import pykka
from . import computenode as cnode
+from .computenode import dispatch
from .config import actor_class
class _ComputeNodeRecord(object):
self.nodes = {}
self.orphans = {}
- def __getitem__(self, key):
- return self.nodes[key]
-
- def __len__(self):
- return len(self.nodes)
+ # Proxy the methods listed below to self.nodes.
+ def _proxy_method(name):
+ method = getattr(dict, name)
+ @functools.wraps(method, ('__name__', '__doc__'))
+ def wrapper(self, *args, **kwargs):
+ return method(self.nodes, *args, **kwargs)
+ return wrapper
- def get(self, key, default=None):
- return self.nodes.get(key, default)
+ for _method_name in ['__contains__', '__getitem__', '__len__', 'get']:
+ locals()[_method_name] = _proxy_method(_method_name)
def record_key(self, record):
return self.item_key(getattr(record, self.RECORD_ATTR))
def __init__(self, server_wishlist_actor, arvados_nodes_actor,
cloud_nodes_actor, cloud_update_actor, timer_actor,
arvados_factory, cloud_factory,
- shutdown_windows, max_nodes,
- poll_stale_after=600, node_stale_after=7200,
- node_setup_class=cnode.ComputeNodeSetupActor,
- node_shutdown_class=cnode.ComputeNodeShutdownActor,
- node_actor_class=cnode.ComputeNodeMonitorActor):
+ shutdown_windows, min_size, min_nodes, max_nodes,
+ poll_stale_after=600,
+ boot_fail_after=1800,
+ node_stale_after=7200,
+ node_setup_class=dispatch.ComputeNodeSetupActor,
+ node_shutdown_class=dispatch.ComputeNodeShutdownActor,
+ node_actor_class=dispatch.ComputeNodeMonitorActor):
super(NodeManagerDaemonActor, self).__init__()
self._node_setup = node_setup_class
self._node_shutdown = node_shutdown_class
self._logger = logging.getLogger('arvnodeman.daemon')
self._later = self.actor_ref.proxy()
self.shutdown_windows = shutdown_windows
+ self.min_cloud_size = min_size
+ self.min_nodes = min_nodes
self.max_nodes = max_nodes
self.poll_stale_after = poll_stale_after
+ self.boot_fail_after = boot_fail_after
self.node_stale_after = node_stale_after
self.last_polls = {}
for poll_name in ['server_wishlist', 'arvados_nodes', 'cloud_nodes']:
break
for key, record in self.cloud_nodes.orphans.iteritems():
record.actor.stop()
+ record.cloud_node = None
self.shutdowns.pop(key, None)
def update_arvados_nodes(self, nodelist):
self._pair_nodes(cloud_rec, arv_node)
break
- def _node_count(self):
- up = sum(len(nodelist) for nodelist in
- [self.cloud_nodes, self.booted, self.booting])
- return up - len(self.shutdowns)
+ def _nodes_up(self):
+ return sum(len(nodelist) for nodelist in
+ [self.cloud_nodes, self.booted, self.booting])
+
+ def _nodes_busy(self):
+ return sum(1 for idle in
+ pykka.get_all(rec.actor.in_state('idle') for rec in
+ self.cloud_nodes.nodes.itervalues())
+ if idle is False)
def _nodes_wanted(self):
- return len(self.last_wishlist) - self._node_count()
+ up_count = self._nodes_up()
+ under_min = self.min_nodes - up_count
+ over_max = up_count - self.max_nodes
+ if over_max >= 0:
+ return -over_max
+ elif under_min > 0:
+ return under_min
+ else:
+ up_count -= len(self.shutdowns) + self._nodes_busy()
+ return len(self.last_wishlist) - up_count
def _nodes_excess(self):
- return -self._nodes_wanted()
+ up_count = self._nodes_up() - len(self.shutdowns)
+ over_min = up_count - self.min_nodes
+ if over_min <= 0:
+ return over_min
+ else:
+ return up_count - self._nodes_busy() - len(self.last_wishlist)
def update_server_wishlist(self, wishlist):
self._update_poll_time('server_wishlist')
- self.last_wishlist = wishlist[:self.max_nodes]
+ self.last_wishlist = wishlist
nodes_wanted = self._nodes_wanted()
if nodes_wanted > 0:
self._later.start_node()
if nodes_wanted < 1:
return None
arvados_node = self.arvados_nodes.find_stale_node(self.node_stale_after)
- cloud_size = self.last_wishlist[nodes_wanted - 1]
+ try:
+ cloud_size = self.last_wishlist[self._nodes_up()]
+ except IndexError:
+ cloud_size = self.min_cloud_size
self._logger.info("Want %s more nodes. Booting a %s node.",
nodes_wanted, cloud_size.name)
new_setup = self._node_setup.start(
if nodes_wanted > 1:
self._later.start_node()
- def _actor_nodes(self, node_actor):
- return pykka.get_all([node_actor.cloud_node, node_actor.arvados_node])
+ def _get_actor_attrs(self, actor, *attr_names):
+ return pykka.get_all([getattr(actor, name) for name in attr_names])
def node_up(self, setup_proxy):
- cloud_node, arvados_node = self._actor_nodes(setup_proxy)
+ cloud_node = setup_proxy.cloud_node.get()
del self.booting[setup_proxy.actor_ref.actor_urn]
setup_proxy.stop()
record = self.cloud_nodes.get(cloud_node.id)
if record is None:
record = self._new_node(cloud_node)
self.booted[cloud_node.id] = record
- self._pair_nodes(record, arvados_node)
+ self._timer.schedule(time.time() + self.boot_fail_after,
+ self._later.shutdown_unpaired_node, cloud_node.id)
@_check_poll_freshness
def stop_booting_node(self):
self._later.stop_booting_node()
break
+ def _begin_node_shutdown(self, node_actor, cancellable):
+ cloud_node_id = node_actor.cloud_node.get().id
+ if cloud_node_id in self.shutdowns:
+ return None
+ shutdown = self._node_shutdown.start(
+ timer_actor=self._timer, cloud_client=self._new_cloud(),
+ node_monitor=node_actor.actor_ref, cancellable=cancellable).proxy()
+ self.shutdowns[cloud_node_id] = shutdown
+ shutdown.subscribe(self._later.node_finished_shutdown)
+
@_check_poll_freshness
def node_can_shutdown(self, node_actor):
- if self._nodes_excess() < 1:
- return None
- cloud_node, arvados_node = self._actor_nodes(node_actor)
- if cloud_node.id in self.shutdowns:
+ if self._nodes_excess() > 0:
+ self._begin_node_shutdown(node_actor, cancellable=True)
+
+ def shutdown_unpaired_node(self, cloud_node_id):
+ for record_dict in [self.cloud_nodes, self.booted]:
+ if cloud_node_id in record_dict:
+ record = record_dict[cloud_node_id]
+ break
+ else:
return None
- shutdown = self._node_shutdown.start(timer_actor=self._timer,
- cloud_client=self._new_cloud(),
- cloud_node=cloud_node).proxy()
- self.shutdowns[cloud_node.id] = shutdown
- shutdown.subscribe(self._later.node_finished_shutdown)
+ if record.arvados_node is None:
+ self._begin_node_shutdown(record.actor, cancellable=False)
def node_finished_shutdown(self, shutdown_actor):
- cloud_node_id = shutdown_actor.cloud_node.get().id
+ success, cloud_node = self._get_actor_attrs(shutdown_actor, 'success',
+ 'cloud_node')
shutdown_actor.stop()
- if cloud_node_id in self.booted:
+ cloud_node_id = cloud_node.id
+ if not success:
+ del self.shutdowns[cloud_node_id]
+ elif cloud_node_id in self.booted:
self.booted.pop(cloud_node_id).actor.stop()
del self.shutdowns[cloud_node_id]