def waitStoreResponses(self, txn_context, tryToResolveConflict):
result = []
append = result.append
resolved_oid_set = set()
update = resolved_oid_set.update
_handleConflicts = self._handleConflicts
queue = txn_context['queue']
conflict_serial_dict = txn_context['conflict_serial_dict']
pending = self.dispatcher.pending
_waitAnyTransactionMessage = self._waitAnyTransactionMessage
while pending(queue) or conflict_serial_dict:
# Note: handler data can be overwritten by _handleConflicts
# so we must set it for each iteration.
_waitAnyTransactionMessage(txn_context)
if conflict_serial_dict:
conflicts = _handleConflicts(txn_context, tryToResolveConflict)
if conflicts:
update(conflicts)
# Check for never-stored objects, and update result for all others
for oid, store_dict in \
txn_context['object_stored_counter_dict'].iteritems():
if not store_dict:
logging.error('tpc_store failed')
raise NEOStorageError('tpc_store failed')
elif oid in resolved_oid_set:
append((oid, ResolvedSerial))
return result
def _run(self):
"""Make sure that the status is sane and start a loop."""
if len(self.name) == 0:
raise RuntimeError, 'cluster name must be non-empty'
# Make a listening port
handler = identification.IdentificationHandler(self)
self.listening_conn = ListeningConnection(self, handler, self.server)
self.server = self.listening_conn.getAddress()
# Connect to a primary master node, verify data, and
# start the operation. This cycle will be executed permanently,
# until the user explicitly requests a shutdown.
self.operational = False
while True:
self.cluster_state = None
if self.master_node is None:
# look for the primary master
self.connectToPrimary()
self.checker = Checker(self)
self.replicator = Replicator(self)
self.tm = TransactionManager(self)
try:
self.initialize()
self.doOperation()
raise RuntimeError, 'should not reach here'
except StoppedOperation, msg:
logging.error('operation stopped: %s', msg)
except PrimaryFailure, msg:
logging.error('primary master is down: %s', msg)
def _connectToPrimaryNode(self):
"""
Lookup for the current primary master node
"""
logging.debug('connecting to primary master...')
self.start()
index = -1
ask = self._ask
handler = self.primary_bootstrap_handler
while 1:
# Get network connection to primary master
while 1:
if self.primary_master_node is not None:
# If I know a primary master node, pinpoint it.
self.trying_master_node = self.primary_master_node
self.primary_master_node = None
else:
# Otherwise, check one by one.
master_list = self.nm.getMasterList()
index = (index + 1) % len(master_list)
self.trying_master_node = master_list[index]
# Connect to master
conn = MTClientConnection(self,
self.notifications_handler,
node=self.trying_master_node,
dispatcher=self.dispatcher)
# Query for primary master node
if conn.getConnector() is None:
# This happens if a connection could not be established.
logging.error('Connection to master node %s failed',
self.trying_master_node)
continue
try:
ask(conn,
Packets.RequestIdentification(NodeTypes.CLIENT,
self.uuid, None,
self.name),
handler=handler)
except ConnectionClosed:
continue
# If we reached the primary master node, mark as connected
if self.primary_master_node is not None and \
self.primary_master_node is self.trying_master_node:
break
logging.info('Connected to %s', self.primary_master_node)
try:
# Request identification and required informations to be
# operational. Might raise ConnectionClosed so that the new
# primary can be looked-up again.
logging.info('Initializing from master')
ask(conn, Packets.AskNodeInformation(), handler=handler)
ask(conn, Packets.AskPartitionTable(), handler=handler)
ask(conn, Packets.AskLastTransaction(), handler=handler)
if self.pt.operational():
break
except ConnectionClosed:
logging.error('Connection to %s lost', self.trying_master_node)
self.primary_master_node = None
logging.info("Connected and ready")
return conn
def waitStoreResponses(self, txn_context, tryToResolveConflict):
result = []
append = result.append
resolved_oid_set = set()
update = resolved_oid_set.update
_handleConflicts = self._handleConflicts
queue = txn_context['queue']
conflict_serial_dict = txn_context['conflict_serial_dict']
pending = self.dispatcher.pending
_waitAnyTransactionMessage = self._waitAnyTransactionMessage
while pending(queue) or conflict_serial_dict:
# Note: handler data can be overwritten by _handleConflicts
# so we must set it for each iteration.
_waitAnyTransactionMessage(txn_context)
if conflict_serial_dict:
conflicts = _handleConflicts(txn_context,
tryToResolveConflict)
if conflicts:
update(conflicts)
# Check for never-stored objects, and update result for all others
for oid, store_dict in \
txn_context['object_stored_counter_dict'].iteritems():
if not store_dict:
logging.error('tpc_store failed')
raise NEOStorageError('tpc_store failed')
elif oid in resolved_oid_set:
append((oid, ResolvedSerial))
return result
def repair(self, weak_app, dry_run):
t = self._repairing
if t and t.is_alive():
logging.error('already repairing')
return
def repair():
l = threading.Lock()
l.acquire()
def finalize():
try:
if data_id_list and not dry_run:
self.commit()
logging.info("repair: deleted %s orphan records",
self._pruneData(data_id_list))
self.commit()
finally:
l.release()
try:
with self._duplicate() as db:
data_id_list = db.getOrphanList()
logging.info("repair: found %s records that may be orphan",
len(data_id_list))
weak_app().em.wakeup(finalize)
l.acquire()
finally:
del self._repairing
logging.info("repair: done")
t = self._repairing = threading.Thread(target=repair)
t.daemon = 1
t.start()
def _connectToPrimaryNode(self):
"""
Lookup for the current primary master node
"""
logging.debug('connecting to primary master...')
self.start()
index = -1
ask = self._ask
handler = self.primary_bootstrap_handler
while 1:
# Get network connection to primary master
while 1:
if self.primary_master_node is not None:
# If I know a primary master node, pinpoint it.
self.trying_master_node = self.primary_master_node
self.primary_master_node = None
else:
# Otherwise, check one by one.
master_list = self.nm.getMasterList()
index = (index + 1) % len(master_list)
self.trying_master_node = master_list[index]
# Connect to master
conn = MTClientConnection(self,
self.notifications_handler,
node=self.trying_master_node,
dispatcher=self.dispatcher)
# Query for primary master node
if conn.getConnector() is None:
# This happens if a connection could not be established.
logging.error('Connection to master node %s failed',
self.trying_master_node)
continue
try:
ask(conn, Packets.RequestIdentification(
NodeTypes.CLIENT, self.uuid, None, self.name),
handler=handler)
except ConnectionClosed:
continue
# If we reached the primary master node, mark as connected
if self.primary_master_node is not None and \
self.primary_master_node is self.trying_master_node:
break
logging.info('Connected to %s', self.primary_master_node)
try:
# Request identification and required informations to be
# operational. Might raise ConnectionClosed so that the new
# primary can be looked-up again.
logging.info('Initializing from master')
ask(conn, Packets.AskNodeInformation(), handler=handler)
ask(conn, Packets.AskPartitionTable(), handler=handler)
ask(conn, Packets.AskLastTransaction(), handler=handler)
if self.pt.operational():
break
except ConnectionClosed:
logging.error('Connection to %s lost', self.trying_master_node)
self.primary_master_node = None
logging.info("Connected and ready")
return conn
def _connectToPrimaryNode(self):
"""
Lookup for the current primary master node
"""
logging.debug('connecting to primary master...')
self.start()
index = -1
fail_count = 0
ask = self._ask
handler = self.primary_bootstrap_handler
while 1:
self.ignore_invalidations = True
# Get network connection to primary master
while fail_count < self.max_reconnection_to_master:
self.nm.reset()
if self.primary_master_node is not None:
# If I know a primary master node, pinpoint it.
node = self.primary_master_node
self.primary_master_node = None
else:
# Otherwise, check one by one.
master_list = self.nm.getMasterList()
index = (index + 1) % len(master_list)
node = master_list[index]
# Connect to master
conn = MTClientConnection(self,
self.notifications_handler,
node=node,
dispatcher=self.dispatcher)
p = Packets.RequestIdentification(NodeTypes.CLIENT, self.uuid,
None, self.name, None)
try:
ask(conn, p, handler=handler)
except ConnectionClosed:
fail_count += 1
else:
self.primary_master_node = node
break
else:
raise NEOPrimaryMasterLost(
"Too many connection failures to the primary master")
logging.info('Connected to %s', self.primary_master_node)
try:
# Request identification and required informations to be
# operational. Might raise ConnectionClosed so that the new
# primary can be looked-up again.
logging.info('Initializing from master')
ask(conn, Packets.AskPartitionTable(), handler=handler)
ask(conn, Packets.AskLastTransaction(), handler=handler)
if self.pt.operational():
break
except ConnectionClosed:
logging.error('Connection to %s lost', self.trying_master_node)
self.primary_master_node = None
fail_count += 1
logging.info("Connected and ready")
return conn
def askStorage(conn, packet):
tid, next_tid, compression, checksum, data, data_tid \
= self._askStorage(conn, packet)
if data or checksum != ZERO_HASH:
if checksum != makeChecksum(data):
logging.error('wrong checksum from %s for oid %s',
conn, dump(oid))
raise NEOStorageReadRetry(False)
return (decompress_list[compression](data),
tid, next_tid, data_tid)
raise NEOStorageCreationUndoneError(dump(oid))
def getDataTID(tid=None, before_tid=None):
tid, data_tid = self._getDataTID(oid, tid, before_tid)
current_tid = tid
while data_tid:
if data_tid < tid:
tid, data_tid = self._getDataTID(oid, data_tid)
if tid is not None:
continue
logging.error("Incorrect data serial for oid %s at tid %s",
oid, current_tid)
return current_tid, current_tid
return current_tid, tid
def getDataTID(tid=None, before_tid=None):
tid, data_tid = self._getDataTID(oid, tid, before_tid)
current_tid = tid
while data_tid:
if data_tid < tid:
tid, data_tid = self._getDataTID(oid, data_tid)
if tid is not None:
continue
logging.error("Incorrect data serial for oid %s at tid %s",
oid, current_tid)
return current_tid, current_tid
return current_tid, tid
def requestIdentification(self, conn, node_type, uuid, address, name):
self.checkClusterName(name)
# reject any incoming connections if not ready
if not self.app.ready:
raise NotReadyError
app = self.app
if uuid is None:
if node_type != NodeTypes.STORAGE:
raise ProtocolError('reject anonymous non-storage node')
handler = StorageOperationHandler(self.app)
conn.setHandler(handler)
else:
if uuid == app.uuid:
raise ProtocolError("uuid conflict or loopback connection")
node = app.nm.getByUUID(uuid)
# If this node is broken, reject it.
if node is not None and node.isBroken():
raise BrokenNodeDisallowedError
# choose the handler according to the node type
if node_type == NodeTypes.CLIENT:
handler = ClientOperationHandler
if node is None:
node = app.nm.createClient(uuid=uuid)
elif node.isConnected():
# This can happen if we haven't processed yet a notification
# from the master, telling us the existing node is not
# running anymore. If we accept the new client, we won't
# know what to do with this late notification.
raise NotReadyError('uuid conflict: retry later')
node.setRunning()
elif node_type == NodeTypes.STORAGE:
if node is None:
logging.error('reject an unknown storage node %s',
uuid_str(uuid))
raise NotReadyError
handler = StorageOperationHandler
else:
raise ProtocolError('reject non-client-or-storage node')
# apply the handler and set up the connection
handler = handler(self.app)
conn.setHandler(handler)
node.setConnection(conn, app.uuid < uuid)
# accept the identification and trigger an event
conn.answer(
Packets.AcceptIdentification(NodeTypes.STORAGE, uuid and app.uuid,
app.pt.getPartitions(),
app.pt.getReplicas(), uuid,
app.master_node.getAddress(), ()))
handler.connectionCompleted(conn)
def requestIdentification(self, conn, node_type,
uuid, address, name):
self.checkClusterName(name)
# reject any incoming connections if not ready
if not self.app.ready:
raise NotReadyError
app = self.app
if uuid is None:
if node_type != NodeTypes.STORAGE:
raise ProtocolError('reject anonymous non-storage node')
handler = StorageOperationHandler(self.app)
conn.setHandler(handler)
else:
if uuid == app.uuid:
raise ProtocolError("uuid conflict or loopback connection")
node = app.nm.getByUUID(uuid)
# If this node is broken, reject it.
if node is not None and node.isBroken():
raise BrokenNodeDisallowedError
# choose the handler according to the node type
if node_type == NodeTypes.CLIENT:
handler = ClientOperationHandler
if node is None:
node = app.nm.createClient(uuid=uuid)
elif node.isConnected():
# This can happen if we haven't processed yet a notification
# from the master, telling us the existing node is not
# running anymore. If we accept the new client, we won't
# know what to do with this late notification.
raise NotReadyError('uuid conflict: retry later')
node.setRunning()
elif node_type == NodeTypes.STORAGE:
if node is None:
logging.error('reject an unknown storage node %s',
uuid_str(uuid))
raise NotReadyError
handler = StorageOperationHandler
else:
raise ProtocolError('reject non-client-or-storage node')
# apply the handler and set up the connection
handler = handler(self.app)
conn.setHandler(handler)
node.setConnection(conn, app.uuid < uuid)
# accept the identification and trigger an event
conn.answer(Packets.AcceptIdentification(NodeTypes.STORAGE, uuid and
app.uuid, app.pt.getPartitions(), app.pt.getReplicas(), uuid,
app.master_node.getAddress(), ()))
handler.connectionCompleted(conn)
def _initNodeConnection(self, node):
"""Init a connection to a given storage node."""
app = self.app
logging.debug('trying to connect to %s - %s', node, node.getState())
conn = MTClientConnection(app, app.storage_event_handler, node,
dispatcher=app.dispatcher)
p = Packets.RequestIdentification(NodeTypes.CLIENT,
app.uuid, None, app.name)
try:
app._ask(conn, p, handler=app.storage_bootstrap_handler)
except ConnectionClosed:
logging.error('Connection to %r failed', node)
except NodeNotReady:
logging.info('%r not ready', node)
else:
logging.info('Connected %r', node)
return conn
self.notifyFailure(node)
def _loadFromStorage(self, oid, at_tid, before_tid):
packet = Packets.AskObject(oid, at_tid, before_tid)
for node, conn in self.cp.iterateForObject(oid, readable=True):
try:
tid, next_tid, compression, checksum, data, data_tid \
= self._askStorage(conn, packet)
except ConnectionClosed:
continue
if data or checksum != ZERO_HASH:
if checksum != makeChecksum(data):
logging.error('wrong checksum from %s for oid %s', conn,
dump(oid))
continue
return (decompress(data) if compression else data, tid,
next_tid, data_tid)
raise NEOStorageCreationUndoneError(dump(oid))
raise NEOStorageError("storage down or corrupted data")
def _loadFromStorage(self, oid, at_tid, before_tid):
packet = Packets.AskObject(oid, at_tid, before_tid)
for node, conn in self.cp.iterateForObject(oid, readable=True):
try:
tid, next_tid, compression, checksum, data, data_tid \
= self._askStorage(conn, packet)
except ConnectionClosed:
continue
if data or checksum != ZERO_HASH:
if checksum != makeChecksum(data):
logging.error('wrong checksum from %s for oid %s',
conn, dump(oid))
continue
return (decompress(data) if compression else data,
tid, next_tid, data_tid)
raise NEOStorageCreationUndoneError(dump(oid))
raise NEOStorageError("storage down or corrupted data")
def _run(self):
"""Make sure that the status is sane and start a loop."""
if len(self.name) == 0:
raise RuntimeError, 'cluster name must be non-empty'
# Make a listening port.
handler = AdminEventHandler(self)
self.listening_conn = ListeningConnection(self, handler, self.server)
while self.cluster_state != ClusterStates.STOPPING:
self.connectToPrimary()
try:
while True:
self.em.poll(1)
except PrimaryFailure:
logging.error('primary master is down')
self.listening_conn.close()
while not self.em.isIdle():
self.em.poll(1)
def _run(self):
"""Make sure that the status is sane and start a loop."""
if len(self.name) == 0:
raise RuntimeError, 'cluster name must be non-empty'
# Make a listening port.
handler = AdminEventHandler(self)
self.listening_conn = ListeningConnection(self, handler, self.server)
while self.cluster_state != ClusterStates.STOPPING:
self.connectToPrimary()
try:
while True:
self.em.poll(1)
except PrimaryFailure:
logging.error('primary master is down')
self.listening_conn.close()
while not self.em.isIdle():
self.em.poll(1)
def _initNodeConnection(self, node):
"""Init a connection to a given storage node."""
app = self.app
if app.master_conn is None:
raise NEOPrimaryMasterLost
conn = MTClientConnection(app, app.storage_event_handler, node,
dispatcher=app.dispatcher)
p = Packets.RequestIdentification(NodeTypes.CLIENT,
app.uuid, None, app.name, app.id_timestamp)
try:
app._ask(conn, p, handler=app.storage_bootstrap_handler)
except ConnectionClosed:
logging.error('Connection to %r failed', node)
except NodeNotReady:
logging.info('%r not ready', node)
else:
logging.info('Connected %r', node)
return conn
self.node_failure_dict[node.getUUID()] = time.time() + MAX_FAILURE_AGE
def tpc_vote(self, transaction, tryToResolveConflict):
"""Store current transaction."""
txn_context = self._txn_container.get(transaction)
result = self.waitStoreResponses(txn_context, tryToResolveConflict)
ttid = txn_context['ttid']
# Store data on each node
assert not txn_context['data_dict'], txn_context
packet = Packets.AskStoreTransaction(ttid, str(transaction.user),
str(transaction.description),
dumps(transaction._extension),
txn_context['cache_dict'])
queue = txn_context['queue']
trans_nodes = []
for node, conn in self.cp.iterateForObject(ttid):
logging.debug("voting transaction %s on %s", dump(ttid),
dump(conn.getUUID()))
try:
conn.ask(packet, queue=queue)
except ConnectionClosed:
continue
trans_nodes.append(node)
# check at least one storage node accepted
if trans_nodes:
involved_nodes = txn_context['involved_nodes']
packet = Packets.AskVoteTransaction(ttid)
for node in involved_nodes.difference(trans_nodes):
conn = self.cp.getConnForNode(node)
if conn is not None:
try:
conn.ask(packet, queue=queue)
except ConnectionClosed:
pass
involved_nodes.update(trans_nodes)
self.waitResponses(queue)
txn_context['voted'] = None
# We must not go further if connection to master was lost since
# tpc_begin, to lower the probability of failing during tpc_finish.
if 'error' in txn_context:
raise NEOStorageError(txn_context['error'])
return result
logging.error('tpc_vote failed')
raise NEOStorageError('tpc_vote failed')
def tpc_vote(self, transaction, tryToResolveConflict):
"""Store current transaction."""
txn_context = self._txn_container.get(transaction)
result = self.waitStoreResponses(txn_context, tryToResolveConflict)
ttid = txn_context['ttid']
# Store data on each node
assert not txn_context['data_dict'], txn_context
packet = Packets.AskStoreTransaction(ttid, str(transaction.user),
str(transaction.description), dumps(transaction._extension),
txn_context['cache_dict'])
queue = txn_context['queue']
trans_nodes = []
for node, conn in self.cp.iterateForObject(ttid):
logging.debug("voting transaction %s on %s", dump(ttid),
dump(conn.getUUID()))
try:
conn.ask(packet, queue=queue)
except ConnectionClosed:
continue
trans_nodes.append(node)
# check at least one storage node accepted
if trans_nodes:
involved_nodes = txn_context['involved_nodes']
packet = Packets.AskVoteTransaction(ttid)
for node in involved_nodes.difference(trans_nodes):
conn = self.cp.getConnForNode(node)
if conn is not None:
try:
conn.ask(packet, queue=queue)
except ConnectionClosed:
pass
involved_nodes.update(trans_nodes)
self.waitResponses(queue)
txn_context['voted'] = None
# We must not go further if connection to master was lost since
# tpc_begin, to lower the probability of failing during tpc_finish.
if 'error' in txn_context:
raise NEOStorageError(txn_context['error'])
return result
logging.error('tpc_vote failed')
raise NEOStorageError('tpc_vote failed')
def _initNodeConnection(self, node):
"""Init a connection to a given storage node."""
app = self.app
logging.debug('trying to connect to %s - %s', node, node.getState())
conn = MTClientConnection(app,
app.storage_event_handler,
node,
dispatcher=app.dispatcher)
p = Packets.RequestIdentification(NodeTypes.CLIENT, app.uuid, None,
app.name)
try:
app._ask(conn, p, handler=app.storage_bootstrap_handler)
except ConnectionClosed:
logging.error('Connection to %r failed', node)
except NodeNotReady:
logging.info('%r not ready', node)
else:
logging.info('Connected %r', node)
return conn
self.notifyFailure(node)
def _connectToStorageNode(self, node):
if self.master_conn is None:
raise NEOPrimaryMasterLost
conn = MTClientConnection(self, self.storage_event_handler, node,
dispatcher=self.dispatcher)
p = Packets.RequestIdentification(NodeTypes.CLIENT,
self.uuid, None, self.name, self.id_timestamp, {})
try:
self._ask(conn, p, handler=self.storage_bootstrap_handler)
except ConnectionClosed:
logging.error('Connection to %r failed', node)
except NodeNotReady:
logging.info('%r not ready', node)
else:
logging.info('Connected %r', node)
# Make sure this node will be considered for the next reads
# even if there was a previous recent failure.
self._node_failure_dict.pop(node.getUUID(), None)
return conn
self._node_failure_dict[node.getUUID()] = time.time() + MAX_FAILURE_AGE
def provideService(self):
logging.info('provide backup')
poll = self.em.poll
app = self.app
pt = app.pt
while True:
app.changeClusterState(ClusterStates.STARTING_BACKUP)
bootstrap = BootstrapManager(self,
NodeTypes.CLIENT,
backup=app.name)
# {offset -> node}
self.primary_partition_dict = {}
# [[tid]]
self.tid_list = tuple([] for _ in xrange(pt.getPartitions()))
try:
while True:
for node in pt.getNodeSet(readable=True):
if not app.isStorageReady(node.getUUID()):
break
else:
break
poll(1)
node, conn = bootstrap.getPrimaryConnection()
try:
app.changeClusterState(ClusterStates.BACKINGUP)
del bootstrap, node
self.ignore_invalidations = True
conn.setHandler(BackupHandler(self))
conn.ask(Packets.AskLastTransaction())
# debug variable to log how big 'tid_list' can be.
self.debug_tid_count = 0
while True:
poll(1)
except PrimaryFailure, msg:
logging.error('upstream master is down: %s', msg)
finally:
app.backup_tid = pt.getBackupTid()
try:
conn.close()
except PrimaryFailure:
pass
try:
del self.pt
except AttributeError:
pass
for node in app.nm.getClientList(True):
node.getConnection().close()
except StateChangedException, e:
if e.args[0] != ClusterStates.STOPPING_BACKUP:
raise
app.changeClusterState(*e.args)
tid = app.backup_tid
# Wait for non-primary partitions to catch up,
# so that all UP_TO_DATE cells are really UP_TO_DATE.
# XXX: Another possibility could be to outdate such cells, and
# they would be quickly updated at the beginning of the
# RUNNING phase. This may simplify code.
# Any unfinished replication from upstream will be truncated.
while pt.getBackupTid(min) < tid:
poll(1)
last_tid = app.getLastTransaction()
handler = EventHandler(app)
if tid < last_tid:
assert tid != ZERO_TID
logging.warning("Truncating at %s (last_tid was %s)",
dump(app.backup_tid), dump(last_tid))
else:
# We will do a dummy truncation, just to leave backup mode,
# so it's fine to start automatically if there's any
# missing storage.
# XXX: Consider using another method to leave backup mode,
# at least when there's nothing to truncate. Because
# in case of StoppedOperation during VERIFYING state,
# this flag will be wrongly set to False.
app._startup_allowed = True
# If any error happened before reaching this line, we'd go back
# to backup mode, which is the right mode to recover.
del app.backup_tid
# Now back to RECOVERY...
return tid
def electPrimary(self):
"""Elect a primary master node.
The difficulty is that a master node must accept connections from
others while attempting to connect to other master nodes at the
same time. Note that storage nodes and client nodes may connect
to self as well as master nodes."""
logging.info('begin the election of a primary master')
client_handler = election.ClientElectionHandler(self)
self.unconnected_master_node_set.clear()
self.negotiating_master_node_set.clear()
self.master_address_dict.clear()
self.listening_conn.setHandler(election.ServerElectionHandler(self))
getByAddress = self.nm.getByAddress
while True:
# handle new connected masters
for node in self.nm.getMasterList():
node.setUnknown()
self.unconnected_master_node_set.add(node.getAddress())
# start the election process
self.primary = None
self.primary_master_node = None
try:
while (self.unconnected_master_node_set or
self.negotiating_master_node_set):
for addr in self.unconnected_master_node_set:
self.negotiating_master_node_set.add(addr)
ClientConnection(self, client_handler,
# XXX: Ugly, but the whole election code will be
# replaced soon
getByAddress(addr))
self.unconnected_master_node_set.clear()
self.em.poll(1)
except ElectionFailure, m:
# something goes wrong, clean then restart
logging.error('election failed: %s', m)
# Ask all connected nodes to reelect a single primary master.
for conn in self.em.getClientList():
conn.notify(Packets.ReelectPrimary())
conn.abort()
# Wait until the connections are closed.
self.primary = None
self.primary_master_node = None
# XXX: Since poll does not wake up anymore every second,
# the following time condition should be reviewed.
# See also playSecondaryRole.
t = time() + 10
while self.em.getClientList() and time() < t:
try:
self.em.poll(1)
except ElectionFailure:
pass
# Close all connections.
for conn in self.em.getClientList() + self.em.getServerList():
conn.close()
else:
# election succeed, stop the process
self.primary = self.primary is None
break
def _nextPartition(self):
app = self.app
def connect(node, uuid=app.uuid, name=app.name):
if node.getUUID() == app.uuid:
return
if node.isConnected(connecting=True):
conn = node.getConnection()
conn.asClient()
else:
conn = ClientConnection(app, StorageOperationHandler(app), node)
conn.ask(Packets.RequestIdentification(
NodeTypes.STORAGE, uuid, app.server, name))
self.conn_dict[conn] = node.isIdentified()
conn_set = set(self.conn_dict)
conn_set.discard(None)
try:
self.conn_dict.clear()
while True:
try:
partition, (name, source), min_tid, max_tid = \
self.queue.popleft()
except IndexError:
return
cell = app.pt.getCell(partition, app.uuid)
if cell is None or cell.isOutOfDate():
msg = "discarded or out-of-date"
else:
try:
for cell in app.pt.getCellList(partition):
# XXX: Ignore corrupted cells for the moment
# because we're still unable to fix them
# (see also AdministrationHandler of master)
if cell.isReadable(): #if not cell.isOutOfDate():
connect(cell.getNode())
if source:
node = app.nm.getByAddress(source)
if name:
source = app.nm.createStorage(address=source) \
if node is None else node
connect(source, None, name)
elif (node.getUUID() == app.uuid or
node.isConnected(connecting=True) and
node.getConnection() in self.conn_dict):
source = node
else:
msg = "unavailable source"
if self.conn_dict:
break
msg = "no replica"
except ConnectionClosed:
msg = "connection closed"
finally:
conn_set.update(self.conn_dict)
self.conn_dict.clear()
logging.error("Failed to start checking partition %u (%s)",
partition, msg)
conn_set.difference_update(self.conn_dict)
finally:
for conn in conn_set:
app.closeClient(conn)
logging.debug("start checking partition %u from %s to %s",
partition, dump(min_tid), dump(max_tid))
self.min_tid = self.next_tid = min_tid
self.max_tid = max_tid
self.next_oid = None
self.partition = partition
self.source = source
def start():
if app.tm.isLockedTid(max_tid):
app.queueEvent(start)
return
args = partition, CHECK_COUNT, min_tid, max_tid
p = Packets.AskCheckTIDRange(*args)
for conn, identified in self.conn_dict.items():
self.conn_dict[conn] = conn.ask(p) if identified else None
self.conn_dict[None] = app.dm.checkTIDRange(*args)
start()
def _notify(self, ask_ids=True):
if ask_ids:
self.askLastIds(self.master_conn)
self.notifying = notifying = {None}
for name, monitor in self.backup_dict.iteritems():
if monitor.operational:
monitor.askLastIds(monitor.conn)
notifying.add(name)
if self.notifying or self.cluster_state is None is not self.master_conn:
return
severity = [], [], []
my_severity = self.severity
severity[my_severity].append(self.name)
changed = set()
if self.monitor_changed:
self.monitor_changed = False
changed.add(self.name)
if self.master_conn is None:
body = NOT_CONNECTED_MESSAGE
else:
upstream, body = self.formatSummary()
body = [body]
for name, backup in self.backup_dict.iteritems():
body += '', name, ' ' + backup.formatSummary(upstream)[1]
severity[backup.severity or backup.lagging].append(name)
if backup.monitor_changed:
backup.monitor_changed = False
changed.add(name)
body = '\n'.join(body)
if changed or self.smtp_retry < time():
logging.debug('monitor notification')
email_list = self.email_list
while email_list: # not a loop
msg = MIMEText(body + (self.smtp_exc or ''))
msg['Date'] = formatdate()
clusters, x = severity[1:]
while 1:
if x:
clusters = clusters + x
x = 'PROBLEM'
elif clusters:
x = 'WARNING'
else:
x = 'OK'
break
clusters = changed.intersection(clusters)
if clusters:
x += ' (%s)' % ', '.join(sorted(clusters))
break
msg['Subject'] = 'NEO monitoring: ' + x
msg['From'] = self.email_from
msg['To'] = ', '.join(email_list)
s = self.SMTP()
try:
s.connect(self.smtp_host)
if self.smtp_tls:
s.starttls()
if self.smtp_login:
s.login(*self.smtp_login)
s.sendmail(None, email_list, msg.as_string())
except Exception:
x = format_exc()
logging.error(x)
if changed or not self.smtp_exc:
self.smtp_exc = (
"\n\nA notification could not be sent at %s:\n\n%s"
% (msg['Date'], x))
retry = self.smtp_retry = time() + 600
else:
self.smtp_exc = None
self.smtp_retry = INF
if not (self.operational and any(
monitor.operational
for monitor in self.backup_dict.itervalues())):
break
retry = time() + 600
finally:
s.close()
self.em.setTimeout(retry, self._notify)
break
neoctl = self.asking_monitor_information
if neoctl:
del severity[my_severity][0]
if self.smtp_exc:
my_severity = 2
body += self.smtp_exc
severity[1].sort()
severity[2].sort()
severity[my_severity].insert(0, None)
p = Packets.AnswerMonitorInformation(severity[1], severity[2],
body)
for conn, msg_id in neoctl:
try:
conn.send(p, msg_id)
except ConnectionClosed:
pass
del self.asking_monitor_information[:]
def _nextPartition(self):
app = self.app
def connect(node, uuid=app.uuid, name=app.name):
if node.getUUID() == app.uuid:
return
if node.isConnected(connecting=True):
conn = node.getConnection()
conn.asClient()
else:
conn = ClientConnection(app, StorageOperationHandler(app),
node)
conn.ask(
Packets.RequestIdentification(NodeTypes.STORAGE, uuid,
app.server, name,
app.id_timestamp, {}))
self.conn_dict[conn] = node.isIdentified()
conn_set = set(self.conn_dict)
conn_set.discard(None)
try:
self.conn_dict.clear()
while True:
try:
partition, (name, source), min_tid, max_tid = \
self.queue.popleft()
except IndexError:
return
cell = app.pt.getCell(partition, app.uuid)
if cell is None or cell.isOutOfDate():
msg = "discarded or out-of-date"
else:
try:
for cell in app.pt.getCellList(partition):
# XXX: Ignore corrupted cells for the moment
# because we're still unable to fix them
# (see also AdministrationHandler of master)
if cell.isReadable(): #if not cell.isOutOfDate():
connect(cell.getNode())
if source:
node = app.nm.getByAddress(source)
if name:
source = app.nm.createStorage(address=source) \
if node is None else node
connect(source, None, name)
elif (node.getUUID() == app.uuid
or node.isConnected(connecting=True)
and node.getConnection() in self.conn_dict):
source = node
else:
msg = "unavailable source"
if self.conn_dict:
break
msg = "no replica"
except ConnectionClosed:
msg = "connection closed"
finally:
conn_set.update(self.conn_dict)
self.conn_dict.clear()
logging.error("Failed to start checking partition %u (%s)",
partition, msg)
conn_set.difference_update(self.conn_dict)
finally:
for conn in conn_set:
app.closeClient(conn)
logging.debug("start checking partition %u from %s to %s", partition,
dump(min_tid), dump(max_tid))
self.min_tid = self.next_tid = min_tid
self.max_tid = max_tid
self.next_oid = None
self.partition = partition
self.source = source
def start():
if app.tm.isLockedTid(max_tid):
app.tm.read_queue.queueEvent(start)
return
args = partition, CHECK_COUNT, min_tid, max_tid
p = Packets.AskCheckTIDRange(*args)
for conn, identified in self.conn_dict.items():
self.conn_dict[conn] = conn.ask(p) if identified else None
self.conn_dict[None] = app.dm.checkTIDRange(*args)
start()
def electPrimary(self):
"""Elect a primary master node.
The difficulty is that a master node must accept connections from
others while attempting to connect to other master nodes at the
same time. Note that storage nodes and client nodes may connect
to self as well as master nodes."""
logging.info('begin the election of a primary master')
client_handler = election.ClientElectionHandler(self)
self.unconnected_master_node_set.clear()
self.negotiating_master_node_set.clear()
self.master_address_dict.clear()
self.listening_conn.setHandler(election.ServerElectionHandler(self))
getByAddress = self.nm.getByAddress
while True:
# handle new connected masters
for node in self.nm.getMasterList():
node.setUnknown()
self.unconnected_master_node_set.add(node.getAddress())
# start the election process
self.primary = None
self.primary_master_node = None
try:
while (self.unconnected_master_node_set
or self.negotiating_master_node_set):
for addr in self.unconnected_master_node_set:
self.negotiating_master_node_set.add(addr)
ClientConnection(
self,
client_handler,
# XXX: Ugly, but the whole election code will be
# replaced soon
getByAddress(addr))
self.unconnected_master_node_set.clear()
self.em.poll(1)
except ElectionFailure, m:
# something goes wrong, clean then restart
logging.error('election failed: %s', m)
# Ask all connected nodes to reelect a single primary master.
for conn in self.em.getClientList():
conn.notify(Packets.ReelectPrimary())
conn.abort()
# Wait until the connections are closed.
self.primary = None
self.primary_master_node = None
# XXX: Since poll does not wake up anymore every second,
# the following time condition should be reviewed.
# See also playSecondaryRole.
t = time() + 10
while self.em.getClientList() and time() < t:
try:
self.em.poll(1)
except ElectionFailure:
pass
# Close all connections.
for conn in self.em.getClientList() + self.em.getServerList():
conn.close()
else:
# election succeed, stop the process
self.primary = self.primary is None
break
def __init__(self, msg=None):
logging.error(msg)
AssertionError.__init__(self, msg)
def __init__(self, msg=None):
logging.error(msg)
AssertionError.__init__(self, msg)
def _run(self):
stdscr = self.stdscr
r, w = os.pipe()
l = threading.Lock()
stdscr.nodelay(1)
input_queue = deque()
def input_read():
x = []
while 1:
c = stdscr.getch()
if c < 0:
if x:
input_queue.append(x)
return input_queue
x.append(c)
def input_thread():
try:
poll = select.poll()
poll.register(0, select.POLLIN)
poll.register(r, select.POLLIN)
while 1:
for fd, _ in poll.poll():
if fd:
return
with l:
empty = not input_queue
if input_read() and empty:
self.em.wakeup()
finally:
os.close(r)
t = threading.Thread(target=input_thread)
t.deamon = True
wait = None
try:
t.start()
self.startCluster()
self.refresh('stress', False)
while 1:
self.failing.clear()
try:
self.connectToPrimary()
self.askLastIDs()
while 1:
self.em.poll(1)
with l:
if input_read():
for x in input_queue:
try:
x, = x
except ValueError:
continue
if x == curses.KEY_RESIZE:
self.refresh()
elif x == curses.KEY_F1:
self.stress()
else:
try:
x = chr(x)
except ValueError:
continue
if x == 'q':
return
input_queue.clear()
except PrimaryFailure:
logging.error('primary master is down')
if self.cluster_state == ClusterStates.STOPPING:
break
self.primaryFailure()
finally:
if self._stress:
self.stress()
wait = time.time()
finally:
os.write(w, '\0')
os.close(w)
t.join()
self.stopCluster(wait)
def provideService(self):
logging.info('provide backup')
poll = self.em.poll
app = self.app
pt = app.pt
while True:
app.changeClusterState(ClusterStates.STARTING_BACKUP)
bootstrap = BootstrapManager(self, self.name, NodeTypes.CLIENT)
# {offset -> node}
self.primary_partition_dict = {}
# [[tid]]
self.tid_list = tuple([] for _ in xrange(pt.getPartitions()))
try:
while True:
for node in pt.getNodeSet(readable=True):
if not app.isStorageReady(node.getUUID()):
break
else:
break
poll(1)
node, conn, uuid, num_partitions, num_replicas = \
bootstrap.getPrimaryConnection()
try:
app.changeClusterState(ClusterStates.BACKINGUP)
del bootstrap, node
if num_partitions != pt.getPartitions():
raise RuntimeError("inconsistent number of partitions")
self.pt = PartitionTable(num_partitions, num_replicas)
conn.setHandler(BackupHandler(self))
conn.ask(Packets.AskNodeInformation())
conn.ask(Packets.AskPartitionTable())
conn.ask(Packets.AskLastTransaction())
# debug variable to log how big 'tid_list' can be.
self.debug_tid_count = 0
while True:
poll(1)
except PrimaryFailure, msg:
logging.error('upstream master is down: %s', msg)
finally:
app.backup_tid = pt.getBackupTid()
try:
conn.close()
except PrimaryFailure:
pass
try:
del self.pt
except AttributeError:
pass
except StateChangedException, e:
if e.args[0] != ClusterStates.STOPPING_BACKUP:
raise
app.changeClusterState(*e.args)
tid = app.backup_tid
# Wait for non-primary partitions to catch up,
# so that all UP_TO_DATE cells are really UP_TO_DATE.
# XXX: Another possibility could be to outdate such cells, and
# they would be quickly updated at the beginning of the
# RUNNING phase. This may simplify code.
# Any unfinished replication from upstream will be truncated.
while pt.getBackupTid(min) < tid:
poll(1)
last_tid = app.getLastTransaction()
handler = EventHandler(app)
if tid < last_tid:
assert tid != ZERO_TID
logging.warning("Truncating at %s (last_tid was %s)",
dump(app.backup_tid), dump(last_tid))
else:
# We will do a dummy truncation, just to leave backup mode,
# so it's fine to start automatically if there's any
# missing storage.
# XXX: Consider using another method to leave backup mode,
# at least when there's nothing to truncate. Because
# in case of StoppedOperation during VERIFYING state,
# this flag will be wrongly set to False.
app._startup_allowed = True
# If any error happened before reaching this line, we'd go back
# to backup mode, which is the right mode to recover.
del app.backup_tid
# Now back to RECOVERY...
return tid
