11002: Merge branch 'master' into 11002-arvput-crash-fix
[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         mock_actor.tell_proxy.return_value = mock_proxy
30
31         self.last_setup = mock_proxy
32         return mock_actor
33
34     def mock_node_shutdown(self, **kwargs):
35         # Make sure that every time the daemon starts a shutdown actor,
36         # it gets a new mock object back.
37         get_cloud_node = mock.MagicMock()
38         if "node_monitor" in kwargs:
39             get_cloud_node.get.return_value = kwargs["node_monitor"].proxy().cloud_node.get()
40         mock_actor = mock.MagicMock()
41         mock_proxy = mock.NonCallableMock(name='shutdown_mock_proxy',
42                                           cloud_node=get_cloud_node,
43                                           actor_ref=mock_actor)
44
45         mock_actor.proxy.return_value = mock_proxy
46         self.last_shutdown = mock_proxy
47
48         return mock_actor
49
50     def make_daemon(self, cloud_nodes=[], arvados_nodes=[], want_sizes=[],
51                     avail_sizes=None,
52                     min_nodes=0, max_nodes=8,
53                     shutdown_windows=[54, 5, 1],
54                     max_total_price=None):
55         for name in ['cloud_nodes', 'arvados_nodes', 'server_wishlist']:
56             setattr(self, name + '_poller', mock.MagicMock(name=name + '_mock'))
57
58         if not avail_sizes:
59             if cloud_nodes or want_sizes:
60                 avail_sizes=[(c.size, {"cores": int(c.id)}) for c in cloud_nodes] + [(s, {"cores": 1}) for s in want_sizes]
61             else:
62                 avail_sizes=[(testutil.MockSize(1), {"cores": 1})]
63
64         self.arv_factory = mock.MagicMock(name='arvados_mock')
65         api_client = mock.MagicMock(name='api_client')
66         api_client.nodes().create().execute.side_effect = [testutil.arvados_node_mock(1),
67                                                            testutil.arvados_node_mock(2)]
68         self.arv_factory.return_value = api_client
69
70         self.cloud_factory = mock.MagicMock(name='cloud_mock')
71         self.cloud_factory().node_start_time.return_value = time.time()
72         self.cloud_updates = mock.MagicMock(name='updates_mock')
73         self.timer = testutil.MockTimer(deliver_immediately=False)
74         self.cloud_factory().node_id.side_effect = lambda node: node.id
75         self.cloud_factory().broken.return_value = False
76
77         self.node_setup = mock.MagicMock(name='setup_mock')
78         self.node_setup.start.side_effect = self.mock_node_start
79         self.node_setup.reset_mock()
80
81         self.node_shutdown = mock.MagicMock(name='shutdown_mock')
82         self.node_shutdown.start.side_effect = self.mock_node_shutdown
83
84         self.daemon = nmdaemon.NodeManagerDaemonActor.start(
85             self.server_wishlist_poller, self.arvados_nodes_poller,
86             self.cloud_nodes_poller, self.cloud_updates, self.timer,
87             self.arv_factory, self.cloud_factory,
88             shutdown_windows, ServerCalculator(avail_sizes),
89             min_nodes, max_nodes, 600, 1800, 3600,
90             self.node_setup, self.node_shutdown,
91             max_total_price=max_total_price).proxy()
92         if arvados_nodes is not None:
93             self.daemon.update_arvados_nodes(arvados_nodes).get(self.TIMEOUT)
94         if cloud_nodes is not None:
95             self.daemon.update_cloud_nodes(cloud_nodes).get(self.TIMEOUT)
96         if want_sizes is not None:
97             self.daemon.update_server_wishlist(want_sizes).get(self.TIMEOUT)
98
99     def monitor_list(self):
100         return pykka.ActorRegistry.get_by_class(ComputeNodeMonitorActor)
101
102     def monitored_arvados_nodes(self):
103         pairings = []
104         for future in [actor.proxy().arvados_node
105                        for actor in self.monitor_list()]:
106             try:
107                 pairings.append(future.get(self.TIMEOUT))
108             except pykka.ActorDeadError:
109                 pass
110         return pairings
111
112     def alive_monitor_count(self):
113         return len(self.monitored_arvados_nodes())
114
115     def assertShutdownCancellable(self, expected=True):
116         self.assertTrue(self.node_shutdown.start.called)
117         self.assertIs(expected,
118                       self.node_shutdown.start.call_args[1]['cancellable'],
119                       "ComputeNodeShutdownActor incorrectly cancellable")
120
121     def test_easy_node_creation(self):
122         size = testutil.MockSize(1)
123         self.make_daemon(want_sizes=[size])
124         self.stop_proxy(self.daemon)
125         self.assertTrue(self.node_setup.start.called)
126
127     def check_monitors_arvados_nodes(self, *arv_nodes):
128         self.assertItemsEqual(arv_nodes, self.monitored_arvados_nodes())
129
130     def test_node_pairing(self):
131         cloud_node = testutil.cloud_node_mock(1)
132         arv_node = testutil.arvados_node_mock(1)
133         self.make_daemon([cloud_node], [arv_node])
134         self.stop_proxy(self.daemon)
135         self.check_monitors_arvados_nodes(arv_node)
136
137     def test_node_pairing_after_arvados_update(self):
138         cloud_node = testutil.cloud_node_mock(2)
139         self.make_daemon([cloud_node],
140                          [testutil.arvados_node_mock(1, ip_address=None)])
141         arv_node = testutil.arvados_node_mock(2)
142         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
143         self.stop_proxy(self.daemon)
144         self.check_monitors_arvados_nodes(arv_node)
145
146     def test_arvados_node_un_and_re_paired(self):
147         # We need to create the Arvados node mock after spinning up the daemon
148         # to make sure it's new enough to pair with the cloud node.
149         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock(3)],
150                          arvados_nodes=None)
151         arv_node = testutil.arvados_node_mock(3)
152         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
153         self.check_monitors_arvados_nodes(arv_node)
154         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
155         self.assertEqual(0, self.alive_monitor_count())
156         self.daemon.update_cloud_nodes([testutil.cloud_node_mock(3)])
157         self.stop_proxy(self.daemon)
158         self.check_monitors_arvados_nodes(arv_node)
159
160     def test_old_arvados_node_not_double_assigned(self):
161         arv_node = testutil.arvados_node_mock(3, age=9000)
162         size = testutil.MockSize(3)
163         self.make_daemon(arvados_nodes=[arv_node],
164                          avail_sizes=[(size, {"cores":1})])
165         self.daemon.update_server_wishlist([size]).get(self.TIMEOUT)
166         self.daemon.update_server_wishlist([size, size]).get(self.TIMEOUT)
167         self.stop_proxy(self.daemon)
168         used_nodes = [call[1].get('arvados_node')
169                       for call in self.node_setup.start.call_args_list]
170         self.assertEqual(2, len(used_nodes))
171         self.assertIn(arv_node, used_nodes)
172         self.assertIn(None, used_nodes)
173
174     def test_node_count_satisfied(self):
175         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock(1)],
176                          want_sizes=[testutil.MockSize(1)])
177         self.stop_proxy(self.daemon)
178         self.assertFalse(self.node_setup.start.called)
179
180     def test_dont_count_missing_as_busy(self):
181         size = testutil.MockSize(1)
182         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock(1, size=size),
183                                       testutil.cloud_node_mock(2, size=size)],
184                          arvados_nodes=[testutil.arvados_node_mock(1),
185                                         testutil.arvados_node_mock(
186                                             2,
187                                             last_ping_at='1970-01-01T01:02:03.04050607Z')],
188                          want_sizes=[size, size])
189         self.stop_proxy(self.daemon)
190         self.assertTrue(self.node_setup.start.called)
191
192     def test_missing_counts_towards_max(self):
193         size = testutil.MockSize(1)
194         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock(1, size=size),
195                                       testutil.cloud_node_mock(2, size=size)],
196                          arvados_nodes=[testutil.arvados_node_mock(1),
197                                         testutil.arvados_node_mock(2, last_ping_at='1970-01-01T01:02:03.04050607Z')],
198                          want_sizes=[size, size],
199                          max_nodes=2)
200         self.stop_proxy(self.daemon)
201         self.assertFalse(self.node_setup.start.called)
202
203     def test_excess_counts_missing(self):
204         size = testutil.MockSize(1)
205         cloud_nodes = [testutil.cloud_node_mock(1, size=size), testutil.cloud_node_mock(2, size=size)]
206         self.make_daemon(cloud_nodes=cloud_nodes,
207                          arvados_nodes=[testutil.arvados_node_mock(1),
208                                         testutil.arvados_node_mock(2, last_ping_at='1970-01-01T01:02:03.04050607Z')],
209                          want_sizes=[size])
210         self.assertEqual(2, self.alive_monitor_count())
211         for mon_ref in self.monitor_list():
212             self.daemon.node_can_shutdown(mon_ref.proxy()).get(self.TIMEOUT)
213         self.assertEqual(1, self.node_shutdown.start.call_count)
214
215     def test_missing_shutdown_not_excess(self):
216         size = testutil.MockSize(1)
217         cloud_nodes = [testutil.cloud_node_mock(1, size=size), testutil.cloud_node_mock(2, size=size)]
218         self.make_daemon(cloud_nodes=cloud_nodes,
219                          arvados_nodes=[testutil.arvados_node_mock(1),
220                                         testutil.arvados_node_mock(2, last_ping_at='1970-01-01T01:02:03.04050607Z')],
221                          want_sizes=[size])
222
223         get_cloud_node = mock.MagicMock(name="get_cloud_node")
224         get_cloud_node.get.return_value = cloud_nodes[1]
225         mock_node_monitor = mock.MagicMock()
226         mock_node_monitor.proxy.return_value = mock.NonCallableMock(cloud_node=get_cloud_node)
227         mock_shutdown = self.node_shutdown.start(node_monitor=mock_node_monitor)
228
229         self.daemon.cloud_nodes.get()[cloud_nodes[1].id].shutdown_actor = mock_shutdown.proxy()
230
231         self.assertEqual(2, self.alive_monitor_count())
232         for mon_ref in self.monitor_list():
233             self.daemon.node_can_shutdown(mon_ref.proxy()).get(self.TIMEOUT)
234         self.assertEqual(1, self.node_shutdown.start.call_count)
235
236     def test_booting_nodes_counted(self):
237         cloud_node = testutil.cloud_node_mock(1)
238         arv_node = testutil.arvados_node_mock(1)
239         server_wishlist = [testutil.MockSize(1)] * 2
240         self.make_daemon([cloud_node], [arv_node], server_wishlist)
241         self.daemon.max_nodes.get(self.TIMEOUT)
242         self.assertTrue(self.node_setup.start.called)
243         self.daemon.update_server_wishlist(server_wishlist).get(self.TIMEOUT)
244         self.stop_proxy(self.daemon)
245         self.assertEqual(1, self.node_setup.start.call_count)
246
247     def test_boot_new_node_when_all_nodes_busy(self):
248         size = testutil.MockSize(2)
249         arv_node = testutil.arvados_node_mock(2, job_uuid=True)
250         self.make_daemon([testutil.cloud_node_mock(2, size=size)], [arv_node],
251                          [size], avail_sizes=[(size, {"cores":1})])
252         self.busywait(lambda: self.node_setup.start.called)
253         self.stop_proxy(self.daemon)
254         self.assertTrue(self.node_setup.start.called)
255
256     def test_boot_new_node_below_min_nodes(self):
257         min_size = testutil.MockSize(1)
258         wish_size = testutil.MockSize(3)
259         avail_sizes = [(min_size, {"cores": 1}),
260                        (wish_size, {"cores": 3})]
261         self.make_daemon([], [], None, avail_sizes=avail_sizes, min_nodes=2)
262         self.daemon.update_server_wishlist([wish_size]).get(self.TIMEOUT)
263         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
264         self.daemon.update_server_wishlist([wish_size]).get(self.TIMEOUT)
265         self.stop_proxy(self.daemon)
266         self.assertEqual([wish_size, min_size],
267                          [call[1].get('cloud_size')
268                           for call in self.node_setup.start.call_args_list])
269
270     def test_no_new_node_when_ge_min_nodes_busy(self):
271         size = testutil.MockSize(2)
272         cloud_nodes = [testutil.cloud_node_mock(n, size=size) for n in range(1, 4)]
273         arv_nodes = [testutil.arvados_node_mock(n, job_uuid=True)
274                      for n in range(1, 4)]
275         self.make_daemon(cloud_nodes, arv_nodes, [], min_nodes=2)
276         self.stop_proxy(self.daemon)
277         self.assertEqual(0, self.node_setup.start.call_count)
278
279     def test_no_new_node_when_max_nodes_busy(self):
280         size = testutil.MockSize(3)
281         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock(3)],
282                          arvados_nodes=[testutil.arvados_node_mock(3, job_uuid=True)],
283                          want_sizes=[size],
284                          max_nodes=1)
285         self.stop_proxy(self.daemon)
286         self.assertFalse(self.node_setup.start.called)
287
288     def start_node_boot(self, cloud_node=None, arv_node=None, id_num=1):
289         if cloud_node is None:
290             cloud_node = testutil.cloud_node_mock(id_num)
291         id_num = int(cloud_node.id)
292         if arv_node is None:
293             arv_node = testutil.arvados_node_mock(id_num)
294         self.make_daemon(want_sizes=[testutil.MockSize(id_num)],
295                          avail_sizes=[(testutil.MockSize(id_num), {"cores":1})])
296         self.daemon.max_nodes.get(self.TIMEOUT)
297         self.assertEqual(1, self.node_setup.start.call_count)
298         self.last_setup.cloud_node.get.return_value = cloud_node
299         self.last_setup.arvados_node.get.return_value = arv_node
300         return self.last_setup
301
302     def test_new_node_when_booted_node_not_usable(self):
303         cloud_node = testutil.cloud_node_mock(4)
304         arv_node = testutil.arvados_node_mock(4, crunch_worker_state='down')
305         setup = self.start_node_boot(cloud_node, arv_node)
306         self.daemon.node_up(setup).get(self.TIMEOUT)
307         self.assertEqual(1, self.alive_monitor_count())
308         self.daemon.update_arvados_nodes([arv_node])
309         self.daemon.update_cloud_nodes([cloud_node])
310         self.monitor_list()[0].proxy().cloud_node_start_time = time.time()-1801
311         self.daemon.update_server_wishlist(
312             [testutil.MockSize(4)]).get(self.TIMEOUT)
313         self.stop_proxy(self.daemon)
314         self.assertEqual(2, self.node_setup.start.call_count)
315
316     def test_no_duplication_when_booting_node_listed_fast(self):
317         # Test that we don't start two ComputeNodeMonitorActors when
318         # we learn about a booting node through a listing before we
319         # get the "node up" message from CloudNodeSetupActor.
320         cloud_node = testutil.cloud_node_mock(1)
321         setup = self.start_node_boot(cloud_node)
322         self.daemon.update_cloud_nodes([cloud_node])
323         self.daemon.node_up(setup).get(self.TIMEOUT)
324         self.assertEqual(1, self.alive_monitor_count())
325
326     def test_no_duplication_when_booted_node_listed(self):
327         cloud_node = testutil.cloud_node_mock(2)
328         setup = self.start_node_boot(cloud_node, id_num=2)
329         self.daemon.node_up(setup)
330         self.daemon.update_cloud_nodes([cloud_node]).get(self.TIMEOUT)
331         self.assertEqual(1, self.alive_monitor_count())
332
333     def test_node_counted_after_boot_with_slow_listing(self):
334         # Test that, after we boot a compute node, we assume it exists
335         # even it doesn't appear in the listing (e.g., because of delays
336         # propagating tags).
337         setup = self.start_node_boot()
338         self.daemon.node_up(setup).get(self.TIMEOUT)
339         self.assertEqual(1, self.alive_monitor_count())
340         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
341         self.assertEqual(1, self.alive_monitor_count())
342
343     def test_booted_unlisted_node_counted(self):
344         setup = self.start_node_boot(id_num=1)
345         self.daemon.node_up(setup)
346         self.daemon.update_server_wishlist(
347             [testutil.MockSize(1)]).get(self.TIMEOUT)
348         self.stop_proxy(self.daemon)
349         self.assertEqual(1, self.node_setup.start.call_count)
350
351     def test_booted_node_can_shutdown(self):
352         setup = self.start_node_boot()
353         self.daemon.node_up(setup).get(self.TIMEOUT)
354         self.assertEqual(1, self.alive_monitor_count())
355         monitor = self.monitor_list()[0].proxy()
356         self.daemon.update_server_wishlist([])
357         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
358         self.stop_proxy(self.daemon)
359         self.assertTrue(self.node_shutdown.start.called,
360                         "daemon did not shut down booted node on offer")
361
362     def test_booted_node_lifecycle(self):
363         cloud_node = testutil.cloud_node_mock(6)
364         setup = self.start_node_boot(cloud_node, id_num=6)
365         self.daemon.node_up(setup).get(self.TIMEOUT)
366         self.assertEqual(1, self.alive_monitor_count())
367         monitor = self.monitor_list()[0].proxy()
368         self.daemon.update_server_wishlist([])
369         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
370         self.assertShutdownCancellable(True)
371         shutdown = self.node_shutdown.start().proxy()
372         shutdown.cloud_node.get.return_value = cloud_node
373         self.daemon.node_finished_shutdown(shutdown).get(self.TIMEOUT)
374         self.daemon.update_cloud_nodes([])
375         self.assertTrue(shutdown.stop.called,
376                         "shutdown actor not stopped after finishing")
377         self.assertTrue(monitor.actor_ref.actor_stopped.wait(self.TIMEOUT),
378                         "monitor for booted node not stopped after shutdown")
379         self.daemon.update_server_wishlist(
380             [testutil.MockSize(2)]).get(self.TIMEOUT)
381         self.stop_proxy(self.daemon)
382         self.assertTrue(self.node_setup.start.called,
383                         "second node not started after booted node stopped")
384
385     def test_booted_node_shut_down_when_never_listed(self):
386         setup = self.start_node_boot()
387         self.cloud_factory().node_start_time.return_value = time.time() - 3601
388         self.daemon.node_up(setup).get(self.TIMEOUT)
389         self.assertEqual(1, self.alive_monitor_count())
390         self.assertFalse(self.node_shutdown.start.called)
391         now = time.time()
392         self.monitor_list()[0].tell_proxy().consider_shutdown()
393         self.busywait(lambda: self.node_shutdown.start.called)
394         self.stop_proxy(self.daemon)
395         self.assertShutdownCancellable(False)
396
397     def test_booted_node_shut_down_when_never_paired(self):
398         cloud_node = testutil.cloud_node_mock(2)
399         setup = self.start_node_boot(cloud_node)
400         self.cloud_factory().node_start_time.return_value = time.time() - 3601
401         self.daemon.node_up(setup).get(self.TIMEOUT)
402         self.assertEqual(1, self.alive_monitor_count())
403         self.daemon.update_cloud_nodes([cloud_node])
404         self.monitor_list()[0].tell_proxy().consider_shutdown()
405         self.busywait(lambda: self.node_shutdown.start.called)
406         self.stop_proxy(self.daemon)
407         self.assertShutdownCancellable(False)
408
409     def test_booted_node_shut_down_when_never_working(self):
410         cloud_node = testutil.cloud_node_mock(4)
411         arv_node = testutil.arvados_node_mock(4, crunch_worker_state='down')
412         setup = self.start_node_boot(cloud_node, arv_node)
413         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
414         self.daemon.node_up(setup).get(self.TIMEOUT)
415         self.assertEqual(1, self.alive_monitor_count())
416         self.monitor_list()[0].proxy().cloud_node_start_time = time.time()-3601
417         self.daemon.update_cloud_nodes([cloud_node])
418         self.busywait(lambda: self.node_shutdown.start.called)
419         self.stop_proxy(self.daemon)
420         self.assertShutdownCancellable(False)
421
422     def test_node_that_pairs_not_considered_failed_boot(self):
423         cloud_node = testutil.cloud_node_mock(3)
424         arv_node = testutil.arvados_node_mock(3)
425         setup = self.start_node_boot(cloud_node, arv_node)
426         self.daemon.node_up(setup).get(self.TIMEOUT)
427         self.assertEqual(1, self.alive_monitor_count())
428         self.daemon.update_cloud_nodes([cloud_node])
429         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
430         self.timer.deliver()
431         self.stop_proxy(self.daemon)
432         self.assertFalse(self.node_shutdown.start.called)
433
434     def test_node_that_pairs_busy_not_considered_failed_boot(self):
435         cloud_node = testutil.cloud_node_mock(5)
436         arv_node = testutil.arvados_node_mock(5, job_uuid=True)
437         setup = self.start_node_boot(cloud_node, arv_node)
438         self.daemon.node_up(setup).get(self.TIMEOUT)
439         self.assertEqual(1, self.alive_monitor_count())
440         self.daemon.update_cloud_nodes([cloud_node])
441         self.daemon.update_arvados_nodes([arv_node]).get(self.TIMEOUT)
442         self.timer.deliver()
443         self.stop_proxy(self.daemon)
444         self.assertFalse(self.node_shutdown.start.called)
445
446     def test_booting_nodes_shut_down(self):
447         self.make_daemon(want_sizes=[testutil.MockSize(1)])
448         self.daemon.update_server_wishlist([]).get(self.TIMEOUT)
449         self.stop_proxy(self.daemon)
450         self.assertTrue(self.last_setup.stop_if_no_cloud_node.called)
451
452     def test_all_booting_nodes_tried_to_shut_down(self):
453         size = testutil.MockSize(2)
454         self.make_daemon(want_sizes=[size], avail_sizes=[(size, {"cores":1})])
455         self.daemon.max_nodes.get(self.TIMEOUT)
456         setup1 = self.last_setup
457         setup1.stop_if_no_cloud_node().get.return_value = False
458         setup1.stop_if_no_cloud_node.reset_mock()
459         self.daemon.update_server_wishlist([size, size]).get(self.TIMEOUT)
460         self.daemon.max_nodes.get(self.TIMEOUT)
461         self.assertIsNot(setup1, self.last_setup)
462         self.last_setup.stop_if_no_cloud_node().get.return_value = True
463         self.last_setup.stop_if_no_cloud_node.reset_mock()
464         self.daemon.update_server_wishlist([]).get(self.TIMEOUT)
465         self.daemon.max_nodes.get(self.TIMEOUT)
466         self.stop_proxy(self.daemon)
467         self.assertEqual(1, self.last_setup.stop_if_no_cloud_node.call_count)
468         self.assertTrue(setup1.stop_if_no_cloud_node.called)
469
470     def test_shutdown_declined_at_wishlist_capacity(self):
471         cloud_node = testutil.cloud_node_mock(1)
472         arv_node = testutil.arvados_node_mock(1)
473         size = testutil.MockSize(1)
474         self.make_daemon(cloud_nodes=[cloud_node], arvados_nodes=[arv_node], want_sizes=[size])
475         self.assertEqual(1, self.alive_monitor_count())
476         monitor = self.monitor_list()[0].proxy()
477         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
478         self.stop_proxy(self.daemon)
479         self.assertFalse(self.node_shutdown.start.called)
480
481     def test_shutdown_declined_below_min_nodes(self):
482         cloud_node = testutil.cloud_node_mock(1)
483         arv_node = testutil.arvados_node_mock(1)
484         self.make_daemon(cloud_nodes=[cloud_node], arvados_nodes=[arv_node], min_nodes=1)
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.stop_proxy(self.daemon)
489         self.assertFalse(self.node_shutdown.start.called)
490
491     def test_shutdown_accepted_below_capacity(self):
492         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock()])
493         self.assertEqual(1, self.alive_monitor_count())
494         monitor = self.monitor_list()[0].proxy()
495         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
496         self.stop_proxy(self.daemon)
497         self.assertTrue(self.node_shutdown.start.called)
498
499     def test_shutdown_declined_when_idle_and_job_queued(self):
500         size = testutil.MockSize(1)
501         cloud_nodes = [testutil.cloud_node_mock(n, size=size) for n in [3, 4]]
502         arv_nodes = [testutil.arvados_node_mock(3, job_uuid=True),
503                      testutil.arvados_node_mock(4, job_uuid=None)]
504         self.make_daemon(cloud_nodes, arv_nodes, [size])
505         self.assertEqual(2, self.alive_monitor_count())
506         for mon_ref in self.monitor_list():
507             monitor = mon_ref.proxy()
508             if monitor.cloud_node.get(self.TIMEOUT) is cloud_nodes[-1]:
509                 break
510         else:
511             self.fail("monitor for idle node not found")
512         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
513         self.stop_proxy(self.daemon)
514         self.assertFalse(self.node_shutdown.start.called)
515
516     def test_node_shutdown_after_cancelled_shutdown(self):
517         cloud_node = testutil.cloud_node_mock(5)
518         self.make_daemon([cloud_node], [testutil.arvados_node_mock(5)])
519         self.assertEqual(1, self.alive_monitor_count())
520         monitor = self.monitor_list()[0].proxy()
521         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
522         self.last_shutdown.success.get.return_value = False
523         self.daemon.node_finished_shutdown(self.last_shutdown).get(self.TIMEOUT)
524         self.assertEqual(1, self.alive_monitor_count())
525
526         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
527         self.last_shutdown.success.get.return_value = True
528         self.last_shutdown.stop.side_effect = lambda: monitor.stop()
529         self.daemon.node_finished_shutdown(self.last_shutdown).get(self.TIMEOUT)
530         self.assertEqual(0, self.alive_monitor_count())
531
532     def test_broken_node_not_counted(self):
533         size = testutil.MockSize(8)
534         cloud_node = testutil.cloud_node_mock(8, size=size)
535         wishlist = [size]
536         self.make_daemon([cloud_node], [testutil.arvados_node_mock(8)],
537                          wishlist, avail_sizes=[(size, {"cores":1})])
538         self.assertEqual(1, self.alive_monitor_count())
539         self.assertFalse(self.node_setup.start.called)
540         monitor = self.monitor_list()[0].proxy()
541         shutdown_proxy = self.node_shutdown.start().proxy
542         shutdown_proxy().cloud_node.get.return_value = cloud_node
543         shutdown_proxy().success.get.return_value = False
544         self.cloud_factory().broken.return_value = True
545         self.daemon.update_server_wishlist([]).get(self.TIMEOUT)
546         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
547         self.daemon.node_finished_shutdown(shutdown_proxy()).get(self.TIMEOUT)
548         self.daemon.update_cloud_nodes([cloud_node]).get(self.TIMEOUT)
549         self.daemon.update_server_wishlist(wishlist).get(self.TIMEOUT)
550         self.stop_proxy(self.daemon)
551         self.assertEqual(1, self.node_setup.start.call_count)
552
553     def test_nodes_shutting_down_replaced_below_max_nodes(self):
554         size = testutil.MockSize(6)
555         cloud_node = testutil.cloud_node_mock(6, size=size)
556         self.make_daemon([cloud_node], [testutil.arvados_node_mock(6, crunch_worker_state='down')],
557                          avail_sizes=[(size, {"cores":1})])
558         self.assertEqual(1, self.alive_monitor_count())
559         monitor = self.monitor_list()[0].proxy()
560         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
561         self.assertTrue(self.node_shutdown.start.called)
562         self.daemon.update_server_wishlist(
563             [testutil.MockSize(6)]).get(self.TIMEOUT)
564         self.stop_proxy(self.daemon)
565         self.assertTrue(self.node_setup.start.called)
566
567     def test_nodes_shutting_down_not_replaced_at_max_nodes(self):
568         cloud_node = testutil.cloud_node_mock(7)
569         self.make_daemon([cloud_node], [testutil.arvados_node_mock(7)],
570                          max_nodes=1)
571         self.assertEqual(1, self.alive_monitor_count())
572         monitor = self.monitor_list()[0].proxy()
573         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
574         self.assertTrue(self.node_shutdown.start.called)
575         self.daemon.update_server_wishlist(
576             [testutil.MockSize(7)]).get(self.TIMEOUT)
577         self.stop_proxy(self.daemon)
578         self.assertFalse(self.node_setup.start.called)
579
580     def test_nodes_shutting_down_count_against_excess(self):
581         size = testutil.MockSize(8)
582         cloud_nodes = [testutil.cloud_node_mock(n, size=size) for n in [8, 9]]
583         arv_nodes = [testutil.arvados_node_mock(n, size=size) for n in [8, 9]]
584         self.make_daemon(cloud_nodes, arv_nodes, [size],
585                          avail_sizes=[(size, {"cores":1})])
586         self.assertEqual(2, self.alive_monitor_count())
587         for mon_ref in self.monitor_list():
588             self.daemon.node_can_shutdown(mon_ref.proxy()).get(self.TIMEOUT)
589         self.assertEqual(1, self.node_shutdown.start.call_count)
590
591     def test_clean_shutdown_waits_for_node_setup_finish(self):
592         new_node = self.start_node_boot()
593         new_node.stop_if_no_cloud_node().get.return_value = False
594         new_node.stop_if_no_cloud_node.reset_mock()
595         self.daemon.shutdown().get(self.TIMEOUT)
596         self.assertTrue(new_node.stop_if_no_cloud_node.called)
597         self.daemon.node_up(new_node).get(self.TIMEOUT)
598         self.assertTrue(new_node.stop.called)
599         self.timer.deliver()
600         self.assertTrue(
601             self.daemon.actor_ref.actor_stopped.wait(self.TIMEOUT))
602
603     def test_wishlist_ignored_after_shutdown(self):
604         new_node = self.start_node_boot()
605         new_node.stop_if_no_cloud_node().get.return_value = False
606         new_node.stop_if_no_cloud_node.reset_mock()
607         self.daemon.shutdown().get(self.TIMEOUT)
608         size = testutil.MockSize(2)
609         self.daemon.update_server_wishlist([size] * 2).get(self.TIMEOUT)
610         self.timer.deliver()
611         self.stop_proxy(self.daemon)
612         self.assertEqual(1, self.node_setup.start.call_count)
613
614     def test_shutdown_actor_stopped_when_cloud_node_delisted(self):
615         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock()])
616         self.assertEqual(1, self.alive_monitor_count())
617         monitor = self.monitor_list()[0].proxy()
618         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
619         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
620         self.stop_proxy(self.daemon)
621         self.assertEqual(
622             1, self.last_shutdown.stop.call_count)
623
624     def test_shutdown_actor_cleanup_copes_with_dead_actors(self):
625         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock()])
626         self.assertEqual(1, self.alive_monitor_count())
627         monitor = self.monitor_list()[0].proxy()
628         self.daemon.node_can_shutdown(monitor).get(self.TIMEOUT)
629         # We're mainly testing that update_cloud_nodes catches and handles
630         # the ActorDeadError.
631         self.last_shutdown.stop.side_effect = pykka.ActorDeadError
632         self.daemon.update_cloud_nodes([]).get(self.TIMEOUT)
633         self.stop_proxy(self.daemon)
634         self.assertEqual(1, self.last_shutdown.stop.call_count)
635
636     def test_node_create_two_sizes(self):
637         small = testutil.MockSize(1)
638         big = testutil.MockSize(2)
639         avail_sizes = [(testutil.MockSize(1), {"cores":1}),
640                         (testutil.MockSize(2), {"cores":2})]
641         self.make_daemon(want_sizes=[small, small, small, big],
642                          avail_sizes=avail_sizes, max_nodes=4)
643
644         # the daemon runs in another thread, so we need to wait and see
645         # if it does all the work we're expecting it to do before stopping it.
646         self.busywait(lambda: self.node_setup.start.call_count == 4)
647         booting = self.daemon.booting.get(self.TIMEOUT)
648         self.stop_proxy(self.daemon)
649         sizecounts = {a[0].id: 0 for a in avail_sizes}
650         for b in booting.itervalues():
651             sizecounts[b.cloud_size.get().id] += 1
652         logging.info(sizecounts)
653         self.assertEqual(3, sizecounts[small.id])
654         self.assertEqual(1, sizecounts[big.id])
655
656     def test_node_max_nodes_two_sizes(self):
657         small = testutil.MockSize(1)
658         big = testutil.MockSize(2)
659         avail_sizes = [(testutil.MockSize(1), {"cores":1}),
660                         (testutil.MockSize(2), {"cores":2})]
661         self.make_daemon(want_sizes=[small, small, small, big],
662                          avail_sizes=avail_sizes, max_nodes=3)
663
664         # the daemon runs in another thread, so we need to wait and see
665         # if it does all the work we're expecting it to do before stopping it.
666         self.busywait(lambda: self.node_setup.start.call_count == 3)
667         booting = self.daemon.booting.get(self.TIMEOUT)
668         self.stop_proxy(self.daemon)
669         sizecounts = {a[0].id: 0 for a in avail_sizes}
670         for b in booting.itervalues():
671             sizecounts[b.cloud_size.get().id] += 1
672         self.assertEqual(2, sizecounts[small.id])
673         self.assertEqual(1, sizecounts[big.id])
674
675     def test_wishlist_reconfigure(self):
676         small = testutil.MockSize(1)
677         big = testutil.MockSize(2)
678         avail_sizes = [(small, {"cores":1}), (big, {"cores":2})]
679
680         self.make_daemon(cloud_nodes=[testutil.cloud_node_mock(1, small),
681                                       testutil.cloud_node_mock(2, small),
682                                       testutil.cloud_node_mock(3, big)],
683                          arvados_nodes=[testutil.arvados_node_mock(1),
684                                         testutil.arvados_node_mock(2),
685                                         testutil.arvados_node_mock(3)],
686                          want_sizes=[small, small, big],
687                          avail_sizes=avail_sizes)
688
689         self.daemon.update_server_wishlist([small, big, big]).get(self.TIMEOUT)
690
691         self.assertEqual(0, self.node_shutdown.start.call_count)
692
693         for c in self.daemon.cloud_nodes.get().nodes.itervalues():
694             self.daemon.node_can_shutdown(c.actor)
695
696         booting = self.daemon.booting.get()
697         cloud_nodes = self.daemon.cloud_nodes.get()
698
699         self.stop_proxy(self.daemon)
700
701         self.assertEqual(1, self.node_setup.start.call_count)
702         self.assertEqual(1, self.node_shutdown.start.call_count)
703
704         # booting a new big node
705         sizecounts = {a[0].id: 0 for a in avail_sizes}
706         for b in booting.itervalues():
707             sizecounts[b.cloud_size.get().id] += 1
708         self.assertEqual(0, sizecounts[small.id])
709         self.assertEqual(1, sizecounts[big.id])
710
711         # shutting down a small node
712         sizecounts = {a[0].id: 0 for a in avail_sizes}
713         for b in cloud_nodes.nodes.itervalues():
714             if b.shutdown_actor is not None:
715                 sizecounts[b.cloud_node.size.id] += 1
716         self.assertEqual(1, sizecounts[small.id])
717         self.assertEqual(0, sizecounts[big.id])
718
719     def test_node_max_price(self):
720         small = testutil.MockSize(1)
721         big = testutil.MockSize(2)
722         avail_sizes = [(testutil.MockSize(1), {"cores":1, "price":1}),
723                         (testutil.MockSize(2), {"cores":2, "price":2})]
724         self.make_daemon(want_sizes=[small, small, small, big],
725                          avail_sizes=avail_sizes,
726                          max_nodes=4,
727                          max_total_price=4)
728         # the daemon runs in another thread, so we need to wait and see
729         # if it does all the work we're expecting it to do before stopping it.
730         self.busywait(lambda: self.node_setup.start.call_count == 3)
731         booting = self.daemon.booting.get()
732         self.stop_proxy(self.daemon)
733
734         sizecounts = {a[0].id: 0 for a in avail_sizes}
735         for b in booting.itervalues():
736             sizecounts[b.cloud_size.get().id] += 1
737         logging.info(sizecounts)
738
739         # Booting 3 small nodes and not booting a big node would also partially
740         # satisfy the wishlist and come in under the price cap, however the way
741         # the update_server_wishlist() currently works effectively results in a
742         # round-robin creation of one node of each size in the wishlist, so
743         # test for that.
744         self.assertEqual(2, sizecounts[small.id])
745         self.assertEqual(1, sizecounts[big.id])