3 from __future__ import absolute_import, print_function
6 from operator import attrgetter
8 import libcloud.common.types as cloud_types
9 from libcloud.compute.base import NodeDriver, NodeAuthSSHKey
11 from ...config import NETWORK_ERRORS
12 from .. import RetryMixin
14 class BaseComputeNodeDriver(RetryMixin):
15 """Abstract base class for compute node drivers.
17 libcloud drivers abstract away many of the differences between
18 cloud providers, but managing compute nodes requires some
19 cloud-specific features (e.g., keeping track of node FQDNs and
20 boot times). Compute node drivers are responsible for translating
21 the node manager's cloud requests to a specific cloud's
24 Subclasses must implement arvados_create_kwargs, sync_node,
25 node_fqdn, and node_start_time.
27 CLOUD_ERRORS = NETWORK_ERRORS + (cloud_types.LibcloudError,)
30 def _create_driver(self, driver_class, **auth_kwargs):
31 return driver_class(**auth_kwargs)
35 self.sizes = {sz.id: sz for sz in self.real.list_sizes()}
37 def __init__(self, auth_kwargs, list_kwargs, create_kwargs,
38 driver_class, retry_wait=1, max_retry_wait=180):
39 """Base initializer for compute node drivers.
42 * auth_kwargs: A dictionary of arguments that are passed into the
43 driver_class constructor to instantiate a libcloud driver.
44 * list_kwargs: A dictionary of arguments that are passed to the
45 libcloud driver's list_nodes method to return the list of compute
47 * create_kwargs: A dictionary of arguments that are passed to the
48 libcloud driver's create_node method to create a new compute node.
49 * driver_class: The class of a libcloud driver to use.
52 super(BaseComputeNodeDriver, self).__init__(retry_wait, max_retry_wait,
53 logging.getLogger(self.__class__.__name__),
56 self.real = self._create_driver(driver_class, **auth_kwargs)
57 self.list_kwargs = list_kwargs
58 self.create_kwargs = create_kwargs
59 # Transform entries in create_kwargs. For each key K, if this class
60 # has an _init_K method, remove the entry and call _init_K with the
61 # corresponding value. If _init_K returns None, the entry stays out
62 # of the dictionary (we expect we're holding the value somewhere
63 # else, like an instance variable). Otherwise, _init_K returns a
64 # key-value tuple pair, and we add that entry to create_kwargs.
65 for key in self.create_kwargs.keys():
66 init_method = getattr(self, '_init_' + key, None)
67 if init_method is not None:
68 new_pair = init_method(self.create_kwargs.pop(key))
69 if new_pair is not None:
70 self.create_kwargs[new_pair[0]] = new_pair[1]
74 def _init_ping_host(self, ping_host):
75 self.ping_host = ping_host
77 def _init_ssh_key(self, filename):
78 with open(filename) as ssh_file:
79 key = NodeAuthSSHKey(ssh_file.read())
82 def search_for(self, term, list_method, key=attrgetter('id'), **kwargs):
83 """Return one matching item from a list of cloud objects.
85 Raises ValueError if the number of matching objects is not exactly 1.
88 * term: The value that identifies a matching item.
89 * list_method: A string that names the method to call on this
90 instance's libcloud driver for a list of objects.
91 * key: A function that accepts a cloud object and returns a
92 value search for a `term` match on each item. Returns the
93 object's 'id' attribute by default.
95 cache_key = (list_method, term)
96 if cache_key not in self.SEARCH_CACHE:
97 items = getattr(self.real, list_method)(**kwargs)
98 results = [item for item in items
102 raise ValueError("{} returned {} results for '{}'".format(
103 list_method, count, term))
104 self.SEARCH_CACHE[cache_key] = results[0]
105 return self.SEARCH_CACHE[cache_key]
107 def list_nodes(self, **kwargs):
108 l = self.list_kwargs.copy()
110 return self.real.list_nodes(**l)
112 def arvados_create_kwargs(self, size, arvados_node):
113 """Return dynamic keyword arguments for create_node.
115 Subclasses must override this method. It should return a dictionary
116 of keyword arguments to pass to the libcloud driver's create_node
117 method. These arguments will extend the static arguments in
121 * size: The node size that will be created (libcloud NodeSize object)
122 * arvados_node: The Arvados node record that will be associated
123 with this cloud node, as returned from the API server.
125 raise NotImplementedError("BaseComputeNodeDriver.arvados_create_kwargs")
127 def broken(self, cloud_node):
128 """Return true if libcloud has indicated the node is in a "broken" state."""
131 def _make_ping_url(self, arvados_node):
132 return 'https://{}/arvados/v1/nodes/{}/ping?ping_secret={}'.format(
133 self.ping_host, arvados_node['uuid'],
134 arvados_node['info']['ping_secret'])
137 def _name_key(cloud_object):
138 return cloud_object.name
140 def create_node(self, size, arvados_node):
142 kwargs = self.create_kwargs.copy()
143 kwargs.update(self.arvados_create_kwargs(size, arvados_node))
144 kwargs['size'] = size
145 return self.real.create_node(**kwargs)
146 except self.CLOUD_ERRORS:
147 # Workaround for bug #6702: sometimes the create node request
148 # succeeds but times out and raises an exception instead of
149 # returning a result. If this happens, we get stuck in a retry
150 # loop forever because subsequent create_node attempts will fail
151 # due to node name collision. So check if the node we intended to
152 # create shows up in the cloud node list and return it if found.
153 node = self.search_for(kwargs['name'], 'list_nodes', self._name_key)
157 # something else went wrong, re-raise the exception
160 def post_create_node(self, cloud_node):
161 # ComputeNodeSetupActor calls this method after the cloud node is
162 # created. Any setup tasks that need to happen afterward (e.g.,
163 # tagging) should be done in this method.
166 def sync_node(self, cloud_node, arvados_node):
167 # When a compute node first pings the API server, the API server
168 # will automatically assign some attributes on the corresponding
169 # node record, like hostname. This method should propagate that
170 # information back to the cloud node appropriately.
171 raise NotImplementedError("BaseComputeNodeDriver.sync_node")
174 def node_fqdn(cls, node):
175 # This method should return the FQDN of the node object argument.
176 # Different clouds store this in different places.
177 raise NotImplementedError("BaseComputeNodeDriver.node_fqdn")
180 def node_start_time(cls, node):
181 # This method should return the time the node was started, in
182 # seconds since the epoch UTC.
183 raise NotImplementedError("BaseComputeNodeDriver.node_start_time")
186 def is_cloud_exception(cls, exception):
187 # libcloud compute drivers typically raise bare Exceptions to
188 # represent API errors. Return True for any exception that is
189 # exactly an Exception, or a better-known higher-level exception.
190 return (isinstance(exception, cls.CLOUD_ERRORS) or
191 type(exception) is Exception)
193 # Now that we've defined all our own methods, delegate generic, public
194 # attributes of libcloud drivers that we haven't defined ourselves.
195 def _delegate_to_real(attr_name):
197 lambda self: getattr(self.real, attr_name),
198 lambda self, value: setattr(self.real, attr_name, value),
199 doc=getattr(getattr(NodeDriver, attr_name), '__doc__', None))
204 raise NotImplementedError("BaseComputeNodeDriver.node_id")
207 for _attr_name in dir(NodeDriver):
208 if (not _attr_name.startswith('_')) and (_attr_name not in _locals):
209 _locals[_attr_name] = _delegate_to_real(_attr_name)