/* * Copyright (c) 2014 Cisco Systems, Inc. and others. All rights reserved. * * This program and the accompanying materials are made available under the * terms of the Eclipse Public License v1.0 which accompanies this distribution, * and is available at http://www.eclipse.org/legal/epl-v10.html */ package org.opendaylight.controller.cluster.raft; import akka.actor.ActorRef; import akka.actor.ActorSelection; import akka.event.Logging; import akka.event.LoggingAdapter; import akka.japi.Procedure; import akka.persistence.RecoveryCompleted; import akka.persistence.SaveSnapshotFailure; import akka.persistence.SaveSnapshotSuccess; import akka.persistence.SnapshotOffer; import akka.persistence.SnapshotSelectionCriteria; import akka.persistence.UntypedPersistentActor; import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot; import org.opendaylight.controller.cluster.raft.base.messages.ApplyState; import org.opendaylight.controller.cluster.raft.base.messages.Replicate; import org.opendaylight.controller.cluster.raft.behaviors.Candidate; import org.opendaylight.controller.cluster.raft.behaviors.Follower; import org.opendaylight.controller.cluster.raft.behaviors.Leader; import org.opendaylight.controller.cluster.raft.behaviors.RaftActorBehavior; import org.opendaylight.controller.cluster.raft.client.messages.AddRaftPeer; import org.opendaylight.controller.cluster.raft.client.messages.FindLeader; import org.opendaylight.controller.cluster.raft.client.messages.FindLeaderReply; import org.opendaylight.controller.cluster.raft.client.messages.RemoveRaftPeer; import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload; import java.io.Serializable; import java.util.ArrayList; import java.util.List; import java.util.Map; /** * RaftActor encapsulates a state machine that needs to be kept synchronized * in a cluster. It implements the RAFT algorithm as described in the paper * * In Search of an Understandable Consensus Algorithm *

* RaftActor has 3 states and each state has a certain behavior associated * with it. A Raft actor can behave as, *

*

*

* A RaftActor MUST be a Leader in order to accept requests from clients to * change the state of it's encapsulated state machine. Once a RaftActor becomes * a Leader it is also responsible for ensuring that all followers ultimately * have the same log and therefore the same state machine as itself. *

*

* The current behavior of a RaftActor determines how election for leadership * is initiated and how peer RaftActors react to request for votes. *

*

* Each RaftActor also needs to know the current election term. It uses this * information for a couple of things. One is to simply figure out who it * voted for in the last election. Another is to figure out if the message * it received to update it's state is stale. *

*

* The RaftActor uses akka-persistence to store it's replicated log. * Furthermore through it's behaviors a Raft Actor determines *

*

*/ public abstract class RaftActor extends UntypedPersistentActor { protected final LoggingAdapter LOG = Logging.getLogger(getContext().system(), this); /** * The current state determines the current behavior of a RaftActor * A Raft Actor always starts off in the Follower State */ private RaftActorBehavior currentBehavior; /** * This context should NOT be passed directly to any other actor it is * only to be consumed by the RaftActorBehaviors */ private RaftActorContext context; /** * The in-memory journal */ private ReplicatedLogImpl replicatedLog = new ReplicatedLogImpl(); public RaftActor(String id, Map peerAddresses) { context = new RaftActorContextImpl(this.getSelf(), this.getContext(), id, new ElectionTermImpl(), -1, -1, replicatedLog, peerAddresses, LOG); } @Override public void onReceiveRecover(Object message) { if (message instanceof SnapshotOffer) { SnapshotOffer offer = (SnapshotOffer) message; Snapshot snapshot = (Snapshot) offer.snapshot(); // Create a replicated log with the snapshot information // The replicated log can be used later on to retrieve this snapshot // when we need to install it on a peer replicatedLog = new ReplicatedLogImpl(snapshot); // Apply the snapshot to the actors state applySnapshot(snapshot.getState()); } else if (message instanceof ReplicatedLogEntry) { replicatedLog.append((ReplicatedLogEntry) message); } else if (message instanceof DeleteEntries) { replicatedLog.removeFrom(((DeleteEntries) message).getFromIndex()); } else if (message instanceof UpdateElectionTerm) { context.getTermInformation().update(((UpdateElectionTerm) message).getCurrentTerm(), ((UpdateElectionTerm) message).getVotedFor()); } else if (message instanceof RecoveryCompleted) { LOG.debug( "Last index in log : " + replicatedLog.lastIndex()); currentBehavior = switchBehavior(RaftState.Follower); } } @Override public void onReceiveCommand(Object message) { if (message instanceof ApplyState){ ApplyState applyState = (ApplyState) message; LOG.debug("Applying state for log index {} data {}", applyState.getReplicatedLogEntry().getIndex(), applyState.getReplicatedLogEntry().getData()); applyState(applyState.getClientActor(), applyState.getIdentifier(), applyState.getReplicatedLogEntry().getData()); } else if(message instanceof ApplySnapshot ) { applySnapshot(((ApplySnapshot) message).getSnapshot()); } else if (message instanceof FindLeader) { getSender().tell( new FindLeaderReply( context.getPeerAddress(currentBehavior.getLeaderId())), getSelf() ); } else if (message instanceof SaveSnapshotSuccess) { SaveSnapshotSuccess success = (SaveSnapshotSuccess) message; // TODO: Not sure if we want to be this aggressive with trimming stuff trimPersistentData(success.metadata().sequenceNr()); } else if (message instanceof SaveSnapshotFailure) { // TODO: Handle failure in saving the snapshot } else if (message instanceof FindLeader){ getSender().tell(new FindLeaderReply( context.getPeerAddress(currentBehavior.getLeaderId())), getSelf()); } else if (message instanceof AddRaftPeer){ // FIXME : Do not add raft peers like this. // When adding a new Peer we have to ensure that the a majority of // the peers know about the new Peer. Doing it this way may cause // a situation where multiple Leaders may emerge AddRaftPeer arp = (AddRaftPeer)message; context.addToPeers(arp.getName(), arp.getAddress()); } else if (message instanceof RemoveRaftPeer){ RemoveRaftPeer rrp = (RemoveRaftPeer)message; context.removePeer(rrp.getName()); } else { RaftState state = currentBehavior.handleMessage(getSender(), message); currentBehavior = switchBehavior(state); } } /** * When a derived RaftActor needs to persist something it must call * persistData. * * @param clientActor * @param identifier * @param data */ protected void persistData(ActorRef clientActor, String identifier, Payload data) { ReplicatedLogEntry replicatedLogEntry = new ReplicatedLogImplEntry( context.getReplicatedLog().lastIndex() + 1, context.getTermInformation().getCurrentTerm(), data); LOG.debug("Persist data {}", replicatedLogEntry); replicatedLog .appendAndPersist(clientActor, identifier, replicatedLogEntry); } protected String getId() { return context.getId(); } /** * Derived actors can call the isLeader method to check if the current * RaftActor is the Leader or not * * @return true it this RaftActor is a Leader false otherwise */ protected boolean isLeader() { return context.getId().equals(currentBehavior.getLeaderId()); } /** * Derived actor can call getLeader if they need a reference to the Leader. * This would be useful for example in forwarding a request to an actor * which is the leader * * @return A reference to the leader if known, null otherwise */ protected ActorSelection getLeader(){ String leaderId = currentBehavior.getLeaderId(); if (leaderId == null) { return null; } String peerAddress = context.getPeerAddress(leaderId); LOG.debug("getLeader leaderId = " + leaderId + " peerAddress = " + peerAddress); return context.actorSelection(peerAddress); } protected RaftState getRaftState() { return currentBehavior.state(); } /** * setPeerAddress sets the address of a known peer at a later time. *

* This is to account for situations where a we know that a peer * exists but we do not know an address up-front. This may also be used in * situations where a known peer starts off in a different location and we * need to change it's address *

* Note that if the peerId does not match the list of peers passed to * this actor during construction an IllegalStateException will be thrown. * * @param peerId * @param peerAddress */ protected void setPeerAddress(String peerId, String peerAddress){ context.setPeerAddress(peerId, peerAddress); } /** * The applyState method will be called by the RaftActor when some data * needs to be applied to the actor's state * * @param clientActor A reference to the client who sent this message. This * is the same reference that was passed to persistData * by the derived actor. clientActor may be null when * the RaftActor is behaving as a follower or during * recovery. * @param identifier The identifier of the persisted data. This is also * the same identifier that was passed to persistData by * the derived actor. identifier may be null when * the RaftActor is behaving as a follower or during * recovery * @param data A piece of data that was persisted by the persistData call. * This should NEVER be null. */ protected abstract void applyState(ActorRef clientActor, String identifier, Object data); /** * This method will be called by the RaftActor when a snapshot needs to be * created. The derived actor should respond with its current state. *

* During recovery the state that is returned by the derived actor will * be passed back to it by calling the applySnapshot method * * @return The current state of the actor */ protected abstract Object createSnapshot(); /** * This method will be called by the RaftActor during recovery to * reconstruct the state of the actor. *

* This method may also be called at any other point during normal * operations when the derived actor is out of sync with it's peers * and the only way to bring it in sync is by applying a snapshot * * @param snapshot A snapshot of the state of the actor */ protected abstract void applySnapshot(Object snapshot); private RaftActorBehavior switchBehavior(RaftState state) { if (currentBehavior != null) { if (currentBehavior.state() == state) { return currentBehavior; } LOG.info("Switching from state " + currentBehavior.state() + " to " + state); try { currentBehavior.close(); } catch (Exception e) { LOG.error(e, "Failed to close behavior : " + currentBehavior.state()); } } else { LOG.info("Switching behavior to " + state); } RaftActorBehavior behavior = null; if (state == RaftState.Candidate) { behavior = new Candidate(context); } else if (state == RaftState.Follower) { behavior = new Follower(context); } else { behavior = new Leader(context); } return behavior; } private void trimPersistentData(long sequenceNumber) { // Trim snapshots // FIXME : Not sure how exactly the SnapshotSelectionCriteria is applied // For now guessing that it is ANDed. deleteSnapshots(new SnapshotSelectionCriteria( sequenceNumber - 100000, 43200000)); // Trim journal deleteMessages(sequenceNumber); } private class ReplicatedLogImpl implements ReplicatedLog { private final List journal; private final Object snapshot; private long snapshotIndex = -1; private long snapshotTerm = -1; public ReplicatedLogImpl(Snapshot snapshot) { this.snapshot = snapshot.getState(); this.snapshotIndex = snapshot.getLastAppliedIndex(); this.snapshotTerm = snapshot.getLastAppliedTerm(); this.journal = new ArrayList<>(snapshot.getUnAppliedEntries()); } public ReplicatedLogImpl() { this.snapshot = null; this.journal = new ArrayList<>(); } @Override public ReplicatedLogEntry get(long index) { int adjustedIndex = adjustedIndex(index); if (adjustedIndex < 0 || adjustedIndex >= journal.size()) { return null; } return journal.get(adjustedIndex); } @Override public ReplicatedLogEntry last() { if (journal.size() == 0) { return null; } return get(journal.size() - 1); } @Override public long lastIndex() { if (journal.size() == 0) { return -1; } return last().getIndex(); } @Override public long lastTerm() { if (journal.size() == 0) { return -1; } return last().getTerm(); } @Override public void removeFrom(long index) { int adjustedIndex = adjustedIndex(index); if (adjustedIndex < 0 || adjustedIndex >= journal.size()) { return; } journal.subList(adjustedIndex , journal.size()).clear(); } @Override public void removeFromAndPersist(long index) { int adjustedIndex = adjustedIndex(index); if (adjustedIndex < 0 || adjustedIndex >= journal.size()) { return; } // FIXME: Maybe this should be done after the command is saved journal.subList(adjustedIndex , journal.size()).clear(); persist(new DeleteEntries(adjustedIndex), new Procedure(){ @Override public void apply(DeleteEntries param) throws Exception { //FIXME : Doing nothing for now } }); } @Override public void append( final ReplicatedLogEntry replicatedLogEntry) { journal.add(replicatedLogEntry); } @Override public List getFrom(long index) { int adjustedIndex = adjustedIndex(index); List entries = new ArrayList<>(100); if (adjustedIndex < 0 || adjustedIndex >= journal.size()) { return entries; } for (int i = adjustedIndex; i < journal.size(); i++) { entries.add(journal.get(i)); } return entries; } @Override public void appendAndPersist( final ReplicatedLogEntry replicatedLogEntry) { appendAndPersist(null, null, replicatedLogEntry); } public void appendAndPersist(final ActorRef clientActor, final String identifier, final ReplicatedLogEntry replicatedLogEntry) { context.getLogger().debug( "Append log entry and persist {} ", replicatedLogEntry); // FIXME : By adding the replicated log entry to the in-memory journal we are not truly ensuring durability of the logs journal.add(replicatedLogEntry); // When persisting events with persist it is guaranteed that the // persistent actor will not receive further commands between the // persist call and the execution(s) of the associated event // handler. This also holds for multiple persist calls in context // of a single command. persist(replicatedLogEntry, new Procedure() { public void apply(ReplicatedLogEntry evt) throws Exception { // FIXME : Tentatively create a snapshot every hundred thousand entries. To be tuned. if (size() > 100000) { ReplicatedLogEntry lastAppliedEntry = get(context.getLastApplied()); long lastAppliedIndex = -1; long lastAppliedTerm = -1; if (lastAppliedEntry != null) { lastAppliedIndex = lastAppliedEntry.getIndex(); lastAppliedTerm = lastAppliedEntry.getTerm(); } saveSnapshot(Snapshot.create(createSnapshot(), getFrom(context.getLastApplied() + 1), lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm)); } // Send message for replication if (clientActor != null) { currentBehavior.handleMessage(getSelf(), new Replicate(clientActor, identifier, replicatedLogEntry) ); } } } ); } @Override public long size() { return journal.size() + snapshotIndex + 1; } @Override public boolean isPresent(long index) { int adjustedIndex = adjustedIndex(index); if (adjustedIndex < 0 || adjustedIndex >= journal.size()) { return false; } return true; } @Override public boolean isInSnapshot(long index) { return index <= snapshotIndex; } @Override public Object getSnapshot() { return snapshot; } @Override public long getSnapshotIndex() { return snapshotIndex; } @Override public long getSnapshotTerm() { return snapshotTerm; } private int adjustedIndex(long index) { if(snapshotIndex < 0){ return (int) index; } return (int) (index - snapshotIndex); } } private static class DeleteEntries implements Serializable { private final int fromIndex; public DeleteEntries(int fromIndex) { this.fromIndex = fromIndex; } public int getFromIndex() { return fromIndex; } } private static class Snapshot implements Serializable { private final Object state; private final List unAppliedEntries; private final long lastIndex; private final long lastTerm; private final long lastAppliedIndex; private final long lastAppliedTerm; private Snapshot(Object state, List unAppliedEntries, long lastIndex, long lastTerm, long lastAppliedIndex, long lastAppliedTerm) { this.state = state; this.unAppliedEntries = unAppliedEntries; this.lastIndex = lastIndex; this.lastTerm = lastTerm; this.lastAppliedIndex = lastAppliedIndex; this.lastAppliedTerm = lastAppliedTerm; } public static Snapshot create(Object state, List entries, long lastIndex, long lastTerm, long lastAppliedIndex, long lastAppliedTerm) { return new Snapshot(state, entries, lastIndex, lastTerm, lastAppliedIndex, lastAppliedTerm); } public Object getState() { return state; } public List getUnAppliedEntries() { return unAppliedEntries; } public long getLastTerm() { return lastTerm; } public long getLastAppliedIndex() { return lastAppliedIndex; } public long getLastAppliedTerm() { return lastAppliedTerm; } } private class ElectionTermImpl implements ElectionTerm { /** * Identifier of the actor whose election term information this is */ private long currentTerm = 0; private String votedFor = null; public long getCurrentTerm() { return currentTerm; } public String getVotedFor() { return votedFor; } @Override public void update(long currentTerm, String votedFor) { LOG.info("Set currentTerm={}, votedFor={}", currentTerm, votedFor); this.currentTerm = currentTerm; this.votedFor = votedFor; } @Override public void updateAndPersist(long currentTerm, String votedFor){ update(currentTerm, votedFor); // FIXME : Maybe first persist then update the state persist(new UpdateElectionTerm(this.currentTerm, this.votedFor), new Procedure(){ @Override public void apply(UpdateElectionTerm param) throws Exception { } }); } } private static class UpdateElectionTerm implements Serializable { private final long currentTerm; private final String votedFor; public UpdateElectionTerm(long currentTerm, String votedFor) { this.currentTerm = currentTerm; this.votedFor = votedFor; } public long getCurrentTerm() { return currentTerm; } public String getVotedFor() { return votedFor; } } }