private void OnTopologyChanged(TopologyChangeCommand tcc)
{
// if we have any removed servers, we need to know let them know that they have
// been removed, we do that by committing the current entry (hopefully they already
// have topology change command, so they know they are being removed from the cluster).
// This is mostly us being nice neighbors, this isn't required, and the cluster will reject
// messages from nodes not considered to be in the cluster.
if (tcc.Previous == null)
return;
var removedNodes = tcc.Previous.AllNodeNames.Except(tcc.Requested.AllNodeNames).ToList();
foreach (var removedNode in removedNodes)
{
var nodeByName = tcc.Previous.GetNodeByName(removedNode);
if (nodeByName == null)
continue;
// try sending the latest updates (which include the topology removal entry)
SendEntriesToPeer(nodeByName);
// at any rate, try sending the disconnection command explicitly, to gracefully shut down the node if we can
Engine.Transport.Send(nodeByName, new DisconnectedFromCluster
{
From = Engine.Name,
ClusterTopologyId = Engine.CurrentTopology.TopologyId,
Term = Engine.PersistentState.CurrentTerm
});
}
}
protected bool FromOurTopology(BaseMessage msg)
{
if (msg.ClusterTopologyId == Engine.CurrentTopology.TopologyId)
return true;
// if we don't have the same topology id, maybe we have _no_ topology, if that is the case,
// we are accepting the new topology id immediately
if (Engine.CurrentTopology.TopologyId == Guid.Empty &&
Engine.CurrentTopology.HasVoters == false)
{
var tcc = new TopologyChangeCommand
{
Requested = new Topology(msg.ClusterTopologyId)
};
Engine.StartTopologyChange(tcc);
Engine.CommitTopologyChange(tcc);
return true;
}
return false;
}
internal void StartTopologyChange(TopologyChangeCommand tcc)
{
Interlocked.Exchange(ref _currentTopology, tcc.Requested);
Interlocked.Exchange(ref _changingTopology, new TaskCompletionSource<object>().Task);
OnTopologyChanging(tcc);
}
private void OnTopologyChanging(TopologyChangeCommand tcc)
{
var handler = TopologyChanging;
if (handler != null)
{
try
{
handler();
}
catch (Exception e)
{
_log.Error("Error on raising TopologyChanging event", e);
}
}
}
protected virtual void OnTopologyChanged(TopologyChangeCommand cmd)
{
_log.Info ("OnTopologyChanged() - " + this.Name);
var handler = TopologyChanged;
if (handler != null)
{
try
{
handler(cmd);
}
catch (Exception e)
{
_log.Error("Error on raising TopologyChanged event", e);
}
}
}
public void CommitTopologyChange(TopologyChangeCommand tcc)
{
//it is logical that _before_ OnTopologyChanged is fired the topology change task will be complete
// - since this task is used to track progress of topoplogy changes in the interface
Interlocked.Exchange(ref _changingTopology, null);
//if no topology was present and TopologyChangeCommand is issued to just
//accept new topology id - then tcc.Previous == null - it means that
//shouldRemainInTopology should be true - because there is no removal from topology actually
var isRemovedFromTopology = tcc.Requested.Contains(Name) == false &&
tcc.Previous != null &&
tcc.Previous.Contains(Name);
if (isRemovedFromTopology)
{
_log.Debug("This node is being removed from topology, setting its state to follower, it will be idle until a leader will join it to the cluster again");
CurrentLeader = null;
SetState(RaftEngineState.Follower);
return;
}
if (_log.IsInfoEnabled)
{
_log.Info("Finished applying new topology: {0}{1}", _currentTopology,
tcc.Previous == null ? ", Previous topology was null - perhaps it is setting topology for the first time?" : String.Empty);
}
OnTopologyChanged(tcc);
}
internal Task ModifyTopology(Topology requested)
{
if (State != RaftEngineState.Leader)
throw new InvalidOperationException("Cannot modify topology from a non leader node, current leader is: " +
(CurrentLeader ?? "no leader"));
var tcc = new TopologyChangeCommand
{
Completion = new TaskCompletionSource<object>(),
Requested = requested,
Previous = _currentTopology,
BufferCommand = false,
};
if (Interlocked.CompareExchange(ref _changingTopology, tcc.Completion.Task, null) != null)
throw new InvalidOperationException("Cannot change the cluster topology while another topology change is in progress");
try
{
_log.Debug("Topology change started on leader");
StartTopologyChange(tcc);
AppendCommand(tcc);
return tcc.Completion.Task;
}
catch (Exception)
{
Interlocked.Exchange(ref _changingTopology, null);
throw;
}
}
public void Handle(DisconnectedFromCluster req)
{
if (FromOurTopology(req) == false)
{
_log.Info("Got a disconnection notification message outside my cluster topology (id: {0}), ignoring", req.ClusterTopologyId);
return;
}
if (req.Term < Engine.PersistentState.CurrentTerm)
{
_log.Info("Got disconnection notification from an older term, ignoring");
return;
}
if (req.From != Engine.CurrentLeader)
{
_log.Info("Got disconnection notification from {0}, who isn't the current leader, ignoring.",
req.From);
return;
}
_log.Warn("Got disconnection notification from the leader, clearing topology and moving to idle follower state");
var tcc = new TopologyChangeCommand
{
Requested = new Topology(req.ClusterTopologyId)
};
Engine.PersistentState.SetCurrentTopology(tcc.Requested, 0L);
Engine.StartTopologyChange(tcc);
Engine.CommitTopologyChange(tcc);
Engine.SetState(RaftEngineState.Follower);
}
public override InstallSnapshotResponse Handle(MessageContext context, InstallSnapshotRequest req, Stream stream)
{
if (_installingSnapshot != null)
{
return new InstallSnapshotResponse
{
Success = false,
Message = "Cannot install snapshot because we are already installing a snapshot",
CurrentTerm = Engine.PersistentState.CurrentTerm,
From = Engine.Name,
ClusterTopologyId = Engine.CurrentTopology.TopologyId,
LastLogIndex = Engine.PersistentState.LastLogEntry().Index
};
}
if (FromOurTopology(req) == false)
{
_log.Info("Got an install snapshot message outside my cluster topology (id: {0}), ignoring", req.ClusterTopologyId);
return new InstallSnapshotResponse
{
Success = false,
Message = "Cannot install snapshot because the cluster topology id doesn't match, mine is: " + Engine.CurrentTopology.TopologyId,
CurrentTerm = Engine.PersistentState.CurrentTerm,
From = Engine.Name,
ClusterTopologyId = Engine.CurrentTopology.TopologyId,
LastLogIndex = Engine.PersistentState.LastLogEntry().Index
};
}
var lastLogEntry = Engine.PersistentState.LastLogEntry();
if (req.Term < lastLogEntry.Term || req.LastIncludedIndex < lastLogEntry.Index)
{
stream.Dispose();
return new InstallSnapshotResponse
{
From = Engine.Name,
ClusterTopologyId = Engine.CurrentTopology.TopologyId,
CurrentTerm = lastLogEntry.Term,
LastLogIndex = lastLogEntry.Index,
Message = string.Format("Snapshot is too old (term {0} index {1}) while we have (term {2} index {3})",
req.Term, req.LastIncludedIndex, lastLogEntry.Term, lastLogEntry.Index),
Success = false
};
}
_log.Info("Received InstallSnapshotRequest from {0} until term {1} / {2}", req.From, req.LastIncludedTerm, req.LastIncludedIndex);
Engine.OnSnapshotInstallationStarted();
// this can be a long running task
_installingSnapshot = Task.Run(() =>
{
try
{
Engine.StateMachine.ApplySnapshot(req.LastIncludedTerm, req.LastIncludedIndex, stream);
Engine.PersistentState.MarkSnapshotFor(req.LastIncludedIndex, req.LastIncludedTerm, int.MaxValue);
Engine.PersistentState.SetCurrentTopology(req.Topology, req.LastIncludedIndex);
var tcc = new TopologyChangeCommand { Requested = req.Topology };
Engine.StartTopologyChange(tcc);
Engine.CommitTopologyChange(tcc);
}
catch (Exception e)
{
_log.Warn(string.Format("Failed to install snapshot term {0} index {1}", req.LastIncludedIndex, req.LastIncludedIndex), e);
context.ExecuteInEventLoop(() =>
{
_installingSnapshot = null;
});
}
// we are doing it this way to ensure that we are single threaded
context.ExecuteInEventLoop(() =>
{
Engine.UpdateCurrentTerm(req.Term, req.From); // implicitly put us in follower state
_log.Info("Updating the commit index to the snapshot last included index of {0}", req.LastIncludedIndex);
Engine.OnSnapshotInstallationEnded(req.Term);
context.Reply(new InstallSnapshotResponse
{
From = Engine.Name,
ClusterTopologyId = Engine.CurrentTopology.TopologyId,
CurrentTerm = req.Term,
LastLogIndex = req.LastIncludedIndex,
Success = true
});
});
});
return null;
}
