astraapi 0.1.15

AstraAPI framework, high performance, easy to learn, fast to code, ready for production

⚡ AstraAPI

High-Performance · Easy to Learn · Fast to Code · Ready for Production

---

AstraAPI is a modern, production-grade Python web framework with a C++20 hot-path core. It is inspired by and fully API-compatible with FastAPI — you get the same beautiful decorator-based routing, automatic OpenAPI docs, Pydantic v2 validation, Depends injection, and WebSocket support, while the inner engine is a compiled C++ extension that handles HTTP parsing, route matching, parameter extraction, JSON serialization, CORS, compression, and response building — with zero Python overhead on the critical path.

"AstraAPI = FastAPI's developer experience + uWebSockets' throughput + Node.js cluster's multi-worker model"

---

✨ Key Features

Feature	Description
🚀 C++20 HTTP Core	HTTP parsing, route matching, param extraction, JSON encode/decode — all in compiled C++ (_astraapi_core.so)
🎯 FastAPI-Compatible API	Same decorators (@app.get, @app.post, …), Depends, Body, Query, Path, Header, Cookie, Form, File, Security
🔄 Zero-Lock Multi-Worker	Each worker is a fully independent OS process with its own GIL, event loop, and memory space. Zero shared state = zero lock contention
🌀 uvloop / winloop Event Loop	Uses uvloop (Linux/macOS) or winloop (Windows) instead of CPython's default asyncio loop — up to 2–4× more I/O throughput
📄 Auto OpenAPI + Swagger/ReDoc	Built-in /docs (Swagger UI) and /redoc (ReDoc) — served directly from C++ pre-built byte buffers
✅ Pydantic v2 Validation	Full request body validation, response model serialization, and model_dump_json — all via Pydantic v2
🛡️ Built-in Middleware	CORS, TrustedHost, GZip/Brotli compression, Rate Limiting — implemented in C++ with zero Python allocation per request
🔌 WebSocket Support	RFC 6455 compliant WebSocket with C++ ring-buffer frame parsing, echo auto-detection, batch frame building, and per-connection backpressure
⚙️ Lazy Imports	Heavy imports (Pydantic, OpenAPI models) are deferred until first access. Cold startup reduced from ~4.7 s → ~1–1.5 s
🔁 Hot Reload	File-watching reloader (watchfiles) restarts workers on code change
🧪 Test Client	Drop-in TestClient based on httpx for sync/async endpoint testing
📁 Static Files + Jinja2 Templates	Plug-and-play static file serving and HTML templating
🔒 Security	OAuth2, HTTP Basic/Bearer, API Key — mirrors FastAPI's security utilities
🪝 Lifespan Events	on_startup / on_shutdown callbacks and async context-manager lifespan
🖥️ Cross-Platform	Linux (SO_REUSEPORT + uvloop), macOS (fork + uvloop), Windows (subprocess + winloop + socket.share)

---

🌟 Inspired by FastAPI

AstraAPI was directly inspired by FastAPI created by Sebastián Ramírez. Every public-facing API — decorators, dependency injection, parameter declarations, OpenAPI generation, security utilities, response models — is intentionally compatible with FastAPI so that existing FastAPI applications can be migrated with minimal changes.

# This is valid AstraAPI code — and also valid FastAPI code
from astraapi import AstraAPI, Depends
from pydantic import BaseModel

app = AstraAPI(title="My API")

class Item(BaseModel):
    name: str
    price: float

@app.post("/items/", response_model=Item)
async def create_item(item: Item):
    return item

What AstraAPI adds on top of the FastAPI mental model:

• A compiled C++20 extension replaces the entire ASGI hot path
• A built-in multi-worker supervisor replaces Gunicorn/Uvicorn
• A uvloop/winloop event loop replaces the default asyncio loop
• Zero-allocation per-request interning for HTTP method strings, header names, status lines

---

📦 Installation

pip install astraapi

Note: The package ships a pre-built _astraapi_core.so (Linux) / .pyd (Windows). To build from source:

cd cpp_core && mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)
cp _astraapi_core*.so ../../astraapi/_astraapi_core.so

---

🚀 Quick Start

from astraapi import AstraAPI

app = AstraAPI()

@app.get("/")
async def root():
    return {"message": "Hello from AstraAPI ⚡"}

# Run with built-in multi-worker server
if __name__ == "__main__":
    app.run(host="0.0.0.0", port=8000, workers=4)

---

🔄 Complete Request Flow

The following diagram shows the full lifecycle of an HTTP request inside AstraAPI:

┌─────────────────────────────────────────────────────────────────────────────┐
│                          CLIENT  (Browser / HTTP Client)                    │
└────────────────────────────────────┬────────────────────────────────────────┘
                                     │  TCP Connection
                                     ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                   MASTER PROCESS  (Supervisor / Accept Thread)              │
│                                                                             │
│  ① listen_sock.accept()  — single blocking accept, no thundering herd      │
│  ② Round-robin dispatch  — SCM_RIGHTS (Linux) / socket.share() (Windows)  │
│  ③ Worker restart        — crashed workers are automatically re-spawned    │
└────────────┬──────────────────────────┬────────────────────────────────────┘
             │ fd / shared socket        │ fd / shared socket
      ┌──────▼──────┐            ┌───────▼──────┐
      │  Worker 0   │            │  Worker 1    │   ... (N workers)
      │  (Process)  │            │  (Process)   │
      │  own GIL    │            │  own GIL     │
      │  own loop   │            │  own loop    │
      └──────┬──────┘            └──────────────┘
             │
             ▼  CppHttpProtocol.data_received()
┌─────────────────────────────────────────────────────────────────────────────┐
│                   C++ HTTP LAYER  (_astraapi_core.so)                       │
│                                                                             │
│  ④ HTTP Parse      — llhttp (Node.js parser, zero-copy view)               │
│  ⑤ Route Match     — Radix trie, O(log n), pre-compiled at startup         │
│  ⑥ Param Extract   — Path params, query string, headers, cookies in C++    │
│  ⑦ JSON Parse      — yyjson (SIMD-accelerated, strict mode)                │
│  ⑧ CORS Check      — case-insensitive origin matching, zero allocation     │
│  ⑨ DI Resolution   — Resolve Depends graph, inject typed params in C++     │
│  ⑩ Compression     — libdeflate (gzip) / Brotli, chosen by Accept-Encoding │
│  ⑪ Rate Limiting   — Sharded mutex counters (16 shards), per-IP           │
└──────────────┬──────────────────────────┬──────────────────────────────────┘
               │                          │
   ┌───────────▼─────────┐    ┌───────────▼─────────────┐
   │  SYNC ENDPOINT      │    │  ASYNC / PYDANTIC        │
   │  (no Python touch)  │    │  ENDPOINT                │
   │                     │    │                          │
   │  C++ builds full    │    │  C++ returns InlineResult│
   │  HTTP response and  │    │  Python: await endpoint  │
   │  calls              │    │  Pydantic validates resp │
   │  transport.write()  │    │  C++ serializes JSON     │
   └─────────────────────┘    └──────────────────────────┘
               │                          │
               └──────────┬───────────────┘
                          ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                      RESPONSE PATH                                          │
│                                                                             │
│  C++ build_response_from_parts() → pre-cached status lines + headers →     │
│  transport.write(bytes)  — single syscall, keep-alive maintained           │
└─────────────────────────────────────────────────────────────────────────────┘
                          │
                          ▼
                      CLIENT  ✅

Request Flow — Step by Step

1. TCP Accept — The master supervisor calls accept() once and dispatches the file descriptor to a worker via Unix socket (SCM_RIGHTS on Linux) or socket.share() (Windows). Workers never compete for connections.
2. data_received — Python's asyncio protocol (CppHttpProtocol) receives raw bytes and calls into _astraapi_core via a single C-extension call.
3. HTTP Parse — C++ uses the llhttp parser (same as Node.js) to parse headers and body. Strings are interned (PyUnicode_InternFromString) once and reused forever.
4. Route Match — A compiled radix trie matches the method + path in O(log n). Routes are frozen before workers fork — shared as read-only COW pages.
5. Parameter Extraction — Path params, query strings, headers, cookies extracted in C++ with zero Python dict construction for sync routes.
6. Endpoint Dispatch — Sync endpoints return directly; async endpoints yield an InlineResult capsule back to Python which awaits the coroutine.
7. Pydantic Validation — Response models call model_dump_json() — result is passed back to C++ for final serialization.
8. Response Write — C++ assembles the full HTTP response from pre-cached byte fragments and calls transport.write() — a single syscall.

---

🏗️ Worker Architecture

┌──────────────────────────────────────────────────────────────────────┐
│                        ASTRAAPI  PROCESS MODEL                       │
├──────────────────────────────────────────────────────────────────────┤
│                                                                      │
│   ┌─────────────────────────────────────────────────────────────┐   │
│   │              MASTER  (Supervisor Process)                    │   │
│   │                                                              │   │
│   │  • listen_sock.accept() ──► round-robin dispatch           │   │
│   │  • Monitors child PIDs via os.waitpid()                    │   │
│   │  • Auto-restarts crashed workers                           │   │
│   │  • Tunes sysctl (somaxconn, tcp_max_syn_backlog)           │   │
│   │  • Raises fd limit (RLIMIT_NOFILE → 65535)                 │   │
│   └──────────┬──────────────┬────────────────┬─────────────────┘   │
│              │              │                │                       │
│    SCM_RIGHTS│   socket.    │  SO_REUSEPORT  │                       │
│    (Linux)   │   share()    │  (kernel does  │  + ─ ─  N workers    │
│              │  (Windows)   │   balancing)   │                       │
│    ┌─────────▼──┐  ┌────────▼──┐  ┌──────────▼──┐                  │
│    │  Worker 0  │  │  Worker 1  │  │  Worker N   │                  │
│    │            │  │            │  │             │                  │
│    │ own GIL    │  │  own GIL   │  │  own GIL    │                  │
│    │ own memory │  │ own memory │  │ own memory  │                  │
│    │ own loop   │  │ own loop   │  │ own loop    │                  │
│    │ uvloop/    │  │ uvloop/    │  │ uvloop/     │                  │
│    │ winloop    │  │ winloop    │  │ winloop     │                  │
│    │            │  │            │  │             │                  │
│    │ CPU pinned │  │ CPU pinned │  │ CPU pinned  │ ← sched_setaff. │
│    │ (core 0)   │  │ (core 1)   │  │ (core N%k)  │                  │
│    └────────────┘  └────────────┘  └─────────────┘                  │
│                                                                      │
│  • Route table frozen before fork → shared as read-only COW pages   │
│  • Zero IPC after startup (SO_REUSEPORT mode)                       │
└──────────────────────────────────────────────────────────────────────┘

Linux — Three Modes (Best → Fallback)

Mode	How it works	Latency
SO_REUSEPORT	Each worker binds its own socket; kernel dispatches SYN packets directly — no master accept thread, no IPC	Lowest
Master-Accept + SCM_RIGHTS	Master calls accept(), sends fd to worker via sendmsg + SCM_RIGHTS over Unix socketpair	Low
Shared socket	Single listen socket shared via socket.share()	Moderate

Windows

Master uses socket.share(pid) + socket.fromshare() over AF_INET socketpairs. Same round-robin guarantee, no thundering herd.

---

⚡ Why AstraAPI is Faster Than Uvicorn + Gunicorn

The table below compares what happens on every single request:

Component	Uvicorn + Gunicorn	AstraAPI
HTTP Parsing	Python ASGI scope dict allocation (7+ keys)	C++ llhttp, zero Python dict
Route Matching	Python regex, dictionary lookup per route	C++ radix trie, O(log n), pre-compiled
Param Extraction	Python re.match() + dict construction	C++ param_extractor, direct struct write
JSON Decode	Python json.loads() / orjson.loads()	C++ yyjson SIMD parser
JSON Encode	Python json.dumps() / orjson.dumps()	C++ yyjson writer, single allocation
CORS	Python middleware, dict lookups, .encode() calls	C++ case-insensitive origin match, no alloc
GZip	Python gzip module (GIL-bound)	C++ libdeflate (2–3× faster than zlib)
Brotli	Requires separate brotli middleware	C++ Brotli native, auto-negotiated
Response Build	Python bytearray + string formatting	C++ pre-cached status lines + single memcpy
Worker Model	Gunicorn master-accept → WSGI/ASGI overhead	Zero-lock independent processes, no arbiter GIL
Event Loop	Default asyncio (libuv-less)	uvloop / winloop (libuv-based, 2–4× faster I/O)
Startup Time	FastAPI ~4–5 s cold import	AstraAPI ~1–1.5 s (lazy imports)
String Interning	.encode() per request for method/headers	Pre-interned PyUnicode_InternFromString at startup
Memory per request	Multiple dict + list allocations in Python	Reused C++ buffer pool (buffer_pool.cpp)

Key Architectural Advantages

1. C++ on the hot path, Python only where needed

Sync endpoint:   TCP bytes → C++ parse → C++ route → C++ params → C++ response  (zero Python)
Async endpoint:  TCP bytes → C++ parse → C++ route → C++ params → Python await → C++ serialize

2. Zero thundering herd

Gunicorn's default pre-fork model has all workers call accept() simultaneously, causing a kernel stampede on connection arrival. AstraAPI's master process accepts once and dispatches — workers consume from a queue with guaranteed even distribution (Node.js cluster SCHED_RR pattern).

3. Per-worker CPU affinity

os.sched_setaffinity(0, {worker_id % cpu_count})

Each worker is pinned to a dedicated CPU core, eliminating L2/L3 cache thrashing and TLB flushes caused by process migration across cores.

4. Pre-warmed buffer pool

A C++ buffer_pool pre-allocates std::vector<char> buffers at startup. Per-request allocation becomes an O(1) list pop — no malloc/free on the critical path.

5. Route table frozen and COW-shared

Before forking, app._sync_routes_to_core() freezes the radix trie. All workers inherit it as read-only copy-on-write pages — never copied, never locked.

6. Lazy imports cut startup time by 3–4×

# Heavy imports (pydantic, openapi, routing) deferred to first use
# `from astraapi import AstraAPI` → ~1-1.5s  (vs ~4.7s eager)

---

🔌 WebSocket Architecture

Client  ──WS Frame──►  C++ ws_frame_parser (ring buffer, RFC 6455)
                            │
                            ▼
                    _WsFastChannel.feed()  ──► Python endpoint await
                            │
                  (echo detected?) ──YES──►  _handle_ws_frames_echo_fd
                            │                 (direct FD write, no Python)
                           NO
                            ▼
                    Python endpoint processes frame
                            │
                    ws.send_text() / send_bytes() / send_json()
                            │
                    C++ ws_build_frame_bytes()  (single allocation)
                            │
                    transport.write()

• Echo auto-detection: if an endpoint echoes received data, AstraAPI switches to a direct-fd echo handler that bypasses Python entirely
• Batch frame building: multiple sends within the same event loop tick are coalesced into a single transport.write() via _ws_build_frames_batch
• Backpressure: pauses transport reading when buffer > 256 messages or 8 MB, resumes at 64 messages / 2 MB
• WebSocket Groups: broadcast to N connections using pre-built frames (_ws_groups.py) — frame built once, sent to all

---

🧩 Project Structure

astraapi/
├── __init__.py              # Lazy-import gateway (startup: ~1.5s)
├── applications.py          # AstraAPI class — main app entrypoint
├── routing.py               # APIRouter, APIRoute, path operation decorators
├── _cpp_server.py           # asyncio.Protocol bridge to C++ core
├── _multiworker.py          # Zero-lock multi-worker supervisor
├── _astraapi_core.so        # Compiled C++20 extension
├── _request.py              # Request object
├── _response.py             # Response classes (JSON, HTML, Stream, File…)
├── _routing_base.py         # Base routing primitives
├── _middleware_impl.py      # Middleware stack builder
├── _websocket.py            # WebSocket wrapper (CppWebSocket)
├── _ws_groups.py            # WebSocket group broadcast
├── param_functions.py       # Body, Query, Path, Header, Cookie, Form, File, Depends, Security
├── dependencies/            # DI resolver
├── openapi/                 # OpenAPI schema generation
└── security/                # OAuth2, HTTP Basic/Bearer, API Key

cpp_core/
├── src/
│   ├── app.cpp              # CoreApp, HTTP handler, InlineResult
│   ├── router.cpp           # Radix trie route matching
│   ├── http_parser.cpp      # llhttp wrapper
│   ├── param_extractor.cpp  # Path/query/header param extraction
│   ├── json_writer.cpp      # yyjson-based JSON serialization
│   ├── json_parser.cpp      # yyjson-based JSON parsing
│   ├── ws_frame_parser.cpp  # RFC 6455 WebSocket frame parser
│   ├── ws_ring_buffer.cpp   # Ring buffer for WS frames
│   ├── middleware_engine.cpp # CORS, TrustedHost in C++
│   ├── body_parser.cpp      # JSON/form body parsing
│   ├── response_pipeline.cpp# Response assembly
│   └── dependency_resolver.cpp # C++ DI graph resolution
├── third_party/
│   ├── llhttp/              # Node.js HTTP parser
│   ├── yyjson/              # Fast JSON (SIMD)
│   └── ryu/                 # Fast float→string (d2s)
└── CMakeLists.txt           # C++20, -O3 -march=native -flto

---

⚙️ Configuration

app = AstraAPI(
    title="My API",
    version="1.0.0",
    description="API description (Markdown supported)",
    docs_url="/docs",          # Swagger UI
    redoc_url="/redoc",        # ReDoc
    openapi_url="/openapi.json",
    root_path="/api/v1",       # For reverse proxy setups
)

app.run(
    host="0.0.0.0",
    port=8000,
    workers=4,         # Number of worker processes
    reload=False,      # Enable hot reload (dev only)
)

---

📚 Dependencies

Package	Role
pydantic ≥ 2.7	Request/response validation and serialization
uvloop (Linux/macOS)	High-performance asyncio event loop
winloop (Windows)	High-performance asyncio event loop for Windows
orjson	Fast JSON fallback for Python-side serialization
watchfiles	File change detection for hot reload
annotated-doc	Rich type annotation documentation

Optional (install via pip install astraapi[standard]): httpx, jinja2, python-multipart, email-validator, pydantic-settings, pydantic-extra-types

---

🤝 Contributing

See CONTRIBUTING.md for how to set up the development environment, build the C++ core, and run the test suite.

---

📄 License

MIT License — see LICENSE.

---

Made with ⚡ by Lumos Labs HQ · Inspired by FastAPI