def __init__(self, clock: Clock, handler: "ReplicationCommandHandler"):
self.clock = clock
self.command_handler = handler
self.last_received_command = self.clock.time_msec()
self.last_sent_command = 0
# When we requested the connection be closed
self.time_we_closed = None # type: Optional[int]
self.received_ping = False # Have we received a ping from the other side
self.state = ConnectionStates.CONNECTING
self.name = "anon" # The name sent by a client.
self.conn_id = random_string(5) # To dedupe in case of name clashes.
# List of pending commands to send once we've established the connection
self.pending_commands = [] # type: List[Command]
# The LoopingCall for sending pings.
self._send_ping_loop = None # type: Optional[task.LoopingCall]
# a logcontext which we use for processing incoming commands. We declare it as a
# background process so that the CPU stats get reported to prometheus.
self._logging_context = BackgroundProcessLoggingContext(
"replication-conn", self.conn_id
)
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# a logcontext which we use for processing incoming commands. We declare it as a
# background process so that the CPU stats get reported to prometheus.
self._logging_context = BackgroundProcessLoggingContext(
"replication_command_handler")
def __init__(self, clock: Clock, handler: "ReplicationCommandHandler"):
self.clock = clock
self.command_handler = handler
self.last_received_command = self.clock.time_msec()
self.last_sent_command = 0
# When we requested the connection be closed
self.time_we_closed: Optional[int] = None
self.received_ping = False # Have we received a ping from the other side
self.state = ConnectionStates.CONNECTING
self.name = "anon" # The name sent by a client.
self.conn_id = random_string(5) # To dedupe in case of name clashes.
# List of pending commands to send once we've established the connection
self.pending_commands: List[Command] = []
# The LoopingCall for sending pings.
self._send_ping_loop: Optional[task.LoopingCall] = None
# a logcontext which we use for processing incoming commands. We declare it as a
# background process so that the CPU stats get reported to prometheus.
with PreserveLoggingContext():
# thanks to `PreserveLoggingContext()`, the new logcontext is guaranteed to
# capture the sentinel context as its containing context and won't prevent
# GC of / unintentionally reactivate what would be the current context.
self._logging_context = BackgroundProcessLoggingContext(
"replication-conn", self.conn_id)
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# a logcontext which we use for processing incoming commands. We declare it as a
# background process so that the CPU stats get reported to prometheus.
with PreserveLoggingContext():
# thanks to `PreserveLoggingContext()`, the new logcontext is guaranteed to
# capture the sentinel context as its containing context and won't prevent
# GC of / unintentionally reactivate what would be the current context.
self._logging_context = BackgroundProcessLoggingContext(
"replication_command_handler")
class BaseReplicationStreamProtocol(LineOnlyReceiver):
"""Base replication protocol shared between client and server.
Reads lines (ignoring blank ones) and parses them into command classes,
asserting that they are valid for the given direction, i.e. server commands
are only sent by the server.
On receiving a new command it calls `on_<COMMAND_NAME>` with the parsed
command before delegating to `ReplicationCommandHandler.on_<COMMAND_NAME>`.
`ReplicationCommandHandler.on_<COMMAND_NAME>` can optionally return a coroutine;
if so, that will get run as a background process.
It also sends `PING` periodically, and correctly times out remote connections
(if they send a `PING` command)
"""
# The transport is going to be an ITCPTransport, but that doesn't have the
# (un)registerProducer methods, those are only on the implementation.
transport = None # type: Connection
delimiter = b"\n"
# Valid commands we expect to receive
VALID_INBOUND_COMMANDS = [] # type: Collection[str]
# Valid commands we can send
VALID_OUTBOUND_COMMANDS = [] # type: Collection[str]
max_line_buffer = 10000
def __init__(self, clock: Clock, handler: "ReplicationCommandHandler"):
self.clock = clock
self.command_handler = handler
self.last_received_command = self.clock.time_msec()
self.last_sent_command = 0
# When we requested the connection be closed
self.time_we_closed = None # type: Optional[int]
self.received_ping = False # Have we received a ping from the other side
self.state = ConnectionStates.CONNECTING
self.name = "anon" # The name sent by a client.
self.conn_id = random_string(5) # To dedupe in case of name clashes.
# List of pending commands to send once we've established the connection
self.pending_commands = [] # type: List[Command]
# The LoopingCall for sending pings.
self._send_ping_loop = None # type: Optional[task.LoopingCall]
# a logcontext which we use for processing incoming commands. We declare it as a
# background process so that the CPU stats get reported to prometheus.
self._logging_context = BackgroundProcessLoggingContext(
"replication-conn", self.conn_id
)
def connectionMade(self):
logger.info("[%s] Connection established", self.id())
self.state = ConnectionStates.ESTABLISHED
connected_connections.append(self) # Register connection for metrics
assert self.transport is not None
self.transport.registerProducer(self, True) # For the *Producing callbacks
self._send_pending_commands()
# Starts sending pings
self._send_ping_loop = self.clock.looping_call(self.send_ping, 5000)
# Always send the initial PING so that the other side knows that they
# can time us out.
self.send_command(PingCommand(self.clock.time_msec()))
self.command_handler.new_connection(self)
def send_ping(self):
"""Periodically sends a ping and checks if we should close the connection
due to the other side timing out.
"""
now = self.clock.time_msec()
if self.time_we_closed:
if now - self.time_we_closed > PING_TIMEOUT_MS:
logger.info(
"[%s] Failed to close connection gracefully, aborting", self.id()
)
assert self.transport is not None
self.transport.abortConnection()
else:
if now - self.last_sent_command >= PING_TIME:
self.send_command(PingCommand(now))
if (
self.received_ping
and now - self.last_received_command > PING_TIMEOUT_MS
):
logger.info(
"[%s] Connection hasn't received command in %r ms. Closing.",
self.id(),
now - self.last_received_command,
)
self.send_error("ping timeout")
def lineReceived(self, line: bytes):
"""Called when we've received a line"""
with PreserveLoggingContext(self._logging_context):
self._parse_and_dispatch_line(line)
def _parse_and_dispatch_line(self, line: bytes):
if line.strip() == "":
# Ignore blank lines
return
linestr = line.decode("utf-8")
try:
cmd = parse_command_from_line(linestr)
except Exception as e:
logger.exception("[%s] failed to parse line: %r", self.id(), linestr)
self.send_error("failed to parse line: %r (%r):" % (e, linestr))
return
if cmd.NAME not in self.VALID_INBOUND_COMMANDS:
logger.error("[%s] invalid command %s", self.id(), cmd.NAME)
self.send_error("invalid command: %s", cmd.NAME)
return
self.last_received_command = self.clock.time_msec()
tcp_inbound_commands_counter.labels(cmd.NAME, self.name).inc()
self.handle_command(cmd)
def handle_command(self, cmd: Command) -> None:
"""Handle a command we have received over the replication stream.
First calls `self.on_<COMMAND>` if it exists, then calls
`self.command_handler.on_<COMMAND>` if it exists (which can optionally
return an Awaitable).
This allows for protocol level handling of commands (e.g. PINGs), before
delegating to the handler.
Args:
cmd: received command
"""
handled = False
# First call any command handlers on this instance. These are for TCP
# specific handling.
cmd_func = getattr(self, "on_%s" % (cmd.NAME,), None)
if cmd_func:
cmd_func(cmd)
handled = True
# Then call out to the handler.
cmd_func = getattr(self.command_handler, "on_%s" % (cmd.NAME,), None)
if cmd_func:
res = cmd_func(self, cmd)
# the handler might be a coroutine: fire it off as a background process
# if so.
if isawaitable(res):
run_as_background_process(
"replication-" + cmd.get_logcontext_id(), lambda: res
)
handled = True
if not handled:
logger.warning("Unhandled command: %r", cmd)
def close(self):
logger.warning("[%s] Closing connection", self.id())
self.time_we_closed = self.clock.time_msec()
assert self.transport is not None
self.transport.loseConnection()
self.on_connection_closed()
def send_error(self, error_string, *args):
"""Send an error to remote and close the connection."""
self.send_command(ErrorCommand(error_string % args))
self.close()
def send_command(self, cmd, do_buffer=True):
"""Send a command if connection has been established.
Args:
cmd (Command)
do_buffer (bool): Whether to buffer the message or always attempt
to send the command. This is mostly used to send an error
message if we're about to close the connection due our buffers
becoming full.
"""
if self.state == ConnectionStates.CLOSED:
logger.debug("[%s] Not sending, connection closed", self.id())
return
if do_buffer and self.state != ConnectionStates.ESTABLISHED:
self._queue_command(cmd)
return
tcp_outbound_commands_counter.labels(cmd.NAME, self.name).inc()
string = "%s %s" % (cmd.NAME, cmd.to_line())
if "\n" in string:
raise Exception("Unexpected newline in command: %r", string)
encoded_string = string.encode("utf-8")
if len(encoded_string) > self.MAX_LENGTH:
raise Exception(
"Failed to send command %s as too long (%d > %d)"
% (cmd.NAME, len(encoded_string), self.MAX_LENGTH)
)
self.sendLine(encoded_string)
self.last_sent_command = self.clock.time_msec()
def _queue_command(self, cmd):
"""Queue the command until the connection is ready to write to again."""
logger.debug("[%s] Queueing as conn %r, cmd: %r", self.id(), self.state, cmd)
self.pending_commands.append(cmd)
if len(self.pending_commands) > self.max_line_buffer:
# The other side is failing to keep up and out buffers are becoming
# full, so lets close the connection.
# XXX: should we squawk more loudly?
logger.error("[%s] Remote failed to keep up", self.id())
self.send_command(ErrorCommand("Failed to keep up"), do_buffer=False)
self.close()
def _send_pending_commands(self):
"""Send any queued commandes"""
pending = self.pending_commands
self.pending_commands = []
for cmd in pending:
self.send_command(cmd)
def on_PING(self, line):
self.received_ping = True
def on_ERROR(self, cmd):
logger.error("[%s] Remote reported error: %r", self.id(), cmd.data)
def pauseProducing(self):
"""This is called when both the kernel send buffer and the twisted
tcp connection send buffers have become full.
We don't actually have any control over those sizes, so we buffer some
commands ourselves before knifing the connection due to the remote
failing to keep up.
"""
logger.info("[%s] Pause producing", self.id())
self.state = ConnectionStates.PAUSED
def resumeProducing(self):
"""The remote has caught up after we started buffering!"""
logger.info("[%s] Resume producing", self.id())
self.state = ConnectionStates.ESTABLISHED
self._send_pending_commands()
def stopProducing(self):
"""We're never going to send any more data (normally because either
we or the remote has closed the connection)
"""
logger.info("[%s] Stop producing", self.id())
self.on_connection_closed()
def connectionLost(self, reason):
logger.info("[%s] Replication connection closed: %r", self.id(), reason)
if isinstance(reason, Failure):
assert reason.type is not None
connection_close_counter.labels(reason.type.__name__).inc()
else:
connection_close_counter.labels(reason.__class__.__name__).inc()
try:
# Remove us from list of connections to be monitored
connected_connections.remove(self)
except ValueError:
pass
# Stop the looping call sending pings.
if self._send_ping_loop and self._send_ping_loop.running:
self._send_ping_loop.stop()
self.on_connection_closed()
def on_connection_closed(self):
logger.info("[%s] Connection was closed", self.id())
self.state = ConnectionStates.CLOSED
self.pending_commands = []
self.command_handler.lost_connection(self)
if self.transport:
self.transport.unregisterProducer()
# mark the logging context as finished
self._logging_context.__exit__(None, None, None)
def __str__(self):
addr = None
if self.transport:
addr = str(self.transport.getPeer())
return "ReplicationConnection<name=%s,conn_id=%s,addr=%s>" % (
self.name,
self.conn_id,
addr,
)
def id(self):
return "%s-%s" % (self.name, self.conn_id)
def lineLengthExceeded(self, line):
"""Called when we receive a line that is above the maximum line length"""
self.send_error("Line length exceeded")
class RedisSubscriber(txredisapi.SubscriberProtocol):
"""Connection to redis subscribed to replication stream.
This class fulfils two functions:
(a) it implements the twisted Protocol API, where it handles the SUBSCRIBEd redis
connection, parsing *incoming* messages into replication commands, and passing them
to `ReplicationCommandHandler`
(b) it implements the IReplicationConnection API, where it sends *outgoing* commands
onto outbound_redis_connection.
Due to the vagaries of `txredisapi` we don't want to have a custom
constructor, so instead we expect the defined attributes below to be set
immediately after initialisation.
Attributes:
synapse_handler: The command handler to handle incoming commands.
synapse_stream_name: The *redis* stream name to subscribe to and publish
from (not anything to do with Synapse replication streams).
synapse_outbound_redis_connection: The connection to redis to use to send
commands.
"""
synapse_handler = None # type: ReplicationCommandHandler
synapse_stream_name = None # type: str
synapse_outbound_redis_connection = None # type: txredisapi.RedisProtocol
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# a logcontext which we use for processing incoming commands. We declare it as a
# background process so that the CPU stats get reported to prometheus.
self._logging_context = BackgroundProcessLoggingContext(
"replication_command_handler")
def connectionMade(self):
logger.info("Connected to redis")
super().connectionMade()
run_as_background_process("subscribe-replication",
self._send_subscribe)
async def _send_subscribe(self):
# it's important to make sure that we only send the REPLICATE command once we
# have successfully subscribed to the stream - otherwise we might miss the
# POSITION response sent back by the other end.
logger.info("Sending redis SUBSCRIBE for %s", self.synapse_stream_name)
await make_deferred_yieldable(self.subscribe(self.synapse_stream_name))
logger.info(
"Successfully subscribed to redis stream, sending REPLICATE command"
)
self.synapse_handler.new_connection(self)
await self._async_send_command(ReplicateCommand())
logger.info("REPLICATE successfully sent")
# We send out our positions when there is a new connection in case the
# other side missed updates. We do this for Redis connections as the
# otherside won't know we've connected and so won't issue a REPLICATE.
self.synapse_handler.send_positions_to_connection(self)
def messageReceived(self, pattern: str, channel: str, message: str):
"""Received a message from redis."""
with PreserveLoggingContext(self._logging_context):
self._parse_and_dispatch_message(message)
def _parse_and_dispatch_message(self, message: str):
if message.strip() == "":
# Ignore blank lines
return
try:
cmd = parse_command_from_line(message)
except Exception:
logger.exception(
"Failed to parse replication line: %r",
message,
)
return
# We use "redis" as the name here as we don't have 1:1 connections to
# remote instances.
tcp_inbound_commands_counter.labels(cmd.NAME, "redis").inc()
self.handle_command(cmd)
def handle_command(self, cmd: Command) -> None:
"""Handle a command we have received over the replication stream.
Delegates to `self.handler.on_<COMMAND>` (which can optionally return an
Awaitable).
Args:
cmd: received command
"""
cmd_func = getattr(self.synapse_handler, "on_%s" % (cmd.NAME, ), None)
if not cmd_func:
logger.warning("Unhandled command: %r", cmd)
return
res = cmd_func(self, cmd)
# the handler might be a coroutine: fire it off as a background process
# if so.
if isawaitable(res):
run_as_background_process("replication-" + cmd.get_logcontext_id(),
lambda: res)
def connectionLost(self, reason):
logger.info("Lost connection to redis")
super().connectionLost(reason)
self.synapse_handler.lost_connection(self)
# mark the logging context as finished
self._logging_context.__exit__(None, None, None)
def send_command(self, cmd: Command):
"""Send a command if connection has been established.
Args:
cmd (Command)
"""
run_as_background_process("send-cmd",
self._async_send_command,
cmd,
bg_start_span=False)
async def _async_send_command(self, cmd: Command):
"""Encode a replication command and send it over our outbound connection"""
string = "%s %s" % (cmd.NAME, cmd.to_line())
if "\n" in string:
raise Exception("Unexpected newline in command: %r", string)
encoded_string = string.encode("utf-8")
# We use "redis" as the name here as we don't have 1:1 connections to
# remote instances.
tcp_outbound_commands_counter.labels(cmd.NAME, "redis").inc()
await make_deferred_yieldable(
self.synapse_outbound_redis_connection.publish(
self.synapse_stream_name, encoded_string))