X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/224f384d411bb1b4cccc7165c55bb64fd5c695ad..dd481ab3359f1b360a22d03c09d8035ef5db1f38:/services/nodemanager/arvnodeman/computenode/driver/__init__.py

diff --git a/services/nodemanager/arvnodeman/computenode/driver/__init__.py b/services/nodemanager/arvnodeman/computenode/driver/__init__.py
index c78f1c6b8d..fb8db51733 100644
--- a/services/nodemanager/arvnodeman/computenode/driver/__init__.py
+++ b/services/nodemanager/arvnodeman/computenode/driver/__init__.py
@@ -1,4 +1,7 @@
 #!/usr/bin/env python
+# Copyright (C) The Arvados Authors. All rights reserved.
+#
+# SPDX-License-Identifier: AGPL-3.0
 
 from __future__ import absolute_import, print_function
 
@@ -6,10 +9,10 @@ import logging
 from operator import attrgetter
 
 import libcloud.common.types as cloud_types
-from libcloud.common.exceptions import BaseHTTPError
 from libcloud.compute.base import NodeDriver, NodeAuthSSHKey
 
-from ...config import NETWORK_ERRORS
+from ...config import CLOUD_ERRORS
+from ...status import tracker
 from .. import RetryMixin
 
 class BaseComputeNodeDriver(RetryMixin):
@@ -25,7 +28,7 @@ class BaseComputeNodeDriver(RetryMixin):
     Subclasses must implement arvados_create_kwargs, sync_node,
     node_fqdn, and node_start_time.
     """
-    CLOUD_ERRORS = NETWORK_ERRORS + (cloud_types.LibcloudError,)
+
 
     @RetryMixin._retry()
     def _create_driver(self, driver_class, **auth_kwargs):
@@ -121,7 +124,11 @@ class BaseComputeNodeDriver(RetryMixin):
     def list_nodes(self, **kwargs):
         l = self.list_kwargs.copy()
         l.update(kwargs)
-        return self.real.list_nodes(**l)
+        try:
+            return self.real.list_nodes(**l)
+        except CLOUD_ERRORS:
+            tracker.counter_add('cloud_errors')
+            raise
 
     def create_cloud_name(self, arvados_node):
         """Return a cloud node name for the given Arvados node record.
@@ -169,7 +176,7 @@ class BaseComputeNodeDriver(RetryMixin):
             kwargs.update(self.arvados_create_kwargs(size, arvados_node))
             kwargs['size'] = size
             return self.real.create_node(**kwargs)
-        except self.CLOUD_ERRORS as create_error:
+        except CLOUD_ERRORS as create_error:
             # Workaround for bug #6702: sometimes the create node request
             # succeeds but times out and raises an exception instead of
             # returning a result.  If this happens, we get stuck in a retry
@@ -179,6 +186,7 @@ class BaseComputeNodeDriver(RetryMixin):
             try:
                 return self.search_for_now(kwargs['name'], 'list_nodes', self._name_key)
             except ValueError:
+                tracker.counter_add('cloud_errors')
                 raise create_error
 
     def post_create_node(self, cloud_node):
@@ -206,18 +214,23 @@ class BaseComputeNodeDriver(RetryMixin):
         # seconds since the epoch UTC.
         raise NotImplementedError("BaseComputeNodeDriver.node_start_time")
 
-    @classmethod
-    def is_cloud_exception(cls, exception):
-        # libcloud compute drivers typically raise bare Exceptions to
-        # represent API errors.  Return True for any exception that is
-        # exactly an Exception, or a better-known higher-level exception.
-        if (type(exception) is BaseHTTPError and
-            exception.message and
-            (exception.message.startswith("InvalidInstanceID.NotFound") or
-             exception.message.startswith("InstanceLimitExceeded"))):
-            return True
-        return (isinstance(exception, cls.CLOUD_ERRORS) or
-                type(exception) is Exception)
+    def destroy_node(self, cloud_node):
+        try:
+            return self.real.destroy_node(cloud_node)
+        except CLOUD_ERRORS:
+            # Sometimes the destroy node request succeeds but times out and
+            # raises an exception instead of returning success.  If this
+            # happens, we get a noisy stack trace.  Check if the node is still
+            # on the node list.  If it is gone, we can declare victory.
+            try:
+                self.search_for_now(cloud_node.id, 'list_nodes')
+            except ValueError:
+                # If we catch ValueError, that means search_for_now didn't find
+                # it, which means destroy_node actually succeeded.
+                return True
+            # The node is still on the list.  Re-raise.
+            tracker.counter_add('cloud_errors')
+            raise
 
     # Now that we've defined all our own methods, delegate generic, public
     # attributes of libcloud drivers that we haven't defined ourselves.