[CI]【Hackathon 10th Spring No.46】Windows Python runtime guards (3/3)

[r-cloudforge]

Motivation

Several FastDeploy Python modules use POSIX-only APIs that crash on Windows:

• /dev/shm paths (unavailable on Windows — must use tempfile.gettempdir())
• os.killpg() (not available on Windows — must use os.kill())
• os.setsid as preexec_fn (not available on Windows)
• multiprocessing.get_context("fork") (fork unavailable — must use "spawn")

This is Part 3 of 3 for 【Hackathon 10th Spring No.46】(Windows Build Support):

• Part 1: C++ #ifndef _WIN32 guards on POSIX-only includes
• Part 2: Build system — setup_ops.py platform-conditional link args + build.bat
• Part 3 (this PR): Python runtime — platform guards for /dev/shm, os.killpg, os.setsid, fork

Modifications

10 Python files modified with sys.platform guards:

File	Guard Added
fastdeploy/engine/common_engine.py	/dev/shm → tempfile.gettempdir() on Windows; preexec_fn=os.setsid → None on Windows
fastdeploy/engine/engine.py	os.killpg → os.kill on Windows; fork → spawn on Windows
fastdeploy/engine/expert_service.py	/dev/shm → tempfile.gettempdir() on Windows
fastdeploy/worker/worker_process.py	fork → spawn on Windows
fastdeploy/cache_manager/cache_messager.py	/dev/shm path guard
fastdeploy/cache_manager/prefix_cache_manager.py	/dev/shm path guard
fastdeploy/eplb/async_expert_loader.py	/dev/shm path guard
fastdeploy/inter_communicator/fmq.py	/dev/shm path guard
fastdeploy/inter_communicator/zmq_client.py	/dev/shm path guard
fastdeploy/inter_communicator/zmq_server.py	/dev/shm path guard

All guards follow the pattern:

if sys.platform == "win32":
    # Windows-safe alternative
else:
    # Original POSIX code

Usage or Command

# On Windows:
import fastdeploy
# Previously: ImportError on os.setsid / /dev/shm access
# Now: gracefully uses Windows-compatible alternatives

No change needed on Linux — guards are win32-only.

Accuracy Tests

Not applicable — platform guard additions only. No algorithmic changes.

Verified via python -c "import fastdeploy" on Windows Server 2022 (no CUDA) —
module loads without POSIX-related ImportError/OSError.

Checklist

• All POSIX-only calls guarded with sys.platform == "win32" check
• Windows alternatives are functionally equivalent
• No behavioral change on Linux
• pre-commit hooks pass (black, isort, flake8, ruff)

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

cloudforge1 seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[paddle-bot]

Thanks for your contribution!

[fastdeploy-bot]

🤖 AI Code Review | 2026-04-11

📋 Review 摘要

PR 概述：为 FastDeploy Python 模块添加 Windows 平台运行时兼容性支持，处理 POSIX-only API（/dev/shm、os.killpg、os.setsid、fork）的跨平台适配

变更范围：10 个 Python 文件添加 sys.platform 平台判断

影响面 Tag：[CI]

PR 规范检查

✅ 标题包含有效 Tag [CI]
✅ 描述包含 Motivation、Modifications、Usage、Checklist 等必要章节

---

问题

级别	文件	概述
🔴 Bug	fastdeploy/inter_communicator/ipc_signal.py:172	Windows 文件锁只锁定 1 字节，破坏互斥性

---

总体评价

PR 整体设计合理，为 Windows 平台兼容性做了全面的平台判断处理。大部分修改（如 /dev/shm 路径替换、os.killpg 替换为 os.kill、fork 替换为 spawn）都是正确的。

但 IPCLock 的 Windows 实现存在严重 bug，在多进程环境下无法正确实现互斥锁功能，需要修复。

[fastdeploy-bot]

🔴 Bug Windows 文件锁实现不正确。

msvcrt.locking(fd, mode, nbytes) 的第三个参数 nbytes 表示要锁定/解锁的字节数。当前代码设置为 1，意味着只锁定文件的第 1 个字节。

这会导致在多进程环境下，多个进程可能同时访问同一个锁文件的不同部分，破坏互斥性。

建议修复：

使用 os.lseek() 获取文件大小，然后锁定整个文件：

def acquire(self) -> None:
    """Acquire the lock (blocking). Uses kernel-level flock for atomicity."""
    if sys.platform == "win32":
        import msvcrt
        # Move to start of file
        os.lseek(self._fd, 0, os.SEEK_SET)
        # Lock the entire file (get file size)
        file_size = os.path.getsize(self._lock_path)
        msvcrt.locking(self._fd, msvcrt.LK_LOCK, file_size)
    else:
        fcntl.flock(self._fd, fcntl.LOCK_EX)

def release(self) -> None:
    """Release the lock."""
    if sys.platform == "win32":
        import msvcrt
        # Move to start of file
        os.lseek(self._fd, 0, os.SEEK_SET)
        # Unlock the entire file
        file_size = os.path.getsize(self._lock_path)
        msvcrt.locking(self._fd, msvcrt.LK_UNLCK, file_size)
    else:
        fcntl.flock(self._fd, fcntl.LOCK_UN)

或者，更简单的方法是使用固定大小（如 1 字节），但确保所有进程都锁定相同位置（从文件开头开始）。