diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index 6dca910..4b33a76 100644
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -14,8 +14,23 @@ Unreleased
 
 **Changed**
 
+* ``Ros.close()`` now blocks until the underlying WebSocket has actually torn
+  down — it waits for the factory's ``"close"`` event (emitted from
+  ``clientConnectionLost``) instead of returning as soon as ``send_close()`` is
+  dispatched on the reactor thread. Downstream callers that immediately
+  construct a new ``Ros`` no longer race the previous socket's cleanup. The
+  ``"closing"`` event semantics are unchanged.
+
 **Fixed**
 
+* ``TwistedEventLoopManager`` no longer leaks one ``PythonLoggingObserver`` per
+  ``Ros`` instance. The observer is now a process-wide singleton — there's only
+  ever one twisted-log → stdlib-logging bridge regardless of how many ``Ros``
+  instances come and go. Long-running pytest sessions that create many
+  short-lived ``Ros`` instances were accumulating one observer per cycle,
+  slowing every log emission linearly and contributing to the occasional
+  ``RosTimeoutError`` flakes that motivated this lifecycle pass.
+
 **Deprecated**
 
 **Removed**
diff --git a/src/roslibpy/comm/comm_autobahn.py b/src/roslibpy/comm/comm_autobahn.py
index 3a830fd..5394f1c 100644
--- a/src/roslibpy/comm/comm_autobahn.py
+++ b/src/roslibpy/comm/comm_autobahn.py
@@ -156,6 +156,37 @@ def set_max_retries(cls, max_retries):
         cls.maxRetries = max_retries
 
 
+# The twisted reactor is a process-wide singleton, so the log observer that
+# fans twisted's log events into the stdlib `logging` module is one too. Any
+# Python process only ever needs one of these registered — registering more
+# means every log event walks an O(n) observer list and ``globalLogPublisher``
+# keeps strong references to every observer for the life of the process.
+#
+# Prior to 2.1, ``TwistedEventLoopManager.__init__`` started a fresh
+# ``PythonLoggingObserver`` per instance, and downstream wrappers that call
+# ``Ros.close()`` (rather than ``Ros.terminate()``, which stops the reactor
+# permanently) never invoked ``manager.terminate()`` and so never stopped the
+# observer. Long-running pytest sessions that create many ``Ros`` instances
+# would accumulate one observer per ``Ros``, slowing every log emission
+# linearly and contributing to occasional ``RosTimeoutError`` flakes on the
+# next ``Ros.run()``.
+_LOG_OBSERVER = None
+_LOG_OBSERVER_LOCK = threading.Lock()
+
+
+def _ensure_log_observer_started():
+    """Lazily start the single process-wide twisted log → stdlib logging bridge."""
+    global _LOG_OBSERVER
+    if _LOG_OBSERVER is not None:
+        return
+    with _LOG_OBSERVER_LOCK:
+        if _LOG_OBSERVER is not None:
+            return
+        observer = log.PythonLoggingObserver()
+        observer.start()
+        _LOG_OBSERVER = observer
+
+
 class TwistedEventLoopManager(object):
     """Manage the main event loop using Twisted reactor.
 
@@ -166,8 +197,11 @@ class TwistedEventLoopManager(object):
     """
 
     def __init__(self):
-        self._log_observer = log.PythonLoggingObserver()
-        self._log_observer.start()
+        # Twisted's reactor is a process singleton — so is the observer that
+        # bridges twisted's log events into stdlib `logging`. See the module-
+        # level comment for the lifecycle details.
+        _ensure_log_observer_started()
+        self._log_observer = _LOG_OBSERVER
 
     def run(self):
         """Kick-starts a non-blocking event loop.
@@ -273,4 +307,13 @@ def terminate(self):
         if self._thread:
             self._thread.join()
 
-        self._log_observer.stop()
+        # The observer is the process-wide singleton (see module top). Stop
+        # it once on terminate — the reactor cannot be restarted after stop,
+        # so no future TwistedEventLoopManager will need the bridge in this
+        # process. Subsequent terminates no-op.
+        global _LOG_OBSERVER
+        with _LOG_OBSERVER_LOCK:
+            if _LOG_OBSERVER is not None:
+                _LOG_OBSERVER.stop()
+                _LOG_OBSERVER = None
+        self._log_observer = None
diff --git a/src/roslibpy/ros.py b/src/roslibpy/ros.py
index 78a29c8..5670c81 100644
--- a/src/roslibpy/ros.py
+++ b/src/roslibpy/ros.py
@@ -76,20 +76,44 @@ def _unset_connecting_flag(*args):
         self.factory.connect()
 
     def close(self, timeout=CONNECTION_TIMEOUT):
-        """Disconnect from ROS."""
-        if self.is_connected:
-            wait_disconnect = threading.Event()
+        """Disconnect from ROS.
+
+        Blocks until the underlying WebSocket has actually torn down
+        (``clientConnectionLost`` fired). Prior to 2.1, ``close()``
+        returned as soon as ``send_close()`` had been dispatched on the
+        reactor thread; callers that immediately constructed a new ``Ros``
+        could race the previous socket's cleanup, occasionally causing
+        the next ``run()`` to time out under contention.
+        """
+        if not self.is_connected:
+            return
+
+        wait_disconnect = threading.Event()
 
-            def _wrapper_callback(proto):
-                self.emit("closing")
-                proto.send_close()
-                wait_disconnect.set()
-                return proto
+        def _on_close(*_args):
+            wait_disconnect.set()
 
-            self.factory.on_ready(_wrapper_callback)
+        # `clientConnectionLost` on the factory emits the "close" event once
+        # the websocket has actually been torn down. Use `once` so the
+        # listener is auto-removed on first fire — protects against the
+        # ReconnectingClientFactory triggering a later non-manual close.
+        self.factory.once("close", _on_close)
+
+        def _wrapper_callback(proto):
+            self.emit("closing")
+            proto.send_close()
+            return proto
+
+        self.factory.on_ready(_wrapper_callback)
 
-            if not wait_disconnect.wait(timeout):
-                raise RosTimeoutError("Failed to disconnect to ROS")
+        if not wait_disconnect.wait(timeout):
+            # Detach the listener so it doesn't fire on a later reconnect
+            # cycle and leak a reference to this Ros instance.
+            try:
+                self.factory.off("close", _on_close)
+            except Exception:
+                pass
+            raise RosTimeoutError("Failed to disconnect to ROS")
 
     def run(self, timeout=CONNECTION_TIMEOUT):
         """Kick-starts a non-blocking event loop.
diff --git a/tests/ros1/test_lifecycle.py b/tests/ros1/test_lifecycle.py
new file mode 100644
index 0000000..c6a2fb3
--- /dev/null
+++ b/tests/ros1/test_lifecycle.py
@@ -0,0 +1,73 @@
+"""Lifecycle regression tests for the 2.1 fixes.
+
+Targets bugs that emerged from downstream pytest sessions (``compas_fab``)
+where many sequential ``Ros`` instances are constructed, run, and closed
+in the same Python process.
+
+If these tests flake, the corresponding fix is regressing — diagnose,
+don't retry-loop.
+"""
+
+from __future__ import print_function
+
+import threading
+
+from roslibpy import Ros
+
+HOST = "127.0.0.1"
+PORT = 9090
+URL = "ws://%s:%d" % (HOST, PORT)
+
+
+def test_log_observer_does_not_leak():
+    """Constructing many ``Ros`` instances must not leak twisted log observers.
+
+    Before A1 every ``TwistedEventLoopManager.__init__`` (created lazily
+    one-per-``Ros``) called ``PythonLoggingObserver().start()`` and never
+    cleaned up unless ``manager.terminate()`` was explicitly called.
+    Downstream wrappers (compas_fab's ``RosClient.__exit__``) call
+    ``close()``, not ``terminate()`` — by design, since ``terminate()``
+    stops the reactor permanently. So observers piled up forever.
+
+    After A1 there's exactly one process-wide observer regardless of how
+    many ``Ros()`` instances come and go.
+    """
+    from twisted.logger import globalLogPublisher
+
+    initial = len(list(globalLogPublisher._observers))
+
+    for _ in range(10):
+        ros = Ros(URL)
+        ros.run()
+        ros.close()
+
+    final = len(list(globalLogPublisher._observers))
+    # Allow a single-observer drift in case the very first cycle is the one
+    # that registers the singleton. The leak case grows by len(cycles).
+    assert final - initial <= 1, (
+        "Leaked %d twisted log observers across 10 lifecycle cycles "
+        "(expected at most 1)" % (final - initial)
+    )
+
+
+def test_close_blocks_until_disconnect():
+    """``Ros.close()`` must not return until the WebSocket is actually closed.
+
+    Before A2 the call returned once ``send_close`` had been dispatched on
+    the reactor thread — the protocol's ``onClose`` would fire shortly
+    after, leaving a brief window where ``is_connected`` could still read
+    True. Now ``close()`` blocks on ``clientConnectionLost`` so the
+    contract matches the name.
+    """
+    ros = Ros(URL)
+    ros.run()
+    assert ros.is_connected
+
+    closed_event = threading.Event()
+    ros.on("close", lambda _: closed_event.set())
+
+    ros.close()
+
+    # The "close" event must have fired by the time close() returned.
+    assert closed_event.is_set(), "close() returned before clientConnectionLost fired"
+    assert not ros.is_connected, "is_connected still True after close()"