11345: Simplify and consolodate retry for API throttling errors.

[arvados.git] / 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 63e4c00dabd92b81caac95be440fa19f820777cd..6d23c2b5a3711ac0c1f50770d12023573ca5fffa 100644 (file)
--- a/services/nodemanager/arvnodeman/computenode/driver/__init__.py
+++ b/services/nodemanager/arvnodeman/computenode/driver/__init__.py
@@ -25,7 +25,7 @@ class BaseComputeNodeDriver(RetryMixin):
     Subclasses must implement arvados_create_kwargs, sync_node,
     node_fqdn, and node_start_time.
     """
-    CLOUD_ERRORS = NETWORK_ERRORS + (cloud_types.LibcloudError, BaseHTTPError)
+    CLOUD_ERRORS = NETWORK_ERRORS + (cloud_types.LibcloudError,)
 
     @RetryMixin._retry()
     def _create_driver(self, driver_class, **auth_kwargs):
@@ -206,13 +206,22 @@ 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.
-        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 self.CLOUD_ERRORS as destroy_error:
+            # 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.
+            raise
 
     # Now that we've defined all our own methods, delegate generic, public
     # attributes of libcloud drivers that we haven't defined ourselves.