from __future__ import absolute_import, print_function
+import logging
+from operator import attrgetter
+
import libcloud.common.types as cloud_types
-from libcloud.compute.base import NodeDriver
+from libcloud.compute.base import NodeDriver, NodeAuthSSHKey
from ...config import NETWORK_ERRORS
+from .. import RetryMixin
-class BaseComputeNodeDriver(object):
+class BaseComputeNodeDriver(RetryMixin):
"""Abstract base class for compute node drivers.
- libcloud abstracts away many of the differences between cloud providers,
- but managing compute nodes requires some cloud-specific features (e.g.,
- on EC2 we use tags to identify compute nodes). Compute node drivers
- are responsible for translating the node manager's cloud requests to a
- specific cloud's vocabulary.
+ libcloud drivers abstract away many of the differences between
+ cloud providers, but managing compute nodes requires some
+ cloud-specific features (e.g., keeping track of node FQDNs and
+ boot times). Compute node drivers are responsible for translating
+ the node manager's cloud requests to a specific cloud's
+ vocabulary.
- Subclasses must implement arvados_create_kwargs (to update node
- creation kwargs with information about the specific Arvados node
- record), sync_node, and node_start_time.
+ Subclasses must implement arvados_create_kwargs, sync_node,
+ node_fqdn, and node_start_time.
"""
CLOUD_ERRORS = NETWORK_ERRORS + (cloud_types.LibcloudError,)
- def __init__(self, auth_kwargs, list_kwargs, create_kwargs, driver_class):
- self.real = driver_class(**auth_kwargs)
+ @RetryMixin._retry()
+ def _create_driver(self, driver_class, **auth_kwargs):
+ return driver_class(**auth_kwargs)
+
+ @RetryMixin._retry()
+ def _set_sizes(self):
+ self.sizes = {sz.id: sz for sz in self.real.list_sizes()}
+
+ def __init__(self, auth_kwargs, list_kwargs, create_kwargs,
+ driver_class, retry_wait=1, max_retry_wait=180):
+ """Base initializer for compute node drivers.
+
+ Arguments:
+ * auth_kwargs: A dictionary of arguments that are passed into the
+ driver_class constructor to instantiate a libcloud driver.
+ * list_kwargs: A dictionary of arguments that are passed to the
+ libcloud driver's list_nodes method to return the list of compute
+ nodes.
+ * create_kwargs: A dictionary of arguments that are passed to the
+ libcloud driver's create_node method to create a new compute node.
+ * driver_class: The class of a libcloud driver to use.
+ """
+
+ super(BaseComputeNodeDriver, self).__init__(retry_wait, max_retry_wait,
+ logging.getLogger(self.__class__.__name__),
+ type(self),
+ None)
+ self.real = self._create_driver(driver_class, **auth_kwargs)
self.list_kwargs = list_kwargs
self.create_kwargs = create_kwargs
+ # Transform entries in create_kwargs. For each key K, if this class
+ # has an _init_K method, remove the entry and call _init_K with the
+ # corresponding value. If _init_K returns None, the entry stays out
+ # of the dictionary (we expect we're holding the value somewhere
+ # else, like an instance variable). Otherwise, _init_K returns a
+ # key-value tuple pair, and we add that entry to create_kwargs.
for key in self.create_kwargs.keys():
init_method = getattr(self, '_init_' + key, None)
if init_method is not None:
if new_pair is not None:
self.create_kwargs[new_pair[0]] = new_pair[1]
+ self._set_sizes()
+
def _init_ping_host(self, ping_host):
self.ping_host = ping_host
- def search_for(self, term, list_method, key=lambda item: item.id):
+ def _init_ssh_key(self, filename):
+ with open(filename) as ssh_file:
+ key = NodeAuthSSHKey(ssh_file.read())
+ return 'auth', key
+
+ def search_for(self, term, list_method, key=attrgetter('id'), **kwargs):
+ """Return one matching item from a list of cloud objects.
+
+ Raises ValueError if the number of matching objects is not exactly 1.
+
+ Arguments:
+ * term: The value that identifies a matching item.
+ * list_method: A string that names the method to call on this
+ instance's libcloud driver for a list of objects.
+ * key: A function that accepts a cloud object and returns a
+ value search for a `term` match on each item. Returns the
+ object's 'id' attribute by default.
+ """
cache_key = (list_method, term)
if cache_key not in self.SEARCH_CACHE:
- results = [item for item in getattr(self.real, list_method)()
+ items = getattr(self.real, list_method)(**kwargs)
+ results = [item for item in items
if key(item) == term]
count = len(results)
if count != 1:
def list_nodes(self):
return self.real.list_nodes(**self.list_kwargs)
- def arvados_create_kwargs(self, arvados_node):
+ def arvados_create_kwargs(self, size, arvados_node):
+ """Return dynamic keyword arguments for create_node.
+
+ Subclasses must override this method. It should return a dictionary
+ of keyword arguments to pass to the libcloud driver's create_node
+ method. These arguments will extend the static arguments in
+ create_kwargs.
+
+ Arguments:
+ * size: The node size that will be created (libcloud NodeSize object)
+ * arvados_node: The Arvados node record that will be associated
+ with this cloud node, as returned from the API server.
+ """
raise NotImplementedError("BaseComputeNodeDriver.arvados_create_kwargs")
+ def broken(self, cloud_node):
+ """Return true if libcloud has indicated the node is in a "broken" state."""
+ return False
+
def _make_ping_url(self, arvados_node):
return 'https://{}/arvados/v1/nodes/{}/ping?ping_secret={}'.format(
self.ping_host, arvados_node['uuid'],
arvados_node['info']['ping_secret'])
+ def find_node(self, name):
+ node = [n for n in self.list_nodes() if n.name == name]
+ if node:
+ return node[0]
+ else:
+ return None
+
def create_node(self, size, arvados_node):
- kwargs = self.create_kwargs.copy()
- kwargs.update(self.arvados_create_kwargs(arvados_node))
- kwargs['size'] = size
- return self.real.create_node(**kwargs)
+ try:
+ kwargs = self.create_kwargs.copy()
+ kwargs.update(self.arvados_create_kwargs(size, arvados_node))
+ kwargs['size'] = size
+ return self.real.create_node(**kwargs)
+ except self.CLOUD_ERRORS:
+ # 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
+ # loop forever because subsequent create_node attempts will fail
+ # due to node name collision. So check if the node we intended to
+ # create shows up in the cloud node list and return it if found.
+ try:
+ node = self.find_node(kwargs['name'])
+ if node:
+ return node
+ except:
+ # Ignore possible exception from find_node in favor of
+ # re-raising the original create_node exception.
+ pass
+ raise
def post_create_node(self, cloud_node):
# ComputeNodeSetupActor calls this method after the cloud node is
# information back to the cloud node appropriately.
raise NotImplementedError("BaseComputeNodeDriver.sync_node")
+ @classmethod
+ def node_fqdn(cls, node):
+ # This method should return the FQDN of the node object argument.
+ # Different clouds store this in different places.
+ raise NotImplementedError("BaseComputeNodeDriver.node_fqdn")
+
@classmethod
def node_start_time(cls, node):
+ # This method should return the time the node was started, in
+ # seconds since the epoch UTC.
raise NotImplementedError("BaseComputeNodeDriver.node_start_time")
@classmethod
# 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
- getattr(exception, '__class__', None) is Exception)
+ type(exception) is Exception)
# Now that we've defined all our own methods, delegate generic, public
# attributes of libcloud drivers that we haven't defined ourselves.