Merge branch '6321-slurm-oserror' closes #6321

diff --git a/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py b/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py
index dfb26bc303d2b310f1186a1fc72baf26438608b1..ec5014e9f9cf1e8848353cf3c755e22875227850 100644 (file)
--- a/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py
+++ b/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py
@@ -34,7 +34,15 @@ class ComputeNodeShutdownActor(ShutdownActorBase):
     def _get_slurm_state(self):
         return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', self._nodename])
 
-    @ShutdownActorBase._retry((subprocess.CalledProcessError,))
+    # The following methods retry on OSError.  This is intended to mitigate bug
+    # #6321 where fork() of node manager raises "OSError: [Errno 12] Cannot
+    # allocate memory" resulting in the untimely death of the shutdown actor
+    # and tends to result in node manager getting into a wedged state where it
+    # won't allocate new nodes or shut down gracefully.  The underlying causes
+    # of the excessive memory usage that result in the "Cannot allocate memory"
+    # error are still being investigated.
+
+    @ShutdownActorBase._retry((subprocess.CalledProcessError, OSError))
     def cancel_shutdown(self):
         if self._nodename:
             if self._get_slurm_state() in self.SLURM_DRAIN_STATES:
@@ -46,15 +54,15 @@ class ComputeNodeShutdownActor(ShutdownActorBase):
                 pass
         return super(ComputeNodeShutdownActor, self).cancel_shutdown()
 
+    @ShutdownActorBase._retry((subprocess.CalledProcessError, OSError))
     @ShutdownActorBase._stop_if_window_closed
-    @ShutdownActorBase._retry((subprocess.CalledProcessError,))
     def issue_slurm_drain(self):
         self._set_node_state('DRAIN', 'Reason=Node Manager shutdown')
         self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
         self._later.await_slurm_drain()
 
+    @ShutdownActorBase._retry((subprocess.CalledProcessError, OSError))
     @ShutdownActorBase._stop_if_window_closed
-    @ShutdownActorBase._retry((subprocess.CalledProcessError,))
     def await_slurm_drain(self):
         output = self._get_slurm_state()
         if output in self.SLURM_END_STATES:

diff --git a/services/nodemanager/tests/test_computenode_dispatch_slurm.py b/services/nodemanager/tests/test_computenode_dispatch_slurm.py
index 2ddf7676c8f8ddc4ca6f84630f38d35ca86abe6a..8648783bac5889f11a328af3b277bd1f21da5665 100644 (file)
--- a/services/nodemanager/tests/test_computenode_dispatch_slurm.py
+++ b/services/nodemanager/tests/test_computenode_dispatch_slurm.py
@@ -73,6 +73,15 @@ class SLURMComputeNodeShutdownActorTestCase(ComputeNodeShutdownActorMixin,
         self.check_success_flag(False, 2)
         self.check_slurm_got_args(proc_mock, 'sinfo', '--noheader', '-o', '%t', '-n', 'compute99')
 
+    def test_cancel_shutdown_retry(self, proc_mock):
+        proc_mock.side_effect = iter([OSError, 'drain\n', OSError, 'idle\n'])
+        self.make_mocks(arvados_node=testutil.arvados_node_mock(job_uuid=True))
+        self.make_actor()
+        self.check_success_flag(False, 2)
+
+    def test_issue_slurm_drain_retry(self, proc_mock):
+        proc_mock.side_effect = iter([OSError, '', OSError, 'drng\n'])
+        self.check_success_after_reset(proc_mock)
 
     def test_arvados_node_cleaned_after_shutdown(self, proc_mock):
         proc_mock.return_value = 'drain\n'