2011-10-26 08:48:16 +02:00
|
|
|
/**
|
|
|
|
|
* Copyright (C) 2009-2011 Typesafe Inc. <http://www.typesafe.com>
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
package akka.remote
|
|
|
|
|
|
|
|
|
|
import akka.actor._
|
|
|
|
|
import akka.actor.Status._
|
2011-10-27 12:46:10 +02:00
|
|
|
import akka.event.Logging
|
2011-11-24 14:56:19 +01:00
|
|
|
import akka.util.Duration
|
2011-11-25 12:02:25 +01:00
|
|
|
import akka.config.ConfigurationException
|
|
|
|
|
import akka.serialization.SerializationExtension
|
|
|
|
|
|
2011-10-26 08:48:16 +02:00
|
|
|
import java.util.concurrent.atomic.AtomicReference
|
2011-11-30 10:20:57 +01:00
|
|
|
import java.util.concurrent.TimeUnit.SECONDS
|
2011-10-27 15:14:15 +02:00
|
|
|
import java.security.SecureRandom
|
2011-10-26 08:48:16 +02:00
|
|
|
import System.{ currentTimeMillis ⇒ newTimestamp }
|
2011-11-25 12:02:25 +01:00
|
|
|
|
2011-10-26 08:48:16 +02:00
|
|
|
import scala.collection.immutable.Map
|
|
|
|
|
import scala.annotation.tailrec
|
2011-11-25 12:02:25 +01:00
|
|
|
|
2011-10-26 08:48:16 +02:00
|
|
|
import com.google.protobuf.ByteString
|
2011-12-11 20:12:55 +01:00
|
|
|
import java.util.concurrent.TimeoutException
|
2011-12-12 22:50:08 +01:00
|
|
|
import akka.dispatch.Await
|
2011-10-26 08:48:16 +02:00
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Interface for node membership change listener.
|
|
|
|
|
*/
|
|
|
|
|
trait NodeMembershipChangeListener {
|
2011-12-11 20:00:26 +01:00
|
|
|
def nodeConnected(node: ParsedTransportAddress)
|
|
|
|
|
def nodeDisconnected(node: ParsedTransportAddress)
|
2011-10-26 08:48:16 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Represents the node state of to gossip, versioned by a vector clock.
|
|
|
|
|
*/
|
|
|
|
|
case class Gossip(
|
|
|
|
|
version: VectorClock,
|
2011-12-11 20:00:26 +01:00
|
|
|
node: ParsedTransportAddress,
|
|
|
|
|
availableNodes: Set[ParsedTransportAddress] = Set.empty[ParsedTransportAddress],
|
|
|
|
|
unavailableNodes: Set[ParsedTransportAddress] = Set.empty[ParsedTransportAddress])
|
2011-10-26 08:48:16 +02:00
|
|
|
|
2011-11-25 12:02:25 +01:00
|
|
|
// ====== START - NEW GOSSIP IMPLEMENTATION ======
|
2011-10-26 08:48:16 +02:00
|
|
|
/*
|
|
|
|
|
case class Gossip(
|
|
|
|
|
version: VectorClock,
|
2011-12-11 20:00:26 +01:00
|
|
|
node: ParsedTransportAddress,
|
|
|
|
|
leader: ParsedTransportAddress, // FIXME leader is always head of 'members', so we probably don't need this field
|
2011-10-26 08:48:16 +02:00
|
|
|
members: SortedSet[Member] = SortetSet.empty[Member](Ordering.fromLessThan[String](_ > _)), // sorted set of members with their status, sorted by name
|
|
|
|
|
seen: Map[Member, VectorClock] = Map.empty[Member, VectorClock], // for ring convergence
|
2011-10-26 14:24:32 +02:00
|
|
|
pendingChanges: Option[Vector[PendingPartitioningChange]] = None, // for handoff
|
2011-10-26 08:48:16 +02:00
|
|
|
meta: Option[Map[String, Array[Byte]]] = None) // misc meta-data
|
|
|
|
|
|
2011-12-11 20:00:26 +01:00
|
|
|
case class Member(address: ParsedTransportAddress, status: MemberStatus)
|
2011-10-26 08:48:16 +02:00
|
|
|
|
|
|
|
|
sealed trait MemberStatus
|
|
|
|
|
object MemberStatus {
|
|
|
|
|
case class Joining(version: VectorClock) extends MemberStatus
|
|
|
|
|
case class Up(version: VectorClock) extends MemberStatus
|
|
|
|
|
case class Leaving(version: VectorClock) extends MemberStatus
|
|
|
|
|
case class Exiting(version: VectorClock) extends MemberStatus
|
|
|
|
|
case class Down(version: VectorClock) extends MemberStatus
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
sealed trait PendingPartitioningStatus
|
|
|
|
|
object PendingPartitioningStatus {
|
|
|
|
|
case object Complete extends PendingPartitioningStatus
|
|
|
|
|
case object Awaiting extends PendingPartitioningStatus
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// FIXME what is this?
|
|
|
|
|
type VNodeMod = AnyRef
|
|
|
|
|
|
|
|
|
|
case class PendingPartitioningChange(
|
2011-12-11 20:00:26 +01:00
|
|
|
owner: ParsedTransportAddress,
|
|
|
|
|
nextOwner: ParsedTransportAddress,
|
2011-10-26 08:48:16 +02:00
|
|
|
changes: Vector[VNodeMod],
|
|
|
|
|
status: PendingPartitioningStatus)
|
|
|
|
|
*/
|
2011-11-25 12:02:25 +01:00
|
|
|
// ====== END - NEW GOSSIP IMPLEMENTATION ======
|
2011-10-26 08:48:16 +02:00
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* This module is responsible for Gossiping cluster information. The abstraction maintains the list of live
|
|
|
|
|
* and dead nodes. Periodically i.e. every 1 second this module chooses a random node and initiates a round
|
|
|
|
|
* of Gossip with it. Whenever it gets gossip updates it updates the Failure Detector with the liveness
|
|
|
|
|
* information.
|
|
|
|
|
* <p/>
|
|
|
|
|
* During each of these runs the node initiates gossip exchange according to following rules (as defined in the
|
|
|
|
|
* Cassandra documentation [http://wiki.apache.org/cassandra/ArchitectureGossip]:
|
|
|
|
|
* <pre>
|
|
|
|
|
* 1) Gossip to random live node (if any)
|
|
|
|
|
* 2) Gossip to random unreachable node with certain probability depending on number of unreachable and live nodes
|
|
|
|
|
* 3) If the node gossiped to at (1) was not seed, or the number of live nodes is less than number of seeds,
|
|
|
|
|
* gossip to random seed with certain probability depending on number of unreachable, seed and live nodes.
|
|
|
|
|
* </pre>
|
|
|
|
|
*/
|
2011-12-11 20:00:26 +01:00
|
|
|
class Gossiper(remote: Remote, system: ActorSystemImpl) {
|
2011-10-26 08:48:16 +02:00
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Represents the state for this Gossiper. Implemented using optimistic lockless concurrency,
|
|
|
|
|
* all state is represented by this immutable case class and managed by an AtomicReference.
|
|
|
|
|
*/
|
|
|
|
|
private case class State(
|
|
|
|
|
currentGossip: Gossip,
|
|
|
|
|
nodeMembershipChangeListeners: Set[NodeMembershipChangeListener] = Set.empty[NodeMembershipChangeListener])
|
|
|
|
|
|
2011-12-08 14:44:05 +01:00
|
|
|
private val remoteSettings = remote.remoteSettings
|
2011-12-11 20:00:26 +01:00
|
|
|
private val serialization = remote.serialization
|
2011-11-18 11:59:43 +01:00
|
|
|
private val log = Logging(system, "Gossiper")
|
2011-10-26 08:48:16 +02:00
|
|
|
private val failureDetector = remote.failureDetector
|
2011-12-11 20:00:26 +01:00
|
|
|
private val connectionManager = new RemoteConnectionManager(system, remote, Map.empty[ParsedTransportAddress, ActorRef])
|
2011-11-25 12:02:25 +01:00
|
|
|
|
|
|
|
|
private val seeds = {
|
2011-12-11 20:00:26 +01:00
|
|
|
val seeds = remoteSettings.SeedNodes flatMap {
|
|
|
|
|
case x: UnparsedTransportAddress ⇒
|
|
|
|
|
x.parse(remote.transports) match {
|
|
|
|
|
case y: ParsedTransportAddress ⇒ Some(y)
|
|
|
|
|
case _ ⇒ None
|
|
|
|
|
}
|
|
|
|
|
case _ ⇒ None
|
|
|
|
|
}
|
2011-11-25 12:02:25 +01:00
|
|
|
if (seeds.isEmpty) throw new ConfigurationException(
|
|
|
|
|
"At least one seed node must be defined in the configuration [akka.cluster.seed-nodes]")
|
|
|
|
|
else seeds
|
|
|
|
|
}
|
2011-10-26 08:48:16 +02:00
|
|
|
|
2011-11-24 16:58:23 +01:00
|
|
|
private val address = remote.remoteAddress
|
2011-10-26 08:48:16 +02:00
|
|
|
private val nodeFingerprint = address.##
|
|
|
|
|
|
2011-10-27 15:14:15 +02:00
|
|
|
private val random = SecureRandom.getInstance("SHA1PRNG")
|
2011-12-08 14:44:05 +01:00
|
|
|
private val initalDelayForGossip = remoteSettings.InitalDelayForGossip
|
|
|
|
|
private val gossipFrequency = remoteSettings.GossipFrequency
|
2011-10-26 08:48:16 +02:00
|
|
|
|
|
|
|
|
private val state = new AtomicReference[State](State(currentGossip = newGossip()))
|
|
|
|
|
|
|
|
|
|
{
|
|
|
|
|
// start periodic gossip and cluster scrutinization - default is run them every second with 1/2 second in between
|
2011-12-02 17:13:46 +01:00
|
|
|
system.scheduler.schedule(Duration(initalDelayForGossip.toSeconds, SECONDS), Duration(gossipFrequency.toSeconds, SECONDS))(initateGossip())
|
|
|
|
|
system.scheduler.schedule(Duration(initalDelayForGossip.toSeconds, SECONDS), Duration(gossipFrequency.toSeconds, SECONDS))(scrutinize())
|
2011-10-26 08:48:16 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Tell the gossiper some gossip news.
|
|
|
|
|
*/
|
|
|
|
|
@tailrec
|
|
|
|
|
final def tell(newGossip: Gossip) {
|
|
|
|
|
val gossipingNode = newGossip.node
|
|
|
|
|
|
|
|
|
|
failureDetector heartbeat gossipingNode // update heartbeat in failure detector
|
|
|
|
|
|
|
|
|
|
val oldState = state.get
|
|
|
|
|
val latestGossip = latestVersionOf(newGossip, oldState.currentGossip)
|
|
|
|
|
val oldAvailableNodes = latestGossip.availableNodes
|
|
|
|
|
val oldUnavailableNodes = latestGossip.unavailableNodes
|
|
|
|
|
|
|
|
|
|
if (!(oldAvailableNodes contains gossipingNode) && !(oldUnavailableNodes contains gossipingNode)) {
|
|
|
|
|
// we have a new node
|
|
|
|
|
val newGossip = latestGossip copy (availableNodes = oldAvailableNodes + gossipingNode)
|
|
|
|
|
val newState = oldState copy (currentGossip = incrementVersionForGossip(newGossip))
|
|
|
|
|
|
|
|
|
|
// if we won the race then update else try again
|
|
|
|
|
if (!state.compareAndSet(oldState, newState)) tell(newGossip) // recur
|
|
|
|
|
else {
|
|
|
|
|
// create connections for all new nodes in the latest gossip
|
|
|
|
|
for {
|
|
|
|
|
node ← oldAvailableNodes
|
|
|
|
|
if connectionManager.connectionFor(node).isEmpty
|
|
|
|
|
} {
|
2011-12-11 20:00:26 +01:00
|
|
|
val connectionFactory = () ⇒ system.actorFor(RootActorPath(RemoteSystemAddress(system.name, gossipingNode)) / "remote")
|
2011-10-26 08:48:16 +02:00
|
|
|
connectionManager.putIfAbsent(node, connectionFactory) // create a new remote connection to the new node
|
|
|
|
|
oldState.nodeMembershipChangeListeners foreach (_ nodeConnected node) // notify listeners about the new nodes
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
} else if (oldUnavailableNodes contains gossipingNode) {
|
|
|
|
|
// gossip from an old former dead node
|
|
|
|
|
|
|
|
|
|
val newUnavailableNodes = oldUnavailableNodes - gossipingNode
|
|
|
|
|
val newAvailableNodes = oldAvailableNodes + gossipingNode
|
|
|
|
|
|
|
|
|
|
val newGossip = latestGossip copy (availableNodes = newAvailableNodes, unavailableNodes = newUnavailableNodes)
|
|
|
|
|
val newState = oldState copy (currentGossip = incrementVersionForGossip(newGossip))
|
|
|
|
|
|
|
|
|
|
// if we won the race then update else try again
|
|
|
|
|
if (!state.compareAndSet(oldState, newState)) tell(newGossip) // recur
|
|
|
|
|
else oldState.nodeMembershipChangeListeners foreach (_ nodeConnected gossipingNode) // notify listeners on successful update of state
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
@tailrec
|
|
|
|
|
final def registerListener(listener: NodeMembershipChangeListener) {
|
|
|
|
|
val oldState = state.get
|
|
|
|
|
val newListeners = oldState.nodeMembershipChangeListeners + listener
|
|
|
|
|
val newState = oldState copy (nodeMembershipChangeListeners = newListeners)
|
|
|
|
|
if (!state.compareAndSet(oldState, newState)) registerListener(listener) // recur
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
@tailrec
|
|
|
|
|
final def unregisterListener(listener: NodeMembershipChangeListener) {
|
|
|
|
|
val oldState = state.get
|
|
|
|
|
val newListeners = oldState.nodeMembershipChangeListeners - listener
|
|
|
|
|
val newState = oldState copy (nodeMembershipChangeListeners = newListeners)
|
|
|
|
|
if (!state.compareAndSet(oldState, newState)) unregisterListener(listener) // recur
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Initates a new round of gossip.
|
|
|
|
|
*/
|
|
|
|
|
private def initateGossip() {
|
|
|
|
|
val oldState = state.get
|
|
|
|
|
val oldGossip = oldState.currentGossip
|
|
|
|
|
|
|
|
|
|
val oldAvailableNodes = oldGossip.availableNodes
|
|
|
|
|
val oldUnavailableNodes = oldGossip.unavailableNodes
|
|
|
|
|
|
|
|
|
|
val oldAvailableNodesSize = oldAvailableNodes.size
|
|
|
|
|
val oldUnavailableNodesSize = oldUnavailableNodes.size
|
|
|
|
|
|
|
|
|
|
// 1. gossip to alive nodes
|
|
|
|
|
val gossipedToSeed =
|
|
|
|
|
if (oldAvailableNodesSize > 0) gossipTo(oldAvailableNodes)
|
|
|
|
|
else false
|
|
|
|
|
|
|
|
|
|
// 2. gossip to dead nodes
|
|
|
|
|
if (oldUnavailableNodesSize > 0) {
|
|
|
|
|
val probability: Double = oldUnavailableNodesSize / (oldAvailableNodesSize + 1)
|
|
|
|
|
if (random.nextDouble() < probability) gossipTo(oldUnavailableNodes)
|
|
|
|
|
}
|
|
|
|
|
|
2011-10-27 15:14:15 +02:00
|
|
|
// 3. gossip to a seed for facilitating partition healing
|
|
|
|
|
if ((!gossipedToSeed || oldAvailableNodesSize < 1) && (seeds.head != address)) {
|
|
|
|
|
if (oldAvailableNodesSize == 0) gossipTo(seeds)
|
|
|
|
|
else {
|
|
|
|
|
val probability = 1.0 / oldAvailableNodesSize + oldUnavailableNodesSize
|
|
|
|
|
if (random.nextDouble() <= probability) gossipTo(seeds)
|
2011-10-26 08:48:16 +02:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Gossips set of nodes passed in as argument. Returns 'true' if it gossiped to a "seed" node.
|
|
|
|
|
*/
|
2011-12-11 20:00:26 +01:00
|
|
|
private def gossipTo(nodes: Set[ParsedTransportAddress]): Boolean = {
|
2011-10-26 08:48:16 +02:00
|
|
|
val peers = nodes filter (_ != address) // filter out myself
|
|
|
|
|
val peer = selectRandomNode(peers)
|
|
|
|
|
val oldState = state.get
|
|
|
|
|
val oldGossip = oldState.currentGossip
|
|
|
|
|
|
|
|
|
|
val connection = connectionManager.connectionFor(peer).getOrElse(
|
|
|
|
|
throw new IllegalStateException("Connection for [" + peer + "] is not set up"))
|
|
|
|
|
|
|
|
|
|
try {
|
2011-12-14 17:26:18 +01:00
|
|
|
val t = remoteSettings.RemoteSystemDaemonAckTimeout
|
2011-12-26 18:23:55 +01:00
|
|
|
Await.result(connection ? (newGossip, t), t) match {
|
2011-12-11 20:12:55 +01:00
|
|
|
case Success(receiver) ⇒ log.debug("Gossip sent to [{}] was successfully received", receiver)
|
|
|
|
|
case Failure(cause) ⇒ log.error(cause, cause.toString)
|
2011-10-26 08:48:16 +02:00
|
|
|
}
|
|
|
|
|
} catch {
|
2011-12-11 20:12:55 +01:00
|
|
|
case e: TimeoutException ⇒ log.error(e, "Gossip to [%s] timed out".format(connection.path))
|
2011-10-26 08:48:16 +02:00
|
|
|
case e: Exception ⇒
|
2011-11-30 23:30:56 +01:00
|
|
|
log.error(e, "Could not gossip to [{}] due to: {}", connection.path, e.toString)
|
2011-10-26 08:48:16 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
seeds exists (peer == _)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Scrutinizes the cluster; marks nodes detected by the failure detector as unavailable, and notifies all listeners
|
|
|
|
|
* of the change in the cluster membership.
|
|
|
|
|
*/
|
|
|
|
|
@tailrec
|
|
|
|
|
final private def scrutinize() {
|
|
|
|
|
val oldState = state.get
|
|
|
|
|
val oldGossip = oldState.currentGossip
|
|
|
|
|
|
|
|
|
|
val oldAvailableNodes = oldGossip.availableNodes
|
|
|
|
|
val oldUnavailableNodes = oldGossip.unavailableNodes
|
2011-10-27 15:14:15 +02:00
|
|
|
val newlyDetectedUnavailableNodes = oldAvailableNodes filterNot failureDetector.isAvailable
|
2011-10-26 08:48:16 +02:00
|
|
|
|
|
|
|
|
if (!newlyDetectedUnavailableNodes.isEmpty) { // we have newly detected nodes marked as unavailable
|
|
|
|
|
val newAvailableNodes = oldAvailableNodes diff newlyDetectedUnavailableNodes
|
|
|
|
|
val newUnavailableNodes = oldUnavailableNodes ++ newlyDetectedUnavailableNodes
|
|
|
|
|
|
|
|
|
|
val newGossip = oldGossip copy (availableNodes = newAvailableNodes, unavailableNodes = newUnavailableNodes)
|
|
|
|
|
val newState = oldState copy (currentGossip = incrementVersionForGossip(newGossip))
|
|
|
|
|
|
|
|
|
|
// if we won the race then update else try again
|
|
|
|
|
if (!state.compareAndSet(oldState, newState)) scrutinize() // recur
|
|
|
|
|
else {
|
|
|
|
|
// notify listeners on successful update of state
|
|
|
|
|
for {
|
|
|
|
|
deadNode ← newUnavailableNodes
|
|
|
|
|
listener ← oldState.nodeMembershipChangeListeners
|
|
|
|
|
} listener nodeDisconnected deadNode
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private def newGossip(): Gossip = Gossip(
|
|
|
|
|
version = VectorClock(),
|
2011-12-11 20:00:26 +01:00
|
|
|
node = address.transport,
|
|
|
|
|
availableNodes = Set(address.transport))
|
2011-10-26 08:48:16 +02:00
|
|
|
|
|
|
|
|
private def incrementVersionForGossip(from: Gossip): Gossip = {
|
|
|
|
|
val newVersion = from.version.increment(nodeFingerprint, newTimestamp)
|
|
|
|
|
from copy (version = newVersion)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private def latestVersionOf(newGossip: Gossip, oldGossip: Gossip): Gossip = {
|
|
|
|
|
(newGossip.version compare oldGossip.version) match {
|
|
|
|
|
case VectorClock.After ⇒ newGossip // gossiped version is newer, use new version
|
|
|
|
|
case VectorClock.Before ⇒ oldGossip // gossiped version is older, use old version
|
|
|
|
|
case VectorClock.Concurrent ⇒ oldGossip // can't establish a causal relationship between two versions => conflict
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-12-11 20:00:26 +01:00
|
|
|
private def selectRandomNode(nodes: Set[ParsedTransportAddress]): ParsedTransportAddress = {
|
2011-10-26 08:48:16 +02:00
|
|
|
nodes.toList(random.nextInt(nodes.size))
|
|
|
|
|
}
|
|
|
|
|
}
|