// 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
}
func setHardState(
ctx context.Context, batch engine.ReadWriter, rangeID roachpb.RangeID, st raftpb.HardState,
) error {
return engine.MVCCPutProto(ctx, batch, nil,
keys.RaftHardStateKey(rangeID),
hlc.ZeroTimestamp, nil, &st)
}
func loadHardState(
ctx context.Context, reader engine.Reader, rangeID roachpb.RangeID,
) (raftpb.HardState, error) {
var hs raftpb.HardState
found, err := engine.MVCCGetProto(ctx, reader,
keys.RaftHardStateKey(rangeID), hlc.ZeroTimestamp, true, nil, &hs)
if !found || err != nil {
return raftpb.HardState{}, err
}
return hs, nil
}
// createRangeData creates sample range data in all possible areas of
// the key space. Returns a slice of the encoded keys of all created
// data.
func createRangeData(t *testing.T, r *Replica) []engine.MVCCKey {
ts0 := hlc.ZeroTimestamp
ts := hlc.Timestamp{WallTime: 1}
desc := r.Desc()
keyTSs := []struct {
key roachpb.Key
ts hlc.Timestamp
}{
{keys.AbortCacheKey(r.RangeID, testTxnID), ts0},
{keys.AbortCacheKey(r.RangeID, testTxnID2), ts0},
{keys.RangeFrozenStatusKey(r.RangeID), ts0},
{keys.RangeLastGCKey(r.RangeID), ts0},
{keys.RaftAppliedIndexKey(r.RangeID), ts0},
{keys.RaftTruncatedStateKey(r.RangeID), ts0},
{keys.LeaseAppliedIndexKey(r.RangeID), ts0},
{keys.RangeStatsKey(r.RangeID), ts0},
{keys.RaftHardStateKey(r.RangeID), ts0},
{keys.RaftLastIndexKey(r.RangeID), ts0},
{keys.RaftLogKey(r.RangeID, 1), ts0},
{keys.RaftLogKey(r.RangeID, 2), ts0},
{keys.RangeLastReplicaGCTimestampKey(r.RangeID), ts0},
{keys.RangeLastVerificationTimestampKey(r.RangeID), ts0},
{keys.RangeDescriptorKey(desc.StartKey), ts},
{keys.TransactionKey(roachpb.Key(desc.StartKey), uuid.NewV4()), ts0},
{keys.TransactionKey(roachpb.Key(desc.StartKey.Next()), uuid.NewV4()), ts0},
{keys.TransactionKey(fakePrevKey(desc.EndKey), uuid.NewV4()), ts0},
// TODO(bdarnell): KeyMin.Next() results in a key in the reserved system-local space.
// Once we have resolved https://github.com/cockroachdb/cockroach/issues/437,
// replace this with something that reliably generates the first valid key in the range.
//{r.Desc().StartKey.Next(), ts},
// The following line is similar to StartKey.Next() but adds more to the key to
// avoid falling into the system-local space.
{append(append([]byte{}, desc.StartKey...), '\x02'), ts},
{fakePrevKey(r.Desc().EndKey), ts},
}
keys := []engine.MVCCKey{}
for _, keyTS := range keyTSs {
if err := engine.MVCCPut(context.Background(), r.store.Engine(), nil, keyTS.key, keyTS.ts, roachpb.MakeValueFromString("value"), nil); err != nil {
t.Fatal(err)
}
keys = append(keys, engine.MVCCKey{Key: keyTS.key, Timestamp: keyTS.ts})
}
return keys
}
// createRangeData creates sample range data in all possible areas of
// the key space. Returns a slice of the encoded keys of all created
// data.
func createRangeData(r *Replica, t *testing.T) []roachpb.EncodedKey {
ts0 := roachpb.ZeroTimestamp
ts := roachpb.Timestamp{WallTime: 1}
keyTSs := []struct {
key roachpb.Key
ts roachpb.Timestamp
}{
{keys.ResponseCacheKey(r.Desc().RangeID, &roachpb.ClientCmdID{WallTime: 1, Random: 1}), ts0},
{keys.ResponseCacheKey(r.Desc().RangeID, &roachpb.ClientCmdID{WallTime: 2, Random: 2}), ts0},
{keys.RaftHardStateKey(r.Desc().RangeID), ts0},
{keys.RaftLogKey(r.Desc().RangeID, 1), ts0},
{keys.RaftLogKey(r.Desc().RangeID, 2), ts0},
{keys.RangeGCMetadataKey(r.Desc().RangeID), ts0},
{keys.RangeLastVerificationTimestampKey(r.Desc().RangeID), ts0},
{keys.RangeStatsKey(r.Desc().RangeID), ts0},
{keys.RangeDescriptorKey(r.Desc().StartKey), ts},
{keys.TransactionKey(roachpb.Key(r.Desc().StartKey), []byte("1234")), ts0},
{keys.TransactionKey(roachpb.Key(r.Desc().StartKey.Next()), []byte("5678")), ts0},
{keys.TransactionKey(fakePrevKey(r.Desc().EndKey), []byte("2468")), ts0},
// TODO(bdarnell): KeyMin.Next() results in a key in the reserved system-local space.
// Once we have resolved https://github.com/cockroachdb/cockroach/issues/437,
// replace this with something that reliably generates the first valid key in the range.
//{r.Desc().StartKey.Next(), ts},
// The following line is similar to StartKey.Next() but adds more to the key to
// avoid falling into the system-local space.
{append(append([]byte{}, r.Desc().StartKey...), '\x01'), ts},
{fakePrevKey(r.Desc().EndKey), ts},
}
keys := []roachpb.EncodedKey{}
for _, keyTS := range keyTSs {
if err := engine.MVCCPut(r.store.Engine(), nil, keyTS.key, keyTS.ts, roachpb.MakeValueFromString("value"), nil); err != nil {
t.Fatal(err)
}
keys = append(keys, engine.MVCCEncodeKey(keyTS.key))
if !keyTS.ts.Equal(ts0) {
keys = append(keys, engine.MVCCEncodeVersionKey(keyTS.key, keyTS.ts))
}
}
return keys
}
// 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
}
// SetHardState implements the multiraft.WriteableGroupStorage interface.
func (r *Range) SetHardState(st raftpb.HardState) error {
return engine.MVCCPutProto(r.rm.Engine(), nil, keys.RaftHardStateKey(r.Desc().RaftID),
proto.ZeroTimestamp, nil, &st)
}
// ApplySnapshot implements the multiraft.WriteableGroupStorage interface.
func (r *Range) ApplySnapshot(snap raftpb.Snapshot) error {
snapData := proto.RaftSnapshotData{}
err := gogoproto.Unmarshal(snap.Data, &snapData)
if err != nil {
return err
}
// First, save the HardState. The HardState must not be changed
// because it may record a previous vote cast by this node.
hardStateKey := keys.RaftHardStateKey(r.Desc().RaftID)
hardState, _, err := engine.MVCCGet(r.rm.Engine(), hardStateKey, proto.ZeroTimestamp, true /* consistent */, nil)
if err != nil {
return err
}
// Extract the updated range descriptor.
desc := snapData.RangeDescriptor
batch := r.rm.Engine().NewBatch()
defer batch.Close()
// Delete everything in the range and recreate it from the snapshot.
for iter := newRangeDataIterator(&desc, r.rm.Engine()); iter.Valid(); iter.Next() {
if err := batch.Clear(iter.Key()); err != nil {
return err
}
}
// Write the snapshot into the range.
for _, kv := range snapData.KV {
if err := batch.Put(kv.Key, kv.Value); err != nil {
return err
}
}
// Restore the saved HardState.
if hardState == nil {
err := engine.MVCCDelete(batch, nil, hardStateKey, proto.ZeroTimestamp, nil)
if err != nil {
return err
}
} else {
err := engine.MVCCPut(batch, nil, hardStateKey, proto.ZeroTimestamp, *hardState, nil)
if err != nil {
return err
}
}
// Read the leader lease.
lease, err := loadLeaderLease(batch, desc.RaftID)
if err != nil {
return err
}
// Copy range stats to new range.
oldStats := r.stats
r.stats, err = newRangeStats(desc.RaftID, batch)
if err != nil {
r.stats = oldStats
return err
}
// The next line sets the persisted last index to the last applied index.
// This is not a correctness issue, but means that we may have just
// transferred some entries we're about to re-request from the leader and
// overwrite.
// However, raft.MultiNode currently expects this behaviour, and the
// performance implications are not likely to be drastic. If our feelings
// about this ever change, we can add a LastIndex field to
// raftpb.SnapshotMetadata.
if err := setLastIndex(batch, r.Desc().RaftID, snap.Metadata.Index); err != nil {
return err
}
if err := batch.Commit(); err != nil {
return err
}
// As outlined above, last and applied index are the same after applying
// the snapshot.
atomic.StoreUint64(&r.lastIndex, snap.Metadata.Index)
atomic.StoreUint64(&r.appliedIndex, snap.Metadata.Index)
// Atomically update the descriptor and lease.
if err := r.setDesc(&desc); err != nil {
return err
}
atomic.StorePointer(&r.lease, unsafe.Pointer(lease))
return nil
}
// setHardState persists the raft HardState.
func (r *Replica) setHardState(batch engine.Engine, st raftpb.HardState) error {
return engine.MVCCPutProto(batch, nil, keys.RaftHardStateKey(r.RangeID),
roachpb.ZeroTimestamp, nil, &st)
}
// applySnapshot updates the replica based on the given snapshot.
// Returns the new last index.
func (r *Replica) applySnapshot(batch engine.Engine, snap raftpb.Snapshot) (uint64, error) {
snapData := roachpb.RaftSnapshotData{}
err := proto.Unmarshal(snap.Data, &snapData)
if err != nil {
return 0, err
}
rangeID := r.RangeID
// First, save the HardState. The HardState must not be changed
// because it may record a previous vote cast by this node. This is
// usually unnecessary because a snapshot is nearly always
// accompanied by a new HardState which incorporates both our former
// state and new information from the leader, but in the event that
// the HardState has not changed, we want to use our own previous
// HardState and not one that was transmitted via the snapshot.
hardStateKey := keys.RaftHardStateKey(rangeID)
hardState, _, err := engine.MVCCGet(batch, hardStateKey, roachpb.ZeroTimestamp, true /* consistent */, nil)
if err != nil {
return 0, err
}
// Extract the updated range descriptor.
desc := snapData.RangeDescriptor
// Delete everything in the range and recreate it from the snapshot.
// We need to delete any old Raft log entries here because any log entries
// that predate the snapshot will be orphaned and never truncated or GC'd.
iter := newReplicaDataIterator(&desc, batch, false /* !replicatedOnly */)
defer iter.Close()
for ; iter.Valid(); iter.Next() {
if err := batch.Clear(iter.Key()); err != nil {
return 0, err
}
}
// Determine the unreplicated key prefix so we can drop any
// unreplicated keys from the snapshot.
unreplicatedPrefix := keys.MakeRangeIDUnreplicatedPrefix(desc.RangeID)
// Write the snapshot into the range.
for _, kv := range snapData.KV {
if bytes.HasPrefix(kv.Key, unreplicatedPrefix) {
continue
}
mvccKey := engine.MVCCKey{
Key: kv.Key,
Timestamp: kv.Timestamp,
}
if err := batch.Put(mvccKey, kv.Value); err != nil {
return 0, err
}
}
// Write the snapshot's Raft log into the range.
if _, err := r.append(batch, 0, snapData.LogEntries); err != nil {
return 0, err
}
// Restore the saved HardState.
if hardState == nil {
err := engine.MVCCDelete(batch, nil, hardStateKey, roachpb.ZeroTimestamp, nil)
if err != nil {
return 0, err
}
} else {
err := engine.MVCCPut(batch, nil, hardStateKey, roachpb.ZeroTimestamp, *hardState, nil)
if err != nil {
return 0, err
}
}
// Read the leader lease.
lease, err := loadLeaderLease(batch, desc.RangeID)
if err != nil {
return 0, err
}
// Load updated range stats. The local newStats variable will be assigned
// to r.stats after the batch commits.
newStats, err := newRangeStats(desc.RangeID, batch)
if err != nil {
return 0, err
}
// The next line sets the persisted last index to the last applied index.
// This is not a correctness issue, but means that we may have just
// transferred some entries we're about to re-request from the leader and
// overwrite.
// However, raft.MultiNode currently expects this behaviour, and the
// performance implications are not likely to be drastic. If our feelings
// about this ever change, we can add a LastIndex field to
// raftpb.SnapshotMetadata.
if err := setLastIndex(batch, rangeID, snap.Metadata.Index); err != nil {
return 0, err
}
batch.Defer(func() {
// Update the range stats.
r.stats.Replace(newStats)
r.mu.Lock()
// As outlined above, last and applied index are the same after applying
// the snapshot.
r.mu.appliedIndex = snap.Metadata.Index
r.mu.leaderLease = lease
r.mu.Unlock()
// Update other fields which are uninitialized or need updating.
// This may not happen if the system config has not yet been loaded.
// While config update will correctly set the fields, there is no order
// guarantee in ApplySnapshot.
// TODO: should go through the standard store lock when adding a replica.
if err := r.updateRangeInfo(&desc); err != nil {
panic(err)
}
// Update the range descriptor. This is done last as this is the step that
// makes the Replica visible in the Store.
if err := r.setDesc(&desc); err != nil {
panic(err)
}
})
return snap.Metadata.Index, nil
}
// SetHardState implements the multiraft.WriteableGroupStorage interface.
func (r *Replica) SetHardState(st raftpb.HardState) error {
return engine.MVCCPutProto(r.store.Engine(), nil, keys.RaftHardStateKey(r.Desc().RangeID),
roachpb.ZeroTimestamp, nil, &st)
}
// applySnapshot updates the replica based on the given snapshot.
func (r *Replica) applySnapshot(snap raftpb.Snapshot) error {
snapData := roachpb.RaftSnapshotData{}
err := proto.Unmarshal(snap.Data, &snapData)
if err != nil {
return err
}
rangeID := r.Desc().RangeID
// First, save the HardState. The HardState must not be changed
// because it may record a previous vote cast by this node.
hardStateKey := keys.RaftHardStateKey(rangeID)
hardState, _, err := engine.MVCCGet(r.store.Engine(), hardStateKey, roachpb.ZeroTimestamp, true /* consistent */, nil)
if err != nil {
return err
}
// Extract the updated range descriptor.
desc := snapData.RangeDescriptor
batch := r.store.Engine().NewBatch()
defer batch.Close()
// Delete everything in the range and recreate it from the snapshot.
iter := newReplicaDataIterator(&desc, r.store.Engine())
defer iter.Close()
for ; iter.Valid(); iter.Next() {
if err := batch.Clear(iter.Key()); err != nil {
return err
}
}
// Write the snapshot into the range.
for _, kv := range snapData.KV {
mvccKey := engine.MVCCKey{
Key: kv.Key,
Timestamp: kv.Timestamp,
}
if err := batch.Put(mvccKey, kv.Value); err != nil {
return err
}
}
// Restore the saved HardState.
if hardState == nil {
err := engine.MVCCDelete(batch, nil, hardStateKey, roachpb.ZeroTimestamp, nil)
if err != nil {
return err
}
} else {
err := engine.MVCCPut(batch, nil, hardStateKey, roachpb.ZeroTimestamp, *hardState, nil)
if err != nil {
return err
}
}
// Read the leader lease.
lease, err := loadLeaderLease(batch, desc.RangeID)
if err != nil {
return err
}
// Load updated range stats. The local newStats variable will be assigned
// to r.stats after the batch commits.
newStats, err := newRangeStats(desc.RangeID, batch)
if err != nil {
return err
}
// The next line sets the persisted last index to the last applied index.
// This is not a correctness issue, but means that we may have just
// transferred some entries we're about to re-request from the leader and
// overwrite.
// However, raft.MultiNode currently expects this behaviour, and the
// performance implications are not likely to be drastic. If our feelings
// about this ever change, we can add a LastIndex field to
// raftpb.SnapshotMetadata.
if err := setLastIndex(batch, rangeID, snap.Metadata.Index); err != nil {
return err
}
if err := batch.Commit(); err != nil {
return err
}
// Update the range stats.
r.stats.Replace(newStats)
// As outlined above, last and applied index are the same after applying
// the snapshot.
atomic.StoreUint64(&r.lastIndex, snap.Metadata.Index)
atomic.StoreUint64(&r.appliedIndex, snap.Metadata.Index)
// Atomically update the descriptor and lease.
if err := r.setDesc(&desc); err != nil {
return err
}
// Update other fields which are uninitialized or need updating.
// This may not happen if the system config has not yet been loaded.
// While config update will correctly set the fields, there is no order
// guarangee in ApplySnapshot.
// TODO: should go through the standard store lock when adding a replica.
if err := r.updateRangeInfo(); err != nil {
return err
}
atomic.StorePointer(&r.lease, unsafe.Pointer(lease))
return nil
}
// setHardState persists the raft HardState.
func (r *Replica) setHardState(batch engine.ReadWriter, st raftpb.HardState) error {
return engine.MVCCPutProto(context.Background(), batch, nil, keys.RaftHardStateKey(r.RangeID),
roachpb.ZeroTimestamp, nil, &st)
}