// In benchmarks, the time taken per AddMutation before was
// plateauing at 2.5 ms with sync per 10 log entries, and increasing
// for sync per 100 log entries (to 3 ms per AddMutation), largely because
// of how index generation was being done.
//
// With this change, the benchmarks perform as good as benchmarks for
// commit.Logger, where the less frequently file sync happens, the faster
// AddMutations run.
//
// PASS
// BenchmarkAddMutations_SyncEveryLogEntry-6 100 24712455 ns/op
// BenchmarkAddMutations_SyncEvery10LogEntry-6 500 2485961 ns/op
// BenchmarkAddMutations_SyncEvery100LogEntry-6 10000 298352 ns/op
// BenchmarkAddMutations_SyncEvery1000LogEntry-6 30000 63544 ns/op
// ok github.com/dgraph-io/dgraph/posting 10.291s
func (l *List) AddMutation(t x.DirectedEdge, op byte) error {
l.wg.Wait()
l.Lock()
defer l.Unlock()
if l.deleteMe {
return E_TMP_ERROR
}
if t.Timestamp.UnixNano() < l.maxMutationTs {
return fmt.Errorf("Mutation ts lower than committed ts.")
}
// Mutation arrives:
// - Check if we had any(SET/DEL) before this, stored in the mutation list.
// - If yes, then replace that mutation. Jump to a)
// a) check if the entity exists in main posting list.
// - If yes, store the mutation.
// - If no, disregard this mutation.
// All edges with a value set, have the same uid. In other words,
// an (entity, attribute) can only have one interface{} value.
if t.Value != nil {
t.ValueId = math.MaxUint64
}
if t.ValueId == 0 {
return fmt.Errorf("ValueId cannot be zero.")
}
mbuf := newPosting(t, op)
uo := flatbuffers.GetUOffsetT(mbuf)
mpost := new(types.Posting)
mpost.Init(mbuf, uo)
glog.WithFields(logrus.Fields{
"uid": mpost.Uid(),
"source": string(mpost.Source()),
"ts": mpost.Ts(),
}).Debug("Add mutation")
l.mergeMutation(mpost)
l.maxMutationTs = t.Timestamp.UnixNano()
if len(l.mindex)+len(l.mlayer) > 0 {
atomic.StoreInt64(&l.dirtyTs, time.Now().UnixNano())
if dirtymap != nil {
dirtymap.Put(l.ghash, true)
}
}
if l.clog == nil {
return nil
}
return l.clog.AddLog(t.Timestamp.UnixNano(), l.hash, mbuf)
}
func addPosting(b *flatbuffers.Builder, p types.Posting) flatbuffers.UOffsetT {
so := b.CreateByteString(p.Source()) // Do this before posting start.
var bo flatbuffers.UOffsetT
if p.ValueLength() > 0 {
bo = b.CreateByteVector(p.ValueBytes())
}
types.PostingStart(b)
types.PostingAddUid(b, p.Uid())
if bo > 0 {
types.PostingAddValue(b, bo)
}
types.PostingAddSource(b, so)
types.PostingAddTs(b, p.Ts())
types.PostingAddOp(b, p.Op())
return types.PostingEnd(b)
}
func (l *List) init(key []byte, pstore *store.Store, clog *commit.Logger) {
l.Lock()
defer l.Unlock()
defer l.wg.Done()
if len(empty) == 0 {
glog.Fatal("empty should have some bytes.")
}
l.key = key
l.pstore = pstore
l.clog = clog
posting := l.getPostingList()
l.maxMutationTs = posting.CommitTs()
l.hash = farm.Fingerprint32(key)
l.ghash = gotomic.IntKey(farm.Fingerprint64(key))
l.mlayer = make(map[int]types.Posting)
if clog == nil {
return
}
glog.Debug("Starting stream entries...")
err := clog.StreamEntries(posting.CommitTs()+1, l.hash,
func(hdr commit.Header, buffer []byte) {
uo := flatbuffers.GetUOffsetT(buffer)
m := new(types.Posting)
m.Init(buffer, uo)
if m.Ts() > l.maxMutationTs {
l.maxMutationTs = m.Ts()
}
glog.WithFields(logrus.Fields{
"uid": m.Uid(),
"source": string(m.Source()),
"ts": m.Ts(),
}).Debug("Got entry from log")
l.mergeMutation(m)
})
if err != nil {
glog.WithError(err).Error("While streaming entries.")
}
glog.Debug("Done streaming entries.")
}