// append the given entries to the raft log. Takes the previous value
// of r.lastIndex and returns a new value. We do this rather than
// modifying r.lastIndex directly because this modification needs to
// be atomic with the commit of the batch.
func (r *Replica) append(batch engine.Engine, prevLastIndex uint64, entries []raftpb.Entry) (uint64, error) {
if len(entries) == 0 {
return prevLastIndex, nil
}
for i := range entries {
ent := &entries[i]
key := keys.RaftLogKey(r.RangeID, ent.Index)
if err := engine.MVCCPutProto(batch, nil, key, roachpb.ZeroTimestamp, nil, ent); err != nil {
return 0, err
}
}
lastIndex := entries[len(entries)-1].Index
// Delete any previously appended log entries which never committed.
for i := lastIndex + 1; i <= prevLastIndex; i++ {
err := engine.MVCCDelete(batch, nil,
keys.RaftLogKey(r.RangeID, i), roachpb.ZeroTimestamp, nil)
if err != nil {
return 0, err
}
}
// Commit the batch and update the last index.
if err := setLastIndex(batch, r.RangeID, lastIndex); err != nil {
return 0, err
}
return lastIndex, nil
}
// append the given entries to the raft log. Takes the previous values of
// r.mu.lastIndex and r.mu.raftLogSize, and returns new values. We do this
// rather than modifying them directly because these modifications need to be
// atomic with the commit of the batch.
func (r *Replica) append(batch engine.ReadWriter, prevLastIndex uint64, prevRaftLogSize int64, entries []raftpb.Entry) (uint64, int64, error) {
if len(entries) == 0 {
return prevLastIndex, prevRaftLogSize, nil
}
var diff enginepb.MVCCStats
ctx := context.Background()
for i := range entries {
ent := &entries[i]
key := keys.RaftLogKey(r.RangeID, ent.Index)
if err := engine.MVCCPutProto(ctx, batch, &diff, key, hlc.ZeroTimestamp, nil /* txn */, ent); err != nil {
return 0, 0, err
}
}
lastIndex := entries[len(entries)-1].Index
// Delete any previously appended log entries which never committed.
for i := lastIndex + 1; i <= prevLastIndex; i++ {
err := engine.MVCCDelete(ctx, batch, &diff, keys.RaftLogKey(r.RangeID, i),
hlc.ZeroTimestamp, nil /* txn */)
if err != nil {
return 0, 0, err
}
}
if err := setLastIndex(batch, r.RangeID, lastIndex); err != nil {
return 0, 0, err
}
raftLogSize := prevRaftLogSize + diff.SysBytes
return lastIndex, raftLogSize, nil
}
// append the given entries to the raft log.
func (r *Replica) append(batch engine.Engine, entries []raftpb.Entry) error {
if len(entries) == 0 {
return nil
}
for _, ent := range entries {
err := engine.MVCCPutProto(batch, nil, keys.RaftLogKey(r.RangeID, ent.Index),
roachpb.ZeroTimestamp, nil, &ent)
if err != nil {
return err
}
}
lastIndex := entries[len(entries)-1].Index
prevLastIndex := atomic.LoadUint64(&r.lastIndex)
// Delete any previously appended log entries which never committed.
for i := lastIndex + 1; i <= prevLastIndex; i++ {
err := engine.MVCCDelete(batch, nil,
keys.RaftLogKey(r.RangeID, i), roachpb.ZeroTimestamp, nil)
if err != nil {
return err
}
}
// Commit the batch and update the last index.
if err := setLastIndex(batch, r.RangeID, lastIndex); err != nil {
return err
}
batch.Defer(func() {
atomic.StoreUint64(&r.lastIndex, lastIndex)
})
return nil
}
// append the given entries to the raft log. Takes the previous value
// of r.lastIndex and returns a new value. We do this rather than
// modifying r.lastIndex directly because this modification needs to
// be atomic with the commit of the batch.
func (r *Replica) append(batch engine.ReadWriter, prevLastIndex uint64, entries []raftpb.Entry) (uint64, error) {
if len(entries) == 0 {
return prevLastIndex, nil
}
for i := range entries {
ent := &entries[i]
key := keys.RaftLogKey(r.RangeID, ent.Index)
if err := engine.MVCCPutProto(context.Background(), batch, nil, key, hlc.ZeroTimestamp, nil, ent); err != nil {
return 0, err
}
}
lastIndex := entries[len(entries)-1].Index
// Delete any previously appended log entries which never committed.
for i := lastIndex + 1; i <= prevLastIndex; i++ {
err := engine.MVCCDelete(context.Background(), batch, nil,
keys.RaftLogKey(r.RangeID, i), hlc.ZeroTimestamp, nil)
if err != nil {
return 0, err
}
}
if err := setLastIndex(batch, r.RangeID, lastIndex); err != nil {
return 0, err
}
return lastIndex, nil
}
// Del removes all abort cache entries for the given transaction.
func (sc *AbortCache) Del(
ctx context.Context,
e engine.Engine,
ms *engine.MVCCStats,
txnID *uuid.UUID,
) error {
key := keys.AbortCacheKey(sc.rangeID, txnID)
return engine.MVCCDelete(ctx, e, ms, key, roachpb.ZeroTimestamp, nil /* txn */)
}
// append the given entries to the raft log. Takes the previous values of
// r.mu.lastIndex and r.mu.raftLogSize, and returns new values. We do this
// rather than modifying them directly because these modifications need to be
// atomic with the commit of the batch.
func (r *Replica) append(
ctx context.Context,
batch engine.ReadWriter,
prevLastIndex uint64,
prevRaftLogSize int64,
entries []raftpb.Entry,
) (uint64, int64, error) {
if len(entries) == 0 {
return prevLastIndex, prevRaftLogSize, nil
}
var diff enginepb.MVCCStats
var value roachpb.Value
for i := range entries {
ent := &entries[i]
key := keys.RaftLogKey(r.RangeID, ent.Index)
if err := value.SetProto(ent); err != nil {
return 0, 0, err
}
value.InitChecksum(key)
var err error
if ent.Index > prevLastIndex {
err = engine.MVCCBlindPut(ctx, batch, &diff, key, hlc.ZeroTimestamp, value, nil /* txn */)
} else {
err = engine.MVCCPut(ctx, batch, &diff, key, hlc.ZeroTimestamp, value, nil /* txn */)
}
if err != nil {
return 0, 0, err
}
}
// Delete any previously appended log entries which never committed.
lastIndex := entries[len(entries)-1].Index
for i := lastIndex + 1; i <= prevLastIndex; i++ {
err := engine.MVCCDelete(ctx, batch, &diff, keys.RaftLogKey(r.RangeID, i),
hlc.ZeroTimestamp, nil /* txn */)
if err != nil {
return 0, 0, err
}
}
if err := setLastIndex(ctx, batch, r.RangeID, lastIndex); err != nil {
return 0, 0, err
}
raftLogSize := prevRaftLogSize + diff.SysBytes
return lastIndex, raftLogSize, nil
}
// Append implements the multiraft.WriteableGroupStorage interface.
func (r *Replica) Append(entries []raftpb.Entry) error {
if len(entries) == 0 {
return nil
}
batch := r.rm.Engine().NewBatch()
defer batch.Close()
rangeID := r.Desc().RangeID
for _, ent := range entries {
err := engine.MVCCPutProto(batch, nil, keys.RaftLogKey(rangeID, ent.Index),
proto.ZeroTimestamp, nil, &ent)
if err != nil {
return err
}
}
lastIndex := entries[len(entries)-1].Index
prevLastIndex := atomic.LoadUint64(&r.lastIndex)
// Delete any previously appended log entries which never committed.
for i := lastIndex + 1; i <= prevLastIndex; i++ {
err := engine.MVCCDelete(batch, nil,
keys.RaftLogKey(rangeID, i), proto.ZeroTimestamp, nil)
if err != nil {
return err
}
}
// Commit the batch and update the last index.
if err := setLastIndex(batch, rangeID, lastIndex); err != nil {
return err
}
if err := batch.Commit(); err != nil {
return err
}
atomic.StoreUint64(&r.lastIndex, lastIndex)
return nil
}
// 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
}
// 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
}
// Delete deletes the key and value specified by key.
func (r *Range) Delete(batch engine.Engine, ms *engine.MVCCStats, args *proto.DeleteRequest, reply *proto.DeleteResponse) {
reply.SetGoError(engine.MVCCDelete(batch, ms, args.Key, args.Timestamp, args.Txn))
}
// Delete deletes the key and value specified by key.
func (r *Range) Delete(batch engine.Engine, ms *engine.MVCCStats, args proto.DeleteRequest) (proto.DeleteResponse, error) {
var reply proto.DeleteResponse
return reply, engine.MVCCDelete(batch, ms, args.Key, args.Timestamp, args.Txn)
}
// 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
}
