8953: Drained SLURM nodes can be eligible for shutdown.
[arvados.git] / services / nodemanager / tests / test_computenode_dispatch_slurm.py
1 #!/usr/bin/env python
2
3 from __future__ import absolute_import, print_function
4
5 import subprocess
6 import time
7 import unittest
8
9 import mock
10
11 import arvnodeman.computenode.dispatch.slurm as slurm_dispatch
12 from . import testutil
13 from .test_computenode_dispatch import ComputeNodeShutdownActorMixin
14
15 @mock.patch('subprocess.check_output')
16 class SLURMComputeNodeShutdownActorTestCase(ComputeNodeShutdownActorMixin,
17                                             unittest.TestCase):
18     ACTOR_CLASS = slurm_dispatch.ComputeNodeShutdownActor
19
20     def check_slurm_got_args(self, proc_mock, *args):
21         self.assertTrue(proc_mock.called)
22         slurm_cmd = proc_mock.call_args[0][0]
23         for s in args:
24             self.assertIn(s, slurm_cmd)
25
26     def check_success_after_reset(self, proc_mock, end_state='drain\n'):
27         self.make_mocks(arvados_node=testutil.arvados_node_mock(63))
28         self.make_actor()
29         self.check_success_flag(None, 0)
30         self.check_success_flag(None, 0)
31         # Order is critical here: if the mock gets called when no return value
32         # or side effect is set, we may invoke a real subprocess.
33         proc_mock.return_value = end_state
34         proc_mock.side_effect = None
35         self.check_success_flag(True, 3)
36         self.check_slurm_got_args(proc_mock, 'compute63')
37
38     def make_wait_state_test(start_state='drng\n', end_state='drain\n'):
39         def test(self, proc_mock):
40             proc_mock.return_value = start_state
41             self.check_success_after_reset(proc_mock, end_state)
42         return test
43
44     for wait_state in ['alloc\n', 'drng\n', 'idle*\n']:
45         locals()['test_wait_while_' + wait_state.strip()
46                  ] = make_wait_state_test(start_state=wait_state)
47
48     for end_state in ['down\n', 'down*\n', 'drain\n', 'fail\n']:
49         locals()['test_wait_until_' + end_state.strip()
50                  ] = make_wait_state_test(end_state=end_state)
51
52     def test_retry_failed_slurm_calls(self, proc_mock):
53         proc_mock.side_effect = subprocess.CalledProcessError(1, ["mock"])
54         self.check_success_after_reset(proc_mock)
55
56     def test_slurm_bypassed_when_no_arvados_node(self, proc_mock):
57         # Test we correctly handle a node that failed to bootstrap.
58         proc_mock.return_value = 'idle\n'
59         self.make_actor(start_time=0)
60         self.check_success_flag(True)
61         self.assertFalse(proc_mock.called)
62
63     def test_node_undrained_when_shutdown_window_closes(self, proc_mock):
64         proc_mock.side_effect = iter(['drng\n', 'idle\n'])
65         self.make_mocks(arvados_node=testutil.arvados_node_mock(job_uuid=True))
66         self.make_actor()
67         self.check_success_flag(False, 2)
68         self.check_slurm_got_args(proc_mock, 'NodeName=compute99', 'State=RESUME')
69
70     def test_alloc_node_undrained_when_shutdown_window_closes(self, proc_mock):
71         proc_mock.side_effect = iter(['alloc\n'])
72         self.make_mocks(arvados_node=testutil.arvados_node_mock(job_uuid=True))
73         self.make_actor()
74         self.check_success_flag(False, 2)
75         self.check_slurm_got_args(proc_mock, 'sinfo', '--noheader', '-o', '%t', '-n', 'compute99')
76
77     def test_cancel_shutdown_retry(self, proc_mock):
78         proc_mock.side_effect = iter([OSError, 'drain\n', OSError, 'idle\n'])
79         self.make_mocks(arvados_node=testutil.arvados_node_mock(job_uuid=True))
80         self.make_actor()
81         self.check_success_flag(False, 2)
82
83     def test_issue_slurm_drain_retry(self, proc_mock):
84         proc_mock.side_effect = iter([OSError, '', OSError, 'drng\n'])
85         self.check_success_after_reset(proc_mock)
86
87     def test_arvados_node_cleaned_after_shutdown(self, proc_mock):
88         proc_mock.return_value = 'drain\n'
89         super(SLURMComputeNodeShutdownActorTestCase,
90               self).test_arvados_node_cleaned_after_shutdown()
91
92 class SLURMComputeNodeMonitorActorTestCase(testutil.ActorTestMixin,
93                                       unittest.TestCase):
94
95     def make_mocks(self, node_num):
96         self.shutdowns = testutil.MockShutdownTimer()
97         self.shutdowns._set_state(False, 300)
98         self.timer = mock.MagicMock(name='timer_mock')
99         self.updates = mock.MagicMock(name='update_mock')
100         self.cloud_mock = testutil.cloud_node_mock(node_num)
101         self.subscriber = mock.Mock(name='subscriber_mock')
102         self.cloud_client = mock.MagicMock(name='cloud_client')
103         self.cloud_client.broken.return_value = False
104
105     def make_actor(self, node_num=1, arv_node=None, start_time=None):
106         if not hasattr(self, 'cloud_mock'):
107             self.make_mocks(node_num)
108         if start_time is None:
109             start_time = time.time()
110         self.node_actor = slurm_dispatch.ComputeNodeMonitorActor.start(
111             self.cloud_mock, start_time, self.shutdowns,
112             testutil.cloud_node_fqdn, self.timer, self.updates, self.cloud_client,
113             arv_node, boot_fail_after=300).proxy()
114         self.node_actor.subscribe(self.subscriber).get(self.TIMEOUT)
115
116     @mock.patch("subprocess.check_output")
117     def test_resume_node(self, check_output):
118         arv_node = testutil.arvados_node_mock()
119         self.make_actor(arv_node=arv_node)
120         check_output.return_value = "drain\n"
121         self.node_actor.resume_node().get(self.TIMEOUT)
122         self.assertIn(mock.call(['sinfo', '--noheader', '-o', '%t', '-n', arv_node['hostname']]), check_output.call_args_list)
123         self.assertIn(mock.call(['scontrol', 'update', 'NodeName=' + arv_node['hostname'], 'State=RESUME']), check_output.call_args_list)
124
125     @mock.patch("subprocess.check_output")
126     def test_no_resume_idle_node(self, check_output):
127         arv_node = testutil.arvados_node_mock()
128         self.make_actor(arv_node=arv_node)
129         check_output.return_value = "idle\n"
130         self.node_actor.resume_node().get(self.TIMEOUT)
131         self.assertIn(mock.call(['sinfo', '--noheader', '-o', '%t', '-n', arv_node['hostname']]), check_output.call_args_list)
132         self.assertNotIn(mock.call(['scontrol', 'update', 'NodeName=' + arv_node['hostname'], 'State=RESUME']), check_output.call_args_list)
133
134     @mock.patch("subprocess.check_output")
135     def test_resume_node_exception(self, check_output):
136         arv_node = testutil.arvados_node_mock()
137         self.make_actor(arv_node=arv_node)
138         check_output.side_effect = Exception()
139         self.node_actor.resume_node().get(self.TIMEOUT)
140         self.assertIn(mock.call(['sinfo', '--noheader', '-o', '%t', '-n', arv_node['hostname']]), check_output.call_args_list)
141         self.assertNotIn(mock.call(['scontrol', 'update', 'NodeName=' + arv_node['hostname'], 'State=RESUME']), check_output.call_args_list)
142
143     @mock.patch("subprocess.check_output")
144     def test_shutdown_down_node(self, check_output):
145         check_output.return_value = "down\n"
146         self.make_actor(arv_node=testutil.arvados_node_mock())
147         self.assertIs(True, self.node_actor.shutdown_eligible().get(self.TIMEOUT))
148
149     @mock.patch("subprocess.check_output")
150     def test_no_shutdown_ineligible_drain_node(self, check_output):
151         check_output.return_value = "drain\n"
152         self.make_actor(arv_node=testutil.arvados_node_mock())
153         self.assertIsNot(True, self.node_actor.shutdown_eligible().get(self.TIMEOUT))
154
155     @mock.patch("subprocess.check_output")
156     def test_shutdown_eligible_drain_node(self, check_output):
157         check_output.return_value = "drain\n"
158         self.make_actor(arv_node=testutil.arvados_node_mock())
159         self.shutdowns._set_state(True, 300)
160         self.assertIs(True, self.node_actor.shutdown_eligible().get(self.TIMEOUT))