Skip to content

LAUNCH_CHECKLIST.md

Latest commit

 

History

History
75 lines (61 loc) · 3.74 KB

LAUNCH_CHECKLIST.md

File metadata and controls

75 lines (61 loc) · 3.74 KB
Raw

ClimateVision — Go-Live Checklist (climatevision.green)

Work top to bottom. Anything in 🔴 Blocker must be done before the site is publicly announced. 🟠 Important should be done launch day. 🟢 Nice-to-have can follow.

🔴 Blockers — do these first

  • Rotate the leaked GEE key. The climatevision-gee@kinos-473422 service-account private key was shared in plaintext. In Google Cloud Console → IAM → Service Accounts → Keys: delete that key, create a new one.
  • Set secrets as runtime env vars, never in the repo. Run ./scripts/setup_render_secrets.sh (Render) or ./scripts/setup_fly_secrets.sh (Fly). Confirm none of these appear in any committed file: GEE_SERVICE_ACCOUNT, GEE_SERVICE_ACCOUNT_KEY_JSON, GEE_PROJECT_ID, ANTHROPIC_API_KEY, API_SECRET_KEY.
  • Confirm CLIMATEVISION_ALLOW_DEV_KEY=0 in production. It is set in render.yaml; verify it on Fly too. A live dev key bypasses API-key auth.
  • Decide the truth-in-labeling posture. Until trained weights are deployed, predictions come from untrained/synthetic models. Either: (a) label the product a technical preview in the UI and API responses, or (b) deploy trained weights first (see notebooks/TRAINING_PIPELINE.md). Do not ship unlabeled demo output to NGOs/agencies as real monitoring.
  • Rotate the teammate Personal Access Tokens stored in .git/config (franchaise, olufemi, and previously hopelyn). They were printed in plaintext.

🟠 Important — launch day

  • CORS matches the domain. render.yaml sets CLIMATEVISION_CORS_ORIGINS=https://climatevision.green,https://www.climatevision.green. Confirm the deployed env has it and that the frontend is built with the right VITE_API_BASE_URL (empty = same-origin, which is correct here).
  • DNS → host. Point climatevision.green (and www) at Render/Fly and verify HTTPS (force_https = true is set in fly.toml).
  • Persistent storage is mounted. fly.toml mounts climatevision_data at /app/outputs and DATABASE_URL=sqlite:///app/outputs/.... Confirm the volume exists so run history + the NGO/alerts DB survive restarts. (On Render free tier, confirm the disk is attached — otherwise SQLite is ephemeral and you lose registrations on redeploy.)
  • Health checks green. Hit /api/health and /api/health/models on the live URL. /api/health/models will tell you whether real weights loaded.
  • Smoke test the core flow. One POST /api/predict and one dashboard load against the live site. Confirm /docs renders.
  • Merge the flood-detection PR (committed under your name) so the live build includes it.

🟢 Nice-to-have — this week

  • Deploy real trained weights for flooding first (best dataset fit: Sen1Floods11), then deforestation and ice. Follow TRAINING_PIPELINE.md.
  • Fill in the README benchmark table (IoU / F1) once a run passes the governance gate — replaces the "in progress" placeholders.
  • Add a status/uptime check and basic error monitoring.
  • Record dataset provenance.json + model cards for each deployed model (credibility for government adoption).
  • Push the 21 unpushed main commits and confirm CI is green on main.

Quick verification commands (run after deploy)

curl -s https://climatevision.green/api/health | jq
curl -s https://climatevision.green/api/health/models | jq   # are real weights loaded?
curl -s https://climatevision.green/api/analysis-types | jq

If /api/health/models shows untrained/demo, you are still in preview mode — which is fine, as long as the UI says so.