8816: Use is_cloud_exception to determine if exception is a "cloud error". Add
[arvados.git] / services / nodemanager / tests / test_computenode_dispatch.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 arvados.errors as arverror
9 import httplib2
10 import mock
11 import pykka
12 import threading
13
14 import arvnodeman.computenode.dispatch as dispatch
15 from . import testutil
16
17 class ComputeNodeSetupActorTestCase(testutil.ActorTestMixin, unittest.TestCase):
18     def make_mocks(self, arvados_effect=None):
19         if arvados_effect is None:
20             arvados_effect = [testutil.arvados_node_mock()]
21         self.arvados_effect = arvados_effect
22         self.timer = testutil.MockTimer()
23         self.api_client = mock.MagicMock(name='api_client')
24         self.api_client.nodes().create().execute.side_effect = arvados_effect
25         self.api_client.nodes().update().execute.side_effect = arvados_effect
26         self.cloud_client = mock.MagicMock(name='cloud_client')
27         self.cloud_client.create_node.return_value = testutil.cloud_node_mock(1)
28
29     def make_actor(self, arv_node=None):
30         if not hasattr(self, 'timer'):
31             self.make_mocks(arvados_effect=[arv_node] if arv_node else None)
32         self.setup_actor = dispatch.ComputeNodeSetupActor.start(
33             self.timer, self.api_client, self.cloud_client,
34             testutil.MockSize(1), arv_node).proxy()
35
36     def assert_node_properties_updated(self, uuid=None,
37                                        size=testutil.MockSize(1)):
38         self.api_client.nodes().update.assert_any_call(
39             uuid=(uuid or self.arvados_effect[-1]['uuid']),
40             body={
41                 'properties': {
42                     'cloud_node': {
43                         'size': size.id,
44                         'price': size.price}}})
45
46     def test_creation_without_arvados_node(self):
47         self.make_actor()
48         finished = threading.Event()
49         self.setup_actor.subscribe(lambda _: finished.set())
50         self.assertEqual(self.arvados_effect[-1],
51                          self.setup_actor.arvados_node.get(self.TIMEOUT))
52         assert(finished.wait(self.TIMEOUT))
53         self.assertEqual(1, self.api_client.nodes().create().execute.call_count)
54         self.assertEqual(1, self.api_client.nodes().update().execute.call_count)
55         self.assert_node_properties_updated()
56         self.assertEqual(self.cloud_client.create_node(),
57                          self.setup_actor.cloud_node.get(self.TIMEOUT))
58
59     def test_creation_with_arvados_node(self):
60         self.make_mocks(arvados_effect=[testutil.arvados_node_mock()]*2)
61         self.make_actor(testutil.arvados_node_mock())
62         finished = threading.Event()
63         self.setup_actor.subscribe(lambda _: finished.set())
64         self.assertEqual(self.arvados_effect[-1],
65                          self.setup_actor.arvados_node.get(self.TIMEOUT))
66         assert(finished.wait(self.TIMEOUT))
67         self.assert_node_properties_updated()
68         self.assertEqual(2, self.api_client.nodes().update().execute.call_count)
69         self.assertEqual(self.cloud_client.create_node(),
70                          self.setup_actor.cloud_node.get(self.TIMEOUT))
71
72     def test_failed_arvados_calls_retried(self):
73         self.make_mocks([
74                 arverror.ApiError(httplib2.Response({'status': '500'}), ""),
75                 testutil.arvados_node_mock(),
76                 ])
77         self.make_actor()
78         self.wait_for_assignment(self.setup_actor, 'arvados_node')
79
80     def test_failed_cloud_calls_retried(self):
81         self.make_mocks()
82         self.cloud_client.create_node.side_effect = [
83             Exception("test cloud creation error"),
84             self.cloud_client.create_node.return_value,
85             ]
86         self.make_actor()
87         self.wait_for_assignment(self.setup_actor, 'cloud_node')
88
89     def test_failed_post_create_retried(self):
90         self.make_mocks()
91         self.cloud_client.post_create_node.side_effect = [
92             Exception("test cloud post-create error"), None]
93         self.make_actor()
94         done = self.FUTURE_CLASS()
95         self.setup_actor.subscribe(done.set)
96         done.get(self.TIMEOUT)
97         self.assertEqual(2, self.cloud_client.post_create_node.call_count)
98
99     def test_stop_when_no_cloud_node(self):
100         self.make_mocks(
101             arverror.ApiError(httplib2.Response({'status': '500'}), ""))
102         self.make_actor()
103         self.assertTrue(
104             self.setup_actor.stop_if_no_cloud_node().get(self.TIMEOUT))
105         self.assertTrue(
106             self.setup_actor.actor_ref.actor_stopped.wait(self.TIMEOUT))
107
108     def test_no_stop_when_cloud_node(self):
109         self.make_actor()
110         self.wait_for_assignment(self.setup_actor, 'cloud_node')
111         self.assertFalse(
112             self.setup_actor.stop_if_no_cloud_node().get(self.TIMEOUT))
113         self.assertTrue(self.stop_proxy(self.setup_actor),
114                         "actor was stopped by stop_if_no_cloud_node")
115
116     def test_subscribe(self):
117         self.make_mocks(
118             arverror.ApiError(httplib2.Response({'status': '500'}), ""))
119         self.make_actor()
120         subscriber = mock.Mock(name='subscriber_mock')
121         self.setup_actor.subscribe(subscriber)
122         retry_resp = [testutil.arvados_node_mock()]
123         self.api_client.nodes().create().execute.side_effect = retry_resp
124         self.api_client.nodes().update().execute.side_effect = retry_resp
125         self.wait_for_assignment(self.setup_actor, 'cloud_node')
126         self.assertEqual(self.setup_actor.actor_ref.actor_urn,
127                          subscriber.call_args[0][0].actor_ref.actor_urn)
128
129     def test_late_subscribe(self):
130         self.make_actor()
131         subscriber = mock.Mock(name='subscriber_mock')
132         self.wait_for_assignment(self.setup_actor, 'cloud_node')
133         self.setup_actor.subscribe(subscriber).get(self.TIMEOUT)
134         self.stop_proxy(self.setup_actor)
135         self.assertEqual(self.setup_actor.actor_ref.actor_urn,
136                          subscriber.call_args[0][0].actor_ref.actor_urn)
137
138
139 class ComputeNodeShutdownActorMixin(testutil.ActorTestMixin):
140     def make_mocks(self, cloud_node=None, arvados_node=None,
141                    shutdown_open=True, node_broken=False):
142         self.timer = testutil.MockTimer()
143         self.shutdowns = testutil.MockShutdownTimer()
144         self.shutdowns._set_state(shutdown_open, 300)
145         self.cloud_client = mock.MagicMock(name='cloud_client')
146         self.cloud_client.broken.return_value = node_broken
147         self.arvados_client = mock.MagicMock(name='arvados_client')
148         self.updates = mock.MagicMock(name='update_mock')
149         if cloud_node is None:
150             cloud_node = testutil.cloud_node_mock()
151         self.cloud_node = cloud_node
152         self.arvados_node = arvados_node
153
154     def make_actor(self, cancellable=True, start_time=None):
155         if not hasattr(self, 'timer'):
156             self.make_mocks()
157         if start_time is None:
158             start_time = time.time()
159         monitor_actor = dispatch.ComputeNodeMonitorActor.start(
160             self.cloud_node, start_time, self.shutdowns,
161             testutil.cloud_node_fqdn, self.timer, self.updates, self.cloud_client,
162             self.arvados_node)
163         self.shutdown_actor = self.ACTOR_CLASS.start(
164             self.timer, self.cloud_client, self.arvados_client, monitor_actor,
165             cancellable).proxy()
166         self.monitor_actor = monitor_actor.proxy()
167
168     def check_success_flag(self, expected, allow_msg_count=1):
169         # allow_msg_count is the number of internal messages that may
170         # need to be handled for shutdown to finish.
171         for try_num in range(1 + allow_msg_count):
172             last_flag = self.shutdown_actor.success.get(self.TIMEOUT)
173             if last_flag is expected:
174                 break
175         else:
176             self.fail("success flag {} is not {}".format(last_flag, expected))
177
178     def test_uncancellable_shutdown(self, *mocks):
179         self.make_mocks(shutdown_open=False)
180         self.cloud_client.destroy_node.return_value = False
181         self.make_actor(cancellable=False)
182         self.check_success_flag(None, 0)
183         self.shutdowns._set_state(True, 600)
184         self.cloud_client.destroy_node.return_value = True
185         self.check_success_flag(True)
186
187     def test_arvados_node_cleaned_after_shutdown(self, *mocks):
188         cloud_node = testutil.cloud_node_mock(62)
189         arv_node = testutil.arvados_node_mock(62)
190         self.make_mocks(cloud_node, arv_node)
191         self.make_actor()
192         self.check_success_flag(True, 3)
193         update_mock = self.arvados_client.nodes().update
194         self.assertTrue(update_mock.called)
195         update_kwargs = update_mock.call_args_list[0][1]
196         self.assertEqual(arv_node['uuid'], update_kwargs.get('uuid'))
197         self.assertIn('body', update_kwargs)
198         for clear_key in ['slot_number', 'hostname', 'ip_address',
199                           'first_ping_at', 'last_ping_at']:
200             self.assertIn(clear_key, update_kwargs['body'])
201             self.assertIsNone(update_kwargs['body'][clear_key])
202         self.assertTrue(update_mock().execute.called)
203
204     def test_arvados_node_not_cleaned_after_shutdown_cancelled(self, *mocks):
205         cloud_node = testutil.cloud_node_mock(61)
206         arv_node = testutil.arvados_node_mock(61)
207         self.make_mocks(cloud_node, arv_node, shutdown_open=False)
208         self.make_actor(cancellable=True)
209         self.check_success_flag(False, 2)
210         self.assertFalse(self.arvados_client.nodes().update.called)
211
212
213 class ComputeNodeShutdownActorTestCase(ComputeNodeShutdownActorMixin,
214                                        unittest.TestCase):
215     ACTOR_CLASS = dispatch.ComputeNodeShutdownActor
216
217     def test_easy_shutdown(self):
218         self.make_actor(start_time=0)
219         self.check_success_flag(True)
220         self.assertTrue(self.cloud_client.destroy_node.called)
221
222     def test_shutdown_cancelled_when_window_closes(self):
223         self.make_mocks(shutdown_open=False)
224         self.make_actor()
225         self.check_success_flag(False, 2)
226         self.assertFalse(self.cloud_client.destroy_node.called)
227         self.assertEqual(self.ACTOR_CLASS.WINDOW_CLOSED,
228                          self.shutdown_actor.cancel_reason.get(self.TIMEOUT))
229
230     def test_shutdown_retries_when_cloud_fails(self):
231         self.make_mocks()
232         self.cloud_client.destroy_node.return_value = False
233         self.make_actor(start_time=0)
234         self.assertIsNone(self.shutdown_actor.success.get(self.TIMEOUT))
235         self.cloud_client.destroy_node.return_value = True
236         self.check_success_flag(True)
237
238     def test_shutdown_cancelled_when_cloud_fails_on_broken_node(self):
239         self.make_mocks(node_broken=True)
240         self.cloud_client.destroy_node.return_value = False
241         self.make_actor(start_time=0)
242         self.check_success_flag(False, 2)
243         self.assertEqual(1, self.cloud_client.destroy_node.call_count)
244         self.assertEqual(self.ACTOR_CLASS.NODE_BROKEN,
245                          self.shutdown_actor.cancel_reason.get(self.TIMEOUT))
246
247     def test_late_subscribe(self):
248         self.make_actor()
249         subscriber = mock.Mock(name='subscriber_mock')
250         self.shutdown_actor.subscribe(subscriber).get(self.TIMEOUT)
251         self.stop_proxy(self.shutdown_actor)
252         self.assertTrue(subscriber.called)
253         self.assertEqual(self.shutdown_actor.actor_ref.actor_urn,
254                          subscriber.call_args[0][0].actor_ref.actor_urn)
255
256
257 class ComputeNodeUpdateActorTestCase(testutil.ActorTestMixin,
258                                      unittest.TestCase):
259     def make_actor(self):
260         self.driver = mock.MagicMock(name='driver_mock')
261         self.updater = dispatch.ComputeNodeUpdateActor.start(self.driver).proxy()
262
263     def test_node_sync(self):
264         self.make_actor()
265         cloud_node = testutil.cloud_node_mock()
266         arv_node = testutil.arvados_node_mock()
267         self.updater.sync_node(cloud_node, arv_node).get(self.TIMEOUT)
268         self.driver().sync_node.assert_called_with(cloud_node, arv_node)
269
270     @testutil.no_sleep
271     def test_node_sync_error(self):
272         self.make_actor()
273         cloud_node = testutil.cloud_node_mock()
274         arv_node = testutil.arvados_node_mock()
275         self.driver().sync_node.side_effect = (IOError, Exception, True)
276         self.updater.sync_node(cloud_node, arv_node).get(self.TIMEOUT)
277         self.updater.sync_node(cloud_node, arv_node).get(self.TIMEOUT)
278         self.updater.sync_node(cloud_node, arv_node).get(self.TIMEOUT)
279         self.driver().sync_node.assert_called_with(cloud_node, arv_node)
280
281 class ComputeNodeMonitorActorTestCase(testutil.ActorTestMixin,
282                                       unittest.TestCase):
283     def make_mocks(self, node_num):
284         self.shutdowns = testutil.MockShutdownTimer()
285         self.shutdowns._set_state(False, 300)
286         self.timer = mock.MagicMock(name='timer_mock')
287         self.updates = mock.MagicMock(name='update_mock')
288         self.cloud_mock = testutil.cloud_node_mock(node_num)
289         self.subscriber = mock.Mock(name='subscriber_mock')
290         self.cloud_client = mock.MagicMock(name='cloud_client')
291         self.cloud_client.broken.return_value = False
292
293     def make_actor(self, node_num=1, arv_node=None, start_time=None):
294         if not hasattr(self, 'cloud_mock'):
295             self.make_mocks(node_num)
296         if start_time is None:
297             start_time = time.time()
298         self.node_actor = dispatch.ComputeNodeMonitorActor.start(
299             self.cloud_mock, start_time, self.shutdowns,
300             testutil.cloud_node_fqdn, self.timer, self.updates, self.cloud_client,
301             arv_node, boot_fail_after=300).proxy()
302         self.node_actor.subscribe(self.subscriber).get(self.TIMEOUT)
303
304     def node_state(self, *states):
305         return self.node_actor.in_state(*states).get(self.TIMEOUT)
306
307     def test_in_state_when_unpaired(self):
308         self.make_actor()
309         self.assertIsNone(self.node_state('idle', 'busy'))
310
311     def test_in_state_when_pairing_stale(self):
312         self.make_actor(arv_node=testutil.arvados_node_mock(
313                 job_uuid=None, age=90000))
314         self.assertIsNone(self.node_state('idle', 'busy'))
315
316     def test_in_state_when_no_state_available(self):
317         self.make_actor(arv_node=testutil.arvados_node_mock(
318                 crunch_worker_state=None))
319         self.assertIsNone(self.node_state('idle', 'busy'))
320
321     def test_in_idle_state(self):
322         self.make_actor(2, arv_node=testutil.arvados_node_mock(job_uuid=None))
323         self.assertTrue(self.node_state('idle'))
324         self.assertFalse(self.node_state('busy'))
325         self.assertTrue(self.node_state('idle', 'busy'))
326
327     def test_in_busy_state(self):
328         self.make_actor(3, arv_node=testutil.arvados_node_mock(job_uuid=True))
329         self.assertFalse(self.node_state('idle'))
330         self.assertTrue(self.node_state('busy'))
331         self.assertTrue(self.node_state('idle', 'busy'))
332
333     def test_init_shutdown_scheduling(self):
334         self.make_actor()
335         self.assertTrue(self.timer.schedule.called)
336         self.assertEqual(300, self.timer.schedule.call_args[0][0])
337
338     def test_shutdown_window_close_scheduling(self):
339         self.make_actor()
340         self.shutdowns._set_state(False, 600)
341         self.timer.schedule.reset_mock()
342         self.node_actor.consider_shutdown().get(self.TIMEOUT)
343         self.stop_proxy(self.node_actor)
344         self.assertTrue(self.timer.schedule.called)
345         self.assertEqual(600, self.timer.schedule.call_args[0][0])
346         self.assertFalse(self.subscriber.called)
347
348     def test_shutdown_subscription(self):
349         self.make_actor(start_time=0)
350         self.shutdowns._set_state(True, 600)
351         self.node_actor.consider_shutdown().get(self.TIMEOUT)
352         self.assertTrue(self.subscriber.called)
353         self.assertEqual(self.node_actor.actor_ref.actor_urn,
354                          self.subscriber.call_args[0][0].actor_ref.actor_urn)
355
356     def test_no_shutdown_booting(self):
357         self.make_actor()
358         self.shutdowns._set_state(True, 600)
359         self.assertTrue(self.node_actor.shutdown_eligible().get(self.TIMEOUT).startswith("node is still booting"))
360
361     def test_shutdown_without_arvados_node(self):
362         self.make_actor(start_time=0)
363         self.shutdowns._set_state(True, 600)
364         self.assertTrue(self.node_actor.shutdown_eligible().get(self.TIMEOUT))
365
366     def test_no_shutdown_missing(self):
367         arv_node = testutil.arvados_node_mock(10, job_uuid=None,
368                                               crunch_worker_state="down",
369                                               last_ping_at='1970-01-01T01:02:03.04050607Z')
370         self.make_actor(10, arv_node)
371         self.shutdowns._set_state(True, 600)
372         self.assertTrue(self.node_actor.shutdown_eligible().get(self.TIMEOUT).startswith("node is not idle."))
373
374     def test_no_shutdown_running_broken(self):
375         arv_node = testutil.arvados_node_mock(12, job_uuid=None,
376                                               crunch_worker_state="down")
377         self.make_actor(12, arv_node)
378         self.shutdowns._set_state(True, 600)
379         self.cloud_client.broken.return_value = True
380         self.assertTrue(self.node_actor.shutdown_eligible().get(self.TIMEOUT).startswith("node is not idle."))
381
382     def test_shutdown_missing_broken(self):
383         arv_node = testutil.arvados_node_mock(11, job_uuid=None,
384                                               crunch_worker_state="down",
385                                               last_ping_at='1970-01-01T01:02:03.04050607Z')
386         self.make_actor(11, arv_node)
387         self.shutdowns._set_state(True, 600)
388         self.cloud_client.broken.return_value = True
389         self.assertTrue(self.node_actor.shutdown_eligible().get(self.TIMEOUT))
390
391     def test_no_shutdown_when_window_closed(self):
392         self.make_actor(3, testutil.arvados_node_mock(3, job_uuid=None))
393         self.assertTrue(self.node_actor.shutdown_eligible().get(self.TIMEOUT).startswith("shutdown window is not open."))
394
395     def test_no_shutdown_when_node_running_job(self):
396         self.make_actor(4, testutil.arvados_node_mock(4, job_uuid=True))
397         self.shutdowns._set_state(True, 600)
398         self.assertTrue(self.node_actor.shutdown_eligible().get(self.TIMEOUT).startswith("node is not idle."))
399
400     def test_no_shutdown_when_node_state_unknown(self):
401         self.make_actor(5, testutil.arvados_node_mock(
402             5, crunch_worker_state=None))
403         self.shutdowns._set_state(True, 600)
404         self.assertTrue(self.node_actor.shutdown_eligible().get(self.TIMEOUT).startswith("node is not idle."))
405
406     def test_no_shutdown_when_node_state_stale(self):
407         self.make_actor(6, testutil.arvados_node_mock(6, age=90000))
408         self.shutdowns._set_state(True, 600)
409         self.assertTrue(self.node_actor.shutdown_eligible().get(self.TIMEOUT).startswith("node is not idle."))
410
411     def test_arvados_node_match(self):
412         self.make_actor(2)
413         arv_node = testutil.arvados_node_mock(
414             2, hostname='compute-two.zzzzz.arvadosapi.com')
415         self.cloud_client.node_id.return_value = '2'
416         pair_id = self.node_actor.offer_arvados_pair(arv_node).get(self.TIMEOUT)
417         self.assertEqual(self.cloud_mock.id, pair_id)
418         self.stop_proxy(self.node_actor)
419         self.updates.sync_node.assert_called_with(self.cloud_mock, arv_node)
420
421     def test_arvados_node_mismatch(self):
422         self.make_actor(3)
423         arv_node = testutil.arvados_node_mock(1)
424         self.assertIsNone(
425             self.node_actor.offer_arvados_pair(arv_node).get(self.TIMEOUT))
426
427     def test_arvados_node_mismatch_first_ping_too_early(self):
428         self.make_actor(4)
429         arv_node = testutil.arvados_node_mock(
430             4, first_ping_at='1971-03-02T14:15:16.1717282Z')
431         self.assertIsNone(
432             self.node_actor.offer_arvados_pair(arv_node).get(self.TIMEOUT))
433
434     def test_update_cloud_node(self):
435         self.make_actor(1)
436         self.make_mocks(2)
437         self.cloud_mock.id = '1'
438         self.node_actor.update_cloud_node(self.cloud_mock)
439         current_cloud = self.node_actor.cloud_node.get(self.TIMEOUT)
440         self.assertEqual([testutil.ip_address_mock(2)],
441                          current_cloud.private_ips)
442
443     def test_missing_cloud_node_update(self):
444         self.make_actor(1)
445         self.node_actor.update_cloud_node(None)
446         current_cloud = self.node_actor.cloud_node.get(self.TIMEOUT)
447         self.assertEqual([testutil.ip_address_mock(1)],
448                          current_cloud.private_ips)
449
450     def test_update_arvados_node(self):
451         self.make_actor(3)
452         job_uuid = 'zzzzz-jjjjj-updatejobnode00'
453         new_arvados = testutil.arvados_node_mock(3, job_uuid)
454         self.node_actor.update_arvados_node(new_arvados)
455         current_arvados = self.node_actor.arvados_node.get(self.TIMEOUT)
456         self.assertEqual(job_uuid, current_arvados['job_uuid'])
457
458     def test_missing_arvados_node_update(self):
459         self.make_actor(4, testutil.arvados_node_mock(4))
460         self.node_actor.update_arvados_node(None)
461         current_arvados = self.node_actor.arvados_node.get(self.TIMEOUT)
462         self.assertEqual(testutil.ip_address_mock(4),
463                          current_arvados['ip_address'])
464
465     def test_update_arvados_node_syncs_when_fqdn_mismatch(self):
466         self.make_mocks(5)
467         self.cloud_mock.extra['testname'] = 'cloudfqdn.zzzzz.arvadosapi.com'
468         self.make_actor()
469         arv_node = testutil.arvados_node_mock(5)
470         self.node_actor.update_arvados_node(arv_node).get(self.TIMEOUT)
471         self.assertEqual(1, self.updates.sync_node.call_count)
472
473     def test_update_arvados_node_skips_sync_when_fqdn_match(self):
474         self.make_mocks(6)
475         arv_node = testutil.arvados_node_mock(6)
476         self.cloud_mock.extra['testname'] ='{n[hostname]}.{n[domain]}'.format(
477             n=arv_node)
478         self.make_actor()
479         self.node_actor.update_arvados_node(arv_node).get(self.TIMEOUT)
480         self.assertEqual(0, self.updates.sync_node.call_count)