// 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
}
// mergeTrigger is called on a successful commit of an AdminMerge
// transaction. It recomputes stats for the receiving range.
func (r *Range) mergeTrigger(batch engine.Engine, merge *proto.MergeTrigger) error {
if !bytes.Equal(r.Desc().StartKey, merge.UpdatedDesc.StartKey) {
return util.Errorf("range and updated range start keys do not match: %s != %s",
r.Desc().StartKey, merge.UpdatedDesc.StartKey)
}
if !r.Desc().EndKey.Less(merge.UpdatedDesc.EndKey) {
return util.Errorf("range end key is not less than the post merge end key: %s >= %s",
r.Desc().EndKey, merge.UpdatedDesc.EndKey)
}
if merge.SubsumedRaftID <= 0 {
return util.Errorf("subsumed raft ID must be provided: %d", merge.SubsumedRaftID)
}
// Copy the subsumed range's response cache to the subsuming one.
if err := r.respCache.CopyFrom(batch, merge.SubsumedRaftID); err != nil {
return util.Errorf("unable to copy response cache to new split range: %s", err)
}
// Compute stats for updated range.
now := r.rm.Clock().Timestamp()
iter := newRangeDataIterator(&merge.UpdatedDesc, batch)
ms, err := engine.MVCCComputeStats(iter, now.WallTime)
iter.Close()
if err != nil {
return util.Errorf("unable to compute stats for the range after merge: %s", err)
}
if err = r.stats.SetMVCCStats(batch, ms); err != nil {
return util.Errorf("unable to write MVCC stats: %s", err)
}
// Clear the timestamp cache. In the case that this replica and the
// subsumed replica each held their respective leader leases, we
// could merge the timestamp caches for efficiency. But it's unlikely
// and not worth the extra logic and potential for error.
r.Lock()
r.tsCache.Clear(r.rm.Clock())
r.Unlock()
batch.Defer(func() {
if err := r.rm.MergeRange(r, merge.UpdatedDesc.EndKey, merge.SubsumedRaftID); err != nil {
// Our in-memory state has diverged from the on-disk state.
log.Fatalf("failed to update store after merging range: %s", err)
}
})
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
}
// splitTrigger is called on a successful commit of an AdminSplit
// transaction. It copies the response cache for the new range and
// recomputes stats for both the existing, updated range and the new
// range.
func (r *Range) splitTrigger(batch engine.Engine, split *proto.SplitTrigger) error {
if !bytes.Equal(r.Desc().StartKey, split.UpdatedDesc.StartKey) ||
!bytes.Equal(r.Desc().EndKey, split.NewDesc.EndKey) {
return util.Errorf("range does not match splits: (%s-%s) + (%s-%s) != %s",
split.UpdatedDesc.StartKey, split.UpdatedDesc.EndKey,
split.NewDesc.StartKey, split.NewDesc.EndKey, r)
}
// Copy the GC metadata.
gcMeta, err := r.GetGCMetadata()
if err != nil {
return util.Errorf("unable to fetch GC metadata: %s", err)
}
if err := engine.MVCCPutProto(batch, nil, keys.RangeGCMetadataKey(split.NewDesc.RaftID), proto.ZeroTimestamp, nil, gcMeta); err != nil {
return util.Errorf("unable to copy GC metadata: %s", err)
}
// Copy the last verification timestamp.
verifyTS, err := r.GetLastVerificationTimestamp()
if err != nil {
return util.Errorf("unable to fetch last verification timestamp: %s", err)
}
if err := engine.MVCCPutProto(batch, nil, keys.RangeLastVerificationTimestampKey(split.NewDesc.RaftID), proto.ZeroTimestamp, nil, &verifyTS); err != nil {
return util.Errorf("unable to copy last verification timestamp: %s", err)
}
// Compute stats for updated range.
now := r.rm.Clock().Timestamp()
iter := newRangeDataIterator(&split.UpdatedDesc, batch)
ms, err := engine.MVCCComputeStats(iter, now.WallTime)
iter.Close()
if err != nil {
return util.Errorf("unable to compute stats for updated range after split: %s", err)
}
if err := r.stats.SetMVCCStats(batch, ms); err != nil {
return util.Errorf("unable to write MVCC stats: %s", err)
}
// Initialize the new range's response cache by copying the original's.
if err = r.respCache.CopyInto(batch, split.NewDesc.RaftID); err != nil {
return util.Errorf("unable to copy response cache to new split range: %s", err)
}
// Add the new split range to the store. This step atomically
// updates the EndKey of the updated range and also adds the
// new range to the store's range map.
newRng, err := NewRange(&split.NewDesc, r.rm)
if err != nil {
return err
}
// Compute stats for new range.
iter = newRangeDataIterator(&split.NewDesc, batch)
ms, err = engine.MVCCComputeStats(iter, now.WallTime)
iter.Close()
if err != nil {
return util.Errorf("unable to compute stats for new range after split: %s", err)
}
if err = newRng.stats.SetMVCCStats(batch, ms); err != nil {
return util.Errorf("unable to write MVCC stats: %s", err)
}
// Copy the timestamp cache into the new range.
r.Lock()
r.tsCache.MergeInto(newRng.tsCache, true /* clear */)
r.Unlock()
batch.Defer(func() {
if err := r.rm.SplitRange(r, newRng); err != nil {
// Our in-memory state has diverged from the on-disk state.
log.Fatalf("failed to update Store after split: %s", err)
}
})
return nil
}