render.yaml

# Render blueprint for the NOVA FastAPI backend.
#
# Deploy: push to GitHub, then on Render choose New + > Blueprint and point it
# at this repository.
#
# WHY DOCKER RATHER THAN runtime: python
# requirements.txt pins torch==2.5.1, and installing that from PyPI pulls the
# CUDA build, roughly 2.5GB of libraries this CPU-only service never uses. That
# alone exceeds the free tier's build and image budget. The Dockerfile installs
# CPU-only torch from download.pytorch.org first, so the rest of the file
# resolves against a wheel that is already satisfied. Building through Docker
# is what makes the free plan possible at all here.
#
# WHY MAX_ROWS IS 6000
# The free instance has 512MB. Measured on 2026-09-04 with
# scripts/measure_memory.py, resident set size:
#
#   torch imported            197 MB
#   + sklearn, pandas, scipy  292 MB
#   + fastapi, uvicorn        306 MB
#   model loaded, idle        334 MB
#   peak during 1000 rows     388 MB
#   peak during 3000 rows     400 MB
#   peak during 6000 rows     446 MB   <- the cap, 66 MB under the limit
#
# So 6000 fits and is not a guess, but the margin is thin: torch alone is more
# than half the budget before a single row is generated. Raising the cap needs
# a larger instance, not a smaller batch.
#
# COLD STARTS
# Free instances sleep after inactivity and take roughly 50 seconds to wake.
# The frontend has to tolerate a slow first request rather than treat it as a
# failure.
services:
  - type: web
    name: nova-backend
    runtime: docker
    plan: free
    rootDir: backend
    dockerfilePath: ./Dockerfile
    healthCheckPath: /api/health
    envVars:
      # Tighten to the Vercel origin once the URL is settled. Left open here so
      # the first deploy is not blocked by a CORS mismatch.
      - key: CORS_ORIGINS
        value: "*"
      - key: MAX_ROWS
        value: "6000"