shareddata 6.85.17__tar.gz → 6.85.18__tar.gz

{shareddata-6.85.17/src/shareddata.egg-info → shareddata-6.85.18}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: shareddata
-Version: 6.85.17
+Version: 6.85.18
 Summary: Memory Mapped / Shared Memory Database with S3 repository
 Home-page: https://shareddata.net
 Author: shareddata

{shareddata-6.85.17 → shareddata-6.85.18}/setup.py

@@ -66,7 +66,7 @@ test_requires = [
 
 setup(
     name='shareddata',
-    version='6.85.17',
+    version='6.85.18',
     description='Memory Mapped / Shared Memory Database with S3 repository',
     author='shareddata',
     author_email='admin@shareddata.net',

{shareddata-6.85.17 → shareddata-6.85.18}/src/SharedData/MultiProc.py

@@ -305,6 +305,9 @@ def io_bound_unordered(thread_func, iterator, args, maxproc=None, maxthreads=1,
             w.join(timeout=5)  # Wait for 5 seconds for each worker to finish
             if w.is_alive():
                 w.terminate()
+                w.join(timeout=5)  # reap it: an unreaped worker stays a zombie,
+                                   # and a zombie holding a table semaphore is
+                                   # unrecoverable until this parent exits
 
         if not input_queue.empty():
             Logger.log.warning('Input queue not totally consumed!')

{shareddata-6.85.17 → shareddata-6.85.18}/src/SharedData/Routines/WorkerLib.py

@@ -1022,16 +1022,16 @@ def _sync_audit(command, audit, state=None, state_path=None):
     # The first healthy start after a warned/refused one says so at the same severity
     # filter operators watch; it is not repeated (the stored record is then healthy).
     recovered = not condition and previous.get('alert') is True
-    level = 'debug'
+    # A healthy pass that changed nothing is not logged at all: the central log
+    # handler ships DEBUG, so even one record per window floods the operator's view.
+    level = None
     if condition and not repeated:
         level = 'warning' if audit['outcome_class'] == 'offline' else 'error'
     elif recovered:
         level = 'warning'
     elif audit.get('changed'):
         level = 'info'
-    if level != 'debug' or not repeated:
-        # Identical no-op starts inside the window are one DEBUG record, not one per start:
-        # the central log handler ships DEBUG, so demotion alone does not quiet the panel.
+    if level:
         getattr(Logger.log, level)('repo_sync %s', json.dumps(dict(
             repo=command['repo'], branch=command.get('branch') or 'main',
             trigger=_sync_trigger(command), **audit, **({'recovered': True} if recovered else {})),

{shareddata-6.85.17 → shareddata-6.85.18}/src/SharedData/SharedData.py

@@ -776,14 +776,34 @@ class SharedData:
         """One monotonic deadline for both layers of a SCHEMA acquisition."""
         return time.monotonic() + 30.0 if SharedData._is_schema_relpath(relpath) else None
 
+    @staticmethod
+    def _holder_alive(pid):
+        """Holder liveness for the acquire() staleness recovery.
+
+        A zombie holder is dead: it exited without release() and will never
+        run again. psutil.pid_exists() alone still reports it alive (a zombie
+        keeps its pid until reaped), and the reaper is usually the parent —
+        which is itself the process stuck in acquire()'s wait loop, so the
+        pid would never disappear. Waiting on pid_exists() therefore hangs
+        forever; only treating the zombie as dead lets the waiter
+        force-reclaim the mutex. Any psutil error we cannot interpret keeps
+        the previous behavior (wait, as if alive)."""
+        try:
+            return psutil.Process(int(pid)).status() != psutil.STATUS_ZOMBIE
+        except psutil.NoSuchProcess:
+            return False
+        except psutil.Error:
+            return True
+
     @staticmethod
     def acquire(mutex, pid, relpath, deadline=None):
         """
         Spin-CAS acquire of a per-table MUTEX's `pid` field (1us sleep loop).
         Atomically sets `mutex.pid = pid` once it observes `mutex.pid == 0`.
         Staleness recovery: if the current holder PID is *dead*
-        (`psutil.pid_exists` is False), force-acquires by CASing
-        `holder_pid -> pid`.
+        (`_holder_alive` is False — includes zombie holders, which
+        `psutil.pid_exists` alone would keep reporting alive), force-acquires
+        by CASing `holder_pid -> pid`.
 
         =====================================================================
         *** NON-REENTRANT ***
@@ -862,7 +882,7 @@ class SharedData:
                 telapsed = time.time() - tini
                 if (telapsed > SharedData.LOCK_WAIT_REPORTED) | ((firstcheck) & (telapsed > 1)):
                     lockingpid = mutex['pid']
-                    if not psutil.pid_exists(lockingpid):
+                    if not SharedData._holder_alive(int(lockingpid)):
                         if cpp.long_compare_and_swap(hdrptr, semseek, lockingpid, pid) != 0:
                             break
                     # A long wait is reported once, when it has lasted LOCK_WAIT_REPORTED

{shareddata-6.85.17 → shareddata-6.85.18/src/shareddata.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: shareddata
-Version: 6.85.17
+Version: 6.85.18
 Summary: Memory Mapped / Shared Memory Database with S3 repository
 Home-page: https://shareddata.net
 Author: shareddata