matchapi
An HTTP service is one function. match replaces the router.
A one-file Python web framework for granian over RSGI, granian's native Rust-backed interface: no router, no middleware, no ASGI adapter. Python 3.11+.
- Fast. On HTTP/1.1 within a few percent of a bare RSGI application written by hand, and 1.16 to 3.1 times the requests per second of Emmett, BlackSheep, Starlette, Litestar and FastAPI under the same granian. Performance.
- Small. One file of about 300 lines that reads in half an hour, pure Python, one dependency: granian.
- Tested. 100% line and branch coverage,
mypy --strict, a real granian in the test suite.
# app.py
import matchapi
@matchapi.app
async def handle(req):
match req.route:
case "GET", "ping":
return "pong"
case "GET", "users", user_id:
return {"id": int(user_id), "name": "ann"}
case "POST", "users":
body = req.json()
return 201, {"id": 7, "name": body["name"]}
case _:
return 404
$ pip install matchapi
$ granian app:handle
[INFO] Starting granian (main PID: 505)
[INFO] Listening at: http://127.0.0.1:8000
[INFO] Spawning worker-1 with PID: 512
[INFO] Started worker-1
[INFO] Started worker-1 runtime-1
$ curl localhost:8000/ping
pong
$ curl localhost:8000/users/42
{"id":42,"name":"ann"}
$ curl -i -X POST localhost:8000/users -d '{"name": "bob"}'
HTTP/1.1 201 Created
content-type: application/json
content-length: 21
server: granian
date: Mon, 12 Oct 2026 01:26:59 GMT
{"id":7,"name":"bob"}
$ curl -i localhost:8000/nothing
HTTP/1.1 404 Not Found
content-type: text/plain; charset=utf-8
content-length: 9
server: granian
date: Mon, 12 Oct 2026 01:26:59 GMT
Not Found
One import and one decorator; no route table, no schema classes, no dependency markers, no middleware stack. The
decorated handle is still the function it was: a test calls it with a request and compares the return value.
- Performance
- Why
- The model
- Running
- Testing
- Recipes
- What matchapi does not do
- Design notes
- Development
Performance
The README's service, written the way each framework writes it, under the same granian, one worker, oha with
64 connections; the multiplier is how many times matchapi's requests per second:
| Requests/s, HTTP/1.1, higher is better | GET /ping |
GET /users/42 |
POST /users |
|---|---|---|---|
| bare RSGI, no framework (the ceiling) | 45,264 | 42,993 | 24,432 |
| matchapi | 45,049 | 43,055 | 23,439 |
| Emmett 2.8 (RSGI) | 38,733 (1.16×) | 30,135 (1.43×) | 17,574 (1.33×) |
| BlackSheep 2.6 | 39,000 (1.16×) | 33,794 (1.27×) | 13,878 (1.69×) |
| Starlette 1.7 | 32,612 (1.38×) | 30,206 (1.43×) | 13,071 (1.79×) |
| Litestar 2.24 | 26,005 (1.73×) | 23,974 (1.80×) | 11,878 (1.97×) |
| FastAPI 0.143 | 17,819 (2.53×) | 13,940 (3.09×) | 8,501 (2.76×) |
matchapi parses nothing the handler does not read: route, query and headers are computed on first use, an
HTTP/1 request without a body is not waited on, and the commonest return values go to granian without an
intermediate object. Over HTTP/2 every request still awaits its body, and there Emmett and BlackSheep are ahead on
GET. bench/ has both protocols, the method and the command that reproduces the tables; one laptop,
Python 3.14, granian 2.8.4.
Why
Every Python web framework, from Flask to FastAPI, is a router with a decorator in front of it and a growing set of concepts behind it: path templates, parameter sources, dependency markers, routers, middleware stacks, response models. A service with three endpoints pays for all of it.
Python has had structural pattern matching since 3.10. A match over (method, *segments) is a router already:
it binds path parameters, it falls through to a default, it nests, and the language, not a library, defines how it
reads. matchapi is what is left of a framework once the router is gone: a Request, a Response, the rules that
turn a return value into one, and a decorator that lets granian
serve the function.
The model
One function
The service is a function that takes a Request and returns something. That is the whole contract. A function is
easy to test, easy to compose, easy to move: two services in one process are a function that calls two others.
@matchapi.app marks the function as the application granian runs. It changes nothing about the function:
handle(req) calls it, handle.__name__ and handle.__doc__ are its own. The one thing that differs is
inspect.iscoroutinefunction(handle), which is False for the decorated object; matchapi itself never asks it,
it looks at what a call returns.
req.route is the subject of match
req.route is a tuple of the method followed by the non-empty path segments: GET /users/42 is
("GET", "users", "42"), and GET / is ("GET",). A flat tuple matches without nesting:
match req.route:
case "GET", "users", user_id: # binds the segment
...
case "GET" | "HEAD", "health": # several methods
...
case _, "api", *rest: # a prefix: delegate the rest
return api(req, rest)
case _:
return 404
Segments are strings, so a numeric id is int(user_id) in the handler; a ValueError there is a 500, so raise
HTTPError(400) yourself when the input is the client's fault. req.method and req.path stay available as
strings for the cases where a plain comparison reads better.
The request is plain data
| Attribute | Type | Example for POST /users/42?q=x |
|---|---|---|
method |
str, upper case |
"POST" |
path |
str |
"/users/42" |
route |
tuple[str, ...] |
("POST", "users", "42") |
query |
dict[str, str], last value wins |
{"q": "x"} |
headers |
dict[str, str], lower-case names, last value wins |
{"content-type": "application/json"} |
body |
bytes, read in full before the call |
b'{"name": "bob"}' |
json() |
the body as JSON; a broken body is a 400 | {"name": "bob"} |
There is no scope or stream underneath: the body is read before the handler runs, so a sync handler needs no
await. Request("POST", "/users", body=b"...")
builds one by hand, which is how a test calls the handler; query takes the raw string or a dict.
Nothing is parsed that the handler does not read: route, query and headers are computed on first use, and
an HTTP/1 request without content-length or transfer-encoding has no body to wait for. On HTTP/1.1
bench/ cannot tell matchapi from a bare RSGI application answering the same routes by hand, and puts
it ahead of Emmett, BlackSheep, Starlette, Litestar and FastAPI under the same granian.
The return value is the response
| Return | Response |
|---|---|
dict or list |
200, application/json |
str |
200, text/plain; charset=utf-8 |
bytes |
200, application/octet-stream |
int (not bool) |
that status, the standard phrase as text: return 404 is Not Found; 1xx, 204 and 304 have no body |
None |
204, no body |
(status, body) |
the body as above with that status: return 201, {"id": 7} |
(status, body, headers) |
the same with extra headers: return 302, None, {"location": "/"} |
Response(status, body, headers) |
as given; a str body is text, an iterable or async iterable of bytes streams |
Anything else, a status outside 100-999 or a header value outside latin-1 is a 500 with the reason in the log.
JSON is compact and keeps non-ASCII text as is: exactly what
json.dumps(value, separators=(",", ":"), ensure_ascii=False) gives. A Response passed in a tuple is copied,
so a module-level NOT_FOUND = Response(404, "no") is safe to reuse.
Errors
HTTPError(status, detail=None) raised anywhere in the handler becomes a JSON response
{"detail": "..."} with that status; detail defaults to the standard phrase. req.json() raises it as a 400.
Any other exception is a 500 with the body Internal Server Error; the traceback goes to the matchapi logger
together with the method and path. Nothing of the exception reaches the client. An exception from the iterator of a streaming body is logged the
same way; the status is out already, so the stream just ends. A client that leaves in the middle of a stream is
granian's business, not an error of the handler.
Sync or async, the same way
def handle(req) and async def handle(req) are both fine. A sync handler runs in a thread, an async one in the
event loop. The kind of handler is read from the result of a call, not from the function, so a
functools.partial, a callable object or a compiled function all work, and a sync function that returns a
coroutine from some branches is awaited on those.
Running
The server is granian, the one dependency of the package; the decorated function is an RSGI application, granian's native interface.
$ granian app:handle # app.py in the working directory, its function `handle`
$ granian api.users:handle # a module in a package
$ granian app:handle --host 0.0.0.0 --port 80 --workers 4
$ granian app:handle --http 2 # HTTP/2 without TLS (h2c)
$ granian app:handle --reload # development; needs `pip install 'granian[reload]'`
$ python -m granian app:handle # the same command
--workers and --reload make every process import the module itself. The rest of the options, TLS, an event
loop, access logs, are granian's: granian --help.
Testing
The handler is a function and Request is a constructor, so a test needs no client and no server:
import asyncio
from matchapi import Request
from app import handle
def test_user() -> None:
assert asyncio.run(handle(Request("GET", "/users/42"))) == {"id": 42, "name": "ann"}
def test_create() -> None:
result = asyncio.run(handle(Request("POST", "/users", body=b'{"name": "bob"}')))
assert result == (201, {"id": 7, "name": "bob"})
def test_unknown() -> None:
assert asyncio.run(handle(Request("DELETE", "/users/42"))) == 404
With pytest-asyncio the tests are async def and await the handler directly. When the headers or the status
line matter, matchapi.respond(value) turns a return value into the Response the server would send.
Recipes
Authorization is a line before the match:
@matchapi.app
async def handle(req):
if req.headers.get("authorization") != f"Bearer {TOKEN}":
return 401
match req.route:
...
Two services in one process are two functions and a prefix:
@matchapi.app
async def handle(req):
match req.route:
case _, "users", *rest:
return await users(req, rest)
case _, "orders", *rest:
return await orders(req, rest)
case _:
return 404
Validation is whatever you like; a dataclass is enough for a simple body:
from dataclasses import dataclass
import matchapi
from matchapi import HTTPError
@dataclass
class NewUser:
name: str
age: int = 0
@matchapi.app
async def handle(req):
match req.route:
case "POST", "users":
try:
user = NewUser(**req.json())
except TypeError as error:
raise HTTPError(400, str(error)) from error
return 201, {"name": user.name}
A redirect, a custom content type, a stream:
return 302, None, {"location": "/login"}
return Response(200, "<h1>hi</h1>", {"content-type": "text/html"})
return Response(200, (chunk for chunk in lines()))
What matchapi does not do
No router or decorators for routes, no middleware stack, no parameter conversion from type hints, no OpenAPI, no cookies or sessions, no templates, no static files, no websockets, no streaming request bodies, no ASGI or WSGI. Each of these is either a few lines in the handler or not matchapi's job: the server is granian and the only interface is RSGI, so there is no middleware layer to plug into.
The line is deliberate. The framework fits in one file of about 300 lines that reads in half an hour, and the handler stays a function that works without a server.
Design notes
The decisions and the alternatives that were rejected are in docs/design.md: why the router went instead of the signature magic, why a decorator rather than a command, why the body is read before the call, why there is only RSGI through granian, and why there is no compiled core. CONTEXT.md is the glossary. bench/ measures matchapi against a bare RSGI application and against other frameworks.
Development
$ uv sync
$ uv run pytest --cov
$ uv run ruff check . && uv run ruff format --check . && uv run mypy
Coverage is kept at 100%. A pull request bumps the version in pyproject.toml and adds a CHANGELOG.md entry;
a merge into master publishes that version to PyPI and tags it.
Metadata
Release files for matchapi 0.3.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| matchapi-0.3.0.tar.gz | 31.3 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| matchapi-0.3.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 43.2 kB
Release files / matchapi-0.3.0.tar.gz
| Download URL | matchapi-0.3.0.tar.gz |
|---|---|
| Size | 31.3 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
e0b2113be8b4f961c40a5c7b79bfa75e00f519437ffb9e9ca33bc47eb28d201b
|
|
BLAKE2b-256 checksum How to use checksums |
f2c886b2860efd550c2968b2973068d674c8f8f2a2fd8db53dc3951f7eb655a3
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 12, 2026.
Transparency logRelease files / matchapi-0.3.0-py3-none-any.whl
| Download URL | matchapi-0.3.0-py3-none-any.whl |
|---|---|
| Size | 11.9 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
0e1723ad88cdc62c92b8ac079507628f32100397441be111ff6c233be821be8b
|
|
BLAKE2b-256 checksum How to use checksums |
8e23885268ddca5b62ff71f9f4b6f45c1ca059abde1c61f307d88823262e4c65
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 12, 2026.
Transparency log