// InitialState implements the raft.Storage interface.
func (r *Range) InitialState() (raftpb.HardState, raftpb.ConfState, error) {
var hs raftpb.HardState
found, err := engine.MVCCGetProto(r.rm.Engine(), keys.RaftHardStateKey(r.Desc().RaftID),
proto.ZeroTimestamp, true, nil, &hs)
if err != nil {
return raftpb.HardState{}, raftpb.ConfState{}, err
}
if !found {
// We don't have a saved HardState, so set up the defaults.
if r.isInitialized() {
// Set the initial log term.
hs.Term = raftInitialLogTerm
hs.Commit = raftInitialLogIndex
atomic.StoreUint64(&r.lastIndex, raftInitialLogIndex)
} else {
// This is a new range we are receiving from another node. Start
// from zero so we will receive a snapshot.
atomic.StoreUint64(&r.lastIndex, 0)
}
}
var cs raftpb.ConfState
// For uninitalized ranges, membership is unknown at this point.
if found || r.isInitialized() {
for _, rep := range r.Desc().Replicas {
cs.Nodes = append(cs.Nodes, uint64(proto.MakeRaftNodeID(rep.NodeID, rep.StoreID)))
}
}
return hs, cs, nil
}
// synthesizeHardState synthesizes a HardState from the given ReplicaState and
// any existing on-disk HardState in the context of a snapshot, while verifying
// that the application of the snapshot does not violate Raft invariants. It
// must be called after the supplied state and ReadWriter have been updated
// with the result of the snapshot.
// If there is an existing HardState, we must respect it and we must not apply
// a snapshot that would move the state backwards.
func synthesizeHardState(
ctx context.Context, eng engine.ReadWriter, s storagebase.ReplicaState, oldHS raftpb.HardState,
) error {
newHS := raftpb.HardState{
Term: s.TruncatedState.Term,
// Note that when applying a Raft snapshot, the applied index is
// equal to the Commit index represented by the snapshot.
Commit: s.RaftAppliedIndex,
}
if oldHS.Commit > newHS.Commit {
return errors.Errorf("can't decrease HardState.Commit from %d to %d",
oldHS.Commit, newHS.Commit)
}
if oldHS.Term > newHS.Term {
// The existing HardState is allowed to be ahead of us, which is
// relevant in practice for the split trigger. We already checked above
// that we're not rewinding the acknowledged index, and we haven't
// updated votes yet.
newHS.Term = oldHS.Term
}
// If the existing HardState voted in this term, remember that.
if oldHS.Term == newHS.Term {
newHS.Vote = oldHS.Vote
}
return errors.Wrapf(setHardState(ctx, eng, s.Desc.RangeID, newHS), "writing HardState %+v", &newHS)
}
func updateHardState(eng engine.ReadWriter, s storagebase.ReplicaState) error {
// Load a potentially existing HardState as we may need to preserve
// information about cast votes. For example, during a Split for which
// another node's new right-hand side has contacted us before our left-hand
// side called in here to create the group.
rangeID := s.Desc.RangeID
oldHS, err := loadHardState(eng, rangeID)
if err != nil {
return err
}
newHS := raftpb.HardState{
Term: s.TruncatedState.Term,
Commit: s.RaftAppliedIndex,
}
if !raft.IsEmptyHardState(oldHS) {
if oldHS.Commit > newHS.Commit {
newHS.Commit = oldHS.Commit
}
if oldHS.Term > newHS.Term {
newHS.Term = oldHS.Term
}
newHS.Vote = oldHS.Vote
}
return setHardState(eng, rangeID, newHS)
}
// writeInitialState bootstraps a new Raft group (i.e. it is called when we
// bootstrap a Range, or when setting up the right hand side of a split).
// Its main task is to persist a consistent Raft (and associated Replica) state
// which does not start from zero but presupposes a few entries already having
// applied.
// The supplied MVCCStats are used for the Stats field after adjusting for
// persisting the state itself, and the updated stats are returned.
func writeInitialState(
eng engine.ReadWriter, ms enginepb.MVCCStats, desc roachpb.RangeDescriptor,
) (enginepb.MVCCStats, error) {
rangeID := desc.RangeID
var s storagebase.ReplicaState
s.TruncatedState = &roachpb.RaftTruncatedState{
Term: raftInitialLogTerm,
Index: raftInitialLogIndex,
}
s.RaftAppliedIndex = s.TruncatedState.Index
s.Desc = &roachpb.RangeDescriptor{
RangeID: rangeID,
}
s.Stats = ms
newMS, err := saveState(eng, s)
if err != nil {
return enginepb.MVCCStats{}, err
}
// Load a potentially existing HardState as we may need to preserve
// information about cast votes. For example, during a Split for which
// another node's new right-hand side has contacted us before our left-hand
// side called in here to create the group.
oldHS, err := loadHardState(eng, rangeID)
if err != nil {
return enginepb.MVCCStats{}, err
}
newHS := raftpb.HardState{
Term: s.TruncatedState.Term,
Commit: s.TruncatedState.Index,
}
if !raft.IsEmptyHardState(oldHS) {
if oldHS.Commit > newHS.Commit {
newHS.Commit = oldHS.Commit
}
if oldHS.Term > newHS.Term {
newHS.Term = oldHS.Term
}
newHS.Vote = oldHS.Vote
}
if err := setHardState(eng, rangeID, newHS); err != nil {
return enginepb.MVCCStats{}, err
}
if err := setLastIndex(eng, rangeID, s.TruncatedState.Index); err != nil {
return enginepb.MVCCStats{}, err
}
return newMS, nil
}
// InitialState implements the raft.Storage interface.
func (r *Replica) InitialState() (raftpb.HardState, raftpb.ConfState, error) {
var hs raftpb.HardState
desc := r.Desc()
found, err := engine.MVCCGetProto(r.store.Engine(), keys.RaftHardStateKey(desc.RangeID),
roachpb.ZeroTimestamp, true, nil, &hs)
if err != nil {
return raftpb.HardState{}, raftpb.ConfState{}, err
}
initialized := r.isInitialized()
if !found {
// We don't have a saved HardState, so set up the defaults.
if initialized {
// Set the initial log term.
hs.Term = raftInitialLogTerm
hs.Commit = raftInitialLogIndex
atomic.StoreUint64(&r.lastIndex, raftInitialLogIndex)
} else {
// This is a new range we are receiving from another node. Start
// from zero so we will receive a snapshot.
atomic.StoreUint64(&r.lastIndex, 0)
}
} else if initialized && hs.Commit == 0 {
// Normally, when the commit index changes, raft gives us a new
// commit index to persist, however, during initialization, which
// occurs entirely in cockroach, raft has no knowledge of this.
// By setting this to the initial log index, we avoid a panic in
// raft caused by this inconsistency.
hs.Commit = raftInitialLogIndex
}
var cs raftpb.ConfState
// For uninitalized ranges, membership is unknown at this point.
if found || initialized {
for _, rep := range desc.Replicas {
cs.Nodes = append(cs.Nodes, uint64(rep.ReplicaID))
}
}
return hs, cs, nil
}
