifacts/<name>.zip.

The supabase-keepalive function fires on its own EventBridge schedule (daily at 06:00 UTC) and queries a real table (user_profiles) so the request registers as database activity. Pinging the PostgREST root (/rest/v1/) does not work: it is served from the schema cache without touching Postgres, and it rejects the anon key with 401 (only the service_role key may call it). It is unrelated to the data pipeline.

Running the pipeline manually

# Start incremental run (skips months already processed)
npm run bsn -- aws stepfunction start

# Force full reprocessing
npm run bsn -- aws stepfunction start --force

# Invoke a single function for debugging
npm run bsn -- aws lambda crime-enricher --month 2024-03

---

CodeBuild artifact pipeline

Terraform provisions a CodeBuild project (burglary-support-lambda-builder) that builds and uploads Lambda zips whenever packages/aws-lambdas/** changes on main.

Flow

1. A push to main touching packages/aws-lambdas/ triggers the GitHub webhook.
2. CodeBuild clones the repo (GitHub PAT stored in var.github_oauth_token).
3. esbuild bundles each Lambda from TypeScript directly — no tsc step. Workspace symlinks from npm install resolve @burglary-support/shared-types.
4. Zips are uploaded to s3://burglary-support-data-eu-west-2/lambda-artifacts/.
5. The next Lambda invocation (or a manual aws lambda update-function-code) picks up the new zip.

Lambda functions are not automatically updated after a build — CodeBuild only publishes the artifact. Use make lambda-deploy or the AWS console to push the new zip to a function.

---

Data CDN (Cloudflare Worker → S3)

A Cloudflare Worker (bsn-data-proxy) serves the S3 data bucket — this replaced the old CloudFront distribution during the dedicated-account migration:

• Domain: data.burglarysupport.com (Cloudflare Universal SSL; data-staging.burglarysupport.com is the pre-cutover verification hostname)
• Origin: the S3 REST endpoint, locked to requests carrying the shared secret the Worker sends (s3-origin-lock.tf)
• CORS: GET/HEAD/OPTIONS allowed from any origin, 24-hour max-age
• Cache TTL: 24 hours for success responses, errors barely cached

# Purge the edge cache after deploying new geodata
make data-cache-purge

---

Running terraform apply

cd terraform
terraform init
terraform plan
terraform apply

Required variables — pass as TF_VAR_* env vars or in a terraform.tfvars file:

Variable	Notes
TF_VAR_cloudflare_api_token	Zone:Read, Zone:Edit, Account:Read, User:Read
TF_VAR_cloudflare_account_id
TF_VAR_github_oauth_token	PAT with repo scope for CodeBuild webhook
TF_VAR_supabase_anon_key	Used by the keepalive Lambda

Cloudflare secrets are stored in SOPS at ~/own/go-setup/k8s/server/secrets.yml. Decrypt with sops -d before running Terraform if you need to pass them as env vars.

Other variables (aws_region, s3_bucket_name, environment, etc.) have sensible defaults in terraform/variables.tf and generally do not need overriding.