diff --git a/CHANGELOG.md b/CHANGELOG.md
index a7805eaca8a53add3889d705b0e58944572cf412..705bbf46f29930cae94c353d72121408278e9417 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,7 +2,7 @@
 
 * 0.92:
 
- * TBA
+ * Added `on_fail` event handler - fired upon connection loss
 
 * v0.91:
   * removed annoying warnings
diff --git a/coolamqp/clustering/cluster.py b/coolamqp/clustering/cluster.py
index 655a676f998e1e400cf0846f868d97f50f4c77a2..f5216f3cadb1305ff63de1091ab8e9eacec9ee61 100644
--- a/coolamqp/clustering/cluster.py
+++ b/coolamqp/clustering/cluster.py
@@ -39,7 +39,8 @@ class Cluster(object):
         """
         :param nodes: list of nodes, or a single node. For now, only one is supported.
         :type nodes: NodeDefinition instance or a list of NodeDefinition instances
-        :param on_fail: callable/0 to call when connection fails. This is a one-shot
+        :param on_fail: callable/0 to call when connection fails in an
+            unclean way. This is a one-shot
         :type on_fail: callable/0
         """
         from coolamqp.objects import NodeDefinition
@@ -175,7 +176,12 @@ class Cluster(object):
         self.snr = SingleNodeReconnector(self.node, self.attache_group, self.listener)
         self.snr.on_fail.add(lambda: self.events.put_nowait(ConnectionLost()))
         if self.on_fail is not None:
-            self.snr.on_fail.add(self.on_fail)
+
+            def nice():
+                if not self.snr.terminating:
+                    self.on_fail()
+
+            self.snr.on_fail.add(nice)
 
         # Spawn a transactional publisher and a noack publisher
         self.pub_tr = Publisher(Publisher.MODE_CNPUB)
diff --git a/tests/test_clustering/test_things.py b/tests/test_clustering/test_things.py
index 27a9ac4a85fc100922d30037e84c901aaa55a1ad..51cf2ff28e136970f97a19d9e79cd58da33df375 100644
--- a/tests/test_clustering/test_things.py
+++ b/tests/test_clustering/test_things.py
@@ -15,13 +15,22 @@ logging.basicConfig(level=logging.DEBUG)
 class TestConnecting(unittest.TestCase):
 
     def test_on_fail(self):
-        q = {}
+        q = {'failed': False}
         c = Cluster([NodeDefinition('127.0.0.1', 'xguest', 'xguest', heartbeat=20)], on_fail=lambda: q.update(failed=True))
         c.start()
         time.sleep(5)
         c.shutdown()
         self.assertTrue(q['failed'])
 
+    def test_on_clean(self):
+        q = {}
+        c = Cluster([NODE], on_fail=lambda: q.update(failed=True))
+        c.start()
+        time.sleep(5)
+        c.shutdown()
+        time.sleep(5)
+        self.assertFalse(q['failed'])
+
     def test_start_called_multiple_times(self):
         c = Cluster([NODE])
         c.start(wait=True)