Skip to content

DEPLOYMENT.md

Latest commit

 

History

History
106 lines (77 loc) · 3.83 KB

DEPLOYMENT.md

File metadata and controls

106 lines (77 loc) · 3.83 KB
Raw

Relay Deployment

Production host

  • Vercel project: relay
  • Production domain: https://www.onrelay.app

Before deploy

pnpm install
pnpm validate:env
pnpm lint
pnpm typecheck
pnpm test
pnpm build

For Project Sources, apply the latest Neon migrations to a dev branch first and smoke test an upload before production:

MIGRATION_DATABASE_URL="<neon-dev-unpooled-url>" node scripts/db-admin.js migrate

Deploy to Vercel

vercel deploy --prod --yes

If the repo is not linked locally, log in with vercel whoami first and confirm the target project is relay.

Required production environment variables

  • DATABASE_URL
  • DATABASE_URL_UNPOOLED
  • NEXT_PUBLIC_RELAY_APP_URL
  • RELAY_INTERNAL_API_SECRET
  • POLAR_ACCESS_TOKEN
  • POLAR_WEBHOOK_SECRET
  • POLAR_PRODUCT_ID_PRO_MONTHLY
  • POLAR_PRODUCT_ID_PRO_ANNUAL
  • RELAY_SOURCE_ENCRYPTION_KEY
  • R2_ENDPOINT
  • R2_ACCESS_KEY_ID
  • R2_SECRET_ACCESS_KEY
  • R2_BUCKET
  • R2_REGION

Project Sources stores encrypted raw files in Cloudflare R2. RELAY_SOURCE_STORAGE_MODE=memory is only for local parser/UI tests and must not be used in production. If RELAY_SOURCE_ENCRYPTION_KEY is unset, source blob encryption falls back to RELAY_CONTENT_ENCRYPTION_KEY.

Neon auth mode also requires:

  • NEON_AUTH_BASE_URL
  • NEXT_PUBLIC_NEON_AUTH_URL
  • NEON_AUTH_COOKIE_SECRET

Local auth mode instead requires:

  • LOCAL_AUTH_SESSION_SECRET

Important note on app URLs

Some current auth/token flows still read NEXT_PUBLIC_APP_URL while most of the app reads NEXT_PUBLIC_RELAY_APP_URL. Set both to the same production origin until the legacy name is fully removed.

Internal drain job

Relay currently exposes POST /api/internal/jobs/drain for digest draining. Production still needs a scheduler to invoke it with RELAY_INTERNAL_API_SECRET.

Recommended options:

  • Vercel Cron calling a small internal route wrapper
  • GitHub Actions scheduled workflow with a protected secret
  • External scheduler hitting the endpoint over HTTPS

Memory v2 — cost-safe async pipeline

Memory writes enqueue durable memory_pipeline_jobs and opportunistically drain a tiny bounded batch. The existing daily /api/internal/jobs/cron is the recovery backstop and processes only queued jobs plus projects marked due for hygiene.

/api/cron/memory-pipeline is manual/operator-only. Do not schedule it at a fixed cadence: frequent wakeups prevent Neon scale-to-zero.

Required env / secrets:

Where Name Notes
Vercel env CRON_SECRET Required in production. Cron returns 401 without Authorization: Bearer <secret>.
Vercel env WORKER_DATABASE_URL Optional. Connection string used by the cron route only. Falls back to DATABASE_URL when unset. Point at relay_worker once that role is provisioned.
Vercel env RELAY_MEMORY_PIPELINE_FULL true → Gemini extractors run. Default false (embed-only). Flip after a quality soak.
Vercel env RELAY_HYGIENE_DRY_RUN true (default) logs proposed transitions; false writes.
Vercel env RELAY_PIPELINE_DAILY_USD_CAP Default 5. In-process circuit-breaker on extractor spend per UTC day.
Vercel env RELAY_EMBED_CANONICAL_ENTITIES Default false. Opt-in vectors for newly created canonical entities only.
Vercel env RELAY_PERSONAL_MEMORY_ITEM_CAP Default 500. Archives oldest non-pinned personal overflow.

Worker role provisioning (optional, future hardening)

docs/memory-v2/roles.sql contains the idempotent SQL block that creates relay_worker + grants. Apply per Neon branch via the Neon SQL editor or mcp__Neon__run_sql, then point WORKER_DATABASE_URL at the new role. Today the app still connects as neondb_owner (RLS is decorative for the app conn — see Memory v2 HANDOFF "Risks"); moving the app to a non-owner role is a separate hardening PR.