org.apache.cassandra.net.MessagingService Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of cassandra-all Show documentation
Show all versions of cassandra-all Show documentation
The Apache Cassandra Project develops a highly scalable second-generation distributed database, bringing together Dynamo's fully distributed design and Bigtable's ColumnFamily-based data model.
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you 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.
*/
package org.apache.cassandra.net;
import java.io.IOException;
import java.nio.channels.ClosedChannelException;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Collections;
import java.util.List;
import java.util.Map;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.TimeoutException;
import java.util.stream.Collectors;
import com.google.common.annotations.VisibleForTesting;
import com.google.common.collect.Lists;
import io.netty.util.concurrent.Future; //checkstyle: permit this import
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.apache.cassandra.concurrent.ScheduledExecutors;
import org.apache.cassandra.concurrent.Stage;
import org.apache.cassandra.config.DatabaseDescriptor;
import org.apache.cassandra.db.SystemKeyspace;
import org.apache.cassandra.exceptions.RequestFailureReason;
import org.apache.cassandra.locator.InetAddressAndPort;
import org.apache.cassandra.locator.Replica;
import org.apache.cassandra.metrics.MessagingMetrics;
import org.apache.cassandra.service.AbstractWriteResponseHandler;
import org.apache.cassandra.utils.ExecutorUtils;
import org.apache.cassandra.utils.FBUtilities;
import org.apache.cassandra.utils.concurrent.AsyncPromise;
import org.apache.cassandra.utils.concurrent.FutureCombiner;
import static java.util.Collections.synchronizedList;
import static java.util.concurrent.TimeUnit.MINUTES;
import static org.apache.cassandra.concurrent.Stage.MUTATION;
import static org.apache.cassandra.config.CassandraRelevantProperties.NON_GRACEFUL_SHUTDOWN;
import static org.apache.cassandra.utils.Clock.Global.nanoTime;
import static org.apache.cassandra.utils.Throwables.maybeFail;
/**
* MessagingService implements all internode communication - with the exception of SSTable streaming (for now).
*
* Specifically, it's responsible for dispatch of outbound messages to other nodes and routing of inbound messages
* to their appropriate {@link IVerbHandler}.
*
* Using MessagingService: sending requests and responses
*
* The are two ways to send a {@link Message}, and you should pick one depending on the desired behaviour:
* 1. To send a request that expects a response back, use
* {@link #sendWithCallback(Message, InetAddressAndPort, RequestCallback)} method. Once a response
* message is received, {@link RequestCallback#onResponse(Message)} method will be invoked on the
* provided callback - in case of a success response. In case of a failure response (see {@link Verb#FAILURE_RSP}),
* or if a response doesn't arrive within verb's configured expiry time,
* {@link RequestCallback#onFailure(InetAddressAndPort, RequestFailureReason)} will be invoked instead.
* 2. To send a response back, or a message that expects no response, use {@link #send(Message, InetAddressAndPort)}
* method.
*
* See also: {@link Message#out(Verb, Object)}, {@link Message#responseWith(Object)},
* and {@link Message#failureResponse(RequestFailureReason)}.
*
* Using MessagingService: handling a request
*
* As described in the previous section, to handle responses you only need to implement {@link RequestCallback}
* interface - so long as your response verb handler is the default {@link ResponseVerbHandler}.
*
* There are two steps you need to perform to implement request handling:
* 1. Create a {@link IVerbHandler} to process incoming requests and responses for the new type (if applicable).
* 2. Add a new {@link Verb} to the enum for the new request type, and, if applicable, one for the response message.
*
* MessagingService will now automatically invoke your handler whenever a {@link Message} with this verb arrives.
*
* Architecture of MessagingService
*
* QOS
*
* Since our messaging protocol is TCP-based, and also doesn't yet support interleaving messages with each other,
* we need a way to prevent head-of-line blocking adversely affecting all messages - in particular, large messages
* being in the way of smaller ones. To achive that (somewhat), we maintain three messaging connections to and
* from each peer:
* - one for large messages - defined as being larger than {@link OutboundConnections#LARGE_MESSAGE_THRESHOLD}
* (65KiB by default)
* - one for small messages - defined as smaller than that threshold
* - and finally, a connection for urgent messages - usually small and/or that are important to arrive
* promptly, e.g. gossip-related ones
*
* Wire format and framing
*
* Small messages are grouped together into frames, and large messages are split over multiple frames.
* Framing provides application-level integrity protection to otherwise raw streams of data - we use
* CRC24 for frame headers and CRC32 for the entire payload. LZ4 is optionally used for compression.
*
* You can find the on-wire format description of individual messages in the comments for
* {@link Message.Serializer}, alongside with format evolution notes.
* For the list and descriptions of available frame decoders see {@link FrameDecoder} comments. You can
* find wire format documented in the javadoc of {@link FrameDecoder} implementations:
* see {@link FrameDecoderCrc} and {@link FrameDecoderLZ4} in particular.
*
* Architecture of outbound messaging
*
* {@link OutboundConnection} is the core class implementing outbound connection logic, with
* {@link OutboundConnection#enqueue(Message)} being its main entry point. The connections are initiated
* by {@link OutboundConnectionInitiator}.
*
* Netty pipeline for outbound messaging connections generally consists of the following handlers:
*
* [(optional) SslHandler] <- [FrameEncoder]
*
* {@link OutboundConnection} handles the entire lifetime of a connection: from the very first handshake
* to any necessary reconnects if necessary.
*
* Message-delivery flow varies depending on the connection type.
*
* For {@link ConnectionType#SMALL_MESSAGES} and {@link ConnectionType#URGENT_MESSAGES},
* {@link Message} serialization and delivery occurs directly on the event loop.
* See {@link OutboundConnection.EventLoopDelivery} for details.
*
* For {@link ConnectionType#LARGE_MESSAGES}, to ensure that servicing large messages doesn't block
* timely service of other requests, message serialization is offloaded to a companion thread pool
* ({@link SocketFactory#synchronousWorkExecutor}). Most of the work will be performed by
* {@link AsyncChannelOutputPlus}. Please see {@link OutboundConnection.LargeMessageDelivery}
* for details.
*
* To prevent fast clients, or slow nodes on the other end of the connection from overwhelming
* a host with enqueued, unsent messages on heap, we impose strict limits on how much memory enqueued,
* undelivered messages can claim.
*
* Every individual connection gets an exclusive permit quota to use - 4MiB by default; every endpoint
* (group of large, small, and urgent connection) is capped at, by default, at 128MiB of undelivered messages,
* and a global limit of 512MiB is imposed on all endpoints combined.
*
* On an attempt to {@link OutboundConnection#enqueue(Message)}, the connection will attempt to allocate
* permits for message-size number of bytes from its exclusive quota; if successful, it will add the
* message to the queue; if unsuccessful, it will need to allocate remainder from both endpoint and lobal
* reserves, and if it fails to do so, the message will be rejected, and its callbacks, if any,
* immediately expired.
*
* For a more detailed description please see the docs and comments of {@link OutboundConnection}.
*
* Architecture of inbound messaging
*
* {@link InboundMessageHandler} is the core class implementing inbound connection logic, paired
* with {@link FrameDecoder}. Inbound connections are initiated by {@link InboundConnectionInitiator}.
* The primary entry points to these classes are {@link FrameDecoder#channelRead(ShareableBytes)}
* and {@link InboundMessageHandler#process(FrameDecoder.Frame)}.
*
* Netty pipeline for inbound messaging connections generally consists of the following handlers:
*
* [(optional) SslHandler] -> [FrameDecoder] -> [InboundMessageHandler]
*
* {@link FrameDecoder} is responsible for decoding incoming frames and work stashing; {@link InboundMessageHandler}
* then takes decoded frames from the decoder and processes the messages contained in them.
*
* The flow differs between small and large messages. Small ones are deserialized immediately, and only
* then scheduled on the right thread pool for the {@link Verb} for execution. Large messages, OTOH,
* aren't deserialized until they are just about to be executed on the appropriate {@link Stage}.
*
* Similarly to outbound handling, inbound messaging imposes strict memory utilisation limits on individual
* endpoints and on global aggregate consumption, and implements simple flow control, to prevent a single
* fast endpoint from overwhelming a host.
*
* Every individual connection gets an exclusive permit quota to use - 4MiB by default; every endpoint
* (group of large, small, and urgent connection) is capped at, by default, at 128MiB of unprocessed messages,
* and a global limit of 512MiB is imposed on all endpoints combined.
*
* On arrival of a message header, the handler will attempt to allocate permits for message-size number
* of bytes from its exclusive quota; if successful, it will proceed to deserializing and processing the message.
* If unsuccessful, the handler will attempt to allocate the remainder from its endpoint and global reserve;
* if either allocation is unsuccessful, the handler will cease any further frame processing, and tell
* {@link FrameDecoder} to stop reading from the network; subsequently, it will put itself on a special
* {@link org.apache.cassandra.net.InboundMessageHandler.WaitQueue}, to be reactivated once more permits
* become available.
*
* For a more detailed description please see the docs and comments of {@link InboundMessageHandler} and
* {@link FrameDecoder}.
*
* Observability
*
* MessagingService exposes diagnostic counters for both outbound and inbound directions - received and sent
* bytes and message counts, overload bytes and message count, error bytes and error counts, and many more.
*
* See {@link org.apache.cassandra.metrics.InternodeInboundMetrics} and
* {@link org.apache.cassandra.metrics.InternodeOutboundMetrics} for JMX-exposed counters.
*
* We also provide {@code system_views.internode_inbound} and {@code system_views.internode_outbound} virtual tables -
* implemented in {@link org.apache.cassandra.db.virtual.InternodeInboundTable} and
* {@link org.apache.cassandra.db.virtual.InternodeOutboundTable} respectively.
*/
public class MessagingService extends MessagingServiceMBeanImpl
{
private static final Logger logger = LoggerFactory.getLogger(MessagingService.class);
// 8 bits version, so don't waste versions
@Deprecated
public static final int VERSION_30 = 10;
@Deprecated
public static final int VERSION_3014 = 11;
public static final int VERSION_40 = 12;
public static final int VERSION_50 = 13; // c14227 TTL overflow, 'uint' timestamps
public static final int minimum_version = VERSION_40;
public static final int current_version = DatabaseDescriptor.getStorageCompatibilityMode().isBefore(5) ? VERSION_40 : VERSION_50;
static AcceptVersions accept_messaging = new AcceptVersions(minimum_version, current_version);
static AcceptVersions accept_streaming = new AcceptVersions(current_version, current_version);
static Map versionOrdinalMap = Arrays.stream(Version.values()).collect(Collectors.toMap(v -> v.value, v -> v.ordinal()));
/**
* This is an optimisation to speed up the translation of the serialization
* version to the {@link Version} enum ordinal.
*
* @param version the serialization version
* @return a {@link Version} ordinal value
*/
public static int getVersionOrdinal(int version)
{
Integer ordinal = versionOrdinalMap.get(version);
if (ordinal == null)
throw new IllegalStateException("Unkown serialization version: " + version);
return ordinal;
}
public enum Version
{
@Deprecated
VERSION_30(10),
@Deprecated
VERSION_3014(11),
VERSION_40(12),
VERSION_50(13);
public final int value;
Version(int value)
{
this.value = value;
}
public static List supportedVersions()
{
List versions = Lists.newArrayList();
for (Version version : values())
if (minimum_version <= version.value)
versions.add(version);
return Collections.unmodifiableList(versions);
}
}
private static class MSHandle
{
public static final MessagingService instance = new MessagingService(false);
}
public static MessagingService instance()
{
return MSHandle.instance;
}
public final SocketFactory socketFactory = new SocketFactory();
public final LatencySubscribers latencySubscribers = new LatencySubscribers();
public final RequestCallbacks callbacks = new RequestCallbacks(this);
// a public hook for filtering messages intended for delivery to this node
public final InboundSink inboundSink = new InboundSink(this);
// the inbound global reserve limits and associated wait queue
private final InboundMessageHandlers.GlobalResourceLimits inboundGlobalReserveLimits = new InboundMessageHandlers.GlobalResourceLimits(
new ResourceLimits.Concurrent(DatabaseDescriptor.getInternodeApplicationReceiveQueueReserveGlobalCapacityInBytes()));
// the socket bindings we accept incoming connections on
private final InboundSockets inboundSockets = new InboundSockets(new InboundConnectionSettings()
.withHandlers(this::getInbound)
.withSocketFactory(socketFactory));
// a public hook for filtering messages intended for delivery to another node
public final OutboundSink outboundSink = new OutboundSink(this::doSend);
final ResourceLimits.Limit outboundGlobalReserveLimit =
new ResourceLimits.Concurrent(DatabaseDescriptor.getInternodeApplicationSendQueueReserveGlobalCapacityInBytes());
private volatile boolean isShuttingDown;
@VisibleForTesting
MessagingService(boolean testOnly)
{
this(testOnly, new EndpointMessagingVersions(), new MessagingMetrics());
}
@VisibleForTesting
MessagingService(boolean testOnly, EndpointMessagingVersions versions, MessagingMetrics metrics)
{
super(testOnly, versions, metrics);
OutboundConnections.scheduleUnusedConnectionMonitoring(this, ScheduledExecutors.scheduledTasks, 1L, TimeUnit.HOURS);
}
public org.apache.cassandra.utils.concurrent.Future> sendWithResult(Message message, InetAddressAndPort to)
{
AsyncPromise> promise = new AsyncPromise<>();
MessagingService.instance().sendWithCallback(message, to, new RequestCallback()
{
@Override
public void onResponse(Message msg)
{
promise.trySuccess(msg);
}
@Override
public void onFailure(InetAddressAndPort from, RequestFailureReason failureReason)
{
promise.tryFailure(new FailureResponseException(from, failureReason));
}
@Override
public boolean invokeOnFailure()
{
return true;
}
});
return promise;
}
public static class FailureResponseException extends IOException
{
private final InetAddressAndPort from;
private final RequestFailureReason failureReason;
public FailureResponseException(InetAddressAndPort from, RequestFailureReason failureReason)
{
super(String.format("Failure from %s: %s", from, failureReason.name()));
this.from = from;
this.failureReason = failureReason;
}
public InetAddressAndPort from()
{
return from;
}
public RequestFailureReason failureReason()
{
return failureReason;
}
}
/**
* Send a non-mutation message to a given endpoint. This method specifies a callback
* which is invoked with the actual response.
*
* @param message message to be sent.
* @param to endpoint to which the message needs to be sent
* @param cb callback interface which is used to pass the responses or
* suggest that a timeout occurred to the invoker of the send().
*/
public void sendWithCallback(Message message, InetAddressAndPort to, RequestCallback cb)
{
sendWithCallback(message, to, cb, null);
}
public void sendWithCallback(Message message, InetAddressAndPort to, RequestCallback cb, ConnectionType specifyConnection)
{
callbacks.addWithExpiration(cb, message, to);
if (cb.invokeOnFailure() && !message.callBackOnFailure())
message = message.withCallBackOnFailure();
send(message, to, specifyConnection);
}
/**
* Send a mutation message or a Paxos Commit to a given endpoint. This method specifies a callback
* which is invoked with the actual response.
* Also holds the message (only mutation messages) to determine if it
* needs to trigger a hint (uses StorageProxy for that).
*
* @param message message to be sent.
* @param to endpoint to which the message needs to be sent
* @param handler callback interface which is used to pass the responses or
* suggest that a timeout occurred to the invoker of the send().
*/
public void sendWriteWithCallback(Message message, Replica to, AbstractWriteResponseHandler handler)
{
assert message.callBackOnFailure();
callbacks.addWithExpiration(handler, message, to);
send(message, to.endpoint(), null);
}
/**
* Send a message to a given endpoint. This method adheres to the fire and forget
* style messaging.
*
* @param message messages to be sent.
* @param to endpoint to which the message needs to be sent
*/
public void send(Message message, InetAddressAndPort to)
{
send(message, to, null);
}
/**
* Send a message to a given endpoint. This method adheres to the fire and forget
* style messaging.
*
* @param message messages to be sent.
* @param response
*/
public void respond(V response, Message message)
{
send(message.responseWith(response), message.respondTo());
}
public void respondWithFailure(RequestFailureReason reason, Message message)
{
send(Message.failureResponse(message.id(), message.expiresAtNanos(), reason), message.respondTo());
}
public void send(Message message, InetAddressAndPort to, ConnectionType specifyConnection)
{
if (logger.isTraceEnabled())
{
logger.trace("{} sending {} to {}@{}", FBUtilities.getBroadcastAddressAndPort(), message.verb(), message.id(), to);
if (to.equals(FBUtilities.getBroadcastAddressAndPort()))
logger.trace("Message-to-self {} going over MessagingService", message);
}
outboundSink.accept(message, to, specifyConnection);
}
private void doSend(Message message, InetAddressAndPort to, ConnectionType specifyConnection)
{
// expire the callback if the message failed to enqueue (failed to establish a connection or exceeded queue capacity)
while (true)
{
OutboundConnections connections = getOutbound(to);
try
{
connections.enqueue(message, specifyConnection);
return;
}
catch (ClosedChannelException e)
{
if (isShuttingDown)
return; // just drop the message, and let others clean up
// remove the connection and try again
channelManagers.remove(to, connections);
}
}
}
void markExpiredCallback(InetAddressAndPort addr)
{
OutboundConnections conn = channelManagers.get(addr);
if (conn != null)
conn.incrementExpiredCallbackCount();
}
/**
* Only to be invoked once we believe the endpoint will never be contacted again.
*
* We close the connection after a five minute delay, to give asynchronous operations a chance to terminate
*/
public void closeOutbound(InetAddressAndPort to)
{
OutboundConnections pool = channelManagers.get(to);
if (pool != null)
pool.scheduleClose(5L, MINUTES, true)
.addListener(future -> channelManagers.remove(to, pool));
}
/**
* Only to be invoked once we believe the connections will never be used again.
*/
void closeOutboundNow(OutboundConnections connections)
{
connections.close(true).addListener(
future -> channelManagers.remove(connections.template().to, connections)
);
}
/**
* Only to be invoked once we believe the connections will never be used again.
*/
public void removeInbound(InetAddressAndPort from)
{
InboundMessageHandlers handlers = messageHandlers.remove(from);
if (null != handlers)
handlers.releaseMetrics();
}
/**
* Closes any current open channel/connection to the endpoint, but does not cause any message loss, and we will
* try to re-establish connections immediately
*/
public void interruptOutbound(InetAddressAndPort to)
{
OutboundConnections pool = channelManagers.get(to);
if (pool != null)
{
pool.interrupt();
logger.info("Interrupted outbound connections to {}", to);
}
}
/**
* Reconnect to the peer using the given {@code addr}. Outstanding messages in each channel will be sent on the
* current channel. Typically this function is used for something like EC2 public IP addresses which need to be used
* for communication between EC2 regions.
*
* @param address IP Address to identify the peer
* @param preferredAddress IP Address to use (and prefer) going forward for connecting to the peer
*/
@SuppressWarnings("UnusedReturnValue")
public Future maybeReconnectWithNewIp(InetAddressAndPort address, InetAddressAndPort preferredAddress)
{
if (!SystemKeyspace.updatePreferredIP(address, preferredAddress))
return null;
OutboundConnections messagingPool = channelManagers.get(address);
if (messagingPool != null)
return messagingPool.reconnectWithNewIp(preferredAddress);
return null;
}
/**
* Wait for callbacks and don't allow anymore to be created (since they could require writing hints)
*/
public void shutdown()
{
if (NON_GRACEFUL_SHUTDOWN.getBoolean())
// this branch is used in unit-tests when we really never restart a node and shutting down means the end of test
shutdownAbrubtly();
else
shutdown(1L, MINUTES, true, true);
}
public void shutdown(long timeout, TimeUnit units, boolean shutdownGracefully, boolean shutdownExecutors)
{
logger.debug("Shutting down: timeout={}s, gracefully={}, shutdownExecutors={}", units.toSeconds(timeout), shutdownGracefully, shutdownExecutors);
if (isShuttingDown)
{
logger.info("Shutdown was already called");
return;
}
isShuttingDown = true;
logger.info("Waiting for messaging service to quiesce");
// We may need to schedule hints on the mutation stage, so it's erroneous to shut down the mutation stage first
assert !MUTATION.executor().isShutdown();
if (shutdownGracefully)
{
callbacks.shutdownGracefully();
List> closing = new ArrayList<>();
for (OutboundConnections pool : channelManagers.values())
closing.add(pool.close(true));
long deadline = nanoTime() + units.toNanos(timeout);
maybeFail(() -> FutureCombiner.nettySuccessListener(closing).get(timeout, units),
() -> {
List inboundExecutors = new ArrayList<>();
inboundSockets.close(synchronizedList(inboundExecutors)::add).get();
ExecutorUtils.awaitTermination(timeout, units, inboundExecutors);
},
() -> {
if (shutdownExecutors)
shutdownExecutors(deadline);
},
() -> callbacks.awaitTerminationUntil(deadline),
inboundSink::clear,
outboundSink::clear);
}
else
{
callbacks.shutdownNow(false);
List> closing = new ArrayList<>();
List inboundExecutors = synchronizedList(new ArrayList());
closing.add(inboundSockets.close(inboundExecutors::add));
for (OutboundConnections pool : channelManagers.values())
closing.add(pool.close(false));
long deadline = nanoTime() + units.toNanos(timeout);
maybeFail(() -> FutureCombiner.nettySuccessListener(closing).get(timeout, units),
() -> {
if (shutdownExecutors)
shutdownExecutors(deadline);
},
() -> ExecutorUtils.awaitTermination(timeout, units, inboundExecutors),
() -> callbacks.awaitTerminationUntil(deadline),
inboundSink::clear,
outboundSink::clear);
}
}
public void shutdownAbrubtly()
{
logger.debug("Shutting down abruptly");
if (isShuttingDown)
{
logger.info("Shutdown was already called");
return;
}
isShuttingDown = true;
logger.info("Waiting for messaging service to quiesce");
// We may need to schedule hints on the mutation stage, so it's erroneous to shut down the mutation stage first
assert !MUTATION.executor().isShutdown();
callbacks.shutdownNow(false);
inboundSockets.close();
for (OutboundConnections pool : channelManagers.values())
pool.close(false);
maybeFail(socketFactory::shutdownNow,
inboundSink::clear,
outboundSink::clear);
}
private void shutdownExecutors(long deadlineNanos) throws TimeoutException, InterruptedException
{
socketFactory.shutdownNow();
socketFactory.awaitTerminationUntil(deadlineNanos);
}
private OutboundConnections getOutbound(InetAddressAndPort to)
{
OutboundConnections connections = channelManagers.get(to);
if (connections == null)
connections = OutboundConnections.tryRegister(channelManagers, to, new OutboundConnectionSettings(to).withDefaults(ConnectionCategory.MESSAGING));
return connections;
}
InboundMessageHandlers getInbound(InetAddressAndPort from)
{
InboundMessageHandlers handlers = messageHandlers.get(from);
if (null != handlers)
return handlers;
return messageHandlers.computeIfAbsent(from, addr ->
new InboundMessageHandlers(FBUtilities.getLocalAddressAndPort(),
addr,
DatabaseDescriptor.getInternodeApplicationReceiveQueueCapacityInBytes(),
DatabaseDescriptor.getInternodeApplicationReceiveQueueReserveEndpointCapacityInBytes(),
inboundGlobalReserveLimits, metrics, inboundSink)
);
}
@VisibleForTesting
boolean isConnected(InetAddressAndPort address, Message messageOut)
{
OutboundConnections pool = channelManagers.get(address);
if (pool == null)
return false;
return pool.connectionFor(messageOut).isConnected();
}
public void listen()
{
inboundSockets.open();
}
public void waitUntilListening() throws InterruptedException
{
inboundSockets.open().await();
}
}
© 2015 - 2024 Weber Informatics LLC | Privacy Policy