Merge branch 'master' into 4486-document-arv
[arvados.git] / services / nodemanager / tests / test_daemon.py
index 0a63222afcb53ad781b7a208f0b92498a2a3ba19..4bffd092a164c75824453f08e00f3ea755315fcb 100644 (file)
@@ -6,13 +6,16 @@ import time
 import unittest
 
 import mock
+import pykka
 
+import arvnodeman.computenode as nmcnode
 import arvnodeman.daemon as nmdaemon
 from . import testutil
 
 class NodeManagerDaemonActorTestCase(testutil.ActorTestMixin,
                                      unittest.TestCase):
-    def make_daemon(self, cloud_nodes=[], arvados_nodes=[], want_sizes=[]):
+    def make_daemon(self, cloud_nodes=[], arvados_nodes=[], want_sizes=[],
+                    min_nodes=0, max_nodes=8):
         for name in ['cloud_nodes', 'arvados_nodes', 'server_wishlist']:
             setattr(self, name + '_poller', mock.MagicMock(name=name + '_mock'))
         self.arv_factory = mock.MagicMock(name='arvados_mock')
@@ -20,15 +23,14 @@ class NodeManagerDaemonActorTestCase(testutil.ActorTestMixin,
         self.cloud_factory().node_start_time.return_value = time.time()
         self.cloud_updates = mock.MagicMock(name='updates_mock')
         self.timer = testutil.MockTimer()
-        self.node_factory = mock.MagicMock(name='factory_mock')
         self.node_setup = mock.MagicMock(name='setup_mock')
         self.node_shutdown = mock.MagicMock(name='shutdown_mock')
         self.daemon = nmdaemon.NodeManagerDaemonActor.start(
             self.server_wishlist_poller, self.arvados_nodes_poller,
             self.cloud_nodes_poller, self.cloud_updates, self.timer,
             self.arv_factory, self.cloud_factory,
-            [54, 5, 1], 8, 600, 3600,
-            self.node_setup, self.node_shutdown, self.node_factory).proxy()
+            [54, 5, 1], min_nodes, max_nodes, 600, 3600,
+            self.node_setup, self.node_shutdown).proxy()
         if cloud_nodes is not None:
             self.daemon.update_cloud_nodes(cloud_nodes).get(self.TIMEOUT)
         if arvados_nodes is not None:
@@ -36,6 +38,12 @@ class NodeManagerDaemonActorTestCase(testutil.ActorTestMixin,
         if want_sizes is not None:
             self.daemon.update_server_wishlist(want_sizes).get(self.TIMEOUT)
 
+    def monitor_list(self):
+        return pykka.ActorRegistry.get_by_class(nmcnode.ComputeNodeMonitorActor)
+
+    def alive_monitor_count(self):
+        return sum(1 for actor in self.monitor_list() if actor.is_alive())
+
     def test_easy_node_creation(self):
         size = testutil.MockSize(1)
         self.make_daemon(want_sizes=[size])
@@ -47,23 +55,22 @@ class NodeManagerDaemonActorTestCase(testutil.ActorTestMixin,
         arv_node = testutil.arvados_node_mock(1)
         self.make_daemon([cloud_node], [arv_node])
         self.stop_proxy(self.daemon)
-        self.node_factory.start().proxy().offer_arvados_pair.assert_called_with(
+        self.assertEqual(1, self.alive_monitor_count())
+        self.assertIs(
+            self.monitor_list()[0].proxy().arvados_node.get(self.TIMEOUT),
             arv_node)
 
     def test_node_pairing_after_arvados_update(self):
         cloud_node = testutil.cloud_node_mock(2)
-        arv_node = testutil.arvados_node_mock(2, ip_address=None)
-        self.make_daemon([cloud_node], None)
-        pair_func = self.node_factory.start().proxy().offer_arvados_pair
-        pair_func().get.return_value = None
-        self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
-        pair_func.assert_called_with(arv_node)
-
-        pair_func().get.return_value = cloud_node.id
-        pair_func.reset_mock()
+        self.make_daemon([cloud_node],
+                         [testutil.arvados_node_mock(2, ip_address=None)])
         arv_node = testutil.arvados_node_mock(2)
         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
-        pair_func.assert_called_with(arv_node)
+        self.stop_proxy(self.daemon)
+        self.assertEqual(1, self.alive_monitor_count())
+        self.assertIs(
+            self.monitor_list()[0].proxy().arvados_node.get(self.TIMEOUT),
+            arv_node)
 
     def test_old_arvados_node_not_double_assigned(self):
         arv_node = testutil.arvados_node_mock(3, age=9000)
@@ -100,22 +107,108 @@ class NodeManagerDaemonActorTestCase(testutil.ActorTestMixin,
         self.stop_proxy(self.daemon)
         self.assertEqual(1, self.node_setup.start.call_count)
 
+    def test_boot_new_node_when_all_nodes_busy(self):
+        arv_node = testutil.arvados_node_mock(2, job_uuid=True)
+        self.make_daemon([testutil.cloud_node_mock(2)], [arv_node],
+                         [testutil.MockSize(2)])
+        self.stop_proxy(self.daemon)
+        self.assertTrue(self.node_setup.start.called)
+
+    def test_no_new_node_when_max_nodes_busy(self):
+        self.make_daemon([testutil.cloud_node_mock(3)],
+                         [testutil.arvados_node_mock(3, job_uuid=True)],
+                         [testutil.MockSize(3)],
+                         max_nodes=1)
+        self.stop_proxy(self.daemon)
+        self.assertFalse(self.node_setup.start.called)
+
+    def mock_setup_actor(self, cloud_node, arv_node):
+        setup = self.node_setup.start().proxy()
+        self.node_setup.reset_mock()
+        setup.actor_urn = cloud_node.id
+        setup.cloud_node.get.return_value = cloud_node
+        setup.arvados_node.get.return_value = arv_node
+        return setup
+
+    def start_node_boot(self, cloud_node=None, arv_node=None, id_num=1):
+        if cloud_node is None:
+            cloud_node = testutil.cloud_node_mock(id_num)
+        if arv_node is None:
+            arv_node = testutil.arvados_node_mock(id_num)
+        self.make_daemon(want_sizes=[testutil.MockSize(id_num)])
+        self.daemon.max_nodes.get(self.TIMEOUT)
+        self.assertEqual(1, self.node_setup.start.call_count)
+        return self.mock_setup_actor(cloud_node, arv_node)
+
     def test_no_duplication_when_booting_node_listed_fast(self):
         # Test that we don't start two ComputeNodeMonitorActors when
         # we learn about a booting node through a listing before we
         # get the "node up" message from CloudNodeSetupActor.
         cloud_node = testutil.cloud_node_mock(1)
-        self.make_daemon(want_sizes=[testutil.MockSize(1)])
-        self.daemon.max_nodes.get(self.TIMEOUT)
-        self.assertEqual(1, self.node_setup.start.call_count)
-        setup = mock.MagicMock(name='setup_node_mock')
-        setup.actor_ref = self.node_setup.start().proxy().actor_ref
-        setup.cloud_node.get.return_value = cloud_node
-        setup.arvados_node.get.return_value = testutil.arvados_node_mock(1)
+        setup = self.start_node_boot(cloud_node)
+        self.daemon.update_cloud_nodes([cloud_node])
+        self.daemon.node_up(setup).get(self.TIMEOUT)
+        self.assertEqual(1, self.alive_monitor_count())
+
+    def test_no_duplication_when_booted_node_listed(self):
+        cloud_node = testutil.cloud_node_mock(2)
+        setup = self.start_node_boot(cloud_node, id_num=2)
+        self.daemon.node_up(setup)
         self.daemon.update_cloud_nodes([cloud_node]).get(self.TIMEOUT)
-        self.assertTrue(self.node_factory.start.called)
+        self.assertEqual(1, self.alive_monitor_count())
+
+    def test_node_counted_after_boot_with_slow_listing(self):
+        # Test that, after we boot a compute node, we assume it exists
+        # even it doesn't appear in the listing (e.g., because of delays
+        # propagating tags).
+        setup = self.start_node_boot()
+        self.daemon.node_up(setup).get(self.TIMEOUT)
+        self.assertEqual(1, self.alive_monitor_count())
+        self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
+        self.assertEqual(1, self.alive_monitor_count())
+
+    def test_booted_unlisted_node_counted(self):
+        setup = self.start_node_boot(id_num=1)
+        self.daemon.node_up(setup)
+        self.daemon.update_server_wishlist(
+            [testutil.MockSize(1)]).get(self.TIMEOUT)
+        self.stop_proxy(self.daemon)
+        self.assertFalse(self.node_setup.start.called,
+                         "daemon did not count booted node toward wishlist")
+
+    def test_booted_node_can_shutdown(self):
+        setup = self.start_node_boot()
+        self.daemon.node_up(setup).get(self.TIMEOUT)
+        self.assertEqual(1, self.alive_monitor_count())
+        monitor = self.monitor_list()[0].proxy()
+        self.daemon.update_server_wishlist([])
+        self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
+        self.stop_proxy(self.daemon)
+        self.assertTrue(self.node_shutdown.start.called,
+                        "daemon did not shut down booted node on offer")
+
+    def test_booted_node_lifecycle(self):
+        cloud_node = testutil.cloud_node_mock(6)
+        setup = self.start_node_boot(cloud_node, id_num=6)
         self.daemon.node_up(setup).get(self.TIMEOUT)
-        self.assertEqual(1, self.node_factory.start.call_count)
+        self.assertEqual(1, self.alive_monitor_count())
+        monitor = self.monitor_list()[0].proxy()
+        self.daemon.update_server_wishlist([])
+        self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
+        self.assertTrue(self.node_shutdown.start.called,
+                        "daemon did not shut down booted node on offer")
+        shutdown = self.node_shutdown.start().proxy()
+        shutdown.cloud_node.get.return_value = cloud_node
+        self.daemon.node_finished_shutdown(shutdown).get(self.TIMEOUT)
+        self.assertTrue(shutdown.stop.called,
+                        "shutdown actor not stopped after finishing")
+        self.assertTrue(monitor.actor_ref.actor_stopped.wait(self.TIMEOUT),
+                        "monitor for booted node not stopped after shutdown")
+        self.daemon.update_server_wishlist(
+            [testutil.MockSize(2)]).get(self.TIMEOUT)
+        self.stop_proxy(self.daemon)
+        self.assertTrue(self.node_setup.start.called,
+                        "second node not started after booted node stopped")
 
     def test_booting_nodes_shut_down(self):
         self.make_daemon(want_sizes=[testutil.MockSize(1)])
@@ -128,23 +221,47 @@ class NodeManagerDaemonActorTestCase(testutil.ActorTestMixin,
         cloud_node = testutil.cloud_node_mock(1)
         size = testutil.MockSize(1)
         self.make_daemon(cloud_nodes=[cloud_node], want_sizes=[size])
-        node_actor = self.node_factory().proxy()
-        self.daemon.node_can_shutdown(node_actor).get(self.TIMEOUT)
+        self.assertEqual(1, self.alive_monitor_count())
+        monitor = self.monitor_list()[0].proxy()
+        self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
+        self.stop_proxy(self.daemon)
+        self.assertFalse(self.node_shutdown.start.called)
+
+    def test_shutdown_declined_below_min_nodes(self):
+        cloud_node = testutil.cloud_node_mock(1)
+        self.make_daemon(cloud_nodes=[cloud_node], min_nodes=1)
+        self.assertEqual(1, self.alive_monitor_count())
+        monitor = self.monitor_list()[0].proxy()
+        self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
         self.stop_proxy(self.daemon)
-        self.assertFalse(node_actor.shutdown.called)
+        self.assertFalse(self.node_shutdown.start.called)
 
     def test_shutdown_accepted_below_capacity(self):
         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock()])
-        node_actor = self.node_factory().proxy()
-        self.daemon.node_can_shutdown(node_actor).get(self.TIMEOUT)
+        self.assertEqual(1, self.alive_monitor_count())
+        monitor = self.monitor_list()[0].proxy()
+        self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
         self.stop_proxy(self.daemon)
         self.assertTrue(self.node_shutdown.start.called)
 
+    def test_shutdown_declined_when_idle_and_job_queued(self):
+        cloud_nodes = [testutil.cloud_node_mock(n) for n in [3, 4]]
+        arv_nodes = [testutil.arvados_node_mock(3, job_uuid=True),
+                     testutil.arvados_node_mock(4, job_uuid=None)]
+        self.make_daemon(cloud_nodes, arv_nodes, [testutil.MockSize(1)])
+        self.assertEqual(2, self.alive_monitor_count())
+        for mon_ref in self.monitor_list():
+            monitor = mon_ref.proxy()
+            if monitor.cloud_node.get(self.TIMEOUT) is cloud_nodes[-1]:
+                break
+        else:
+            self.fail("monitor for idle node not found")
+        self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
+        self.stop_proxy(self.daemon)
+        self.assertFalse(self.node_shutdown.start.called)
+
     def test_clean_shutdown_waits_for_node_setup_finish(self):
-        self.make_daemon(want_sizes=[testutil.MockSize(1)])
-        self.daemon.max_nodes.get(self.TIMEOUT)
-        self.assertTrue(self.node_setup.start.called)
-        new_node = self.node_setup.start().proxy()
+        new_node = self.start_node_boot()
         self.daemon.shutdown().get(self.TIMEOUT)
         self.assertTrue(new_node.stop_if_no_cloud_node.called)
         self.daemon.node_up(new_node).get(self.TIMEOUT)