[bug] Windows: log rotation crashes with WinError 32 because fcntl-based cross-process lock is a no-op

[superafun]

Bug description

On Windows, Xinference's log rotation crashes every midnight (and on size-based rotation), emitting hundreds of --- Logging error --- tracebacks and disabling log rotation entirely.

Environment

• OS: Windows 10 / 11 (NT)
• Xinference: 3.1.0
• Python: 3.12
• Deployment: xinference-local with multiple models loaded (bge-m3 + bge-reranker-v2-m3), i.e. several xoscar sub-pool processes each holding the log file open.

Symptom

At midnight (or when a size threshold is hit) the logs fill with:

--- Logging error ---
Traceback (most recent call last):
  ...
  File "...\logging\handlers.py", line ..., in doRollover
    self.rotate(self.baseFilename, dfn)
  File "...\logging\handlers.py", line ..., in rotate
    os.rename(source, dest)
PermissionError: [WinError 32] The process cannot access the file because it is being used by another process.

On a busy host (Hindsight driving high-frequency embedding calls, each triggering a sentence_transformers deprecation warning) this turns into tens of MB of tracebacks per night. The active log file stops rotating / grows unbounded. The service itself keeps running fine.

Root cause

xinference/deploy/utils.py defines SafeRotatingFileHandler / SafeTimedRotatingFileHandler / SafeTimedAndSizeRotatingFileHandler. Their doRollover is serialized across processes via a file lock:

if sys.platform == "win32":
    fcntl = None          # utils.py line ~28
else:
    import fcntl

def _acquire_rotation_lock(self):
    if fcntl is None:
        return None       # <-- on Windows the lock is a NO-OP
    ...

Because fcntl is None on Windows, _acquire_rotation_lock() returns None, so every xoscar sub-pool calls doRollover concurrently with no serialization. The standard-library rotation then does os.rename(baseFilename, dfn), which fails on Windows whenever any sibling process still holds the log file open → WinError 32. Python's logging module catches that inside its own error handler and prints a full traceback on every subsequent log record, which is what produces the spew.

Suggested fix (cross-platform)

1. Cross-process lock that works on Windows. Replace the fcntl-only lock with a cross-platform one — on Windows use msvcrt.locking (or portalocker) on the existing *.rotate.lock file that the handlers already create.
2. Windows-safe rotate(). When os.rename raises PermissionError / WinError 32, fall back to shutil.copy2(source, dest) + truncate(0) the source in place. The source keeps its inode, so every process that already had it open keeps appending to the (now empty) live log — no data loss, no crash.
3. Per-day rotation marker. Add a <base>.rotated.<YYYY-MM-DD> marker so only one process performs the midnight rotation; a racing second process skips instead of overwriting the freshly-written archive with an empty copy (the copy+truncate fallback alone would otherwise let a delayed process clobber the archive).

Steps to reproduce

1. Launch xinference-local on Windows with ≥2 models loaded (so ≥2 xoscar sub-pools hold the log file open).
2. Either wait for midnight, or force a quick size rotation by setting a very small maxBytes / backupCount.
3. Observe WinError 32 tracebacks in the log and a log file that never rotates.

[github-actions]

This issue is stale because it has been open for 14 days with no activity.