[ironic] Minimize window for a resource provider to be lost

This patch is based upon a downstream patch which came up in discussion
amongst the ironic community when some operators began discussing a case
where resource providers had disappeared from a running deployment with
several thousand baremetal nodes.

Discussion amongst operators and developers ensued and we were able
to determine that this was still an issue in the current upstream code
and that time difference between collecting data and then reconciling
the records was a source of the issue. Per Arun, they have been running
this change downstream and had not seen any reoccurances of the issue
since the patch was applied.

This patch was originally authored by Arun S A G, and below is his
original commit mesage.

An instance could be launched and scheduled to a compute node between
get_uuids_by_host() call and _get_node_list() call. If that happens
the ironic node.instance_uuid may not be None but the instance_uuid
will be missing from the instance list returned by get_uuids_by_host()
method. This is possible because _get_node_list() takes several minutes to return
in large baremetal clusters and a lot can happen in that time.

This causes the compute node to be orphaned and associated resource
provider to be deleted from placement. Once the resource provider is
deleted it is never created again until the service restarts. Since
resource provider is deleted subsequent boots/rebuilds to the same
host will fail.

This behaviour is visibile in VMbooter nodes because it constantly
launches and deletes instances there by increasing the likelihood
of this race condition happening in large ironic clusters.

To reduce the chance of this race condition we call _get_node_list()
first followed by get_uuids_by_host() method.

Change-Id: I55bde8dd33154e17bbdb3c4b0e7a83a20e8487e8
Co-Authored-By: Arun S A G <[email protected]>
Related-Bug: #1841481

Author: juliakreger

nova/tests/unit/virt/ironic/test_driver.py

@@ -3038,6 +3038,9 @@ def _test__refresh_cache(self, instances, nodes, hosts, mock_instances,
         mock_instances.return_value = instances
         mock_nodes.return_value = nodes
         mock_hosts.side_effect = hosts
+        parent_mock = mock.MagicMock()
+        parent_mock.attach_mock(mock_nodes, 'get_node_list')
+        parent_mock.attach_mock(mock_instances, 'get_uuids_by_host')
         if not can_send_146:
             mock_can_send.side_effect = (
                 exception.IronicAPIVersionNotAvailable(version='1.46'))
@@ -3050,6 +3053,15 @@ def _test__refresh_cache(self, instances, nodes, hosts, mock_instances,
 
         self.driver._refresh_cache()
 
+        # assert if get_node_list() is called before get_uuids_by_host()
+        parent_mock.assert_has_calls(
+            [
+                mock.call.get_node_list(fields=ironic_driver._NODE_FIELDS,
+                                        **kwargs),
+                mock.call.get_uuids_by_host(mock.ANY, self.host)
+            ]
+        )
+
         mock_hash_ring.assert_called_once_with(mock.ANY)
         mock_instances.assert_called_once_with(mock.ANY, self.host)
         mock_nodes.assert_called_once_with(fields=ironic_driver._NODE_FIELDS,

nova/virt/ironic/driver.py

@@ -733,10 +733,15 @@ def _refresh_hash_ring(self, ctxt):
     def _refresh_cache(self):
         ctxt = nova_context.get_admin_context()
         self._refresh_hash_ring(ctxt)
-        instances = objects.InstanceList.get_uuids_by_host(ctxt, CONF.host)
         node_cache = {}
 
         def _get_node_list(**kwargs):
+            # NOTE(TheJulia): This call can take a substantial amount
+            # of time as it may be attempting to retrieve thousands of
+            # baremetal nodes. Depending on the version of Ironic,
+            # this can be as long as 2-10 seconds per every thousand
+            # nodes, and this call may retrieve all nodes in a deployment,
+            # depending on if any filter paramters are applied.
             return self._get_node_list(fields=_NODE_FIELDS, **kwargs)
 
         # NOTE(jroll) if partition_key is set, we need to limit nodes that
@@ -760,6 +765,15 @@ def _get_node_list(**kwargs):
         else:
             nodes = _get_node_list()
 
+        # NOTE(saga): As _get_node_list() will take a long
+        # time to return in large clusters we need to call it before
+        # get_uuids_by_host() method. Otherwise the instances list we get from
+        # get_uuids_by_host() method will become stale.
+        # A stale instances list can cause a node that is managed by this
+        # compute host to be excluded in error and cause the compute node
+        # to be orphaned and associated resource provider to be deleted.
+        instances = objects.InstanceList.get_uuids_by_host(ctxt, CONF.host)
+
         for node in nodes:
             # NOTE(jroll): we always manage the nodes for instances we manage
             if node.instance_uuid in instances:

releasenotes/notes/minimize-bug-1841481-race-window-f76912d4985770ad.yaml

@@ -0,0 +1,13 @@
+---
+fixes:
+  - |
+    Minimizes a race condition window when using the ``ironic`` virt driver
+    where the data generated for the Resource Tracker may attempt to compare
+    potentially stale instance information with the latest known baremetal
+    node information. While this doesn't completely prevent nor resolve the
+    underlying race condition identified in
+    `bug 1841481 <https://bugs.launchpad.net/nova/+bug/1841481>`_,
+    this change allows Nova to have the latest state information, as opposed
+    to state information which may be out of date due to the time which it may
+    take to retrieve the status from Ironic. This issue was most observable
+    on baremetal clusters with several thousand physical nodes.