ORM Benchmarks

A benchmark suite comparing the end-to-end HTTP performance of several PostgreSQL access strategies from Node.js:

• Prisma 8 RC (@prisma/orm-postgres)
• Prisma 7 (@prisma/client)
• Drizzle (drizzle-orm)
• Raw pg + raw SQL (baseline)

Each implementation exposes the same 13 HTTP endpoints (paginated lookups, single-record lookups, full-text search, and relational reads) against the same PostgreSQL database. k6 drives the load; results are combined into data.json and rendered by visualization.html.

This repository is a fork of drizzle-team/drizzle-benchmarks.

Requirements

• Docker — to run PostgreSQL
• k6 — load generator
• DuckDB — used by the result-preparation pipeline
• Node.js ≥ 22.18 (required by Prisma 8; the project is developed against Node 24) and pnpm

Setup

pnpm install

1. Start PostgreSQL

pnpm start:docker

The container's port is configured in src/docker.ts (default 5432). Make sure .env points at it:

DATABASE_URL="postgres://postgres:postgres@localhost:5432/postgres"

2. Seed the database

pnpm start:seed

Dataset size is configured in src/seed.ts (nano | micro).

3. Generate the shared request list

pnpm start:generate

This writes ./data/requests.json — the deterministic list of HTTP requests that k6 replays against every server.

Running each benchmark

Each server is clustered across all CPU cores and listens on its own port, so only one at a time can be running. For every implementation you want to benchmark, repeat the following loop:

1. Start the server (table below).
2. From the load-generator machine (or another shell on the same machine), run pnpm tsx bench/index against it.
3. Stop the server.

Prisma 8 RC — port 3002

# Server (one-time contract codegen, then start)
pnpm prepare:prisma-next
pnpm start:prisma-next

# Benchmark
pnpm tsx bench/index --host http://<host>:3002 --name prisma-next --folder results

Prisma 7 — port 3001

# Server (one-time codegen, then start)
pnpm prepare:prisma
pnpm start:prisma

# Benchmark
pnpm tsx bench/index --host http://<host>:3001 --name prisma7 --folder results

Drizzle — port 3000

# Server
pnpm start:drizzle

# Benchmark
pnpm tsx bench/index --host http://<host>:3000 --name drizzle --folder results

Raw pg + raw SQL — port 3003

# Server
pnpm start:pg

# Benchmark
pnpm tsx bench/index --host http://<host>:3003 --name pg --folder results

The --name argument is the key used in data.json and in visualization.html. Use exactly drizzle, pg, prisma7, or prisma-next if you want runs to show up in the bundled dashboard.

Each run takes ~10 minutes (k6 ramps from 200 → 3000 VUs across the stages defined in bench/bench.js) and produces two files in results/:

• results/<name>.parquet — k6 metrics
• results/cpu-usage-<name>.csv — server CPU samples

Combine results and view

After all the runs you want are in results/, merge them into a single data.json:

pnpm tsx bench/prepare --folder results

Then open visualization.html from any static file server in this directory, for example:

npx serve .
# then open http://localhost:3000/visualization.html

The dashboard reads data.json and renders one series per implementation configured in the colors map inside visualization.html.

Notes

• For maximum accuracy, run the server and the load generator on two separate machines connected over a low-latency link so the load generator doesn't compete for CPU with the server under test.
• Only one server should be running at a time — each forks across every available CPU core.