Skip to content


Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
Fetching contributors…

Cannot retrieve contributors at this time

1104 lines (925 sloc) 43.879 kb
# spread requests among connections to multiple nodes in a Cassandra cluster
Quick start:
>>> my_seed_nodes = ['', '', '']
>>> mypool = CassandraClusterPool(my_seed_nodes, keyspace='MyKeyspace', pool_size=10)
>>> mypool.startService()
>>> mypool.get('Key12345', 'SomeCF')
<Deferred at 0x1b2b248>
CassandraClusterPool will respond to all the methods on CassandraClient, but
if you prefer to have separate CassandraClient instances, set your pool object
as their manager.
Some of the most useful additional methods on CassandraClusterPool:
adjustPoolSize(newsize) # change the size of the connection pool,
# without interrupting ongoing requests
addNode((address, port)) # manually add another node to the pool.
# Normally this shouldn't be necessary; once
# the pool can connect to one or more of your
# your seed nodes, it can inspect the ring
# and find the rest of the nodes.
removeNode((address, port)) # manually remove a node from the pool. It will
# be re-added later if it shows up in the
# ring with a subsequent connection, though.
set_keyspace(ksname) # change the keyspace used for future requests
# on this pool
* check cluster name on connecting to each new host, to make sure it's
actually in the same cluster
* take node error/connection history into account with add_connection_score
* remove nodes that have been missing or unconnectable for too long
* when seed node list is shofter than requested pool size, don't try to
fill the pool completely until after a seed node is contacted and an
initial live-node list collected
import sys
import random
import socket
from time import time
from itertools import izip, groupby
from warnings import warn
from twisted.application import service
from twisted.internet import defer, protocol, error
from twisted.python import failure, log
from thrift import Thrift
from thrift.transport import TTwisted, TTransport
from thrift.protocol import TBinaryProtocol
from telephus.protocol import (ManagedThriftRequest, ClientBusy,
from telephus.cassandra.c08 import Cassandra as Cassandra08
from telephus.cassandra.ttypes import *
from telephus.client import CassandraClient
from telephus.translate import (thrift_api_ver_to_cassandra_ver, translateArgs,
noop = lambda *a, **kw: None
class NoKeyspacesAvailable(UserWarning):
Indicates CassandraClusterPool could not collect information about the
cluster ring, in order to automatically add nodes to the pool.
When Cassandra's thrift interface allows specifying null for describe_ring
(like the underlying java interface already does), we can remove this.
class NoNodesAvailable(Exception):
Indicates there are nodes to which we are allowed to make another immediate
connection. The argument to this exception should be the expected number
of seconds before a node /will/ be available.
This should be handled internally; user code is not expected to see or
handle this type of exception.
def lame_log_insufficient_nodes(poolsize, pooltarget, pending_reqs, waittime):
msg = '(No candidate nodes to expand pool to target size %d from %d;' \
' there are %d pending requests.' % (pooltarget, poolsize, pending_reqs)
if waittime is None:
msg += ')'
msg += ' Expected candidate node retry in %.1f seconds.)' % waittime
class CassandraPoolParticipantClient(TTwisted.ThriftClientProtocol):
thriftFactory = TBinaryProtocol.TBinaryProtocolAcceleratedFactory
def __init__(self):
TTwisted.ThriftClientProtocol.__init__(self, Cassandra08.Client,
def connectionMade(self):
def connectionLost(self, reason):
# the TTwisted version of this call does not account for the
# possibility of other things happening during the errback.
tex = TTransport.TTransportException(
message='Connection closed (%s)' % reason)
while self.client._reqs:
k = iter(self.client._reqs).next()
v = self.client._reqs.pop(k)
del self.client._reqs
del self.client
class CassandraPoolReconnectorFactory(protocol.ClientFactory):
protocol = CassandraPoolParticipantClient
connector = None
last_error = None
noisy = False
# store the keyspace this connection is set to. we will take thrift
# requests along with the keyspace in which they expect to be made, and
# change keyspaces if necessary. this is done this way to avoid having
# another layer of queueing for requests in this class (in addition to the
# queue in CassandraClusterPool), or special logic here to pass on
# set_keyspace calls from the service at the right time (so already-queued
# requests still get made in their right keyspaces).
keyspace = None
def __init__(self, node, service, api_version=None):
self.node = node
# if self.service is None, don't bother doing anything. nobody loves us.
self.service = service
self.my_proto = None
self.job_d = self.jobphase = None
self.api_version = api_version
def clientConnectionMade(self, proto):
assert self.my_proto is None
assert self.jobphase is None, 'jobphase=%r' % (self.jobphase,)
if self.service is None:
self.my_proto = proto
def clientConnectionFailed(self, connector, reason):
assert self.my_proto is None
assert self.jobphase is None, 'jobphase=%r' % (self.jobphase,)
self.my_proto = None
if self.service is not None:
self.connector = connector
self.service.client_conn_failed(reason, self)
def clientConnectionLost(self, connector, reason):
p = self.my_proto
self.my_proto = None
if p is not None and self.service is not None:
self.connector = connector
self.service.client_conn_lost(self, reason)
def stopFactory(self):
# idempotent
if self.connector:
except error.NotConnectingError:
self.connector = None
p = self.my_proto
self.my_proto = None
if p is not None and p.transport is not None:
def isConnecting(self):
if self.connector is None:
if self.my_proto is None:
# initial connection attempt
return True
# initial connection succeeded and hasn't died
return False
return self.connector.state == 'connecting'
def retry(self):
Retry this factory's connection. It is assumed that a previous
connection was attempted and failed- either before or after a
successful connection.
if self.connector is None:
raise ValueError("No connector to retry")
if self.service is None:
def prep_connection(self, creds=None, keyspace=None):
Do login and set_keyspace tasks as necessary, and also check this
node's idea of the Cassandra ring. Expects that our connection is
Return a Deferred that will fire with the ring information, or be
errbacked if something goes wrong.
d = self.my_describe_version()
def check_version(thrift_ver):
cassver = thrift_api_ver_to_cassandra_ver(thrift_ver)
if self.api_version is None:
self.api_version = cassver
elif self.api_version != cassver:
raise APIMismatch("%s is exposing thrift protocol version %s -> "
"Cassandra version %s, but %s was expected"
% (self.node, thrift_ver, cassver, self.api_version))
if creds is not None:
d.addCallback(lambda _: self.my_login(creds))
if keyspace is not None:
d.addCallback(lambda _: self.my_set_keyspace(keyspace))
d.addCallback(lambda _: self.my_describe_ring(keyspace))
return d
# The following my_* methods are for internal use, to facilitate the
# management of the pool and the queries we get. The user should make
# use of the methods on CassandraClient.
def my_login(self, creds):
return self.execute(
ManagedThriftRequest('login', AuthenticationRequest(credentials=creds))
def my_set_keyspace(self, keyspace):
return self.execute(ManagedThriftRequest('set_keyspace', keyspace))
def my_describe_ring(self, keyspace=None):
if keyspace is None or keyspace == 'system':
d = self.my_pick_non_system_keyspace()
d = defer.succeed(keyspace)
d.addCallback(lambda k: self.execute(ManagedThriftRequest('describe_ring', k)))
def suppress_no_keyspaces_error(f):
return ()
return d
def my_describe_version(self):
return self.execute(ManagedThriftRequest('describe_version'))
def my_describe_keyspaces(self):
return self.execute(ManagedThriftRequest('describe_keyspaces'))
def my_pick_non_system_keyspace(self):
Find a keyspace in the cluster which is not 'system', for the purpose
of getting a valid ring view. Can't use 'system' or null.
d = self.my_describe_keyspaces()
def pick_non_system(klist):
for k in klist:
if != 'system':
err = NoKeyspacesAvailable("Can't gather information about the "
"Cassandra ring; no non-system "
"keyspaces available")
raise err
return d
def store_successful_keyspace_set(self, val, ksname):
self.keyspace = ksname
return val
def execute(self, req, keyspace=None):
if self.my_proto is None:
return defer.errback(error.ConnectionClosed(
'Lost connection before %s request could be made'
% (req.method,)))
method = getattr(self.my_proto.client, req.method, None)
if method is None:
raise InvalidThriftRequest("don't understand %s request" % req.method)
d = defer.succeed(0)
if req.method == 'set_keyspace':
newksname = req.args[0]
d.addCallback(lambda _: method(newksname))
d.addCallback(self.store_successful_keyspace_set, newksname)
if keyspace is not None and keyspace != self.keyspace:
d.addCallback(lambda _: self.my_set_keyspace(keyspace))
args = translateArgs(req, self.api_version)
d.addCallback(lambda _: method(*args))
d.addCallback(lambda results: postProcess(results, req.method))
return d
def clear_job(self, x):
self.logstate('clear job %s (result %r)' % (self.jobphase, x))
self.jobphase = None
self.job_d = None
return x
def job(self, _name, _func, *args, **kw):
self.logstate('start job %s' % _name)
if self.jobphase is not None:
raise ClientBusy('Tried to start job phase %s while in %s'
% (_name, self.jobphase))
self.jobphase = _name
d = defer.maybeDeferred(_func, *args, **kw)
self.job_d = d
return d
def process_request_error(self, err, req, keyspace, req_d, retries):
self.logstate('process_request_error: %s, retries=%d' % (err, retries))
self.last_error = err
if retries > 0 and self.service is not None \
and err.check(*self.service.retryables):
self.logstate('-- resubmit --')
assert self.jobphase is None, \
'Factory might retry its own fatal error'
self.service.resubmit(req, keyspace, req_d, retries - 1)
self.logstate('-- giving up [retries=%d service=%s err=%s] --'
% (retries, self.service, err.value))
def work_on_request(self, reqtuple):
req, keyspace, req_d, retries = reqtuple
if req_d.called:
# cancelled while pending in the queue
d = self.job('pending_request', self.execute, req, keyspace)
d.addErrback(self.process_request_error, req, keyspace, req_d, retries)
return d
def maybe_do_more_work(self, _, q):
# it's possible this is being called as part of the immediate callback
# chain after the protocol's connectionLost errbacking. if so, our own
# connectionLost hasn't been called yet. allow all current processing
# to finish before deciding whether we get to do more.
def _real_maybe_do_more_work():
if not self.keep_working:
elif self.service is not None:
if self.service is not None:
self.service.reactor.callLater(0, _real_maybe_do_more_work)
def scream_like_a_little_girl(self, fail):
if self.service is not None:
complain = self.service.err
complain = log.err
complain(fail, "Factory for connection to %s had problems dealing with"
" the queue" % (self.node,))
# don't process more requests
def work_on_queue(self, q):
self.keep_working = True
d = self.job('queue_getter', q.get)
d.addCallback(self.maybe_do_more_work, q)
d.addErrback(lambda f: f.trap(defer.CancelledError))
return d
def stop_working_on_queue(self):
self.logstate('stop_working_on_queue [jobphase=%s]' % self.jobphase)
self.keep_working = False
if self.jobphase == 'queue_getter':
def finish_and_die(self):
If there is a request pending, let it finish and be handled, then
disconnect and die. If not, cancel any pending queue requests and
just die.
if self.jobphase != 'pending_request':
def logstate(self, msg):
if getattr(self, 'debugging', False):
log.msg('CPRF 0x%x (node %s) [%s]: %s'
% (id(self), self.node, self.jobphase, msg))
class CassandraKeyspaceConnection:
Glue class which acts as a manager for CassandraClient but passes requests
on to a CassandraClusterPool- in the case where you want all requests
through this manager to be guaranteed to go to the same keyspace,
regardless of what other consumers of the CassandraClusterPool might do.
def __init__(self, pool, keyspace):
self.pool = pool
self.keyspace = keyspace
def pushRequest(self, req, retries=None):
return self.pool.pushRequest(req, retries=retries, keyspace=self.keyspace)
def set_keyspace(self, keyspace):
raise RuntimeError("Don't call set_keyspace on a CassandraKeyspaceConnection")
def login(self, credentials):
return self.pool.login(credentials)
class CassandraNode:
Represent a Cassandra node, in the same sense Cassandra uses.
Keep track of connection success and failure history for some time, so
that smarter decisions can be made about where to make new connections
within a pool.
Implement exponential backoff in reconnect time when connections fail.
@ivar history_interval: Keep history entries for at least this many seconds
@ivar max_delay: Forced delay between connection attempts will not exceed
this value (although actual connection attempts may be farther apart
than this, if the pool has enough connections without it)
history_interval = 86400
max_delay = 180
initial_delay = 0.05
# NIST backoff factors
factor = protocol.ReconnectingClientFactory.factor
jitter = protocol.ReconnectingClientFactory.jitter
def __init__(self, host, port): = host
self.port = port
self.reconnect_delay = self.initial_delay
self.can_reconnect_at = 0
# a list of (timestamp, error) tuples, least recent first.
# (timestamp, None) tuples will be inserted on a successful connection.
self.history = []
def record_hist(self, value):
now = time()
if self.history and self.history[0][0] < (now - self.history_interval * 2):
# it has been 2x history_interval; prune history
cutoff = now - self.history_interval
for n, (tstamp, hval) in enumerate(self.history):
if tstamp > cutoff:
self.history = self.history[n:]
self.history.append((now, value))
def conn_success(self):
self.reconnect_delay = self.initial_delay
self.can_reconnect_at = 0
def conn_fail(self, reason):
# these tend to come in clusters. if the same error was received
# recently (before the reconnect delay expired), return False to
# indicate the event is not 'notable', and don't bump the delay
# to a higher level.
is_notable = self.is_failure_notable(reason)
if is_notable:
newdelay = min(self.reconnect_delay * self.factor, self.max_delay)
if self.jitter:
newdelay = random.normalvariate(newdelay, newdelay * self.jitter)
self.reconnect_delay = newdelay
self.can_reconnect_at = time() + newdelay
# reset but use the same delay
self.can_reconnect_at = time() + self.reconnect_delay
return is_notable
def is_failure_notable(self, reason):
tstamp, last_err = self.history[-1]
except IndexError:
if type(last_err) is type(reason.value):
if time() < self.can_reconnect_at:
return False
return True
def seconds_until_connect_ok(self):
return self.can_reconnect_at - time()
def __str__(self):
return '<%s %s:%s @0x%x>' % (self.__class__.__name__,, self.port, id(self))
__repr__ = __str__
def __eq__(self, other):
return self.__class__ == other.__class__ \
and == \
and self.port == other.port
def __cmp__(self, other):
return cmp((, self.port), (, other.port))
def __hash__(self):
return hash((self.__class__,, self.port))
class CassandraClusterPool(service.Service, object):
Manage a pool of connections to nodes in a Cassandra cluster.
Requests made to the pool will go to whichever host is the least loaded (as
far as this class can tell). If the requests specify multiple retries, the
retries will be executed on different hosts if possible.
Will periodically check an unparticular connection to see if new nodes can
be found, and add them to the pool.
Note that like most Services, the pool will not start until startService is
called. If you have a parent Service (like a
L{twisted.service.application.Application} instance), set that to be this
service's parent:
>>> cluster_pool.setServiceParent(application)
and the startService() and stopService() methods will be called when
@ivar default_cassandra_thrift_port: just what it says on the tin
@ivar max_connections_per_node: do our best not not to exceed this many
connections to a single Cassandra endpoint
@type max_connections_per_node: int
@ivar on_insufficient_nodes: if set to a callback, this will be called
in the event that there are no valid places to connect to expand
the pool to its target size. Regardless of actions taken by this
callback, the service will wait until a node is expected to be
available and then check again.
@type on_insufficient_nodes: callback taking four arguments: the current
size of the connection pool, the target size of the pool, the
number of pending requests, and the number of seconds before a
candidate node will be available to try connecting (or None, if no
candidate is in sight).
@ivar on_insufficient_conns: if set to a callback, this will be called
when a request is made and all current connections are busy. The
request will still be expected to go through, once another connection
is available, but it may be helpful to know how often this is
happening and possibly expand the pool
@type on_insufficient_conns: callback taking two arguments: the current
size of the connection pool, and the number of requests which are
pending in the CassandraClusterPool queue
@ivar request_retries: the default number of retries which will be
performed for requests when the retry number is unspecified
@type request_retries: int
@ivar retryables: A list of Exception types which, if they are raised in
the course of a Cassandra Thrift operation, mean both that (a) the
request can be tried again on another connection, and that (b) if the
connection was lost right after this error, it can be retried
default_cassandra_thrift_port = 9160
max_connections_per_node = 25
on_insufficient_nodes = staticmethod(lame_log_insufficient_nodes)
on_insufficient_conns = staticmethod(noop)
request_retries = 4
conn_factory = CassandraPoolReconnectorFactory
retryables = (IOError, socket.error, Thrift.TException,
TimedOutException, UnavailableException,
def __init__(self, seed_list, keyspace=None, creds=None, thrift_port=None,
pool_size=None, conn_timeout=10, bind_address=None,
log_cb=log.msg, reactor=None, require_api_version=None):
Initialize a CassandraClusterPool.
@param keyspace: If given and not None, determines the keyspace to
which all connections in this pool will be made.
@param creds: Credentials to use to authenticate Cassandra connections
@type creds: A dict (or other mapping) of strings to strings
@param seed_list: An initial set of host addresses which, if
connectable, are part of this cluster.
@type seed_list: iterable
@param thrift_port: The port to use for connections to Cassandra nodes
@param pool_size: The target size for the connection pool. Naturally,
the actual size may be higher or lower as nodes connect and
disconnect, but an effort will be made to adjust toward this size.
@type pool_size: int
@param conn_timeout: The number of seconds before a pending connection
is deemed unsuccessful and aborted. Of course, when a connection
error can be detected before this time, the connection will be
aborted appropriately.
@type conn_timeout: float
@param bind_address: The local interface to which to bind when making
outbound connections. Default: determined by the system's socket
@type bind_address: str
@param log_cb: A callable which is expected to work like
L{twisted.python.log.msg}. Will be used when certain connection
and disconnection events occur. The default is log.msg.
@param reactor: The reactor instance to use when starting thrift
connections or setting timers.
@param require_api_version: If not None, Telephus will require that
all connections conform to the API for the given Cassandra version.
Possible values are "0.7", "0.8", "1.0", etc.
If None, Telephus will consider all supported API versions to be
If the api version reported by a remote node is not compatible, the
connection to that node will be aborted. Default: None
self.seed_list = list(seed_list)
if thrift_port is None:
thrift_port = self.default_cassandra_thrift_port
self.thrift_port = thrift_port
if pool_size is None:
pool_size = len(self.seed_list)
self.target_pool_size = pool_size
self.log = log_cb
self.conn_timeout = conn_timeout
self.bind_address = bind_address
self.keyspace = keyspace
self.creds = creds
self.request_queue = defer.DeferredQueue()
self.require_api_version = require_api_version
self.future_fill_pool = None
self.removed_nodes = set()
self._client_instance = CassandraClient(self)
if reactor is None:
from twisted.internet import reactor
self.reactor = reactor
# A set of CassandraNode instances representing known nodes. This
# includes nodes from the initial seed list, nodes seen in
# describe_ring calls to existing nodes, and nodes explicitly added
# by the addNode() method. Nodes are only removed from this set if
# no connections have been successful in self.forget_node_interval
# seconds, or by an explicit call to removeNode().
self.nodes = set()
# A set of CassandraPoolReconnectorFactory instances corresponding to
# connections which are either live or pending. Failed attempts to
# connect will remove a connector from this set. When connections are
# lost, an immediate reconnect will be attempted.
self.connectors = set()
# A collection of objects from self.connectors corresponding to
# existing, working (as far as we know) connections. This will be
# derivable from self.connectors, but hopefully will be maintained to
# present a good snapshot of what is alive, now, and what is not.
# This is stored in a deque so that it can be efficiently rotated
# to distribute requests.
self.good_conns = set()
# A set of CassandraPoolReconnectorFactory instances, formerly in
# self.connectors, the connections for which are draining. No new
# requests should be fed to these instances; they are tracked only so
# that they can be terminated more fully in case this service is shut
# down before they finish.
self.dying_conns = set()
def startService(self):
for addr in self.seed_list:
if isinstance(addr, tuple) and len(addr) == 2:
self.addNode((addr, self.thrift_port))
def stopService(self):
if self.future_fill_pool is not None and
for factory in self.connectors.copy():
factory.service = None
self.connectors = set()
self.good_conns = set()
self.dying_conns = set()
def addNode(self, node):
if not isinstance(node, CassandraNode):
node = CassandraNode(*node)
if node in self.nodes:
raise ValueError("%s is already known" % (node,))
if node in self.removed_nodes:
def removeNode(self, node):
if not isinstance(node, CassandraNode):
node = CassandraNode(*node)
for f in self.all_connectors_to(node):
for f in self.dying_conns.copy():
if f.node == node:
def err(self, _stuff=None, _why=None, **kw):
if _stuff is None:
_stuff = failure.Failure()
kw['isError'] = True
kw['why'] = _why
if isinstance(_stuff, failure.Failure):
self.log(failure=_stuff, **kw)
elif isinstance(_stuff, Exception):
self.log(failure=failure.Failure(_stuff), **kw)
self.log(repr(_stuff), **kw)
# methods for inspecting current connection state
def all_connectors(self):
return self.connectors.copy()
def num_connectors(self):
Return the total number of current connectors, including both live and
pending connections.
return len(self.connectors)
def all_connectors_to(self, node):
return [f for f in self.connectors if f.node == node]
def num_connectors_to(self, host):
return len(self.all_connectors_to(host))
def all_active_conns(self):
return self.good_conns.copy()
def num_active_conns(self):
return len(self.good_conns)
def all_active_conns_to(self, node):
return [f for f in self.good_conns if f.node == node]
def num_active_conns_to(self, node):
return len(self.all_active_conns_to(node))
def all_working_conns(self):
return [f for f in self.good_conns if f.jobphase == 'pending_request']
def num_working_conns(self):
return len(self.all_working_conns())
def all_pending_conns(self):
return self.connectors - self.good_conns
def num_pending_conns(self):
return len(self.all_pending_conns())
def all_pending_conns_to(self, node):
return [f for f in self.all_pending_conns() if f.node == node]
def num_pending_conns_to(self, node):
return len(self.all_pending_conns_to(node))
def add_connection_score(self, node):
Return a numeric value that determines this node's score for adding
a new connection. A negative value indicates that no connections
should be made to this node for at least that number of seconds.
A value of -inf indicates no connections should be made to this
node for the foreseeable future.
This score should ideally take into account the connectedness of
available nodes, so that those with less current connections will
get more.
# TODO: this should ideally take node history into account
conntime = node.seconds_until_connect_ok()
if conntime > 0:
self.log("not considering %r for new connection; has %r left on "
"connect blackout" % (node, conntime))
return -conntime
numconns = self.num_connectors_to(node)
if numconns >= self.max_connections_per_node:
return float('-Inf')
return sys.maxint - numconns
def adjustPoolSize(self, newsize):
Change the target pool size. If we have too many connections already,
ask some to finish what they're doing and die (preferring to kill
connections to the node that already has the most connections). If
we have too few, create more.
if newsize < 0:
raise ValueError("pool size must be nonnegative")
self.log("Adjust pool size from %d to %d." % (self.target_pool_size, newsize))
self.target_pool_size = newsize
def update_known_nodes(self, ring):
for tokenrange in ring:
for addr in tokenrange.endpoints:
if ':' in addr:
addr, port = addr.split(':', 1)
port = int(port)
port = self.thrift_port
node = CassandraNode(addr, port)
if node not in self.removed_nodes and node not in self.nodes:
def choose_nodes_to_connect(self):
while True:
nodes = list(self.nodes)
scores = map(self.add_connection_score, nodes)
bestscore, bestnode = max(zip(scores, nodes))
if bestscore < 0:
raise NoNodesAvailable(-bestscore)
yield bestnode
def choose_pending_conns_to_kill(self):
# prefer to junk pending conns to most-redundantly-connected node
while True:
pending_conns = self.all_pending_conns()
if len(pending_conns) == 0:
yield max(pending_conns, key=lambda f: self.num_connectors_to(f.node))
def choose_conns_to_kill(self):
nodegetter = lambda f: f.node
# prefer to junk conns to most-redundantly-connected node
while True:
active_conns = self.all_active_conns()
if len(active_conns) == 0:
nodes_and_conns = groupby(sorted(active_conns, key=nodegetter), nodegetter)
nodes_and_counts = ((n, len(list(conns))) for (n, conns) in nodes_and_conns)
bestnode, bestcount = max(nodes_and_counts, key=lambda (n,count): count)
# should be safe from IndexError
yield self.all_active_conns_to(bestnode)[0]
def kill_excess_pending_conns(self):
killnum = self.num_connectors() - self.target_pool_size
if killnum <= 0:
for n, f in izip(xrange(killnum), self.choose_pending_conns_to_kill()):
self.log("Aborting pending conn to %r" % (f.node,))
def kill_excess_conns(self):
need_to_kill = self.num_active_conns() - self.target_pool_size
if need_to_kill <= 0:
for n, f in izip(xrange(need_to_kill), self.choose_conns_to_kill()):
self.log("Draining conn to %r" % (f.node,))
def fill_pool(self):
Add connections as necessary to meet the target pool size. If there
are no nodes to connect to (because we maxed out connections-per-node
on all active connections and any unconnected nodes have pending
reconnect timers), call the on_insufficient_nodes callback.
need = self.target_pool_size - self.num_connectors()
if need <= 0:
for num, node in izip(xrange(need), self.choose_nodes_to_connect()):
except NoNodesAvailable, e:
waittime = e.args[0]
pending_requests = len(self.request_queue.pending)
if self.on_insufficient_nodes:
waittime if waittime != float('Inf') else None)
if self.num_connectors() == 0 and pending_requests > 0:
if self.on_insufficient_conns:
def schedule_future_fill_pool(self, seconds):
if seconds == float('Inf'):
if self.future_fill_pool is None or not
self.future_fill_pool = self.reactor.callLater(seconds, self.fill_pool)
def make_conn(self, node):
self.log('Adding connection to %s' % (node,))
f = self.conn_factory(node, self, self.require_api_version)
if bindaddr is not None and isinstance(bindaddr, str):
bindaddr = (bindaddr, 0)
self.reactor.connectTCP(, node.port, f,
def remove_good_conn(self, f):
except KeyError:
def remove_connector(self, f):
except KeyError:
except KeyError:
def client_conn_failed(self, reason, f):
is_notable = f.node.conn_fail(reason)
if is_notable:
self.err(reason, 'Thrift pool connection to %s failed' % (f.node,))
def client_conn_made(self, f):
d = f.prep_connection(self.creds, self.keyspace)
d.addCallback(self.client_ready, f)
d.addErrback(self.client_conn_failed, f)
def client_ready(self, ring, f):
self.log('Successfully added connection to %s to the pool' % (f.node,))
def client_conn_lost(self, f, reason):
if reason.check(error.ConnectionDone):
self.log('Thrift pool connection to %s failed (cleanly)' % (f.node,))
self.err(reason, 'Thrift pool connection to %s was lost' % (f.node,))
if f.last_error is None or f.last_error.check(*self.retryables):
self.log('Retrying connection right away')
def pushRequest(self, req, retries=None, keyspace=None):
if keyspace is None:
keyspace = self.keyspace
retries = retries if retries is not None else self.request_retries
req_d = defer.Deferred()
self.pushRequest_really(req, keyspace, req_d, retries)
return req_d
def pushRequest_really(self, req, keyspace, req_d, retries):
call_insuff_conns = False
if len(self.request_queue.waiting) == 0:
# no workers are immediately available
if self.on_insufficient_conns:
call_insuff_conns = True
self.request_queue.put((req, keyspace, req_d, retries))
if call_insuff_conns:
def resubmit(self, req, keyspace, req_d, retries):
Push this request to the front of the line, just to be a jerk.
self.log('resubmitting %s request' % (req.method,))
self.pushRequest_really(req, keyspace, req_d, retries)
self.request_queue.pending.remove((req, keyspace, req_d, retries))
except ValueError:
# it's already been scooped up
self.request_queue.pending.insert(0, (req, keyspace, req_d, retries))
def set_keyspace(self, keyspace):
Change the keyspace which will be used for subsequent requests to this
CassandraClusterPool, and return a Deferred that will fire once it can
be verified that connections can successfully use that keyspace.
If something goes wrong trying to change a connection to that keyspace,
the Deferred will errback, and the keyspace to be used for future
requests will not be changed.
Requests made between the time this method is called and the time that
the returned Deferred is fired may be made in either the previous
keyspace or the new keyspace. If you may need to make use of multiple
keyspaces at the same time in the same app, consider using the
specialized CassandraKeyspaceConnection interface provided by the
keyspaceConnection method.
# push a real set_keyspace on some (any) connection; the idea is that
# if it succeeds there, it is likely to succeed everywhere, and vice
# versa. don't bother waiting for all connections to change- some of
# them may be doing long blocking tasks and by the time they're done,
# the keyspace might be changed again anyway
d = self.pushRequest(ManagedThriftRequest('set_keyspace', keyspace))
def store_keyspace(_):
self.keyspace = keyspace
return d
def __getattr__(self, name):
Make CassandraClusterPool act like its own CassandraClient when
the user wants to use it that way
return getattr(self._client_instance, name)
def get_consistency(self):
return self._client_instance.consistency
def set_consistency(self, value):
self._client_instance.consistency = value
consistency = property(get_consistency, set_consistency)
def keyspaceConnection(self, keyspace, consistency=ConsistencyLevel.ONE):
Return a CassandraClient instance which uses this CassandraClusterPool
by way of a CassandraKeyspaceConnection, so that all requests made
through it are guaranteed to go to the given keyspace, no matter what
other consumers of this pool may do.
return CassandraClient(CassandraKeyspaceConnection(self, keyspace),
def __str__(self):
return '<%s: [%d nodes known] [%d connections]>' \
% (self.__class__.__name__, len(self.nodes), self.num_active_conns())
__repr__ = __str__
Jump to Line
Something went wrong with that request. Please try again.