Benchmarks¶

Django-Bolt is the fastest Python web framework we have measured: ~311,000 requests/second on a JSON hello-world from a single 12-core desktop, ~187,000 req/s for a 10 KB JSON body, and 21,000–27,000 req/s for a 10-row Django ORM query. Every number below states its conditions and can be reproduced with one command.

Reference setup¶

Because bombardier shares the 12 cores with the 8 server processes, the load generator is part of the system under test. Read the numbers as a floor, not a ceiling.

Results by endpoint¶

The ORM rows are bounded by SQLite's single-writer file lock and by the ORM's own Python cost, not by Bolt; on PostgreSQL the ORM executor uses more threads. The middleware row runs the entire settings.MIDDLEWARE chain (sessions, CSRF, auth, messages, CSP) in Python for every request — it is there to show that nothing is stripped, and what the full stack costs.

Streaming: Server-Sent Events¶

10,000 concurrent SSE clients held open for 60 seconds:

Against JavaScript runtimes¶

Same JSON payloads, same machine, same load generator, byte-identical response bodies checked on every run. C=100, N=100000.

8 processes each (Bolt: --processes 8; Hono/Node: node:cluster; Bun: 8 processes with reusePort):

Single process:

Reading: Elysia leads on tiny payloads at one process (1.2×) and ties Bolt at eight (1.05×, inside the noise). Bolt leads on 10 KB at both scales (1.17× / 1.27×) — once serialization and body handling dominate, msgspec plus the zero-copy response path win. A Python framework beating Bun-native servers on realistic payload sizes is the headline here.

Against Python frameworks¶

A separate harness, python-api-frameworks-benchmark, runs Django-Bolt, Litestar, and FastAPI against the same seven endpoints and the same PostgreSQL data (2026-08-30, same machine, Python 3.14, one process each). Each framework runs alone in a Docker container with a 750 MB memory limit and no CPU limit. bombardier runs on the host with C=100 for 10 s per endpoint. Each framework starts twice; the table shows the median.

Django-Bolt uses the Django ORM with psycopg 3 on a bounded thread pool. Litestar and FastAPI use SQLAlchemy with asyncpg. All three use the same pool size (8). Litestar and Django-Bolt serialize with msgspec; FastAPI serializes with Pydantic. The ORM rows compare two database stacks as well as two frameworks. Read them with that in mind.

Reproduce¶

git clone https://github.com/dj-bolt/django-bolt.git && cd django-bolt
uv sync && just build
go install github.com/codesenberg/bombardier@latest

just save-bench                    # full suite → python/benchmark/BENCHMARK.md
just save-bench 127.0.0.1 8001 100 100000 8   # host port C N processes

Cross-framework table (Docker, PostgreSQL, one process per framework):

git clone https://github.com/FarhanAliRaza/python-api-frameworks-benchmark.git && cd python-api-frameworks-benchmark
./run_all.sh --docker --frameworks bolt litestar-uvicorn fastapi-uvicorn --restarts 2

The app code for each framework is in the repository root: django_project/api.py (Bolt), litestar_app.py, and fastapi_app.py.

How to quote these numbers¶

• ✅ "Django-Bolt: ~311k req/s JSON hello-world, 8 processes, C=100, Ryzen 5 5600G, loopback bombardier."
• ✅ "Django-Bolt beats Bun/Elysia by 1.27× on 10 KB JSON at 8 processes (same machine)."
• ❌ "Django-Bolt does 300k RPS" with no conditions.
• ❌ Comparing to numbers measured on a different machine or day.