Merge branch '3137-arv-mount-stats' closes #3137
[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.jobqueue import ServerCalculator
13 from arvnodeman.computenode.dispatch import ComputeNodeMonitorActor
14 from . import testutil
15 import logging
16
17 class NodeManagerDaemonActorTestCase(testutil.ActorTestMixin,
18                                      unittest.TestCase):
19     def mock_node_start(self, **kwargs):
20         # Make sure that every time the daemon starts a setup actor,
21         # it gets a new mock object back.
22         get_cloud_size = mock.MagicMock()
23         get_cloud_size.get.return_value = kwargs["cloud_size"]
24         mock_actor = mock.MagicMock()
25         mock_proxy = mock.NonCallableMock(name='setup_mock_proxy',
26                                           cloud_size=get_cloud_size,
27                                           actor_ref=mock_actor)
28         mock_actor.proxy.return_value = mock_proxy
29
30         self.last_setup = mock_proxy
31         return mock_actor
32
33     def mock_node_shutdown(self, **kwargs):
34         # Make sure that every time the daemon starts a shutdown actor,
35         # it gets a new mock object back.
36         get_cloud_node = mock.MagicMock()
37         if "node_monitor" in kwargs:
38             get_cloud_node.get.return_value = kwargs["node_monitor"].proxy().cloud_node.get()
39         mock_actor = mock.MagicMock()
40         mock_proxy = mock.NonCallableMock(name='shutdown_mock_proxy',
41                                           cloud_node=get_cloud_node,
42                                           actor_ref=mock_actor)
43
44         mock_actor.proxy.return_value = mock_proxy
45         self.last_shutdown = mock_proxy
46
47         return mock_actor
48
49     def make_daemon(self, cloud_nodes=[], arvados_nodes=[], want_sizes=[],
50                     avail_sizes=[(testutil.MockSize(1), {"cores": 1})],
51                     min_nodes=0, max_nodes=8,
52                     shutdown_windows=[54, 5, 1],
53                     max_total_price=None):
54         for name in ['cloud_nodes', 'arvados_nodes', 'server_wishlist']:
55             setattr(self, name + '_poller', mock.MagicMock(name=name + '_mock'))
56         self.arv_factory = mock.MagicMock(name='arvados_mock')
57         self.cloud_factory = mock.MagicMock(name='cloud_mock')
58         self.cloud_factory().node_start_time.return_value = time.time()
59         self.cloud_updates = mock.MagicMock(name='updates_mock')
60         self.timer = testutil.MockTimer(deliver_immediately=False)
61
62         self.node_setup = mock.MagicMock(name='setup_mock')
63         self.node_setup.start.side_effect = self.mock_node_start
64         self.node_setup.reset_mock()
65
66         self.node_shutdown = mock.MagicMock(name='shutdown_mock')
67         self.node_shutdown.start.side_effect = self.mock_node_shutdown
68
69         self.daemon = nmdaemon.NodeManagerDaemonActor.start(
70             self.server_wishlist_poller, self.arvados_nodes_poller,
71             self.cloud_nodes_poller, self.cloud_updates, self.timer,
72             self.arv_factory, self.cloud_factory,
73             shutdown_windows, ServerCalculator(avail_sizes),
74             min_nodes, max_nodes, 600, 1800, 3600,
75             self.node_setup, self.node_shutdown,
76             max_total_price=max_total_price).proxy()
77         if cloud_nodes is not None:
78             self.daemon.update_cloud_nodes(cloud_nodes).get(self.TIMEOUT)
79         if arvados_nodes is not None:
80             self.daemon.update_arvados_nodes(arvados_nodes).get(self.TIMEOUT)
81         if want_sizes is not None:
82             self.daemon.update_server_wishlist(want_sizes).get(self.TIMEOUT)
83
84     def monitor_list(self):
85         return pykka.ActorRegistry.get_by_class(ComputeNodeMonitorActor)
86
87     def monitored_arvados_nodes(self):
88         pairings = []
89         for future in [actor.proxy().arvados_node
90                        for actor in self.monitor_list()]:
91             try:
92                 pairings.append(future.get(self.TIMEOUT))
93             except pykka.ActorDeadError:
94                 pass
95         return pairings
96
97     def alive_monitor_count(self):
98         return len(self.monitored_arvados_nodes())
99
100     def assertShutdownCancellable(self, expected=True):
101         self.assertTrue(self.node_shutdown.start.called)
102         self.assertIs(expected,
103                       self.node_shutdown.start.call_args[1]['cancellable'],
104                       "ComputeNodeShutdownActor incorrectly cancellable")
105
106     def test_easy_node_creation(self):
107         size = testutil.MockSize(1)
108         self.make_daemon(want_sizes=[size])
109         self.stop_proxy(self.daemon)
110         self.assertTrue(self.node_setup.start.called)
111
112     def check_monitors_arvados_nodes(self, *arv_nodes):
113         self.assertItemsEqual(arv_nodes, self.monitored_arvados_nodes())
114
115     def test_node_pairing(self):
116         cloud_node = testutil.cloud_node_mock(1)
117         arv_node = testutil.arvados_node_mock(1)
118         self.make_daemon([cloud_node], [arv_node])
119         self.stop_proxy(self.daemon)
120         self.check_monitors_arvados_nodes(arv_node)
121
122     def test_node_pairing_after_arvados_update(self):
123         cloud_node = testutil.cloud_node_mock(2)
124         self.make_daemon([cloud_node],
125                          [testutil.arvados_node_mock(2, ip_address=None)])
126         arv_node = testutil.arvados_node_mock(2)
127         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
128         self.stop_proxy(self.daemon)
129         self.check_monitors_arvados_nodes(arv_node)
130
131     def test_arvados_node_un_and_re_paired(self):
132         # We need to create the Arvados node mock after spinning up the daemon
133         # to make sure it's new enough to pair with the cloud node.
134         self.make_daemon([testutil.cloud_node_mock(3)], arvados_nodes=None)
135         arv_node = testutil.arvados_node_mock(3)
136         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
137         self.check_monitors_arvados_nodes(arv_node)
138         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
139         self.assertEqual(0, self.alive_monitor_count())
140         self.daemon.update_cloud_nodes([testutil.cloud_node_mock(3)])
141         self.stop_proxy(self.daemon)
142         self.check_monitors_arvados_nodes(arv_node)
143
144     def test_old_arvados_node_not_double_assigned(self):
145         arv_node = testutil.arvados_node_mock(3, age=9000)
146         size = testutil.MockSize(3)
147         self.make_daemon(arvados_nodes=[arv_node], avail_sizes=[(size, {"cores":1})])
148         self.daemon.update_server_wishlist([size]).get(self.TIMEOUT)
149         self.daemon.update_server_wishlist([size, size]).get(self.TIMEOUT)
150         self.stop_proxy(self.daemon)
151         used_nodes = [call[1].get('arvados_node')
152                       for call in self.node_setup.start.call_args_list]
153         self.assertEqual(2, len(used_nodes))
154         self.assertIn(arv_node, used_nodes)
155         self.assertIn(None, used_nodes)
156
157     def test_node_count_satisfied(self):
158         self.make_daemon([testutil.cloud_node_mock()],
159                          want_sizes=[testutil.MockSize(1)])
160         self.stop_proxy(self.daemon)
161         self.assertFalse(self.node_setup.start.called)
162
163     def test_dont_count_missing_as_busy(self):
164         size = testutil.MockSize(1)
165         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock(1),
166                                       testutil.cloud_node_mock(2)],
167                          arvados_nodes=[testutil.arvados_node_mock(1),
168                                       testutil.arvados_node_mock(2, last_ping_at='1970-01-01T01:02:03.04050607Z')],
169                          want_sizes=[size, size])
170         self.stop_proxy(self.daemon)
171         self.assertTrue(self.node_setup.start.called)
172
173     def test_missing_counts_towards_max(self):
174         size = testutil.MockSize(1)
175         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock(1),
176                                       testutil.cloud_node_mock(2)],
177                          arvados_nodes=[testutil.arvados_node_mock(1),
178                                         testutil.arvados_node_mock(2, last_ping_at='1970-01-01T01:02:03.04050607Z')],
179                          want_sizes=[size, size],
180                          max_nodes=2)
181         self.stop_proxy(self.daemon)
182         self.assertFalse(self.node_setup.start.called)
183
184     def test_excess_counts_missing(self):
185         size = testutil.MockSize(1)
186         cloud_nodes = [testutil.cloud_node_mock(1), testutil.cloud_node_mock(2)]
187         self.make_daemon(cloud_nodes=cloud_nodes,
188                          arvados_nodes=[testutil.arvados_node_mock(1),
189                                         testutil.arvados_node_mock(2, last_ping_at='1970-01-01T01:02:03.04050607Z')],
190                          want_sizes=[size])
191         self.assertEqual(2, self.alive_monitor_count())
192         for mon_ref in self.monitor_list():
193             self.daemon.node_can_shutdown(mon_ref.proxy()).get(self.TIMEOUT)
194         self.assertEqual(1, self.node_shutdown.start.call_count)
195
196     def test_missing_shutdown_not_excess(self):
197         size = testutil.MockSize(1)
198         cloud_nodes = [testutil.cloud_node_mock(1), testutil.cloud_node_mock(2)]
199         self.make_daemon(cloud_nodes=cloud_nodes,
200                          arvados_nodes=[testutil.arvados_node_mock(1),
201                                         testutil.arvados_node_mock(2, last_ping_at='1970-01-01T01:02:03.04050607Z')],
202                          want_sizes=[size])
203
204         get_cloud_node = mock.MagicMock(name="get_cloud_node")
205         get_cloud_node.get.return_value = cloud_nodes[1]
206         mock_node_monitor = mock.MagicMock()
207         mock_node_monitor.proxy.return_value = mock.NonCallableMock(cloud_node=get_cloud_node)
208         mock_shutdown = self.node_shutdown.start(node_monitor=mock_node_monitor)
209
210         self.daemon.shutdowns.get()[cloud_nodes[1].id] = mock_shutdown.proxy()
211
212         self.assertEqual(2, self.alive_monitor_count())
213         for mon_ref in self.monitor_list():
214             self.daemon.node_can_shutdown(mon_ref.proxy()).get(self.TIMEOUT)
215         self.assertEqual(1, self.node_shutdown.start.call_count)
216
217     def test_booting_nodes_counted(self):
218         cloud_node = testutil.cloud_node_mock(1)
219         arv_node = testutil.arvados_node_mock(1)
220         server_wishlist = [testutil.MockSize(1)] * 2
221         self.make_daemon([cloud_node], [arv_node], server_wishlist)
222         self.daemon.max_nodes.get(self.TIMEOUT)
223         self.assertTrue(self.node_setup.start.called)
224         self.daemon.update_server_wishlist(server_wishlist).get(self.TIMEOUT)
225         self.stop_proxy(self.daemon)
226         self.assertEqual(1, self.node_setup.start.call_count)
227
228     def test_boot_new_node_when_all_nodes_busy(self):
229         size = testutil.MockSize(2)
230         arv_node = testutil.arvados_node_mock(2, job_uuid=True)
231         self.make_daemon([testutil.cloud_node_mock(2, size=size)], [arv_node],
232                          [size], avail_sizes=[(size, {"cores":1})])
233         self.stop_proxy(self.daemon)
234         self.assertTrue(self.node_setup.start.called)
235
236     def test_boot_new_node_below_min_nodes(self):
237         min_size = testutil.MockSize(1)
238         wish_size = testutil.MockSize(3)
239         avail_sizes = [(min_size, {"cores": 1}),
240                        (wish_size, {"cores": 3})]
241         self.make_daemon([], [], None, avail_sizes=avail_sizes, min_nodes=2)
242         self.daemon.update_server_wishlist([wish_size]).get(self.TIMEOUT)
243         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
244         self.daemon.update_server_wishlist([wish_size]).get(self.TIMEOUT)
245         self.stop_proxy(self.daemon)
246         self.assertEqual([wish_size, min_size],
247                          [call[1].get('cloud_size')
248                           for call in self.node_setup.start.call_args_list])
249
250     def test_no_new_node_when_ge_min_nodes_busy(self):
251         cloud_nodes = [testutil.cloud_node_mock(n) for n in range(1, 4)]
252         arv_nodes = [testutil.arvados_node_mock(n, job_uuid=True)
253                      for n in range(1, 4)]
254         self.make_daemon(cloud_nodes, arv_nodes, [], min_nodes=2)
255         self.stop_proxy(self.daemon)
256         self.assertEqual(0, self.node_setup.start.call_count)
257
258     def test_no_new_node_when_max_nodes_busy(self):
259         self.make_daemon([testutil.cloud_node_mock(3)],
260                          [testutil.arvados_node_mock(3, job_uuid=True)],
261                          [testutil.MockSize(3)],
262                          max_nodes=1)
263         self.stop_proxy(self.daemon)
264         self.assertFalse(self.node_setup.start.called)
265
266     def start_node_boot(self, cloud_node=None, arv_node=None, id_num=1):
267         if cloud_node is None:
268             cloud_node = testutil.cloud_node_mock(id_num)
269         if arv_node is None:
270             arv_node = testutil.arvados_node_mock(id_num)
271         self.make_daemon(want_sizes=[testutil.MockSize(id_num)],
272                          avail_sizes=[(testutil.MockSize(id_num), {"cores":1})])
273         self.daemon.max_nodes.get(self.TIMEOUT)
274         self.assertEqual(1, self.node_setup.start.call_count)
275         self.last_setup.cloud_node.get.return_value = cloud_node
276         self.last_setup.arvados_node.get.return_value = arv_node
277         return self.last_setup
278
279     def test_no_new_node_when_booted_node_not_usable(self):
280         cloud_node = testutil.cloud_node_mock(4)
281         arv_node = testutil.arvados_node_mock(4, crunch_worker_state='down')
282         setup = self.start_node_boot(cloud_node, arv_node)
283         self.daemon.node_up(setup).get(self.TIMEOUT)
284         self.assertEqual(1, self.alive_monitor_count())
285         self.daemon.update_cloud_nodes([cloud_node])
286         self.daemon.update_arvados_nodes([arv_node])
287         self.daemon.update_server_wishlist(
288             [testutil.MockSize(1)]).get(self.TIMEOUT)
289         self.stop_proxy(self.daemon)
290         self.assertEqual(1, self.node_setup.start.call_count)
291
292     def test_no_duplication_when_booting_node_listed_fast(self):
293         # Test that we don't start two ComputeNodeMonitorActors when
294         # we learn about a booting node through a listing before we
295         # get the "node up" message from CloudNodeSetupActor.
296         cloud_node = testutil.cloud_node_mock(1)
297         setup = self.start_node_boot(cloud_node)
298         self.daemon.update_cloud_nodes([cloud_node])
299         self.daemon.node_up(setup).get(self.TIMEOUT)
300         self.assertEqual(1, self.alive_monitor_count())
301
302     def test_no_duplication_when_booted_node_listed(self):
303         cloud_node = testutil.cloud_node_mock(2)
304         setup = self.start_node_boot(cloud_node, id_num=2)
305         self.daemon.node_up(setup)
306         self.daemon.update_cloud_nodes([cloud_node]).get(self.TIMEOUT)
307         self.assertEqual(1, self.alive_monitor_count())
308
309     def test_node_counted_after_boot_with_slow_listing(self):
310         # Test that, after we boot a compute node, we assume it exists
311         # even it doesn't appear in the listing (e.g., because of delays
312         # propagating tags).
313         setup = self.start_node_boot()
314         self.daemon.node_up(setup).get(self.TIMEOUT)
315         self.assertEqual(1, self.alive_monitor_count())
316         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
317         self.assertEqual(1, self.alive_monitor_count())
318
319     def test_booted_unlisted_node_counted(self):
320         setup = self.start_node_boot(id_num=1)
321         self.daemon.node_up(setup)
322         self.daemon.update_server_wishlist(
323             [testutil.MockSize(1)]).get(self.TIMEOUT)
324         self.stop_proxy(self.daemon)
325         self.assertEqual(1, self.node_setup.start.call_count)
326
327     def test_booted_node_can_shutdown(self):
328         setup = self.start_node_boot()
329         self.daemon.node_up(setup).get(self.TIMEOUT)
330         self.assertEqual(1, self.alive_monitor_count())
331         monitor = self.monitor_list()[0].proxy()
332         self.daemon.update_server_wishlist([])
333         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
334         self.stop_proxy(self.daemon)
335         self.assertTrue(self.node_shutdown.start.called,
336                         "daemon did not shut down booted node on offer")
337
338     def test_booted_node_lifecycle(self):
339         cloud_node = testutil.cloud_node_mock(6)
340         setup = self.start_node_boot(cloud_node, id_num=6)
341         self.daemon.node_up(setup).get(self.TIMEOUT)
342         self.assertEqual(1, self.alive_monitor_count())
343         monitor = self.monitor_list()[0].proxy()
344         self.daemon.update_server_wishlist([])
345         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
346         self.assertShutdownCancellable(True)
347         shutdown = self.node_shutdown.start().proxy()
348         shutdown.cloud_node.get.return_value = cloud_node
349         self.daemon.node_finished_shutdown(shutdown).get(self.TIMEOUT)
350         self.assertTrue(shutdown.stop.called,
351                         "shutdown actor not stopped after finishing")
352         self.assertTrue(monitor.actor_ref.actor_stopped.wait(self.TIMEOUT),
353                         "monitor for booted node not stopped after shutdown")
354         self.daemon.update_server_wishlist(
355             [testutil.MockSize(2)]).get(self.TIMEOUT)
356         self.stop_proxy(self.daemon)
357         self.assertTrue(self.node_setup.start.called,
358                         "second node not started after booted node stopped")
359
360     def test_booted_node_shut_down_when_never_listed(self):
361         setup = self.start_node_boot()
362         self.daemon.node_up(setup).get(self.TIMEOUT)
363         self.assertEqual(1, self.alive_monitor_count())
364         self.assertFalse(self.node_shutdown.start.called)
365         self.timer.deliver()
366         self.stop_proxy(self.daemon)
367         self.assertShutdownCancellable(False)
368
369     def test_booted_node_shut_down_when_never_paired(self):
370         cloud_node = testutil.cloud_node_mock(2)
371         setup = self.start_node_boot(cloud_node)
372         self.daemon.node_up(setup).get(self.TIMEOUT)
373         self.assertEqual(1, self.alive_monitor_count())
374         self.daemon.update_cloud_nodes([cloud_node])
375         self.timer.deliver()
376         self.stop_proxy(self.daemon)
377         self.assertShutdownCancellable(False)
378
379     def test_booted_node_shut_down_when_never_working(self):
380         cloud_node = testutil.cloud_node_mock(4)
381         arv_node = testutil.arvados_node_mock(4, crunch_worker_state='down')
382         setup = self.start_node_boot(cloud_node, arv_node)
383         self.daemon.node_up(setup).get(self.TIMEOUT)
384         self.assertEqual(1, self.alive_monitor_count())
385         self.daemon.update_cloud_nodes([cloud_node])
386         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
387         self.timer.deliver()
388         self.stop_proxy(self.daemon)
389         self.assertShutdownCancellable(False)
390
391     def test_node_that_pairs_not_considered_failed_boot(self):
392         cloud_node = testutil.cloud_node_mock(3)
393         arv_node = testutil.arvados_node_mock(3)
394         setup = self.start_node_boot(cloud_node, arv_node)
395         self.daemon.node_up(setup).get(self.TIMEOUT)
396         self.assertEqual(1, self.alive_monitor_count())
397         self.daemon.update_cloud_nodes([cloud_node])
398         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
399         self.timer.deliver()
400         self.stop_proxy(self.daemon)
401         self.assertFalse(self.node_shutdown.start.called)
402
403     def test_node_that_pairs_busy_not_considered_failed_boot(self):
404         cloud_node = testutil.cloud_node_mock(5)
405         arv_node = testutil.arvados_node_mock(5, job_uuid=True)
406         setup = self.start_node_boot(cloud_node, arv_node)
407         self.daemon.node_up(setup).get(self.TIMEOUT)
408         self.assertEqual(1, self.alive_monitor_count())
409         self.daemon.update_cloud_nodes([cloud_node])
410         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
411         self.timer.deliver()
412         self.stop_proxy(self.daemon)
413         self.assertFalse(self.node_shutdown.start.called)
414
415     def test_booting_nodes_shut_down(self):
416         self.make_daemon(want_sizes=[testutil.MockSize(1)])
417         self.daemon.update_server_wishlist([]).get(self.TIMEOUT)
418         self.stop_proxy(self.daemon)
419         self.assertTrue(self.last_setup.stop_if_no_cloud_node.called)
420
421     def test_all_booting_nodes_tried_to_shut_down(self):
422         size = testutil.MockSize(2)
423         self.make_daemon(want_sizes=[size], avail_sizes=[(size, {"cores":1})])
424         self.daemon.max_nodes.get(self.TIMEOUT)
425         setup1 = self.last_setup
426         setup1.stop_if_no_cloud_node().get.return_value = False
427         setup1.stop_if_no_cloud_node.reset_mock()
428         self.daemon.update_server_wishlist([size, size]).get(self.TIMEOUT)
429         self.daemon.max_nodes.get(self.TIMEOUT)
430         self.assertIsNot(setup1, self.last_setup)
431         self.last_setup.stop_if_no_cloud_node().get.return_value = True
432         self.last_setup.stop_if_no_cloud_node.reset_mock()
433         self.daemon.update_server_wishlist([]).get(self.TIMEOUT)
434         self.daemon.max_nodes.get(self.TIMEOUT)
435         self.stop_proxy(self.daemon)
436         self.assertEqual(1, self.last_setup.stop_if_no_cloud_node.call_count)
437         self.assertTrue(setup1.stop_if_no_cloud_node.called)
438
439     def test_shutdown_declined_at_wishlist_capacity(self):
440         cloud_node = testutil.cloud_node_mock(1)
441         size = testutil.MockSize(1)
442         self.make_daemon(cloud_nodes=[cloud_node], want_sizes=[size])
443         self.assertEqual(1, self.alive_monitor_count())
444         monitor = self.monitor_list()[0].proxy()
445         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
446         self.stop_proxy(self.daemon)
447         self.assertFalse(self.node_shutdown.start.called)
448
449     def test_shutdown_declined_below_min_nodes(self):
450         cloud_node = testutil.cloud_node_mock(1)
451         self.make_daemon(cloud_nodes=[cloud_node], min_nodes=1)
452         self.assertEqual(1, self.alive_monitor_count())
453         monitor = self.monitor_list()[0].proxy()
454         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
455         self.stop_proxy(self.daemon)
456         self.assertFalse(self.node_shutdown.start.called)
457
458     def test_shutdown_accepted_below_capacity(self):
459         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock()])
460         self.assertEqual(1, self.alive_monitor_count())
461         monitor = self.monitor_list()[0].proxy()
462         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
463         self.stop_proxy(self.daemon)
464         self.assertTrue(self.node_shutdown.start.called)
465
466     def test_shutdown_declined_when_idle_and_job_queued(self):
467         cloud_nodes = [testutil.cloud_node_mock(n) for n in [3, 4]]
468         arv_nodes = [testutil.arvados_node_mock(3, job_uuid=True),
469                      testutil.arvados_node_mock(4, job_uuid=None)]
470         self.make_daemon(cloud_nodes, arv_nodes, [testutil.MockSize(1)])
471         self.assertEqual(2, self.alive_monitor_count())
472         for mon_ref in self.monitor_list():
473             monitor = mon_ref.proxy()
474             if monitor.cloud_node.get(self.TIMEOUT) is cloud_nodes[-1]:
475                 break
476         else:
477             self.fail("monitor for idle node not found")
478         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
479         self.stop_proxy(self.daemon)
480         self.assertFalse(self.node_shutdown.start.called)
481
482     def test_node_shutdown_after_cancelled_shutdown(self):
483         cloud_node = testutil.cloud_node_mock(5)
484         self.make_daemon([cloud_node], [testutil.arvados_node_mock(5)])
485         self.assertEqual(1, self.alive_monitor_count())
486         monitor = self.monitor_list()[0].proxy()
487         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
488         self.last_shutdown.success.get.return_value = False
489         self.daemon.node_finished_shutdown(self.last_shutdown).get(self.TIMEOUT)
490         self.assertEqual(1, self.alive_monitor_count())
491
492         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
493         self.last_shutdown.success.get.return_value = True
494         self.last_shutdown.stop.side_effect = lambda: monitor.stop()
495         self.daemon.node_finished_shutdown(self.last_shutdown).get(self.TIMEOUT)
496         self.assertEqual(0, self.alive_monitor_count())
497
498     def test_broken_node_blackholed_after_cancelled_shutdown(self):
499         size = testutil.MockSize(8)
500         cloud_node = testutil.cloud_node_mock(8, size=size)
501         wishlist = [size]
502         self.make_daemon([cloud_node], [testutil.arvados_node_mock(8)],
503                          wishlist, avail_sizes=[(size, {"cores":1})])
504         self.assertEqual(1, self.alive_monitor_count())
505         self.assertFalse(self.node_setup.start.called)
506         monitor = self.monitor_list()[0].proxy()
507         shutdown_proxy = self.node_shutdown.start().proxy
508         shutdown_proxy().cloud_node.get.return_value = cloud_node
509         shutdown_proxy().success.get.return_value = False
510         shutdown_proxy().cancel_reason.get.return_value = self.node_shutdown.NODE_BROKEN
511         self.daemon.update_server_wishlist([]).get(self.TIMEOUT)
512         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
513         self.daemon.node_finished_shutdown(shutdown_proxy()).get(self.TIMEOUT)
514         self.daemon.update_cloud_nodes([cloud_node]).get(self.TIMEOUT)
515         self.daemon.update_server_wishlist(wishlist).get(self.TIMEOUT)
516         self.stop_proxy(self.daemon)
517         self.assertEqual(1, self.node_setup.start.call_count)
518
519     def test_nodes_shutting_down_replaced_below_max_nodes(self):
520         size = testutil.MockSize(6)
521         cloud_node = testutil.cloud_node_mock(6, size=size)
522         self.make_daemon([cloud_node], [testutil.arvados_node_mock(6)],
523                          avail_sizes=[(size, {"cores":1})])
524         self.assertEqual(1, self.alive_monitor_count())
525         monitor = self.monitor_list()[0].proxy()
526         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
527         self.assertTrue(self.node_shutdown.start.called)
528         self.daemon.update_server_wishlist(
529             [testutil.MockSize(6)]).get(self.TIMEOUT)
530         self.stop_proxy(self.daemon)
531         self.assertTrue(self.node_setup.start.called)
532
533     def test_nodes_shutting_down_not_replaced_at_max_nodes(self):
534         cloud_node = testutil.cloud_node_mock(7)
535         self.make_daemon([cloud_node], [testutil.arvados_node_mock(7)],
536                          max_nodes=1)
537         self.assertEqual(1, self.alive_monitor_count())
538         monitor = self.monitor_list()[0].proxy()
539         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
540         self.assertTrue(self.node_shutdown.start.called)
541         self.daemon.update_server_wishlist(
542             [testutil.MockSize(7)]).get(self.TIMEOUT)
543         self.stop_proxy(self.daemon)
544         self.assertFalse(self.node_setup.start.called)
545
546     def test_nodes_shutting_down_count_against_excess(self):
547         size = testutil.MockSize(8)
548         cloud_nodes = [testutil.cloud_node_mock(n, size=size) for n in [8, 9]]
549         arv_nodes = [testutil.arvados_node_mock(n, size=size) for n in [8, 9]]
550         self.make_daemon(cloud_nodes, arv_nodes, [size],
551                          avail_sizes=[(size, {"cores":1})])
552         self.assertEqual(2, self.alive_monitor_count())
553         for mon_ref in self.monitor_list():
554             self.daemon.node_can_shutdown(mon_ref.proxy()).get(self.TIMEOUT)
555         self.assertEqual(1, self.node_shutdown.start.call_count)
556
557     def test_clean_shutdown_waits_for_node_setup_finish(self):
558         new_node = self.start_node_boot()
559         new_node.stop_if_no_cloud_node().get.return_value = False
560         new_node.stop_if_no_cloud_node.reset_mock()
561         self.daemon.shutdown().get(self.TIMEOUT)
562         self.assertTrue(new_node.stop_if_no_cloud_node.called)
563         self.daemon.node_up(new_node).get(self.TIMEOUT)
564         self.assertTrue(new_node.stop.called)
565         self.timer.deliver()
566         self.assertTrue(
567             self.daemon.actor_ref.actor_stopped.wait(self.TIMEOUT))
568
569     def test_wishlist_ignored_after_shutdown(self):
570         new_node = self.start_node_boot()
571         new_node.stop_if_no_cloud_node().get.return_value = False
572         new_node.stop_if_no_cloud_node.reset_mock()
573         self.daemon.shutdown().get(self.TIMEOUT)
574         size = testutil.MockSize(2)
575         self.daemon.update_server_wishlist([size] * 2).get(self.TIMEOUT)
576         self.timer.deliver()
577         self.stop_proxy(self.daemon)
578         self.assertEqual(1, self.node_setup.start.call_count)
579
580     def test_shutdown_actor_stopped_when_cloud_node_delisted(self):
581         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock()])
582         self.assertEqual(1, self.alive_monitor_count())
583         monitor = self.monitor_list()[0].proxy()
584         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
585         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
586         self.stop_proxy(self.daemon)
587         self.assertEqual(
588             1, self.last_shutdown.stop.call_count)
589
590     def test_shutdown_actor_cleanup_copes_with_dead_actors(self):
591         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock()])
592         self.assertEqual(1, self.alive_monitor_count())
593         monitor = self.monitor_list()[0].proxy()
594         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
595         # We're mainly testing that update_cloud_nodes catches and handles
596         # the ActorDeadError.
597         self.last_shutdown.stop.side_effect = pykka.ActorDeadError
598         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
599         self.stop_proxy(self.daemon)
600         self.assertEqual(1, self.last_shutdown.stop.call_count)
601
602     def busywait(self, f):
603         n = 0
604         while not f() and n < 10:
605             time.sleep(.1)
606             n += 1
607         self.assertTrue(f())
608
609     def test_node_create_two_sizes(self):
610         small = testutil.MockSize(1)
611         big = testutil.MockSize(2)
612         avail_sizes = [(testutil.MockSize(1), {"cores":1}),
613                         (testutil.MockSize(2), {"cores":2})]
614         self.make_daemon(want_sizes=[small, small, small, big],
615                          avail_sizes=avail_sizes, max_nodes=4)
616
617         # the daemon runs in another thread, so we need to wait and see
618         # if it does all the work we're expecting it to do before stopping it.
619         self.busywait(lambda: self.node_setup.start.call_count == 4)
620         booting = self.daemon.booting.get(self.TIMEOUT)
621         self.stop_proxy(self.daemon)
622         sizecounts = {a[0].id: 0 for a in avail_sizes}
623         for b in booting.itervalues():
624             sizecounts[b.cloud_size.get().id] += 1
625         logging.info(sizecounts)
626         self.assertEqual(3, sizecounts[small.id])
627         self.assertEqual(1, sizecounts[big.id])
628
629     def test_node_max_nodes_two_sizes(self):
630         small = testutil.MockSize(1)
631         big = testutil.MockSize(2)
632         avail_sizes = [(testutil.MockSize(1), {"cores":1}),
633                         (testutil.MockSize(2), {"cores":2})]
634         self.make_daemon(want_sizes=[small, small, small, big],
635                          avail_sizes=avail_sizes, max_nodes=3)
636
637         # the daemon runs in another thread, so we need to wait and see
638         # if it does all the work we're expecting it to do before stopping it.
639         self.busywait(lambda: self.node_setup.start.call_count == 3)
640         booting = self.daemon.booting.get(self.TIMEOUT)
641         self.stop_proxy(self.daemon)
642         sizecounts = {a[0].id: 0 for a in avail_sizes}
643         for b in booting.itervalues():
644             sizecounts[b.cloud_size.get().id] += 1
645         self.assertEqual(2, sizecounts[small.id])
646         self.assertEqual(1, sizecounts[big.id])
647
648     def test_wishlist_reconfigure(self):
649         small = testutil.MockSize(1)
650         big = testutil.MockSize(2)
651         avail_sizes = [(small, {"cores":1}), (big, {"cores":2})]
652
653         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock(1, small),
654                                       testutil.cloud_node_mock(2, small),
655                                       testutil.cloud_node_mock(3, big)],
656                          arvados_nodes=[testutil.arvados_node_mock(1),
657                                         testutil.arvados_node_mock(2),
658                                         testutil.arvados_node_mock(3)],
659                          want_sizes=[small, small, big],
660                          avail_sizes=avail_sizes)
661
662         self.daemon.update_server_wishlist([small, big, big]).get(self.TIMEOUT)
663
664         self.assertEqual(0, self.node_shutdown.start.call_count)
665
666         for c in self.daemon.cloud_nodes.get().nodes.itervalues():
667             self.daemon.node_can_shutdown(c.actor)
668
669         booting = self.daemon.booting.get()
670         shutdowns = self.daemon.shutdowns.get()
671
672         self.stop_proxy(self.daemon)
673
674         self.assertEqual(1, self.node_setup.start.call_count)
675         self.assertEqual(1, self.node_shutdown.start.call_count)
676
677         # booting a new big node
678         sizecounts = {a[0].id: 0 for a in avail_sizes}
679         for b in booting.itervalues():
680             sizecounts[b.cloud_size.get().id] += 1
681         self.assertEqual(0, sizecounts[small.id])
682         self.assertEqual(1, sizecounts[big.id])
683
684         # shutting down a small node
685         sizecounts = {a[0].id: 0 for a in avail_sizes}
686         for b in shutdowns.itervalues():
687             sizecounts[b.cloud_node.get().size.id] += 1
688         self.assertEqual(1, sizecounts[small.id])
689         self.assertEqual(0, sizecounts[big.id])
690
691     def test_node_max_price(self):
692         small = testutil.MockSize(1)
693         big = testutil.MockSize(2)
694         avail_sizes = [(testutil.MockSize(1), {"cores":1, "price":1}),
695                         (testutil.MockSize(2), {"cores":2, "price":2})]
696         self.make_daemon(want_sizes=[small, small, small, big],
697                          avail_sizes=avail_sizes,
698                          max_nodes=4,
699                          max_total_price=4)
700         # the daemon runs in another thread, so we need to wait and see
701         # if it does all the work we're expecting it to do before stopping it.
702         self.busywait(lambda: self.node_setup.start.call_count == 3)
703         booting = self.daemon.booting.get()
704         self.stop_proxy(self.daemon)
705
706         sizecounts = {a[0].id: 0 for a in avail_sizes}
707         for b in booting.itervalues():
708             sizecounts[b.cloud_size.get().id] += 1
709         logging.info(sizecounts)
710
711         # Booting 3 small nodes and not booting a big node would also partially
712         # satisfy the wishlist and come in under the price cap, however the way
713         # the update_server_wishlist() currently works effectively results in a
714         # round-robin creation of one node of each size in the wishlist, so
715         # test for that.
716         self.assertEqual(2, sizecounts[small.id])
717         self.assertEqual(1, sizecounts[big.id])