# Copyright (C) 2011-2014 OpenStack Foundation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or # implied. # # See the License for the specific language governing permissions and # limitations under the License. import contextlib import logging import math import os import os.path import socket import threading import time import uuid from kazoo import exceptions as kze from nodepool import exceptions from nodepool import provider_manager from nodepool import stats from nodepool import config as nodepool_config from nodepool.zk import zookeeper as zk from nodepool.zk import ZooKeeperClient from nodepool.zk.components import LauncherComponent, PoolComponent from nodepool.driver.utils import QuotaInformation, QuotaSupport from nodepool.logconfig import get_annotated_logger from nodepool.version import get_version_string MINS = 60 HOURS = 60 * MINS # Interval between checking if new servers needed WATERMARK_SLEEP = 10 # When to delete node request lock znodes LOCK_CLEANUP = 8 * HOURS # How long to wait between checks for ZooKeeper connectivity if it disappears. SUSPEND_WAIT_TIME = 30 class PoolWorker(threading.Thread, stats.StatsReporter): ''' Class that manages node requests for a single provider pool. The NodePool thread will instantiate a class of this type for each provider pool found in the nodepool configuration file. If the pool or provider to which this thread is assigned is removed from the configuration file, then that will be recognized and this thread will shut itself down. ''' def __init__(self, nodepool, provider_name, pool_name): threading.Thread.__init__( self, name='PoolWorker.%s-%s' % (provider_name, pool_name) ) self.log = logging.getLogger("nodepool.%s" % self.name) self.nodepool = nodepool self.provider_name = provider_name self.pool_name = pool_name self.running = False self.stop_event = threading.Event() self.paused_handler = None self.request_handlers = [] self.watermark_sleep = nodepool.watermark_sleep self.zk = self.getZK() self.launcher_id = "%s-%s-%s" % (socket.getfqdn(), self.name, uuid.uuid4().hex) stats.StatsReporter.__init__(self) def getPriority(self): pool = self.getPoolConfig() provider = self.getProviderConfig() if pool.priority is not None: priority = pool.priority elif provider.priority is not None: priority = provider.priority else: priority = 100 return priority # --------------------------------------------------------------- # Private methods # --------------------------------------------------------------- def _assignHandlers(self, timeout=15): ''' For each request we can grab, create a NodeRequestHandler for it. The NodeRequestHandler object will kick off any threads needed to satisfy the request, then return. We will need to periodically poll the handler for completion. If exceeds the timeout it stops further iteration and returns False in order to give us time to call _removeCompletedHandlers. Otherwise it returns True to signal that it is finished for now. ''' start = time.monotonic() provider = self.getProviderConfig() if not provider: self.log.info("Missing config. Deleted provider?") return True if provider.max_concurrency == 0: return True # Get the launchers which are currently online. This may # become out of date as the loop progresses, but it should be # good enough to determine whether we should process requests # which express a preference for a specific provider. launcher_pools = self.zk.getRegisteredPools() pm = self.getProviderManager() has_quota_support = isinstance(pm, QuotaSupport) if has_quota_support: # The label quota limits will be used for the whole loop since we # don't want to accept lower priority requests when a label becomes # available after we've already deferred higher priority requests. label_quota = pm.getLabelQuota() pool = self.getPoolConfig() pool_labels = set(pool.labels) def _sort_key(request): missing_labels = set(request.node_types) - pool_labels # Return a tuple with the number of labels that are *not* served by # this pool and the request priority. This will sort the requests # that we can serve (no missing labels) before the requests that we # need to decline (missing labels > 0). return len(missing_labels), request.priority requests = sorted(self.zk.nodeRequestIterator(), key=_sort_key) for req in requests: if not self.running: return True if self.paused_handler: self.log.debug("Handler is now paused") return True # Get active threads for all pools for this provider active_threads = sum([ w.activeThreads() for w in self.nodepool.getPoolWorkers(self.provider_name) ]) # Short-circuit for limited request handling if (provider.max_concurrency > 0 and active_threads >= provider.max_concurrency): self.log.debug("Request handling limited: %s active threads ", "with max concurrency of %s", active_threads, provider.max_concurrency) return True req = self.zk.getNodeRequest(req.id) if not req: continue # Only interested in unhandled requests if req.state != zk.REQUESTED: continue # Skip it if we've already declined if self.launcher_id in req.declined_by: continue log = get_annotated_logger(self.log, event_id=req.event_id, node_request_id=req.id) # Get the candidate launchers for these nodes candidate_launcher_pools = set( x for x in launcher_pools if (set(x.supported_labels).issuperset(set(req.node_types)) and x.id not in req.declined_by) ) # Skip this request if it is requesting another provider # which is online if req.provider and req.provider != self.provider_name: # The request is asking for a specific provider launcher_pool_ids_for_provider = set( x.id for x in candidate_launcher_pools if x.provider_name == req.provider ) if launcher_pool_ids_for_provider: # There is a launcher online which can satisfy the # request that has not yet declined the request, # so yield to it. log.debug("Yielding request to provider %s %s", req.provider, launcher_pool_ids_for_provider) continue priority = self.getPriority() launcher_pool_ids_with_higher_priority = set( x.id for x in candidate_launcher_pools if x.priority < priority and not x.paused ) if launcher_pool_ids_with_higher_priority: log.debug("Yielding request to higher priority providers %s", launcher_pool_ids_with_higher_priority) continue if has_quota_support and not all(label_quota.get(l, math.inf) > 0 for l in req.node_types): # Defer the request as we can't provide the required labels at # the moment. log.debug("Deferring request because labels are unavailable") continue # check tenant quota if the request has a tenant associated # and there are resource limits configured for this tenant check_tenant_quota = req.tenant_name and req.tenant_name \ in self.nodepool.config.tenant_resource_limits \ and has_quota_support if check_tenant_quota and not self._hasTenantQuota(req, pm): # Defer request for it to be handled and fulfilled at a later # run. log.debug("Deferring request %s because it would " "exceed tenant quota", req) continue log.debug("Locking request") try: self.zk.lockNodeRequest(req, blocking=False) except exceptions.ZKLockException: log.debug("Request is locked by someone else") continue # Make sure the state didn't change on us after getting the lock if req.state != zk.REQUESTED: self.zk.unlockNodeRequest(req) log.debug("Request is in state %s", req.state) continue # Got a lock, so assign it log.info("Assigning node request %s" % req) if has_quota_support: # Adjust the label quota so we don't accept more requests # than we have labels available. for label in req.node_types: with contextlib.suppress(KeyError): label_quota[label] -= 1 rh = pm.getRequestHandler(self, req) rh.run() if rh.paused: self.paused_handler = rh self.request_handlers.append(rh) # if we exceeded the timeout stop iterating here if time.monotonic() - start > timeout: return False return True def _removeCompletedHandlers(self): ''' Poll handlers to see which have completed. ''' active_handlers = [] for r in self.request_handlers: log = get_annotated_logger(self.log, event_id=r.request.event_id, node_request_id=r.request.id) try: if not r.poll(): active_handlers.append(r) if r.paused: self.paused_handler = r else: log.debug("Removing request handler") except kze.SessionExpiredError: # If we lost our ZooKeeper session, we've lost our NodeRequest # lock so it's no longer active log.error("Removing request handler (lost zookeeper session)") continue except Exception: # If we fail to poll a request handler log it but move on # and process the other handlers. We keep this handler around # and will try again later. log.exception("Error polling request handler") active_handlers.append(r) self.request_handlers = active_handlers active_reqs = [r.request.id for r in self.request_handlers] self.log.debug("Active requests: %s", active_reqs) def _hasTenantQuota(self, request, provider_manager): ''' Checks if a tenant has enough quota to handle a list of nodes. This takes into account the all currently existing nodes as reported by zk. :param request: the node request in question :param provider_manager: the provider manager for the request :return: True if there is enough quota for the tenant, False otherwise ''' log = get_annotated_logger(self.log, event_id=request.event_id, node_request_id=request.id) tenant_name = request.tenant_name needed_quota = QuotaInformation() pool = self.getPoolConfig() for ntype in request.node_types: # can not determine needed quota for ntype if label not in pool # therefore just skip them here to avoid errors in # 'quotaNeededByLabel' if ntype not in pool.labels: continue needed_quota.add(provider_manager.quotaNeededByLabel(ntype, pool)) used_quota = self._getUsedQuotaForTenant(tenant_name) tenant_quota = QuotaInformation( default=math.inf, **self.nodepool.config.tenant_resource_limits[tenant_name]) tenant_quota.subtract(used_quota) log.debug("Current tenant quota for %s: %s", tenant_name, tenant_quota) tenant_quota.subtract(needed_quota) log.debug("Predicted remaining tenant quota for %s: %s", tenant_name, tenant_quota) return tenant_quota.non_negative() def _getUsedQuotaForTenant(self, tenant_name): used_quota = QuotaInformation() for node in self.zk.nodeIterator(cached_ids=True): if not node.resources: continue if node.tenant_name == tenant_name: resources = QuotaInformation(**node.resources) used_quota.add(resources) return used_quota # --------------------------------------------------------------- # Public methods # --------------------------------------------------------------- def activeThreads(self): ''' Return the number of alive threads in use by this provider. This is an approximate, top-end number for alive threads, since some threads obviously may have finished by the time we finish the calculation. ''' total = 0 for r in self.request_handlers: total += r.alive_thread_count return total def getZK(self): return self.nodepool.getZK() def getProviderConfig(self): return self.nodepool.config.providers.get(self.provider_name) def getPoolConfig(self): provider = self.getProviderConfig() if provider: return provider.pools[self.pool_name] else: return None def getProviderManager(self): return self.nodepool.getProviderManager(self.provider_name) def run(self): self.running = True # Make sure we're always registered with ZK hostname = socket.gethostname() self.component_info = PoolComponent( self.zk.client, hostname, version=get_version_string()) labels = set() for prov_cfg in self.nodepool.config.providers.values(): labels.update(prov_cfg.getSupportedLabels()) self.component_info.content.update({ 'id': self.launcher_id, 'provider_name': self.provider_name, 'supported_labels': list(labels), 'state': self.component_info.RUNNING, 'priority': self.getPriority(), }) self.component_info.register() while self.running: try: # Don't do work if we've lost communication with the ZK cluster did_suspend = False while self.zk and (self.zk.suspended or self.zk.lost): did_suspend = True self.log.info("ZooKeeper suspended. Waiting") time.sleep(SUSPEND_WAIT_TIME) if did_suspend: self.log.info("ZooKeeper available. Resuming") labels = set() for prov_cfg in self.nodepool.config.providers.values(): labels.update(prov_cfg.getSupportedLabels()) self.component_info.supported_labels = list(labels) self.component_info.priority = self.getPriority() self.updateProviderLimits( self.nodepool.config.providers.get(self.provider_name)) self.updateTenantLimits( self.nodepool.config.tenant_resource_limits) if not self.paused_handler: self.component_info.paused = False while not self._assignHandlers(): # _assignHandlers can take quite some time on a busy # system so sprinkle _removeCompletedHandlers in # between such that we have a chance to fulfill # requests that already have all nodes. self._removeCompletedHandlers() else: self.component_info.paused = True # If we are paused, one request handler could not # satisfy its assigned request, so give it # another shot. Unpause ourselves if it completed. self.paused_handler.run() if not self.paused_handler.paused: self.paused_handler = None self._removeCompletedHandlers() except Exception: self.log.exception("Error in PoolWorker:") self.stop_event.wait(self.watermark_sleep) # Cleanup on exit if self.paused_handler: self.paused_handler.unlockNodeSet(clear_allocation=True) def stop(self): ''' Shutdown the PoolWorker thread. Do not wait for the request handlers to finish. Any nodes that are in the process of launching will be cleaned up on a restart. They will be unlocked and BUILDING in ZooKeeper. ''' self.log.info("%s received stop" % self.name) self.running = False self.component_info.unregister() self.stop_event.set() class BaseCleanupWorker(threading.Thread): def __init__(self, nodepool, interval, name): threading.Thread.__init__(self, name=name) self._nodepool = nodepool self._interval = interval self._running = False self._stop_event = threading.Event() def run(self): self.log.info("Starting") self._running = True while self._running: # Don't do work if we've lost communication with the ZK cluster did_suspend = False zk_conn = self._nodepool.getZK() while zk_conn and (zk_conn.suspended or zk_conn.lost): did_suspend = True self.log.info("ZooKeeper suspended. Waiting") time.sleep(SUSPEND_WAIT_TIME) if did_suspend: self.log.info("ZooKeeper available. Resuming") self.log.info('Starting cleanup') self._run() self.log.info('Finished cleanup') self._stop_event.wait(self._interval) self.log.info("Stopped") def stop(self): self._running = False self._stop_event.set() self.join() class CleanupWorker(BaseCleanupWorker): def __init__(self, nodepool, interval): super(CleanupWorker, self).__init__( nodepool, interval, name='CleanupWorker') self.log = logging.getLogger("nodepool.CleanupWorker") # List of callable tasks we want to perform during cleanup, and a # brief description of the task. self._tasks = [ (self._cleanupNodeRequestLocks, 'node request lock cleanup'), (self._cleanupLeakedInstances, 'leaked instance cleanup'), (self._cleanupLostRequests, 'lost request cleanup'), (self._cleanupMaxReadyAge, 'max ready age cleanup'), (self._cleanupMaxHoldAge, 'max hold age cleanup'), (self._cleanupEmptyNodes, 'empty node cleanup'), ] def _resetLostRequest(self, zk_conn, req): ''' Reset the request state and deallocate nodes. :param ZooKeeper zk_conn: A ZooKeeper connection object. :param NodeRequest req: The lost NodeRequest object. ''' # Double check the state after the lock req = zk_conn.getNodeRequest(req.id) if req.state != zk.PENDING: return log = get_annotated_logger(self.log, event_id=req.event_id, node_request_id=req.id) for node in zk_conn.nodeIterator(): if node.allocated_to == req.id: try: zk_conn.lockNode(node) except exceptions.ZKLockException: log.warning( "Unable to grab lock to deallocate node %s from " "request", node.id) return # If the node is in state init then the launcher that worked # on the lost request has been interrupted between creating # the znode and locking/setting to building. In this case the # znode is leaked and we should delete the node instead of # just deallocating it. if node.state == zk.INIT: node.state = zk.DELETING node.allocated_to = None try: zk_conn.storeNode(node) log.debug("Deallocated node %s for lost request", node.id) except Exception: log.exception( "Unable to deallocate node %s from request:", node.id) zk_conn.unlockNode(node) req.state = zk.REQUESTED req.nodes = [] zk_conn.storeNodeRequest(req) log.info("Reset lost request") def _cleanupLostRequests(self): ''' Look for lost requests and reset them. A lost request is a node request that was left in the PENDING state when nodepool exited. We need to look for these (they'll be unlocked) and disassociate any nodes we've allocated to the request and reset the request state to REQUESTED so it will be processed again. ''' zk_conn = self._nodepool.getZK() for req in zk_conn.nodeRequestIterator(): log = get_annotated_logger(self.log, event_id=req.event_id, node_request_id=req.id) if req.state == zk.PENDING: try: zk_conn.lockNodeRequest(req, blocking=False) except exceptions.ZKLockException: continue try: self._resetLostRequest(zk_conn, req) except Exception: log.exception("Error resetting lost request:") zk_conn.unlockNodeRequest(req) def _cleanupNodeRequestLocks(self): ''' Remove request locks where the request no longer exists. Because the node request locks are not direct children of the request znode, we need to remove the locks separately after the request has been processed. Only remove them after LOCK_CLEANUP seconds have passed. This helps reduce chances of the scenario where a request could go away _while_ a lock is currently held for processing and the cleanup thread attempts to delete it. The delay should reduce the chance that we delete a currently held lock. ''' zk = self._nodepool.getZK() requests = zk.getNodeRequests() now = time.time() for lock_stat in zk.nodeRequestLockStatsIterator(): if lock_stat.lock_id in requests: continue if (now - lock_stat.stat.mtime / 1000) > LOCK_CLEANUP: zk.deleteNodeRequestLock(lock_stat.lock_id) def _cleanupLeakedInstances(self): ''' Allow each provider manager a chance to cleanup resources. ''' for provider in self._nodepool.config.providers.values(): manager = self._nodepool.getProviderManager(provider.name) if manager: try: manager.cleanupLeakedResources() except Exception: self.log.exception( "Failure during resource cleanup for provider %s", provider.name) def _cleanupMaxReadyAge(self): ''' Delete any server past their max-ready-age. Remove any servers which are longer than max-ready-age in ready state. ''' # first get all labels with max_ready_age > 0 label_names = [] for label_name in self._nodepool.config.labels: if self._nodepool.config.labels[label_name].max_ready_age > 0: label_names.append(label_name) zk_conn = self._nodepool.getZK() ready_nodes = zk_conn.getReadyNodesOfTypes(label_names) for label_name in ready_nodes: # get label from node label = self._nodepool.config.labels[label_name] for node in ready_nodes[label_name]: # Can't do anything if we aren't configured for this provider. if node.provider not in self._nodepool.config.providers: continue # check state time against now now = int(time.time()) if (now - node.state_time) < label.max_ready_age: continue try: zk_conn.lockNode(node, blocking=False) except exceptions.ZKLockException: continue # Double check the state now that we have a lock since it # may have changed on us. if node.state != zk.READY: zk_conn.unlockNode(node) continue self.log.debug("Node %s exceeds max ready age: %s >= %s", node.id, now - node.state_time, label.max_ready_age) try: node.state = zk.DELETING zk_conn.storeNode(node) except Exception: self.log.exception( "Failure marking aged node %s for delete:", node.id) finally: zk_conn.unlockNode(node) def _cleanupMaxHoldAge(self): ''' Delete any held server past their max-hold-age. Remove any servers which are longer than max-hold-age in hold state. ''' self.log.debug('Cleaning up held nodes...') zk_conn = self._nodepool.getZK() held_nodes = [n for n in zk_conn.nodeIterator() if n.state == zk.HOLD] for node in held_nodes: # Can't do anything if we aren't configured for this provider. if node.provider not in self._nodepool.config.providers: continue if node.hold_expiration is not None and node.hold_expiration > 0: expiration = node.hold_expiration else: expiration = math.inf max_uptime = min(expiration, self._nodepool.config.max_hold_age) if math.isinf(max_uptime): continue # check state time against now now = int(time.time()) if (now - node.state_time) < max_uptime: continue try: zk_conn.lockNode(node, blocking=False) except exceptions.ZKLockException: continue # Double check the state now that we have a lock since it # may have changed on us. if node.state != zk.HOLD: zk_conn.unlockNode(node) continue self.log.debug("Node %s exceeds max hold age (%s): %s >= %s", node.id, ("manual setting" and node.hold_expiration == max_uptime or "configuration setting"), now - node.state_time, max_uptime) try: node.state = zk.DELETING zk_conn.storeNode(node) except Exception: self.log.exception( "Failure marking aged node %s for delete:", node.id) finally: zk_conn.unlockNode(node) def _cleanupEmptyNodes(self): ''' Remove any Node znodes that may be totally empty. ''' self.log.debug('Cleaning up empty nodes...') zk_conn = self._nodepool.getZK() # We cannot use nodeIterator() here since that does not yield us # empty nodes. for node_id in zk_conn.getNodes(): node = zk_conn.getNode(node_id) if node is None: self.log.debug("Removing empty node %s", node_id) zk_conn.deleteRawNode(node_id) def _run(self): ''' Catch exceptions individually so that other cleanup routines may have a chance. ''' for task, description in self._tasks: try: task() except Exception: self.log.exception( "Exception in %s (%s)", self.name, description) class DeletedNodeWorker(BaseCleanupWorker): ''' Class for deleting all nodes with state of DELETING. ''' def __init__(self, nodepool, interval): super(DeletedNodeWorker, self).__init__( nodepool, interval, name='DeletedNodeWorker') self.log = logging.getLogger("nodepool.DeletedNodeWorker") def _deleteInstance(self, node): ''' Delete an instance from a provider. A thread will be spawned to delete the actual instance from the provider. :param Node node: A Node object representing the instance to delete. ''' self.log.info("Deleting %s instance %s from %s", node.state, node.external_id, node.provider) try: pm = self._nodepool.getProviderManager(node.provider) pm.startNodeCleanup(node) except Exception: self.log.exception("Could not delete instance %s on provider %s", node.external_id, node.provider) def _cleanupNodes(self): ''' Delete instances from providers and nodes entries from ZooKeeper. ''' cleanup_states = (zk.USED, zk.IN_USE, zk.BUILDING, zk.FAILED, zk.DELETING, zk.DELETED, zk.ABORTED) zk_conn = self._nodepool.getZK() for node in zk_conn.nodeIterator(): # If a ready node has been allocated to a request, but that # request is now missing, deallocate it. if (node.state == zk.READY and node.allocated_to and not zk_conn.getNodeRequest(node.allocated_to)): try: zk_conn.lockNode(node, blocking=False) except exceptions.ZKLockException: pass else: # Double check node conditions after lock if node.state == zk.READY and node.allocated_to: node.allocated_to = None try: zk_conn.storeNode(node) self.log.debug( "Deallocated node %s with missing request %s", node.id, node.allocated_to) except Exception: self.log.exception( "Failed to deallocate node %s for missing " "request %s:", node.id, node.allocated_to) zk_conn.unlockNode(node) # Can't do anything if we aren't configured for this provider. if node.provider not in self._nodepool.config.providers: continue # Any nodes in these states that are unlocked can be deleted. if node.state in cleanup_states: try: zk_conn.lockNode(node, blocking=False) except exceptions.ZKLockException: continue if (node.state == zk.DELETED or node.provider is None): # The node has been deleted out from under us -- # we only obtained the lock because in the # recursive delete, the lock is deleted first and # we locked the node between the time of the lock # delete and the node delete. We need to clean up # the mess. try: # This should delete the lock as well zk_conn.deleteNode(node) except Exception: self.log.exception( "Error deleting already deleted znode:") try: zk_conn.unlockNode(node) except Exception: self.log.exception( "Error unlocking already deleted znode:") continue # Double check the state now that we have a lock since it # may have changed on us. if node.state not in cleanup_states: zk_conn.unlockNode(node) continue self.log.debug( "Marking for deletion unlocked node %s " "(state: %s, allocated_to: %s)", node.id, node.state, node.allocated_to) # The NodeDeleter thread will unlock and remove the # node from ZooKeeper if it succeeds. try: self._deleteInstance(node) except Exception: self.log.exception( "Failure deleting node %s in cleanup state %s:", node.id, node.state) zk_conn.unlockNode(node) def _run(self): try: self._cleanupNodes() except Exception: self.log.exception("Exception in DeletedNodeWorker:") class StatsWorker(BaseCleanupWorker, stats.StatsReporter): def __init__(self, nodepool, interval): super().__init__(nodepool, interval, name='StatsWorker') self.log = logging.getLogger('nodepool.StatsWorker') self.stats_event = threading.Event() self.election = None def stop(self): self._running = False if self.election is not None: self.log.debug('Cancel leader election') self.election.cancel() self.stats_event.set() super().stop() def _run(self): try: stats.StatsReporter.__init__(self) if not self._statsd: return if self.election is None: zk = self._nodepool.getZK() identifier = "%s-%s" % (socket.gethostname(), os.getpid()) self.election = zk.getStatsElection(identifier) if not self._running: return self.election.run(self._run_stats) except Exception: self.log.exception('Exception in StatsWorker:') def _run_stats(self): self.log.info('Won stats reporter election') # enable us getting events zk = self._nodepool.getZK() zk.setNodeStatsEvent(self.stats_event) while self._running: signaled = self.stats_event.wait() if not self._running: break if not signaled: continue self.stats_event.clear() try: self.updateNodeStats(zk) except Exception: self.log.exception("Exception while reporting stats:") time.sleep(1) # Unregister from node stats events zk.setNodeStatsEvent(None) class NodePool(threading.Thread): log = logging.getLogger("nodepool.NodePool") def __init__(self, securefile, configfile, watermark_sleep=WATERMARK_SLEEP): threading.Thread.__init__(self, name='NodePool') self.securefile = securefile self.configfile = configfile self.watermark_sleep = watermark_sleep self.cleanup_interval = 60 self.delete_interval = 5 self.stats_interval = 5 self._stopped = False self._stop_event = threading.Event() self.config = None self.zk = None self.statsd = stats.get_client() self._pool_threads = {} self._cleanup_thread = None self._delete_thread = None self._stats_thread = None self._submittedRequests = {} self.ready = False def stop(self): self._stopped = True self._stop_event.set() # Our run method can start new threads, so make sure it has # completed before we continue the shutdown. if self.is_alive(): self.join() if self._cleanup_thread: self._cleanup_thread.stop() self._cleanup_thread.join() if self._delete_thread: self._delete_thread.stop() self._delete_thread.join() if self._stats_thread: self._stats_thread.stop() self._stats_thread.join() # Don't let stop() return until all pool threads have been # terminated. self.log.debug("Stopping pool threads") for thd in self._pool_threads.values(): if thd.is_alive(): thd.stop() self.log.debug("Waiting for %s" % thd.name) thd.join() # Stop providers after all the cleanup threads have stopped. if self.config: provider_manager.ProviderManager.stopProviders(self.config) if self.zk: self.zk.disconnect() self.log.debug("Finished stopping") def loadConfig(self): config = nodepool_config.loadConfig(self.configfile) if self.securefile: nodepool_config.loadSecureConfig(config, self.securefile) return config def reconfigureZooKeeper(self, config): if self.config: running = self.config.zookeeper_servers else: running = None configured = config.zookeeper_servers if running == configured: return if not self.zk and configured: self.log.debug("Connecting to ZooKeeper servers") self.zk_client = ZooKeeperClient( configured, tls_cert=config.zookeeper_tls_cert, tls_key=config.zookeeper_tls_key, tls_ca=config.zookeeper_tls_ca, timeout=config.zookeeper_timeout, ) self.zk_client.connect() self.zk = zk.ZooKeeper(self.zk_client) hostname = socket.gethostname() self.component_info = LauncherComponent( self.zk_client, hostname, version=get_version_string()) self.component_info.register() else: self.log.debug("Detected ZooKeeper server changes") self.zk_client.resetHosts(configured) def setConfig(self, config): self.config = config def getZK(self): return self.zk def getProviderManager(self, provider_name): return self.config.provider_managers.get(provider_name) def getPoolWorkers(self, provider_name): return [t for t in self._pool_threads.values() if t.provider_name == provider_name] def updateConfig(self): config = self.loadConfig() self.reconfigureZooKeeper(config) provider_manager.ProviderManager.reconfigure(self.config, config, self.getZK()) for provider_name in list(config.providers.keys()): if provider_name not in config.provider_managers: del config.providers[provider_name] self.setConfig(config) def removeCompletedRequests(self): ''' Remove (locally and in ZK) fulfilled node requests. We also must reset the allocated_to attribute for each Node assigned to our request, since we are deleting the request. ''' # Use a copy of the labels because we modify _submittedRequests # within the loop below. Note that keys() returns an iterator in # py3, so we need to explicitly make a new list. requested_labels = list(self._submittedRequests.keys()) for label in requested_labels: label_requests = self._submittedRequests[label] active_requests = [] for req in label_requests: req = self.zk.getNodeRequest(req.id) if not req: continue log = get_annotated_logger(self.log, event_id=req.event_id, node_request_id=req.id) if req.state == zk.FULFILLED: # Reset node allocated_to for node_id in req.nodes: node = self.zk.getNode(node_id) node.allocated_to = None # NOTE: locking shouldn't be necessary since a node # with allocated_to set should not be locked except # by the creator of the request (us). self.zk.storeNode(node) self.zk.deleteNodeRequest(req) elif req.state == zk.FAILED: log.debug("min-ready node request failed: %s", req) self.zk.deleteNodeRequest(req) else: active_requests.append(req) if active_requests: self._submittedRequests[label] = active_requests else: self.log.debug( "No more active min-ready requests for label %s", label) del self._submittedRequests[label] def labelImageIsAvailable(self, label): ''' Check if the image associated with a label is ready in any provider. :param Label label: The label config object. :returns: True if image associated with the label is uploaded and ready in at least one provider. False otherwise. ''' for pool in label.pools: if not pool.provider.manage_images: # Provider doesn't manage images, assuming label is ready return True for pool_label in pool.labels.values(): if pool_label.diskimage: if self.zk.getMostRecentImageUpload( pool_label.diskimage.name, pool.provider.name): return True else: manager = self.getProviderManager(pool.provider.name) if manager.labelReady(pool_label): return True return False def createMinReady(self): ''' Create node requests to make the minimum amount of ready nodes. Since this method will be called repeatedly, we need to take care to note when we have already submitted node requests to satisfy min-ready. Requests we've already submitted are stored in the _submittedRequests dict, keyed by label. ''' def createRequest(label_name): req = zk.NodeRequest() req.state = zk.REQUESTED req.requestor = "NodePool:min-ready" req.node_types.append(label_name) req.reuse = False # force new node launches self.zk.storeNodeRequest(req, priority="900") if label_name not in self._submittedRequests: self._submittedRequests[label_name] = [] self._submittedRequests[label_name].append(req) # Since we could have already submitted node requests, do not # resubmit a request for a type if a request for that type is # still in progress. self.removeCompletedRequests() label_names = list(self.config.labels.keys()) requested_labels = list(self._submittedRequests.keys()) needed_labels = list(set(label_names) - set(requested_labels)) # Note we explicitly don't use the cache here because otherwise we can # end up creating more min-ready nodes than we want. ready_nodes = self.zk.getReadyNodesOfTypes(needed_labels, cached=False) for label in self.config.labels.values(): if label.name not in needed_labels: continue min_ready = label.min_ready if min_ready <= 0: continue # disabled # Calculate how many nodes of this type we need created need = 0 if label.name not in ready_nodes: need = label.min_ready elif len(ready_nodes[label.name]) < min_ready: need = min_ready - len(ready_nodes[label.name]) if need and self.labelImageIsAvailable(label): # Create requests for 1 node at a time. This helps to split # up requests across providers, and avoids scenario where a # single provider might fail the entire request because of # quota (e.g., min-ready=2, but max-servers=1). self.log.info("Creating requests for %d %s nodes", need, label.name) for i in range(0, need): createRequest(label.name) def run(self): ''' Start point for the NodePool thread. ''' self.log.info("Nodepool launcher %s starting", get_version_string()) while not self._stopped: try: self.updateConfig() # Don't do work if we've lost communication with the ZK cluster did_suspend = False while self.zk and (self.zk.suspended or self.zk.lost): did_suspend = True self.log.info("ZooKeeper suspended. Waiting") time.sleep(SUSPEND_WAIT_TIME) if did_suspend: self.log.info("ZooKeeper available. Resuming") if self.component_info.state != self.component_info.RUNNING: self.component_info.state = self.component_info.RUNNING self.createMinReady() if not self._cleanup_thread: self._cleanup_thread = CleanupWorker( self, self.cleanup_interval) self._cleanup_thread.start() if not self._delete_thread: self._delete_thread = DeletedNodeWorker( self, self.delete_interval) self._delete_thread.start() if not self._stats_thread: self._stats_thread = StatsWorker(self, self.stats_interval) self._stats_thread.start() # Stop any PoolWorker threads if the pool was removed # from the config. pool_keys = set() for provider in self.config.providers.values(): for pool in provider.pools.values(): pool_keys.add(provider.name + '-' + pool.name) new_pool_threads = {} for key in self._pool_threads.keys(): if key not in pool_keys: self._pool_threads[key].stop() else: new_pool_threads[key] = self._pool_threads[key] self._pool_threads = new_pool_threads # Start (or restart) provider threads for each provider in # the config. Removing a provider from the config and then # adding it back would cause a restart. for provider in self.config.providers.values(): for pool in provider.pools.values(): key = provider.name + '-' + pool.name if key not in self._pool_threads: t = PoolWorker(self, provider.name, pool.name) self.log.info("Starting %s" % t.name) t.start() self._pool_threads[key] = t elif not self._pool_threads[key].is_alive(): self._pool_threads[key].stop() self._pool_threads[key].join() t = PoolWorker(self, provider.name, pool.name) self.log.info("Restarting %s" % t.name) t.start() self._pool_threads[key] = t except Exception: self.log.exception("Exception in main loop:") # At this point all providers are registered and fully functional # so we can mark nodepool as ready. self.ready = True self._stop_event.wait(self.watermark_sleep)