2 # Copyright (C) The Arvados Authors. All rights reserved.
4 # SPDX-License-Identifier: AGPL-3.0
6 from __future__ import absolute_import, print_function
14 from . import computenode as cnode
16 from .computenode import dispatch
17 from .config import actor_class
19 class _ComputeNodeRecord(object):
20 def __init__(self, actor=None, cloud_node=None, arvados_node=None,
21 assignment_time=float('-inf')):
23 self.cloud_node = cloud_node
24 self.arvados_node = arvados_node
25 self.assignment_time = assignment_time
26 self.shutdown_actor = None
28 class _BaseNodeTracker(object):
33 # Proxy the methods listed below to self.nodes.
34 def _proxy_method(name):
35 method = getattr(dict, name)
36 @functools.wraps(method, ('__name__', '__doc__'))
37 def wrapper(self, *args, **kwargs):
38 return method(self.nodes, *args, **kwargs)
41 for _method_name in ['__contains__', '__getitem__', '__len__', 'get']:
42 locals()[_method_name] = _proxy_method(_method_name)
44 def record_key(self, record):
45 return self.item_key(getattr(record, self.RECORD_ATTR))
47 def add(self, record):
48 self.nodes[self.record_key(record)] = record
50 def update_record(self, key, item):
51 setattr(self.nodes[key], self.RECORD_ATTR, item)
53 def update_from(self, response):
54 unseen = set(self.nodes.iterkeys())
56 key = self.item_key(item)
59 self.update_record(key, item)
62 self.orphans = {key: self.nodes.pop(key) for key in unseen}
65 return (record for record in self.nodes.itervalues()
66 if getattr(record, self.PAIR_ATTR) is None)
69 class _CloudNodeTracker(_BaseNodeTracker):
70 RECORD_ATTR = 'cloud_node'
71 PAIR_ATTR = 'arvados_node'
72 item_key = staticmethod(lambda cloud_node: cloud_node.id)
75 class _ArvadosNodeTracker(_BaseNodeTracker):
76 RECORD_ATTR = 'arvados_node'
77 PAIR_ATTR = 'cloud_node'
78 item_key = staticmethod(lambda arvados_node: arvados_node['uuid'])
80 def find_stale_node(self, stale_time):
81 # Try to select a stale node record that have an assigned slot first
82 for record in sorted(self.nodes.itervalues(),
83 key=lambda r: r.arvados_node['slot_number'],
85 node = record.arvados_node
86 if (not cnode.timestamp_fresh(cnode.arvados_node_mtime(node),
88 not cnode.timestamp_fresh(record.assignment_time,
94 class NodeManagerDaemonActor(actor_class):
95 """Node Manager daemon.
97 This actor subscribes to all information polls about cloud nodes,
98 Arvados nodes, and the job queue. It creates a ComputeNodeMonitorActor
99 for every cloud node, subscribing them to poll updates
100 appropriately. It creates and destroys cloud nodes based on job queue
101 demand, and stops the corresponding ComputeNode actors when their work
104 def __init__(self, server_wishlist_actor, arvados_nodes_actor,
105 cloud_nodes_actor, cloud_update_actor, timer_actor,
106 arvados_factory, cloud_factory,
107 shutdown_windows, server_calculator,
108 min_nodes, max_nodes,
109 poll_stale_after=600,
110 boot_fail_after=1800,
111 node_stale_after=7200,
112 node_setup_class=dispatch.ComputeNodeSetupActor,
113 node_shutdown_class=dispatch.ComputeNodeShutdownActor,
114 node_actor_class=dispatch.ComputeNodeMonitorActor,
116 super(NodeManagerDaemonActor, self).__init__()
117 self._node_setup = node_setup_class
118 self._node_shutdown = node_shutdown_class
119 self._node_actor = node_actor_class
120 self._cloud_updater = cloud_update_actor
121 self._timer = timer_actor
122 self._new_arvados = arvados_factory
123 self._new_cloud = cloud_factory
124 self._cloud_driver = self._new_cloud()
125 self._later = self.actor_ref.tell_proxy()
126 self.shutdown_windows = shutdown_windows
127 self.server_calculator = server_calculator
128 self.min_cloud_size = self.server_calculator.cheapest_size()
129 self.min_nodes = min_nodes
130 self.max_nodes = max_nodes
131 self.node_quota = max_nodes
132 self.max_total_price = max_total_price
133 self.poll_stale_after = poll_stale_after
134 self.boot_fail_after = boot_fail_after
135 self.node_stale_after = node_stale_after
137 for poll_name in ['server_wishlist', 'arvados_nodes', 'cloud_nodes']:
138 poll_actor = locals()[poll_name + '_actor']
139 poll_actor.subscribe(getattr(self._later, 'update_' + poll_name))
140 setattr(self, '_{}_actor'.format(poll_name), poll_actor)
141 self.last_polls[poll_name] = -self.poll_stale_after
142 self.cloud_nodes = _CloudNodeTracker()
143 self.arvados_nodes = _ArvadosNodeTracker()
144 self.booting = {} # Actor IDs to ComputeNodeSetupActors
145 self.sizes_booting = {} # Actor IDs to node size
148 self._logger = logging.getLogger("%s.%s" % (self.__class__.__name__, self.actor_urn[33:]))
149 self._logger.debug("Daemon started")
151 def _update_poll_time(self, poll_key):
152 self.last_polls[poll_key] = time.time()
154 def _pair_nodes(self, node_record, arvados_node):
155 self._logger.info("Cloud node %s is now paired with Arvados node %s with hostname %s",
156 node_record.cloud_node.name, arvados_node['uuid'], arvados_node['hostname'])
157 self._arvados_nodes_actor.subscribe_to(
158 arvados_node['uuid'], node_record.actor.update_arvados_node)
159 node_record.arvados_node = arvados_node
160 self.arvados_nodes.add(node_record)
162 def _new_node(self, cloud_node):
163 start_time = self._cloud_driver.node_start_time(cloud_node)
164 shutdown_timer = cnode.ShutdownTimer(start_time,
165 self.shutdown_windows)
166 actor = self._node_actor.start(
167 cloud_node=cloud_node,
168 cloud_node_start_time=start_time,
169 shutdown_timer=shutdown_timer,
170 update_actor=self._cloud_updater,
171 timer_actor=self._timer,
173 poll_stale_after=self.poll_stale_after,
174 node_stale_after=self.node_stale_after,
175 cloud_client=self._cloud_driver,
176 boot_fail_after=self.boot_fail_after)
177 actorTell = actor.tell_proxy()
178 actorTell.subscribe(self._later.node_can_shutdown)
179 self._cloud_nodes_actor.subscribe_to(cloud_node.id,
180 actorTell.update_cloud_node)
181 record = _ComputeNodeRecord(actor.proxy(), cloud_node)
184 def _register_cloud_node(self, node):
185 rec = self.cloud_nodes.get(node.id)
187 self._logger.info("Registering new cloud node %s", node.id)
188 record = self._new_node(node)
189 self.cloud_nodes.add(record)
191 rec.cloud_node = node
193 def update_cloud_nodes(self, nodelist):
194 self._update_poll_time('cloud_nodes')
195 for _, node in self.cloud_nodes.update_from(nodelist):
196 self._register_cloud_node(node)
200 for record in self.cloud_nodes.orphans.itervalues():
201 if record.shutdown_actor:
203 record.shutdown_actor.stop()
204 except pykka.ActorDeadError:
206 record.shutdown_actor = None
208 # A recently booted node is a node that successfully completed the
209 # setup actor but has not yet appeared in the cloud node list.
210 # This will have the tag _nodemanager_recently_booted on it, which
211 # means (if we're not shutting it down) we want to put it back into
212 # the cloud node list. Once it really appears in the cloud list,
213 # the object in record.cloud_node will be replaced by a new one
214 # that lacks the "_nodemanager_recently_booted" tag.
215 if hasattr(record.cloud_node, "_nodemanager_recently_booted"):
216 self.cloud_nodes.add(record)
218 # Node disappeared from the cloud node list. If it's paired,
219 # remove its idle time counter.
220 if record.arvados_node:
221 status.tracker.idle_out(record.arvados_node.get('hostname'))
222 # Stop the monitor actor if necessary and forget about the node.
226 except pykka.ActorDeadError:
229 record.cloud_node = None
231 def _register_arvados_node(self, key, arv_node):
232 self._logger.info("Registering new Arvados node %s", key)
233 record = _ComputeNodeRecord(arvados_node=arv_node)
234 self.arvados_nodes.add(record)
236 def update_arvados_nodes(self, nodelist):
237 self._update_poll_time('arvados_nodes')
238 for key, node in self.arvados_nodes.update_from(nodelist):
239 self._register_arvados_node(key, node)
242 def try_pairing(self):
243 for record in self.cloud_nodes.unpaired():
244 for arv_rec in self.arvados_nodes.unpaired():
245 if record.actor is not None and record.actor.offer_arvados_pair(arv_rec.arvados_node).get():
246 self._pair_nodes(record, arv_rec.arvados_node)
249 def _nodes_booting(self, size):
251 for c in self.booting.iterkeys()
252 if size is None or self.sizes_booting[c].id == size.id)
255 def _node_states(self, size):
258 for rec in self.cloud_nodes.nodes.itervalues():
259 if size is None or rec.cloud_node.size.id == size.id:
260 if rec.shutdown_actor is None and rec.actor is not None:
261 proxy_states.append(rec.actor.get_state())
263 states.append("shutdown")
264 return states + pykka.get_all(proxy_states)
266 def _update_tracker(self):
269 for k in status.tracker.keys()
270 if k.startswith('nodes_')
272 for s in self._node_states(size=None):
273 updates.setdefault('nodes_'+s, 0)
274 updates['nodes_'+s] += 1
275 updates['nodes_wish'] = len(self.last_wishlist)
276 updates['node_quota'] = self.node_quota
277 status.tracker.update(updates)
279 def _state_counts(self, size):
280 states = self._node_states(size)
282 "booting": self._nodes_booting(size),
291 counts[s] = counts[s] + 1
294 def _nodes_up(self, counts):
295 up = counts["booting"] + counts["unpaired"] + counts["idle"] + counts["busy"]
298 def _total_price(self):
300 cost += sum(self.sizes_booting[c].price
301 for c in self.booting.iterkeys())
302 cost += sum(c.cloud_node.size.price
303 for c in self.cloud_nodes.nodes.itervalues())
306 def _size_wishlist(self, size):
307 return sum(1 for c in self.last_wishlist if c.id == size.id)
309 def _nodes_wanted(self, size):
310 total_node_count = self._nodes_booting(None) + len(self.cloud_nodes)
311 under_min = self.min_nodes - total_node_count
312 over_max = total_node_count - self.node_quota
313 total_price = self._total_price()
315 counts = self._state_counts(size)
317 up_count = self._nodes_up(counts)
318 busy_count = counts["busy"]
319 wishlist_count = self._size_wishlist(size)
321 size_name = size.name
323 size_name += ' (preemptable)'
324 self._logger.info("%s: wishlist %i, up %i (booting %i, unpaired %i, idle %i, busy %i), down %i, shutdown %i", size_name,
331 counts["down"]+counts["fail"],
336 elif under_min > 0 and size.id == self.min_cloud_size.id:
339 wanted = wishlist_count - (up_count - busy_count)
340 if wanted > 0 and self.max_total_price and ((total_price + (size.price*wanted)) > self.max_total_price):
341 can_boot = int((self.max_total_price - total_price) / size.price)
343 self._logger.info("Not booting %s (price %s) because with it would exceed max_total_price of %s (current total_price is %s)",
344 size.name, size.price, self.max_total_price, total_price)
349 def _nodes_excess(self, size):
350 counts = self._state_counts(size)
351 up_count = self._nodes_up(counts)
352 if size.id == self.min_cloud_size.id:
353 up_count -= self.min_nodes
354 return up_count - (counts["busy"] + self._size_wishlist(size))
356 def update_server_wishlist(self, wishlist):
357 self._update_poll_time('server_wishlist')
358 requestable_nodes = self.node_quota - (self._nodes_booting(None) + len(self.cloud_nodes))
359 self.last_wishlist = wishlist[:requestable_nodes]
360 for size in reversed(self.server_calculator.cloud_sizes):
362 nodes_wanted = self._nodes_wanted(size)
364 self._later.start_node(size)
365 elif (nodes_wanted < 0) and self.booting:
366 self._later.stop_booting_node(size)
368 self._logger.exception("while calculating nodes wanted for size %s", getattr(size, "id", "(id not available)"))
370 self._update_tracker()
372 self._logger.exception("while updating tracker")
374 def _check_poll_freshness(orig_func):
375 """Decorator to inhibit a method when poll information is stale.
377 This decorator checks the timestamps of all the poll information the
378 daemon has received. The decorated method is only called if none
379 of the timestamps are considered stale.
381 @functools.wraps(orig_func)
382 def wrapper(self, *args, **kwargs):
384 if all(now - t < self.poll_stale_after
385 for t in self.last_polls.itervalues()):
386 return orig_func(self, *args, **kwargs)
391 @_check_poll_freshness
392 def start_node(self, cloud_size):
393 nodes_wanted = self._nodes_wanted(cloud_size)
396 arvados_node = self.arvados_nodes.find_stale_node(self.node_stale_after)
397 self._logger.info("Want %i more %s nodes. Booting a node.",
398 nodes_wanted, cloud_size.name)
399 new_setup = self._node_setup.start(
400 timer_actor=self._timer,
401 arvados_client=self._new_arvados(),
402 arvados_node=arvados_node,
403 cloud_client=self._new_cloud(),
404 cloud_size=self.server_calculator.find_size(cloud_size.id)).proxy()
405 self.booting[new_setup.actor_ref.actor_urn] = new_setup
406 self.sizes_booting[new_setup.actor_ref.actor_urn] = cloud_size
408 if arvados_node is not None:
409 self.arvados_nodes[arvados_node['uuid']].assignment_time = (
411 new_setup.subscribe(self._later.node_setup_finished)
413 self._later.start_node(cloud_size)
415 def _get_actor_attrs(self, actor, *attr_names):
416 return pykka.get_all([getattr(actor, name) for name in attr_names])
418 def node_setup_finished(self, setup_proxy):
419 # Called when a SetupActor has completed.
420 cloud_node, arvados_node, error = self._get_actor_attrs(
421 setup_proxy, 'cloud_node', 'arvados_node', 'error')
424 if cloud_node is None:
425 # If cloud_node is None then the node create wasn't successful.
426 if error == dispatch.QuotaExceeded:
427 # We've hit a quota limit, so adjust node_quota to stop trying to
428 # boot new nodes until the node count goes down.
429 self.node_quota = len(self.cloud_nodes)
430 self._logger.warning("After quota exceeded error setting node quota to %s", self.node_quota)
432 # Node creation succeeded. Update cloud node list.
433 cloud_node._nodemanager_recently_booted = True
434 self._register_cloud_node(cloud_node)
436 # Different quota policies may in force depending on the cloud
437 # provider, account limits, and the specific mix of nodes sizes
438 # that are already created. If we are right at the quota limit,
439 # we want to probe to see if the last quota still applies or if we
440 # are allowed to create more nodes.
442 # For example, if the quota is actually based on core count, the
443 # quota might be 20 single-core machines or 10 dual-core machines.
444 # If we previously set node_quota to 10 dual core machines, but are
445 # now booting single core machines (actual quota 20), we want to
446 # allow the quota to expand so we don't get stuck at 10 machines
448 if len(self.cloud_nodes) >= self.node_quota:
449 self.node_quota = len(self.cloud_nodes)+1
450 self._logger.warning("After successful boot setting node quota to %s", self.node_quota)
452 self.node_quota = min(self.node_quota, self.max_nodes)
453 del self.booting[setup_proxy.actor_ref.actor_urn]
454 del self.sizes_booting[setup_proxy.actor_ref.actor_urn]
456 @_check_poll_freshness
457 def stop_booting_node(self, size):
458 nodes_excess = self._nodes_excess(size)
459 if (nodes_excess < 1) or not self.booting:
461 for key, node in self.booting.iteritems():
462 if node and node.cloud_size.get().id == size.id and node.stop_if_no_cloud_node().get():
463 del self.booting[key]
464 del self.sizes_booting[key]
467 self._later.stop_booting_node(size)
470 def _begin_node_shutdown(self, node_actor, cancellable):
471 cloud_node_obj = node_actor.cloud_node.get()
472 cloud_node_id = cloud_node_obj.id
473 record = self.cloud_nodes[cloud_node_id]
474 if record.shutdown_actor is not None:
476 shutdown = self._node_shutdown.start(
477 timer_actor=self._timer, cloud_client=self._new_cloud(),
478 arvados_client=self._new_arvados(),
479 node_monitor=node_actor.actor_ref, cancellable=cancellable)
480 record.shutdown_actor = shutdown.proxy()
481 shutdown.tell_proxy().subscribe(self._later.node_finished_shutdown)
483 @_check_poll_freshness
484 def node_can_shutdown(self, node_actor):
486 if self._nodes_excess(node_actor.cloud_node.get().size) > 0:
487 self._begin_node_shutdown(node_actor, cancellable=True)
488 elif self.cloud_nodes.nodes.get(node_actor.cloud_node.get().id).arvados_node is None:
489 # Node is unpaired, which means it probably exceeded its booting
490 # grace period without a ping, so shut it down so we can boot a new
492 self._begin_node_shutdown(node_actor, cancellable=False)
493 elif node_actor.in_state('down', 'fail').get():
494 # Node is down and unlikely to come back.
495 self._begin_node_shutdown(node_actor, cancellable=False)
496 except pykka.ActorDeadError as e:
497 # The monitor actor sends shutdown suggestions every time the
498 # node's state is updated, and these go into the daemon actor's
499 # message queue. It's possible that the node has already been shut
500 # down (which shuts down the node monitor actor). In that case,
501 # this message is stale and we'll get ActorDeadError when we try to
502 # access node_actor. Log the error.
503 self._logger.debug("ActorDeadError in node_can_shutdown: %s", e)
505 def node_finished_shutdown(self, shutdown_actor):
507 cloud_node, success = self._get_actor_attrs(
508 shutdown_actor, 'cloud_node', 'success')
509 except pykka.ActorDeadError:
511 cloud_node_id = cloud_node.id
514 shutdown_actor.stop()
515 except pykka.ActorDeadError:
519 record = self.cloud_nodes[cloud_node_id]
521 # Cloud node was already removed from the cloud node list
522 # supposedly while the destroy_node call was finishing its
525 record.shutdown_actor = None
530 # Shutdown was successful, so stop the monitor actor, otherwise it
531 # will keep offering the node as a candidate for shutdown.
535 # If the node went from being booted to being shut down without ever
536 # appearing in the cloud node list, it will have the
537 # _nodemanager_recently_booted tag, so get rid of it so that the node
538 # can be forgotten completely.
539 if hasattr(record.cloud_node, "_nodemanager_recently_booted"):
540 del record.cloud_node._nodemanager_recently_booted
543 self._logger.info("Shutting down after signal.")
544 self.poll_stale_after = -1 # Inhibit starting/stopping nodes
547 self._server_wishlist_actor.stop()
548 self._arvados_nodes_actor.stop()
549 self._cloud_nodes_actor.stop()
551 # Clear cloud node list
552 self.update_cloud_nodes([])
554 # Stop setup actors unless they are in the middle of setup.
555 setup_stops = {key: node.stop_if_no_cloud_node()
556 for key, node in self.booting.iteritems()}
557 self.booting = {key: self.booting[key]
558 for key in setup_stops if not setup_stops[key].get()}
559 self._later.await_shutdown()
561 def await_shutdown(self):
563 self._timer.schedule(time.time() + 1, self._later.await_shutdown)