NoLoop Dev Log — Migrating the Backend from NestJS to FastAPI

Big day on NoLoop. Here's everything that shipped.

Kicked Off the Backend Migration: NestJS → FastAPI

I decided to move the entire backend to Python — FastAPI + SQLAlchemy 2.0 async — so the whole data + AI layer lives in one ecosystem. One rule for the migration: the frontends must not notice. Same routes, same response shapes, same JWTJWTJSON Web Token — a compact, signed token carrying claims (such as a user ID and an expiry time) that a server can verify without a session lookup. contract.

Four PRs merged today, each verified against the live system before merging:

• PR #1 — Scaffold: FastAPI app, pydantic-settings config, async engine on Supabase's pooler, and a /health endpoint with a real DB round-trip.
• PR #2 — Models: all 12 tables + 11 enums mapped 1:1 onto the existing Prisma-created schema — zero DB migration, verified by querying every table live.
• PR #3 — Auth: signup/login/me, JWT with an identical payload, and passlib verifying the old bcryptjs hashes so existing users keep logging in. Error bodies are NestJS-shaped.
• PR #4 — Org + Admin: ran both backends against the same DB with the same tokens and diffed responses — 8/8 byte-identical, including 401/403 error shapes.

PR #5 (claims + AI client) is code-complete and pushed — read-parity passed 7/7 (all 56 claims, full detail shapes, public tracking) — held open until the mutation tests run.

Gotchas Earned Today

• pgbouncerPgBouncerA lightweight connection pooler for PostgreSQL that lets many clients share a few database connections. In transaction-pooling mode, session features such as prepared statements need care. in transaction mode needs NullPool and unique prepared-statement names — statement_cache_size=0 alone is not enough.
• Naive datetimes without a trailing Z silently shift every timestamp by 5½ hours in the browser.
• Error bodies are a contract. My frontend reads data.message, so FastAPI now speaks Nest's error dialect.

Planned the Whole Road Ahead — 14 Detailed GitHub Issues

Everything from finishing the migration to production hardening is now on the board:

• Async task queue — retries, idempotency, DLQs
• The RAGRAGRetrieval-Augmented Generation — fetching relevant passages from a document store at query time and handing them to a language model as context, so its answer is grounded in those sources. centerpiece — pgvectorpgvectorA PostgreSQL extension that adds a vector column type and similarity search, so embeddings can live alongside ordinary relational data., hybrid retrieval, citation-grounded answers, an eval set
• Render + Vercel deploy, observability, rate limiting, CI/CD
• DB scaling + a backup restore drill, k6 load testing, Redis caching

Deployed the Admin Dashboard 🚀

Live on Vercel with a custom domain set up via the CLI: noloop-admin.sandeepp.in. Proper README + MIT license added. All changes shipped through 3 small branch → PR → merge cycles.

The Multi-Repo Scare

Also survived a classic: I accidentally worked in a fork of the team's org repo instead of my personal one. Caught it, swapped to the right repo, redid the work cleanly — and wrote down the rule:

Check git remote -v before pushing. Always pass --repo to gh pr create.