Merge branch 'master' into 4233-graph-job-stats
[arvados.git] / services / nodemanager / tests / test_daemon.py
1 #!/usr/bin/env python
2
3 from __future__ import absolute_import, print_function
4
5 import time
6 import unittest
7
8 import mock
9 import pykka
10
11 import arvnodeman.daemon as nmdaemon
12 from arvnodeman.computenode.dispatch import ComputeNodeMonitorActor
13 from . import testutil
14
15 class NodeManagerDaemonActorTestCase(testutil.ActorTestMixin,
16                                      unittest.TestCase):
17     def make_daemon(self, cloud_nodes=[], arvados_nodes=[], want_sizes=[],
18                     min_nodes=0, max_nodes=8):
19         for name in ['cloud_nodes', 'arvados_nodes', 'server_wishlist']:
20             setattr(self, name + '_poller', mock.MagicMock(name=name + '_mock'))
21         self.arv_factory = mock.MagicMock(name='arvados_mock')
22         self.cloud_factory = mock.MagicMock(name='cloud_mock')
23         self.cloud_factory().node_start_time.return_value = time.time()
24         self.cloud_updates = mock.MagicMock(name='updates_mock')
25         self.timer = testutil.MockTimer()
26         self.node_setup = mock.MagicMock(name='setup_mock')
27         self.node_shutdown = mock.MagicMock(name='shutdown_mock')
28         self.daemon = nmdaemon.NodeManagerDaemonActor.start(
29             self.server_wishlist_poller, self.arvados_nodes_poller,
30             self.cloud_nodes_poller, self.cloud_updates, self.timer,
31             self.arv_factory, self.cloud_factory,
32             [54, 5, 1], min_nodes, max_nodes, 600, 3600,
33             self.node_setup, self.node_shutdown).proxy()
34         if cloud_nodes is not None:
35             self.daemon.update_cloud_nodes(cloud_nodes).get(self.TIMEOUT)
36         if arvados_nodes is not None:
37             self.daemon.update_arvados_nodes(arvados_nodes).get(self.TIMEOUT)
38         if want_sizes is not None:
39             self.daemon.update_server_wishlist(want_sizes).get(self.TIMEOUT)
40
41     def monitor_list(self):
42         return pykka.ActorRegistry.get_by_class(ComputeNodeMonitorActor)
43
44     def alive_monitor_count(self):
45         return sum(1 for actor in self.monitor_list() if actor.is_alive())
46
47     def test_easy_node_creation(self):
48         size = testutil.MockSize(1)
49         self.make_daemon(want_sizes=[size])
50         self.stop_proxy(self.daemon)
51         self.assertTrue(self.node_setup.start.called)
52
53     def test_node_pairing(self):
54         cloud_node = testutil.cloud_node_mock(1)
55         arv_node = testutil.arvados_node_mock(1)
56         self.make_daemon([cloud_node], [arv_node])
57         self.stop_proxy(self.daemon)
58         self.assertEqual(1, self.alive_monitor_count())
59         self.assertIs(
60             self.monitor_list()[0].proxy().arvados_node.get(self.TIMEOUT),
61             arv_node)
62
63     def test_node_pairing_after_arvados_update(self):
64         cloud_node = testutil.cloud_node_mock(2)
65         self.make_daemon([cloud_node],
66                          [testutil.arvados_node_mock(2, ip_address=None)])
67         arv_node = testutil.arvados_node_mock(2)
68         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
69         self.stop_proxy(self.daemon)
70         self.assertEqual(1, self.alive_monitor_count())
71         self.assertIs(
72             self.monitor_list()[0].proxy().arvados_node.get(self.TIMEOUT),
73             arv_node)
74
75     def test_old_arvados_node_not_double_assigned(self):
76         arv_node = testutil.arvados_node_mock(3, age=9000)
77         size = testutil.MockSize(3)
78         self.make_daemon(arvados_nodes=[arv_node])
79         setup_ref = self.node_setup.start().proxy().actor_ref
80         setup_ref.actor_urn = 0
81         self.node_setup.start.reset_mock()
82         self.daemon.update_server_wishlist([size]).get(self.TIMEOUT)
83         self.daemon.max_nodes.get(self.TIMEOUT)
84         setup_ref.actor_urn += 1
85         self.daemon.update_server_wishlist([size, size]).get(self.TIMEOUT)
86         self.stop_proxy(self.daemon)
87         used_nodes = [call[1].get('arvados_node')
88                       for call in self.node_setup.start.call_args_list]
89         self.assertEqual(2, len(used_nodes))
90         self.assertIn(arv_node, used_nodes)
91         self.assertIn(None, used_nodes)
92
93     def test_node_count_satisfied(self):
94         self.make_daemon([testutil.cloud_node_mock()],
95                          want_sizes=[testutil.MockSize(1)])
96         self.stop_proxy(self.daemon)
97         self.assertFalse(self.node_setup.called)
98
99     def test_booting_nodes_counted(self):
100         cloud_node = testutil.cloud_node_mock(1)
101         arv_node = testutil.arvados_node_mock(1)
102         server_wishlist = [testutil.MockSize(1)] * 2
103         self.make_daemon([cloud_node], [arv_node], server_wishlist)
104         self.daemon.max_nodes.get(self.TIMEOUT)
105         self.assertTrue(self.node_setup.start.called)
106         self.daemon.update_server_wishlist(server_wishlist).get(self.TIMEOUT)
107         self.stop_proxy(self.daemon)
108         self.assertEqual(1, self.node_setup.start.call_count)
109
110     def test_boot_new_node_when_all_nodes_busy(self):
111         arv_node = testutil.arvados_node_mock(2, job_uuid=True)
112         self.make_daemon([testutil.cloud_node_mock(2)], [arv_node],
113                          [testutil.MockSize(2)])
114         self.stop_proxy(self.daemon)
115         self.assertTrue(self.node_setup.start.called)
116
117     def test_no_new_node_when_max_nodes_busy(self):
118         self.make_daemon([testutil.cloud_node_mock(3)],
119                          [testutil.arvados_node_mock(3, job_uuid=True)],
120                          [testutil.MockSize(3)],
121                          max_nodes=1)
122         self.stop_proxy(self.daemon)
123         self.assertFalse(self.node_setup.start.called)
124
125     def mock_setup_actor(self, cloud_node, arv_node):
126         setup = self.node_setup.start().proxy()
127         self.node_setup.reset_mock()
128         setup.actor_urn = cloud_node.id
129         setup.cloud_node.get.return_value = cloud_node
130         setup.arvados_node.get.return_value = arv_node
131         return setup
132
133     def start_node_boot(self, cloud_node=None, arv_node=None, id_num=1):
134         if cloud_node is None:
135             cloud_node = testutil.cloud_node_mock(id_num)
136         if arv_node is None:
137             arv_node = testutil.arvados_node_mock(id_num)
138         self.make_daemon(want_sizes=[testutil.MockSize(id_num)])
139         self.daemon.max_nodes.get(self.TIMEOUT)
140         self.assertEqual(1, self.node_setup.start.call_count)
141         return self.mock_setup_actor(cloud_node, arv_node)
142
143     def test_no_duplication_when_booting_node_listed_fast(self):
144         # Test that we don't start two ComputeNodeMonitorActors when
145         # we learn about a booting node through a listing before we
146         # get the "node up" message from CloudNodeSetupActor.
147         cloud_node = testutil.cloud_node_mock(1)
148         setup = self.start_node_boot(cloud_node)
149         self.daemon.update_cloud_nodes([cloud_node])
150         self.daemon.node_up(setup).get(self.TIMEOUT)
151         self.assertEqual(1, self.alive_monitor_count())
152
153     def test_no_duplication_when_booted_node_listed(self):
154         cloud_node = testutil.cloud_node_mock(2)
155         setup = self.start_node_boot(cloud_node, id_num=2)
156         self.daemon.node_up(setup)
157         self.daemon.update_cloud_nodes([cloud_node]).get(self.TIMEOUT)
158         self.assertEqual(1, self.alive_monitor_count())
159
160     def test_node_counted_after_boot_with_slow_listing(self):
161         # Test that, after we boot a compute node, we assume it exists
162         # even it doesn't appear in the listing (e.g., because of delays
163         # propagating tags).
164         setup = self.start_node_boot()
165         self.daemon.node_up(setup).get(self.TIMEOUT)
166         self.assertEqual(1, self.alive_monitor_count())
167         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
168         self.assertEqual(1, self.alive_monitor_count())
169
170     def test_booted_unlisted_node_counted(self):
171         setup = self.start_node_boot(id_num=1)
172         self.daemon.node_up(setup)
173         self.daemon.update_server_wishlist(
174             [testutil.MockSize(1)]).get(self.TIMEOUT)
175         self.stop_proxy(self.daemon)
176         self.assertFalse(self.node_setup.start.called,
177                          "daemon did not count booted node toward wishlist")
178
179     def test_booted_node_can_shutdown(self):
180         setup = self.start_node_boot()
181         self.daemon.node_up(setup).get(self.TIMEOUT)
182         self.assertEqual(1, self.alive_monitor_count())
183         monitor = self.monitor_list()[0].proxy()
184         self.daemon.update_server_wishlist([])
185         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
186         self.stop_proxy(self.daemon)
187         self.assertTrue(self.node_shutdown.start.called,
188                         "daemon did not shut down booted node on offer")
189
190     def test_booted_node_lifecycle(self):
191         cloud_node = testutil.cloud_node_mock(6)
192         setup = self.start_node_boot(cloud_node, id_num=6)
193         self.daemon.node_up(setup).get(self.TIMEOUT)
194         self.assertEqual(1, self.alive_monitor_count())
195         monitor = self.monitor_list()[0].proxy()
196         self.daemon.update_server_wishlist([])
197         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
198         self.assertTrue(self.node_shutdown.start.called,
199                         "daemon did not shut down booted node on offer")
200         shutdown = self.node_shutdown.start().proxy()
201         shutdown.cloud_node.get.return_value = cloud_node
202         self.daemon.node_finished_shutdown(shutdown).get(self.TIMEOUT)
203         self.assertTrue(shutdown.stop.called,
204                         "shutdown actor not stopped after finishing")
205         self.assertTrue(monitor.actor_ref.actor_stopped.wait(self.TIMEOUT),
206                         "monitor for booted node not stopped after shutdown")
207         self.daemon.update_server_wishlist(
208             [testutil.MockSize(2)]).get(self.TIMEOUT)
209         self.stop_proxy(self.daemon)
210         self.assertTrue(self.node_setup.start.called,
211                         "second node not started after booted node stopped")
212
213     def test_booting_nodes_shut_down(self):
214         self.make_daemon(want_sizes=[testutil.MockSize(1)])
215         self.daemon.update_server_wishlist([]).get(self.TIMEOUT)
216         self.stop_proxy(self.daemon)
217         self.assertTrue(
218             self.node_setup.start().proxy().stop_if_no_cloud_node.called)
219
220     def test_shutdown_declined_at_wishlist_capacity(self):
221         cloud_node = testutil.cloud_node_mock(1)
222         size = testutil.MockSize(1)
223         self.make_daemon(cloud_nodes=[cloud_node], want_sizes=[size])
224         self.assertEqual(1, self.alive_monitor_count())
225         monitor = self.monitor_list()[0].proxy()
226         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
227         self.stop_proxy(self.daemon)
228         self.assertFalse(self.node_shutdown.start.called)
229
230     def test_shutdown_declined_below_min_nodes(self):
231         cloud_node = testutil.cloud_node_mock(1)
232         self.make_daemon(cloud_nodes=[cloud_node], min_nodes=1)
233         self.assertEqual(1, self.alive_monitor_count())
234         monitor = self.monitor_list()[0].proxy()
235         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
236         self.stop_proxy(self.daemon)
237         self.assertFalse(self.node_shutdown.start.called)
238
239     def test_shutdown_accepted_below_capacity(self):
240         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock()])
241         self.assertEqual(1, self.alive_monitor_count())
242         monitor = self.monitor_list()[0].proxy()
243         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
244         self.stop_proxy(self.daemon)
245         self.assertTrue(self.node_shutdown.start.called)
246
247     def test_shutdown_declined_when_idle_and_job_queued(self):
248         cloud_nodes = [testutil.cloud_node_mock(n) for n in [3, 4]]
249         arv_nodes = [testutil.arvados_node_mock(3, job_uuid=True),
250                      testutil.arvados_node_mock(4, job_uuid=None)]
251         self.make_daemon(cloud_nodes, arv_nodes, [testutil.MockSize(1)])
252         self.assertEqual(2, self.alive_monitor_count())
253         for mon_ref in self.monitor_list():
254             monitor = mon_ref.proxy()
255             if monitor.cloud_node.get(self.TIMEOUT) is cloud_nodes[-1]:
256                 break
257         else:
258             self.fail("monitor for idle node not found")
259         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
260         self.stop_proxy(self.daemon)
261         self.assertFalse(self.node_shutdown.start.called)
262
263     def test_node_shutdown_after_cancelled_shutdown(self):
264         cloud_node = testutil.cloud_node_mock(5)
265         self.make_daemon([cloud_node], [testutil.arvados_node_mock(5)])
266         self.assertEqual(1, self.alive_monitor_count())
267         monitor = self.monitor_list()[0].proxy()
268         shutdown_proxy = self.node_shutdown.start().proxy
269         shutdown_proxy().cloud_node.get.return_value = cloud_node
270         shutdown_proxy().success.get.return_value = False
271         shutdown_proxy.reset_mock()
272         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
273         self.assertTrue(shutdown_proxy.called)
274         self.daemon.node_finished_shutdown(shutdown_proxy()).get(self.TIMEOUT)
275         shutdown_proxy().success.get.return_value = True
276         shutdown_proxy.reset_mock()
277         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
278         self.assertTrue(shutdown_proxy.called)
279
280     def test_nodes_shutting_down_replaced_below_max_nodes(self):
281         cloud_node = testutil.cloud_node_mock(6)
282         self.make_daemon([cloud_node], [testutil.arvados_node_mock(6)])
283         self.assertEqual(1, self.alive_monitor_count())
284         monitor = self.monitor_list()[0].proxy()
285         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
286         self.assertTrue(self.node_shutdown.start.called)
287         self.daemon.update_server_wishlist(
288             [testutil.MockSize(6)]).get(self.TIMEOUT)
289         self.stop_proxy(self.daemon)
290         self.assertTrue(self.node_setup.start.called)
291
292     def test_nodes_shutting_down_not_replaced_at_max_nodes(self):
293         cloud_node = testutil.cloud_node_mock(7)
294         self.make_daemon([cloud_node], [testutil.arvados_node_mock(7)],
295                          max_nodes=1)
296         self.assertEqual(1, self.alive_monitor_count())
297         monitor = self.monitor_list()[0].proxy()
298         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
299         self.assertTrue(self.node_shutdown.start.called)
300         self.daemon.update_server_wishlist(
301             [testutil.MockSize(7)]).get(self.TIMEOUT)
302         self.stop_proxy(self.daemon)
303         self.assertFalse(self.node_setup.start.called)
304
305     def test_nodes_shutting_down_count_against_excess(self):
306         cloud_nodes = [testutil.cloud_node_mock(n) for n in [8, 9]]
307         arv_nodes = [testutil.arvados_node_mock(n) for n in [8, 9]]
308         self.make_daemon(cloud_nodes, arv_nodes, [testutil.MockSize(8)])
309         self.assertEqual(2, self.alive_monitor_count())
310         for mon_ref in self.monitor_list():
311             self.daemon.node_can_shutdown(mon_ref.proxy()).get(self.TIMEOUT)
312         self.assertEqual(1, self.node_shutdown.start.call_count)
313
314     def test_clean_shutdown_waits_for_node_setup_finish(self):
315         new_node = self.start_node_boot()
316         self.daemon.shutdown().get(self.TIMEOUT)
317         self.assertTrue(new_node.stop_if_no_cloud_node.called)
318         self.daemon.node_up(new_node).get(self.TIMEOUT)
319         self.assertTrue(new_node.stop.called)
320         self.assertTrue(
321             self.daemon.actor_ref.actor_stopped.wait(self.TIMEOUT))
322
323     def test_wishlist_ignored_after_shutdown(self):
324         size = testutil.MockSize(2)
325         self.make_daemon(want_sizes=[size])
326         self.daemon.shutdown().get(self.TIMEOUT)
327         self.daemon.update_server_wishlist([size] * 2).get(self.TIMEOUT)
328         self.stop_proxy(self.daemon)
329         self.assertEqual(1, self.node_setup.start.call_count)