void ReplicatedLog::OnLearnChosen(PaxosMessage& imsg)
{
uint64_t runID;
#ifdef RLOG_DEBUG_MESSAGES
Log_Debug("OnLearnChosen begin");
#endif
if (context->GetDatabase()->IsCommiting())
{
#ifdef RLOG_DEBUG_MESSAGES
Log_Debug("Database is commiting, dropping Paxos message");
#endif
return;
}
Log_Trace();
if (imsg.paxosID > paxosID)
{
RequestChosen(imsg.nodeID); // I am lagging and need to catch-up
return;
}
else if (imsg.paxosID < paxosID)
return;
if (imsg.type == PAXOS_LEARN_VALUE)
{
runID = imsg.runID;
acceptor.state.accepted = true;
acceptor.state.acceptedValue.Write(imsg.value);
}
else if (imsg.type == PAXOS_LEARN_PROPOSAL && acceptor.state.accepted &&
acceptor.state.acceptedProposalID == imsg.proposalID)
{
runID = acceptor.state.acceptedRunID;
}
else
{
RequestChosen(imsg.nodeID);
return;
}
ProcessLearnChosen(imsg.nodeID, runID);
#ifdef RLOG_DEBUG_MESSAGES
Log_Debug("OnLearnChosen end");
#endif
}
bool ReplicatedLog::OnLearnChosen(PaxosMessage& imsg)
{
uint64_t runID;
#ifdef RLOG_DEBUG_MESSAGES
Log_Debug("OnLearnChosen begin");
#endif
if (context->GetDatabase()->IsCommitting())
{
#ifdef RLOG_DEBUG_MESSAGES
Log_Debug("Database is commiting, dropping Paxos message");
#endif
return true;
}
if (waitingOnAppend)
{
#ifdef RLOG_DEBUG_MESSAGES
Log_Debug("Waiting OnAppend, dropping Paxos message");
#endif
return true;
}
// if I was the primary, lost the lease in the middle of round 1000, another node replicated some rounds 1001-1100,
// I get the lease back, and then receive the learn message for my previous round (1000)
// then multi will be turned back on, but then subsequent rounds (1001-1100), where I will
// receive learn messages will not succeed, and due to the code above
// I would throw it away any learn messages sent to me
//if (imsg.nodeID != MY_NODEID && proposer.state.multi)
//{
// Log_Debug("Received learn message from %U, but I'm in multi paxos mode", imsg.nodeID);
// return true;
//}
// it's valid for me to be the primary and be lagging by one round * at the beginning of my lease *
// this can happen if the old primary completes a round of replication, fails
// and I get the lease before its OnLearnChosen arrives
// if I throw away the OnLearnChosen, but the others in the quorum do not
// they will advance their paxosID, I will not advance mine
// and I will not be able to replicate => read-only cluster
//if (imsg.nodeID != MY_NODEID && context->IsLeaseOwner())
//{
// Log_Debug("Received learn message from %U, but I'm the lease owner", imsg.nodeID);
// return true;
//}
Log_Trace();
if (imsg.paxosID > paxosID)
{
RequestChosen(imsg.nodeID); // I am lagging and need to catch-up
return true;
}
else if (imsg.paxosID < paxosID)
return true;
if (imsg.type == PAXOS_LEARN_VALUE)
{
runID = 0;
// in the PAXOS_LEARN_VALUE case (and only in this case) runID is 0
// for legacy reasons the PAXOS_LEARN_VALUE message also includes a runID,
// which is always set to 0
acceptor.state.accepted = true;
acceptor.state.acceptedValue.Write(imsg.value);
acceptor.WriteState();
}
else if (imsg.type == PAXOS_LEARN_PROPOSAL && acceptor.state.accepted &&
acceptor.state.acceptedProposalID == imsg.proposalID)
{
runID = acceptor.state.acceptedRunID;
}
else
{
RequestChosen(imsg.nodeID);
return true;
}
ProcessLearnChosen(imsg.nodeID, runID);
#ifdef RLOG_DEBUG_MESSAGES
Log_Debug("OnLearnChosen end");
#endif
return false;
}