# Copyright 2013-2015 DataStax, Inc.
#
# 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.
from itertools import islice, cycle, groupby, repeat
import logging
from random import randint
from threading import Lock
import six
from cassandra import ConsistencyLevel
from six.moves import range
log = logging.getLogger(__name__)
[docs]class HostDistance(object):
"""
A measure of how "distant" a node is from the client, which
may influence how the load balancer distributes requests
and how many connections are opened to the node.
"""
IGNORED = -1
"""
A node with this distance should never be queried or have
connections opened to it.
"""
LOCAL = 0
"""
Nodes with ``LOCAL`` distance will be preferred for operations
under some load balancing policies (such as :class:`.DCAwareRoundRobinPolicy`)
and will have a greater number of connections opened against
them by default.
This distance is typically used for nodes within the same
datacenter as the client.
"""
REMOTE = 1
"""
Nodes with ``REMOTE`` distance will be treated as a last resort
by some load balancing policies (such as :class:`.DCAwareRoundRobinPolicy`)
and will have a smaller number of connections opened against
them by default.
This distance is typically used for nodes outside of the
datacenter that the client is running in.
"""
class HostStateListener(object):
def on_up(self, host):
""" Called when a node is marked up. """
raise NotImplementedError()
def on_down(self, host):
""" Called when a node is marked down. """
raise NotImplementedError()
def on_add(self, host):
"""
Called when a node is added to the cluster. The newly added node
should be considered up.
"""
raise NotImplementedError()
def on_remove(self, host):
""" Called when a node is removed from the cluster. """
raise NotImplementedError()
[docs]class LoadBalancingPolicy(HostStateListener):
"""
Load balancing policies are used to decide how to distribute
requests among all possible coordinator nodes in the cluster.
In particular, they may focus on querying "near" nodes (those
in a local datacenter) or on querying nodes who happen to
be replicas for the requested data.
You may also use subclasses of :class:`.LoadBalancingPolicy` for
custom behavior.
"""
_hosts_lock = None
def __init__(self):
self._hosts_lock = Lock()
[docs] def distance(self, host):
"""
Returns a measure of how remote a :class:`~.pool.Host` is in
terms of the :class:`.HostDistance` enums.
"""
raise NotImplementedError()
[docs] def populate(self, cluster, hosts):
"""
This method is called to initialize the load balancing
policy with a set of :class:`.Host` instances before its
first use. The `cluster` parameter is an instance of
:class:`.Cluster`.
"""
raise NotImplementedError()
[docs] def make_query_plan(self, working_keyspace=None, query=None):
"""
Given a :class:`~.query.Statement` instance, return a iterable
of :class:`.Host` instances which should be queried in that
order. A generator may work well for custom implementations
of this method.
Note that the `query` argument may be :const:`None` when preparing
statements.
`working_keyspace` should be the string name of the current keyspace,
as set through :meth:`.Session.set_keyspace()` or with a ``USE``
statement.
"""
raise NotImplementedError()
[docs] def check_supported(self):
"""
This will be called after the cluster Metadata has been initialized.
If the load balancing policy implementation cannot be supported for
some reason (such as a missing C extension), this is the point at
which it should raise an exception.
"""
pass
[docs]class RoundRobinPolicy(LoadBalancingPolicy):
"""
A subclass of :class:`.LoadBalancingPolicy` which evenly
distributes queries across all nodes in the cluster,
regardless of what datacenter the nodes may be in.
This load balancing policy is used by default.
"""
_live_hosts = frozenset(())
def populate(self, cluster, hosts):
self._live_hosts = frozenset(hosts)
if len(hosts) <= 1:
self._position = 0
else:
self._position = randint(0, len(hosts) - 1)
def distance(self, host):
return HostDistance.LOCAL
def make_query_plan(self, working_keyspace=None, query=None):
# not thread-safe, but we don't care much about lost increments
# for the purposes of load balancing
pos = self._position
self._position += 1
hosts = self._live_hosts
length = len(hosts)
if length:
pos %= length
return islice(cycle(hosts), pos, pos + length)
else:
return []
def on_up(self, host):
with self._hosts_lock:
self._live_hosts = self._live_hosts.union((host, ))
def on_down(self, host):
with self._hosts_lock:
self._live_hosts = self._live_hosts.difference((host, ))
def on_add(self, host):
with self._hosts_lock:
self._live_hosts = self._live_hosts.union((host, ))
def on_remove(self, host):
with self._hosts_lock:
self._live_hosts = self._live_hosts.difference((host, ))
[docs]class DCAwareRoundRobinPolicy(LoadBalancingPolicy):
"""
Similar to :class:`.RoundRobinPolicy`, but prefers hosts
in the local datacenter and only uses nodes in remote
datacenters as a last resort.
"""
local_dc = None
used_hosts_per_remote_dc = 0
def __init__(self, local_dc='', used_hosts_per_remote_dc=0):
"""
The `local_dc` parameter should be the name of the datacenter
(such as is reported by ``nodetool ring``) that should
be considered local. If not specified, the driver will choose
a local_dc based on the first host among :attr:`.Cluster.contact_points`
having a valid DC. If relying on this mechanism, all specified
contact points should be nodes in a single, local DC.
`used_hosts_per_remote_dc` controls how many nodes in
each remote datacenter will have connections opened
against them. In other words, `used_hosts_per_remote_dc` hosts
will be considered :attr:`~.HostDistance.REMOTE` and the
rest will be considered :attr:`~.HostDistance.IGNORED`.
By default, all remote hosts are ignored.
"""
self.local_dc = local_dc
self.used_hosts_per_remote_dc = used_hosts_per_remote_dc
self._dc_live_hosts = {}
self._position = 0
self._contact_points = []
LoadBalancingPolicy.__init__(self)
def _dc(self, host):
return host.datacenter or self.local_dc
def populate(self, cluster, hosts):
for dc, dc_hosts in groupby(hosts, lambda h: self._dc(h)):
self._dc_live_hosts[dc] = tuple(set(dc_hosts))
if not self.local_dc:
self._contact_points = cluster.contact_points
self._position = randint(0, len(hosts) - 1) if hosts else 0
def distance(self, host):
dc = self._dc(host)
if dc == self.local_dc:
return HostDistance.LOCAL
if not self.used_hosts_per_remote_dc:
return HostDistance.IGNORED
else:
dc_hosts = self._dc_live_hosts.get(dc)
if not dc_hosts:
return HostDistance.IGNORED
if host in list(dc_hosts)[:self.used_hosts_per_remote_dc]:
return HostDistance.REMOTE
else:
return HostDistance.IGNORED
def make_query_plan(self, working_keyspace=None, query=None):
# not thread-safe, but we don't care much about lost increments
# for the purposes of load balancing
pos = self._position
self._position += 1
local_live = self._dc_live_hosts.get(self.local_dc, ())
pos = (pos % len(local_live)) if local_live else 0
for host in islice(cycle(local_live), pos, pos + len(local_live)):
yield host
# the dict can change, so get candidate DCs iterating over keys of a copy
other_dcs = [dc for dc in self._dc_live_hosts.copy().keys() if dc != self.local_dc]
for dc in other_dcs:
remote_live = self._dc_live_hosts.get(dc, ())
for host in remote_live[:self.used_hosts_per_remote_dc]:
yield host
def on_up(self, host):
# not worrying about threads because this will happen during
# control connection startup/refresh
if not self.local_dc and host.datacenter:
if host.address in self._contact_points:
self.local_dc = host.datacenter
log.info("Using datacenter '%s' for DCAwareRoundRobinPolicy (via host '%s'); "
"if incorrect, please specify a local_dc to the constructor, "
"or limit contact points to local cluster nodes" %
(self.local_dc, host.address))
del self._contact_points
dc = self._dc(host)
with self._hosts_lock:
current_hosts = self._dc_live_hosts.get(dc, ())
if host not in current_hosts:
self._dc_live_hosts[dc] = current_hosts + (host, )
def on_down(self, host):
dc = self._dc(host)
with self._hosts_lock:
current_hosts = self._dc_live_hosts.get(dc, ())
if host in current_hosts:
hosts = tuple(h for h in current_hosts if h != host)
if hosts:
self._dc_live_hosts[dc] = hosts
else:
del self._dc_live_hosts[dc]
def on_add(self, host):
self.on_up(host)
def on_remove(self, host):
self.on_down(host)
[docs]class TokenAwarePolicy(LoadBalancingPolicy):
"""
A :class:`.LoadBalancingPolicy` wrapper that adds token awareness to
a child policy.
This alters the child policy's behavior so that it first attempts to
send queries to :attr:`~.HostDistance.LOCAL` replicas (as determined
by the child policy) based on the :class:`.Statement`'s
:attr:`~.Statement.routing_key`. Once those hosts are exhausted, the
remaining hosts in the child policy's query plan will be used.
If no :attr:`~.Statement.routing_key` is set on the query, the child
policy's query plan will be used as is.
"""
_child_policy = None
_cluster_metadata = None
def __init__(self, child_policy):
self._child_policy = child_policy
def populate(self, cluster, hosts):
self._cluster_metadata = cluster.metadata
self._child_policy.populate(cluster, hosts)
def check_supported(self):
if not self._cluster_metadata.can_support_partitioner():
raise Exception(
'%s cannot be used with the cluster partitioner (%s) because '
'the relevant C extension for this driver was not compiled. '
'See the installation instructions for details on building '
'and installing the C extensions.' %
(self.__class__.__name__, self._cluster_metadata.partitioner))
def distance(self, *args, **kwargs):
return self._child_policy.distance(*args, **kwargs)
def make_query_plan(self, working_keyspace=None, query=None):
if query and query.keyspace:
keyspace = query.keyspace
else:
keyspace = working_keyspace
child = self._child_policy
if query is None:
for host in child.make_query_plan(keyspace, query):
yield host
else:
routing_key = query.routing_key
if routing_key is None or keyspace is None:
for host in child.make_query_plan(keyspace, query):
yield host
else:
replicas = self._cluster_metadata.get_replicas(keyspace, routing_key)
for replica in replicas:
if replica.is_up and \
child.distance(replica) == HostDistance.LOCAL:
yield replica
for host in child.make_query_plan(keyspace, query):
# skip if we've already listed this host
if host not in replicas or \
child.distance(host) == HostDistance.REMOTE:
yield host
def on_up(self, *args, **kwargs):
return self._child_policy.on_up(*args, **kwargs)
def on_down(self, *args, **kwargs):
return self._child_policy.on_down(*args, **kwargs)
def on_add(self, *args, **kwargs):
return self._child_policy.on_add(*args, **kwargs)
def on_remove(self, *args, **kwargs):
return self._child_policy.on_remove(*args, **kwargs)
[docs]class WhiteListRoundRobinPolicy(RoundRobinPolicy):
"""
A subclass of :class:`.RoundRobinPolicy` which evenly
distributes queries across all nodes in the cluster,
regardless of what datacenter the nodes may be in, but
only if that node exists in the list of allowed nodes
This policy is addresses the issue described in
https://datastax-oss.atlassian.net/browse/JAVA-145
Where connection errors occur when connection
attempts are made to private IP addresses remotely
"""
def __init__(self, hosts):
"""
The `hosts` parameter should be a sequence of hosts to permit
connections to.
"""
self._allowed_hosts = hosts
RoundRobinPolicy.__init__(self)
def populate(self, cluster, hosts):
self._live_hosts = frozenset(h for h in hosts if h.address in self._allowed_hosts)
if len(hosts) <= 1:
self._position = 0
else:
self._position = randint(0, len(hosts) - 1)
def distance(self, host):
if host.address in self._allowed_hosts:
return HostDistance.LOCAL
else:
return HostDistance.IGNORED
def on_up(self, host):
if host.address in self._allowed_hosts:
RoundRobinPolicy.on_up(self, host)
def on_add(self, host):
if host.address in self._allowed_hosts:
RoundRobinPolicy.on_add(self, host)
[docs]class ConvictionPolicy(object):
"""
A policy which decides when hosts should be considered down
based on the types of failures and the number of failures.
If custom behavior is needed, this class may be subclassed.
"""
def __init__(self, host):
"""
`host` is an instance of :class:`.Host`.
"""
self.host = host
[docs] def add_failure(self, connection_exc):
"""
Implementations should return :const:`True` if the host should be
convicted, :const:`False` otherwise.
"""
raise NotImplementedError()
[docs] def reset(self):
"""
Implementations should clear out any convictions or state regarding
the host.
"""
raise NotImplementedError()
[docs]class SimpleConvictionPolicy(ConvictionPolicy):
"""
The default implementation of :class:`ConvictionPolicy`,
which simply marks a host as down after the first failure
of any kind.
"""
def add_failure(self, connection_exc):
return True
def reset(self):
pass
[docs]class ReconnectionPolicy(object):
"""
This class and its subclasses govern how frequently an attempt is made
to reconnect to nodes that are marked as dead.
If custom behavior is needed, this class may be subclassed.
"""
[docs] def new_schedule(self):
"""
This should return a finite or infinite iterable of delays (each as a
floating point number of seconds) inbetween each failed reconnection
attempt. Note that if the iterable is finite, reconnection attempts
will cease once the iterable is exhausted.
"""
raise NotImplementedError()
[docs]class ConstantReconnectionPolicy(ReconnectionPolicy):
"""
A :class:`.ReconnectionPolicy` subclass which sleeps for a fixed delay
inbetween each reconnection attempt.
"""
def __init__(self, delay, max_attempts=64):
"""
`delay` should be a floating point number of seconds to wait inbetween
each attempt.
`max_attempts` should be a total number of attempts to be made before
giving up, or :const:`None` to continue reconnection attempts forever.
The default is 64.
"""
if delay < 0:
raise ValueError("delay must not be negative")
if max_attempts is not None and max_attempts < 0:
raise ValueError("max_attempts must not be negative")
self.delay = delay
self.max_attempts = max_attempts
def new_schedule(self):
if self.max_attempts:
return repeat(self.delay, self.max_attempts)
return repeat(self.delay)
[docs]class ExponentialReconnectionPolicy(ReconnectionPolicy):
"""
A :class:`.ReconnectionPolicy` subclass which exponentially increases
the length of the delay inbetween each reconnection attempt up to
a set maximum delay.
"""
def __init__(self, base_delay, max_delay):
"""
`base_delay` and `max_delay` should be in floating point units of
seconds.
"""
if base_delay < 0 or max_delay < 0:
raise ValueError("Delays may not be negative")
if max_delay < base_delay:
raise ValueError("Max delay must be greater than base delay")
self.base_delay = base_delay
self.max_delay = max_delay
def new_schedule(self):
return (min(self.base_delay * (2 ** i), self.max_delay) for i in range(64))
[docs]class WriteType(object):
"""
For usage with :class:`.RetryPolicy`, this describe a type
of write operation.
"""
SIMPLE = 0
"""
A write to a single partition key. Such writes are guaranteed to be atomic
and isolated.
"""
BATCH = 1
"""
A write to multiple partition keys that used the distributed batch log to
ensure atomicity.
"""
UNLOGGED_BATCH = 2
"""
A write to multiple partition keys that did not use the distributed batch
log. Atomicity for such writes is not guaranteed.
"""
COUNTER = 3
"""
A counter write (for one or multiple partition keys). Such writes should
not be replayed in order to avoid overcount.
"""
BATCH_LOG = 4
"""
The initial write to the distributed batch log that Cassandra performs
internally before a BATCH write.
"""
CAS = 5
"""
A lighweight-transaction write, such as "DELETE ... IF EXISTS".
"""
WriteType.name_to_value = {
'SIMPLE': WriteType.SIMPLE,
'BATCH': WriteType.BATCH,
'UNLOGGED_BATCH': WriteType.UNLOGGED_BATCH,
'COUNTER': WriteType.COUNTER,
'BATCH_LOG': WriteType.BATCH_LOG,
'CAS': WriteType.CAS
}
[docs]class RetryPolicy(object):
"""
A policy that describes whether to retry, rethrow, or ignore coordinator
timeout and unavailable failures. These are failures reported from the
server side. Timeouts are configured by
`settings in cassandra.yaml <https://github.com/apache/cassandra/blob/cassandra-2.1.4/conf/cassandra.yaml#L568-L584>`_.
Unavailable failures occur when the coordinator cannot acheive the consistency
level for a request. For further information see the method descriptions
below.
To specify a default retry policy, set the
:attr:`.Cluster.default_retry_policy` attribute to an instance of this
class or one of its subclasses.
To specify a retry policy per query, set the :attr:`.Statement.retry_policy`
attribute to an instance of this class or one of its subclasses.
If custom behavior is needed for retrying certain operations,
this class may be subclassed.
"""
RETRY = 0
"""
This should be returned from the below methods if the operation
should be retried on the same connection.
"""
RETHROW = 1
"""
This should be returned from the below methods if the failure
should be propagated and no more retries attempted.
"""
IGNORE = 2
"""
This should be returned from the below methods if the failure
should be ignored but no more retries should be attempted.
"""
[docs] def on_read_timeout(self, query, consistency, required_responses,
received_responses, data_retrieved, retry_num):
"""
This is called when a read operation times out from the coordinator's
perspective (i.e. a replica did not respond to the coordinator in time).
It should return a tuple with two items: one of the class enums (such
as :attr:`.RETRY`) and a :class:`.ConsistencyLevel` to retry the
operation at or :const:`None` to keep the same consistency level.
`query` is the :class:`.Statement` that timed out.
`consistency` is the :class:`.ConsistencyLevel` that the operation was
attempted at.
The `required_responses` and `received_responses` parameters describe
how many replicas needed to respond to meet the requested consistency
level and how many actually did respond before the coordinator timed
out the request. `data_retrieved` is a boolean indicating whether
any of those responses contained data (as opposed to just a digest).
`retry_num` counts how many times the operation has been retried, so
the first time this method is called, `retry_num` will be 0.
By default, operations will be retried at most once, and only if
a sufficient number of replicas responded (with data digests).
"""
if retry_num != 0:
return (self.RETHROW, None)
elif received_responses >= required_responses and not data_retrieved:
return (self.RETRY, consistency)
else:
return (self.RETHROW, None)
[docs] def on_write_timeout(self, query, consistency, write_type,
required_responses, received_responses, retry_num):
"""
This is called when a write operation times out from the coordinator's
perspective (i.e. a replica did not respond to the coordinator in time).
`query` is the :class:`.Statement` that timed out.
`consistency` is the :class:`.ConsistencyLevel` that the operation was
attempted at.
`write_type` is one of the :class:`.WriteType` enums describing the
type of write operation.
The `required_responses` and `received_responses` parameters describe
how many replicas needed to acknowledge the write to meet the requested
consistency level and how many replicas actually did acknowledge the
write before the coordinator timed out the request.
`retry_num` counts how many times the operation has been retried, so
the first time this method is called, `retry_num` will be 0.
By default, failed write operations will retried at most once, and
they will only be retried if the `write_type` was
:attr:`~.WriteType.BATCH_LOG`.
"""
if retry_num != 0:
return (self.RETHROW, None)
elif write_type == WriteType.BATCH_LOG:
return (self.RETRY, consistency)
else:
return (self.RETHROW, None)
[docs] def on_unavailable(self, query, consistency, required_replicas, alive_replicas, retry_num):
"""
This is called when the coordinator node determines that a read or
write operation cannot be successful because the number of live
replicas are too low to meet the requested :class:`.ConsistencyLevel`.
This means that the read or write operation was never forwared to
any replicas.
`query` is the :class:`.Statement` that failed.
`consistency` is the :class:`.ConsistencyLevel` that the operation was
attempted at.
`required_replicas` is the number of replicas that would have needed to
acknowledge the operation to meet the requested consistency level.
`alive_replicas` is the number of replicas that the coordinator
considered alive at the time of the request.
`retry_num` counts how many times the operation has been retried, so
the first time this method is called, `retry_num` will be 0.
By default, no retries will be attempted and the error will be re-raised.
"""
return (self.RETHROW, None)
[docs]class FallthroughRetryPolicy(RetryPolicy):
"""
A retry policy that never retries and always propagates failures to
the application.
"""
def on_read_timeout(self, *args, **kwargs):
return (self.RETHROW, None)
def on_write_timeout(self, *args, **kwargs):
return (self.RETHROW, None)
def on_unavailable(self, *args, **kwargs):
return (self.RETHROW, None)
[docs]class DowngradingConsistencyRetryPolicy(RetryPolicy):
"""
A retry policy that sometimes retries with a lower consistency level than
the one initially requested.
**BEWARE**: This policy may retry queries using a lower consistency
level than the one initially requested. By doing so, it may break
consistency guarantees. In other words, if you use this retry policy,
there are cases (documented below) where a read at :attr:`~.QUORUM`
*may not* see a preceding write at :attr:`~.QUORUM`. Do not use this
policy unless you have understood the cases where this can happen and
are ok with that. It is also recommended to subclass this class so
that queries that required a consistency level downgrade can be
recorded (so that repairs can be made later, etc).
This policy implements the same retries as :class:`.RetryPolicy`,
but on top of that, it also retries in the following cases:
* On a read timeout: if the number of replicas that responded is
greater than one but lower than is required by the requested
consistency level, the operation is retried at a lower consistency
level.
* On a write timeout: if the operation is an :attr:`~.UNLOGGED_BATCH`
and at least one replica acknowledged the write, the operation is
retried at a lower consistency level. Furthermore, for other
write types, if at least one replica acknowledged the write, the
timeout is ignored.
* On an unavailable exception: if at least one replica is alive, the
operation is retried at a lower consistency level.
The reasoning behind this retry policy is as follows: if, based
on the information the Cassandra coordinator node returns, retrying the
operation with the initially requested consistency has a chance to
succeed, do it. Otherwise, if based on that information we know the
initially requested consistency level cannot be achieved currently, then:
* For writes, ignore the exception (thus silently failing the
consistency requirement) if we know the write has been persisted on at
least one replica.
* For reads, try reading at a lower consistency level (thus silently
failing the consistency requirement).
In other words, this policy implements the idea that if the requested
consistency level cannot be achieved, the next best thing for writes is
to make sure the data is persisted, and that reading something is better
than reading nothing, even if there is a risk of reading stale data.
"""
def _pick_consistency(self, num_responses):
if num_responses >= 3:
return (self.RETRY, ConsistencyLevel.THREE)
elif num_responses >= 2:
return (self.RETRY, ConsistencyLevel.TWO)
elif num_responses >= 1:
return (self.RETRY, ConsistencyLevel.ONE)
else:
return (self.RETHROW, None)
def on_read_timeout(self, query, consistency, required_responses,
received_responses, data_retrieved, retry_num):
if retry_num != 0:
return (self.RETHROW, None)
elif received_responses < required_responses:
return self._pick_consistency(received_responses)
elif not data_retrieved:
return (self.RETRY, consistency)
else:
return (self.RETHROW, None)
def on_write_timeout(self, query, consistency, write_type,
required_responses, received_responses, retry_num):
if retry_num != 0:
return (self.RETHROW, None)
if write_type in (WriteType.SIMPLE, WriteType.BATCH, WriteType.COUNTER):
if received_responses > 0:
# persisted on at least one replica
return (self.IGNORE, None)
else:
return (self.RETHROW, None)
elif write_type == WriteType.UNLOGGED_BATCH:
return self._pick_consistency(received_responses)
elif write_type == WriteType.BATCH_LOG:
return (self.RETRY, consistency)
return (self.RETHROW, None)
def on_unavailable(self, query, consistency, required_replicas, alive_replicas, retry_num):
if retry_num != 0:
return (self.RETHROW, None)
else:
return self._pick_consistency(alive_replicas)