class Manager(object):
"""
Manage a collection of ClusterMonitors.
Subscribe to ceph/cluster events, and create a ClusterMonitor
for any FSID we haven't seen before.
"""
def __init__(self):
self._complete = gevent.event.Event()
self._rpc_thread = RpcThread(self)
self._discovery_thread = TopLevelEvents(self)
self._process_monitor = ProcessMonitorThread()
self.notifier = NotificationThread()
if sqlalchemy is not None:
try:
# Prepare persistence
engine = create_engine(config.get('cthulhu', 'db_path'))
Session.configure(bind=engine)
self.persister = Persister()
except sqlalchemy.exc.ArgumentError as e:
log.error("Database error: %s" % e)
raise
else:
class NullPersister(object):
def start(self):
pass
def stop(self):
pass
def join(self):
pass
def __getattribute__(self, item):
if item.startswith('_'):
return object.__getattribute__(self, item)
else:
try:
return object.__getattribute__(self, item)
except AttributeError:
def blackhole(*args, **kwargs):
pass
return blackhole
self.persister = NullPersister()
# Remote operations
self.requests = RequestCollection(self)
self._request_ticker = Ticker(request_collection.TICK_PERIOD,
lambda: self.requests.tick())
# FSID to ClusterMonitor
self.clusters = {}
# Generate events on state changes
self.eventer = Eventer(self)
# Handle all ceph/server messages
self.servers = ServerMonitor(self.persister, self.eventer,
self.requests)
def delete_cluster(self, fs_id):
"""
Note that the cluster will pop right back again if it's
still sending heartbeats.
"""
victim = self.clusters[fs_id]
victim.stop()
victim.done.wait()
del self.clusters[fs_id]
self._expunge(fs_id)
def stop(self):
log.info("%s stopping" % self.__class__.__name__)
for monitor in self.clusters.values():
monitor.stop()
self._rpc_thread.stop()
self._discovery_thread.stop()
self._process_monitor.stop()
self.notifier.stop()
self.eventer.stop()
self._request_ticker.stop()
def _expunge(self, fsid):
session = Session()
session.query(SyncObject).filter_by(fsid=fsid).delete()
session.commit()
def _recover(self):
if sqlalchemy is None:
return
session = Session()
for server in session.query(Server).all():
log.debug("Recovered server %s" % server.fqdn)
assert server.boot_time is None or server.boot_time.tzinfo is not None # expect timezone-aware DB backend
self.servers.inject_server(
ServerState(fqdn=server.fqdn,
hostname=server.hostname,
managed=server.managed,
last_contact=server.last_contact,
boot_time=server.boot_time,
ceph_version=server.ceph_version))
for service in session.query(Service).all():
if service.server:
server = session.query(Server).get(service.server)
else:
server = None
log.debug("Recovered service %s/%s/%s on %s" %
(service.fsid, service.service_type, service.service_id,
server.fqdn if server else None))
self.servers.inject_service(
ServiceState(fsid=service.fsid,
service_type=service.service_type,
service_id=service.service_id),
server.fqdn if server else None)
# I want the most recent version of every sync_object
fsids = [(row[0], row[1]) for row in session.query(
SyncObject.fsid, SyncObject.cluster_name).distinct(SyncObject.fsid)
]
for fsid, name in fsids:
cluster_monitor = ClusterMonitor(fsid, name, self.notifier,
self.persister, self.servers,
self.eventer, self.requests)
self.clusters[fsid] = cluster_monitor
object_types = [
row[0]
for row in session.query(SyncObject.sync_type).filter_by(
fsid=fsid).distinct()
]
for sync_type in object_types:
latest_record = session.query(SyncObject).filter_by(
fsid=fsid,
sync_type=sync_type).order_by(SyncObject.version.desc(),
SyncObject.when.desc())[0]
# FIXME: bit of a hack because records persisted only store their 'version'
# if it's a real counter version, underlying problem is that we have
# underlying data (health, pg_brief) without usable version counters.
def md5(raw):
hasher = hashlib.md5()
hasher.update(raw)
return hasher.hexdigest()
if latest_record.version:
version = latest_record.version
else:
version = md5(latest_record.data)
when = latest_record.when
when = when.replace(tzinfo=tzutc())
if cluster_monitor.update_time is None or when > cluster_monitor.update_time:
cluster_monitor.update_time = when
cluster_monitor.inject_sync_object(
None, sync_type, version,
msgpack.unpackb(latest_record.data))
for monitor in self.clusters.values():
log.info("Recovery: Cluster %s with update time %s" %
(monitor.fsid, monitor.update_time))
monitor.start()
def start(self):
log.info("%s starting" % self.__class__.__name__)
# Before we start listening to the outside world, recover
# our last known state from persistent storage
try:
self._recover()
except:
log.exception("Recovery failed")
os._exit(-1)
self._rpc_thread.bind()
self._rpc_thread.start()
self._discovery_thread.start()
self._process_monitor.start()
self.notifier.start()
self.persister.start()
self.eventer.start()
self._request_ticker.start()
self.servers.start()
def join(self):
log.info("%s joining" % self.__class__.__name__)
self._rpc_thread.join()
self._discovery_thread.join()
self._process_monitor.join()
self.notifier.join()
self.persister.join()
self.eventer.join()
self._request_ticker.join()
self.servers.join()
for monitor in self.clusters.values():
monitor.join()
def on_discovery(self, minion_id, heartbeat_data):
log.info("on_discovery: {0}/{1}".format(minion_id,
heartbeat_data['fsid']))
cluster_monitor = ClusterMonitor(heartbeat_data['fsid'],
heartbeat_data['name'], self.notifier,
self.persister, self.servers,
self.eventer, self.requests)
self.clusters[heartbeat_data['fsid']] = cluster_monitor
# Run before passing on the heartbeat, because otherwise the
# syncs resulting from the heartbeat might not be received
# by the monitor.
cluster_monitor.start()
# Wait for ClusterMonitor to start accepting events before asking it
# to do anything
cluster_monitor.ready()
cluster_monitor.on_heartbeat(minion_id, heartbeat_data)
class Manager(object):
"""
Manage a collection of ClusterMonitors.
Subscribe to ceph/cluster events, and create a ClusterMonitor
for any FSID we haven't seen before.
"""
def __init__(self):
self._complete = gevent.event.Event()
self._rpc_thread = RpcThread(self)
self._discovery_thread = TopLevelEvents(self)
self._process_monitor = ProcessMonitorThread()
db_path = config.get('cthulhu', 'db_path')
if sqlalchemy is not None and db_path:
try:
# Prepare persistence
engine = create_engine(config.get('cthulhu', 'db_path')) # noqa
Session.configure(bind=engine)
self.persister = Persister()
except sqlalchemy.exc.ArgumentError as e:
log.error("Database error: %s" % e)
raise
else:
class NullPersister(object):
def start(self):
pass
def stop(self):
pass
def join(self):
pass
def __getattribute__(self, item):
if item.startswith('_'):
return object.__getattribute__(self, item)
else:
try:
return object.__getattribute__(self, item)
except AttributeError:
def blackhole(*args, **kwargs):
pass
return blackhole
self.persister = NullPersister()
# Remote operations
self.requests = RequestCollection(self)
self._request_ticker = Ticker(request_collection.TICK_PERIOD,
lambda: self.requests.tick())
# FSID to ClusterMonitor
self.clusters = {}
# Generate events on state changes
self.eventer = Eventer(self)
# Handle all ceph/server messages
self.servers = ServerMonitor(self.persister, self.eventer, self.requests)
def delete_cluster(self, fs_id):
"""
Note that the cluster will pop right back again if it's
still sending heartbeats.
"""
victim = self.clusters[fs_id]
victim.stop()
victim.done.wait()
del self.clusters[fs_id]
self._expunge(fs_id)
def stop(self):
log.info("%s stopping" % self.__class__.__name__)
for monitor in self.clusters.values():
monitor.stop()
self._rpc_thread.stop()
self._discovery_thread.stop()
self._process_monitor.stop()
self.eventer.stop()
self._request_ticker.stop()
def _expunge(self, fsid):
if sqlalchemy is None:
return
session = Session()
session.query(SyncObject).filter_by(fsid=fsid).delete()
session.commit()
def _recover(self):
if sqlalchemy is None:
return
session = Session()
for server in session.query(Server).all():
log.debug("Recovered server %s" % server.fqdn)
assert server.boot_time is None or server.boot_time.tzinfo is not None # expect timezone-aware DB backend
self.servers.inject_server(ServerState(
fqdn=server.fqdn,
hostname=server.hostname,
managed=server.managed,
last_contact=server.last_contact,
boot_time=server.boot_time,
ceph_version=server.ceph_version
))
for service in session.query(Service).all():
if service.server:
server = session.query(Server).get(service.server)
else:
server = None
log.debug("Recovered service %s/%s/%s on %s" % (
service.fsid, service.service_type, service.service_id, server.fqdn if server else None
))
self.servers.inject_service(ServiceState(
fsid=service.fsid,
service_type=service.service_type,
service_id=service.service_id
), server.fqdn if server else None)
# I want the most recent version of every sync_object
fsids = [(row[0], row[1]) for row in session.query(SyncObject.fsid, SyncObject.cluster_name).distinct(SyncObject.fsid)]
for fsid, name in fsids:
cluster_monitor = ClusterMonitor(fsid, name, self.persister, self.servers,
self.eventer, self.requests)
self.clusters[fsid] = cluster_monitor
object_types = [row[0] for row in session.query(SyncObject.sync_type).filter_by(fsid=fsid).distinct()]
for sync_type in object_types:
latest_record = session.query(SyncObject).filter_by(
fsid=fsid, sync_type=sync_type).order_by(
SyncObject.version.desc(), SyncObject.when.desc())[0]
# FIXME: bit of a hack because records persisted only store their 'version'
# if it's a real counter version, underlying problem is that we have
# underlying data (health, pg_brief) without usable version counters.
def md5(raw):
hasher = hashlib.md5()
hasher.update(raw)
return hasher.hexdigest()
if latest_record.version:
version = latest_record.version
else:
version = md5(latest_record.data)
when = latest_record.when
when = when.replace(tzinfo=tzutc())
if cluster_monitor.update_time is None or when > cluster_monitor.update_time:
cluster_monitor.update_time = when
cluster_monitor.inject_sync_object(None, sync_type, version, msgpack.unpackb(latest_record.data))
for monitor in self.clusters.values():
log.info("Recovery: Cluster %s with update time %s" % (monitor.fsid, monitor.update_time))
monitor.start()
def start(self):
log.info("%s starting" % self.__class__.__name__)
self._rpc_thread.bind()
self._rpc_thread.start()
self._discovery_thread.start()
self._process_monitor.start()
self.persister.start()
self.eventer.start()
self._request_ticker.start()
self.servers.start()
return True
def join(self):
log.info("%s joining" % self.__class__.__name__)
self._rpc_thread.join()
self._discovery_thread.join()
self._process_monitor.join()
self.persister.join()
self.eventer.join()
self._request_ticker.join()
self.servers.join()
for monitor in self.clusters.values():
monitor.join()
def on_discovery(self, minion_id, heartbeat_data):
log.info("on_discovery: {0}/{1}".format(minion_id, heartbeat_data['fsid']))
cluster_monitor = ClusterMonitor(heartbeat_data['fsid'], heartbeat_data['name'],
self.persister, self.servers, self.eventer, self.requests)
self.clusters[heartbeat_data['fsid']] = cluster_monitor
# Run before passing on the heartbeat, because otherwise the
# syncs resulting from the heartbeat might not be received
# by the monitor.
cluster_monitor.start()
# Wait for ClusterMonitor to start accepting events before asking it
# to do anything
cluster_monitor.ready()
cluster_monitor.on_heartbeat(minion_id, heartbeat_data)
