func ExternalId(uid uint64) (xid string, rerr error) {
key := posting.Key(uid, "_xid_") // uid -> "_xid_" -> xid
pl := posting.GetOrCreate(key, uidStore)
if pl.Length() == 0 {
return "", errors.New("NO external id")
}
if pl.Length() > 1 {
glog.WithField("uid", uid).Fatal("This shouldn't be happening.")
return "", errors.New("Multiple external ids for this uid.")
}
var p types.Posting
if ok := pl.Get(&p, 0); !ok {
glog.WithField("uid", uid).Error("While retrieving posting")
return "", errors.New("While retrieving posting")
}
if p.Uid() != math.MaxUint64 {
glog.WithField("uid", uid).Fatal("Value uid must be MaxUint64.")
}
var t interface{}
rerr = posting.ParseValue(&t, p.ValueBytes())
xid = t.(string)
return xid, rerr
}
// Caller must hold at least a read lock.
func (l *List) lePostingIndex(maxUid uint64) (int, uint64) {
posting := l.getPostingList()
left, right := 0, posting.PostingsLength()-1
sofar := -1
p := new(types.Posting)
for left <= right {
pos := (left + right) / 2
if ok := posting.Postings(p, pos); !ok {
glog.WithField("idx", pos).Fatal("Unable to parse posting from list.")
}
val := p.Uid()
if val > maxUid {
right = pos - 1
continue
}
if val == maxUid {
return pos, val
}
sofar = pos
left = pos + 1
}
if sofar == -1 {
return -1, 0
}
if ok := posting.Postings(p, sofar); !ok {
glog.WithField("idx", sofar).Fatal("Unable to parse posting from list.")
}
return sofar, p.Uid()
}
func assignNew(pl *posting.List, xid string, instanceIdx uint64,
numInstances uint64) (uint64, error) {
entry := lmgr.newOrExisting(xid)
entry.Lock()
entry.ts = time.Now()
defer entry.Unlock()
if pl.Length() > 1 {
glog.Fatalf("We shouldn't have more than 1 uid for xid: %v\n", xid)
} else if pl.Length() > 0 {
var p types.Posting
if ok := pl.Get(&p, 0); !ok {
return 0, errors.New("While retrieving entry from posting list.")
}
return p.Uid(), nil
}
// No current id exists. Create one.
uid, err := allocateUniqueUid(xid, instanceIdx, numInstances)
if err != nil {
return 0, err
}
t := x.DirectedEdge{
ValueId: uid,
Source: "_assigner_",
Timestamp: time.Now(),
}
rerr := pl.AddMutation(t, posting.Set)
return uid, rerr
}
// 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 samePosting(a *types.Posting, b *types.Posting) bool {
if a.Uid() != b.Uid() {
return false
}
if a.ValueLength() != b.ValueLength() {
return false
}
if !bytes.Equal(a.ValueBytes(), b.ValueBytes()) {
return false
}
if !bytes.Equal(a.Source(), b.Source()) {
return false
}
return true
}
func checkUids(t *testing.T, l *List, uids ...uint64) error {
if l.Length() != len(uids) {
return fmt.Errorf("Expected: %d. Length: %d", len(uids), l.Length())
}
for i := 0; i < len(uids); i++ {
var p types.Posting
if ok := l.Get(&p, i); !ok {
return fmt.Errorf("Unable to retrieve posting.")
}
if p.Uid() != uids[i] {
return fmt.Errorf("Expected: %v. Got: %v", uids[i], p.Uid())
}
}
return nil
}
func (l *List) GetUids() []uint64 {
l.wg.Wait()
l.RLock()
defer l.RUnlock()
result := make([]uint64, l.length())
result = result[:0]
var p types.Posting
for i := 0; i < l.length(); i++ {
if ok := l.get(&p, i); !ok || p.Uid() == math.MaxUint64 {
break
}
result = append(result, p.Uid())
}
return result
}
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) Value() (result []byte, rerr error) {
l.wg.Wait()
l.RLock()
defer l.RUnlock()
if l.length() == 0 {
return result, fmt.Errorf("No value found")
}
var p types.Posting
if ok := l.get(&p, l.length()-1); !ok {
return result, fmt.Errorf("Unable to get last posting")
}
if p.Uid() != math.MaxUint64 {
return result, fmt.Errorf("No value found")
}
return p.ValueBytes(), nil
}
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.")
}
func GetOrAssign(xid string, instanceIdx uint64, numInstances uint64) (uid uint64, rerr error) {
key := stringKey(xid)
pl := posting.GetOrCreate(key)
if pl.Length() == 0 {
return assignNew(pl, xid, instanceIdx, numInstances)
} else if pl.Length() > 1 {
glog.Fatalf("We shouldn't have more than 1 uid for xid: %v\n", xid)
} else {
// We found one posting.
var p types.Posting
if ok := pl.Get(&p, 0); !ok {
return 0, errors.New("While retrieving entry from posting list")
}
return p.Uid(), nil
}
return 0, errors.New("Some unhandled route lead me here." +
" Wake the stupid developer up.")
}
// mutationIndex (mindex) is useful to avoid having to parse the entire
// postinglist upto idx, for every Get(*types.Posting, idx), which has a
// complexity of O(idx). Iteration over N size posting list would this push
// us into O(N^2) territory, without this technique.
//
// Using this technique,
// we can overlay mutation layers over immutable posting list, to allow for
// O(m) lookups, where m = size of mutation list. Obviously, the size of
// mutation list should be much smaller than the size of posting list, except
// in tiny posting lists, where performance wouldn't be such a concern anyways.
//
// Say we have this data:
// Posting List (plist, immutable):
// idx: 0 1 2 3 4 5
// value: 2 5 9 10 13 15
//
// Mutation List (mlist):
// idx: 0 1 2
// value: 7 10 13' // posting uid is 13 but other values vary.
// Op: SET DEL SET
// Effective: ADD DEL REP (REP = replace)
//
// ----------------------------------------------------------------------------
// regenerateIndex would generate these:
// mlayer (layer just above posting list contains only replace instructions)
// idx: 4
// value: 13'
// Op: SET
// Effective: REP (REP = replace)
//
// mindex:
// idx: 2 4
// value: 7 10
// moveidx: -1 +1
// Effective: ADD DEL
//
// Now, let's see how the access would work:
// idx: get --> calculation [idx, served from, value]
// idx: 0 --> 0 [0, plist, 2]
// idx: 1 --> 1 [1, plist, 5]
// idx: 2 --> ADD from mindex
// --> [2, mindex, 7] // also has moveidx = -1
// idx: 3 --> 3 + moveidx=-1 = 2 [2, plist, 9]
// idx: 4 --> DEL from mindex
// --> 4 + moveidx=-1 + moveidx=+1 = 4 [4, mlayer, 13']
// idx: 5 --> 5 + moveidx=-1 + moveidx=+1 = 5 [5, plist, 15]
//
// Thus we can provide mutation layers over immutable posting list, while
// still ensuring fast lookup access.
//
// NOTE: This function expects the caller to hold a RW Lock.
// Update: With mergeMutation function, we're adding mutations with a cost
// of O(log M + log N), where M = number of previous mutations, and N =
// number of postings in the immutable posting list.
func (l *List) mergeMutation(mp *types.Posting) {
curUid := mp.Uid()
pi, puid := l.lePostingIndex(curUid) // O(log N)
mi, muid := l.leMutationIndex(curUid) // O(log M)
inPlist := puid == curUid
// O(1) follows, but any additions or deletions from mindex would
// be O(M) due to element shifting. In terms of benchmarks, this performs
// a LOT better than when I was running O(N + M), re-generating mutation
// flatbuffers, linked lists etc.
mlink := new(MutationLink)
mlink.posting = mp
if mp.Op() == Del {
if muid == curUid { // curUid found in mindex.
if inPlist { // In plist, so replace previous instruction in mindex.
mlink.moveidx = 1
mlink.idx = pi + mi
l.mindex[mi] = mlink
} else { // Not in plist, so delete previous instruction in mindex.
l.mdelta -= 1
l.mindexDeleteAt(mi)
}
} else { // curUid not found in mindex.
if inPlist { // In plist, so insert in mindex.
mlink.moveidx = 1
l.mdelta -= 1
mlink.idx = pi + mi + 1
l.mindexInsertAt(mlink, mi+1)
} else {
// Not found in plist, and not found in mindex. So, ignore.
}
}
} else if mp.Op() == Set {
if muid == curUid { // curUid found in mindex.
if inPlist { // In plist, so delete previous instruction, set in mlayer.
l.mindexDeleteAt(mi)
l.mlayer[pi] = *mp
} else { // Not in plist, so replace previous set instruction in mindex.
// NOTE: This prev instruction couldn't have been a Del instruction.
mlink.idx = pi + 1 + mi
mlink.moveidx = -1
l.mindex[mi] = mlink
}
} else { // curUid not found in mindex.
if inPlist { // In plist, so just set it in mlayer.
// If this posting matches what we already have in posting list,
// we don't need to `dirty` this by adding to mlayer.
plist := l.getPostingList()
var cp types.Posting
if ok := plist.Postings(&cp, pi); ok {
if samePosting(&cp, mp) {
return // do nothing.
}
}
l.mlayer[pi] = *mp
} else { // not in plist, not in mindex, so insert in mindex.
mlink.moveidx = -1
l.mdelta += 1
mlink.idx = pi + 1 + mi + 1 // right of pi, and right of mi.
l.mindexInsertAt(mlink, mi+1)
}
}
} else {
glog.WithField("op", mp.Op()).Fatal("Invalid operation.")
}
}
func TestAddMutation(t *testing.T) {
// logrus.SetLevel(logrus.DebugLevel)
l := NewList()
key := Key(1, "name")
dir, err := ioutil.TempDir("", "storetest_")
if err != nil {
t.Error(err)
return
}
defer os.RemoveAll(dir)
ps := new(store.Store)
ps.Init(dir)
clog := commit.NewLogger(dir, "mutations", 50<<20)
clog.Init()
defer clog.Close()
l.init(key, ps, clog)
edge := x.DirectedEdge{
ValueId: 9,
Source: "testing",
Timestamp: time.Now(),
}
if err := l.AddMutation(edge, Set); err != nil {
t.Error(err)
}
/*
if err := l.CommitIfDirty(); err != nil {
t.Error(err)
}
*/
if l.Length() != 1 {
t.Error("Unable to find added elements in posting list")
}
var p types.Posting
if ok := l.Get(&p, 0); !ok {
t.Error("Unable to retrieve posting at 1st iter")
t.Fail()
}
if p.Uid() != 9 {
t.Errorf("Expected 9. Got: %v", p.Uid)
}
if string(p.Source()) != "testing" {
t.Errorf("Expected testing. Got: %v", string(p.Source()))
}
// return // Test 1.
// Add another edge now.
edge.ValueId = 81
l.AddMutation(edge, Set)
// l.CommitIfDirty()
if l.Length() != 2 {
t.Errorf("Length: %d", l.Length())
t.Fail()
}
var uid uint64
uid = 1
for i := 0; i < l.Length(); i++ {
if ok := l.Get(&p, i); !ok {
t.Error("Unable to retrieve posting at 2nd iter")
}
uid *= 9
if p.Uid() != uid {
t.Logf("Expected: %v. Got: %v", uid, p.Uid())
}
}
// return // Test 2.
// Add another edge, in between the two above.
uids := []uint64{
9, 49, 81,
}
edge.ValueId = 49
if err := l.AddMutation(edge, Set); err != nil {
t.Error(err)
}
/*
if err := l.CommitIfDirty(); err != nil {
t.Error(err)
}
*/
if err := checkUids(t, l, uids...); err != nil {
t.Error(err)
}
// return // Test 3.
// Delete an edge, add an edge, replace an edge
edge.ValueId = 49
if err := l.AddMutation(edge, Del); err != nil {
t.Error(err)
}
edge.ValueId = 69
if err := l.AddMutation(edge, Set); err != nil {
t.Error(err)
}
edge.ValueId = 9
edge.Source = "anti-testing"
if err := l.AddMutation(edge, Set); err != nil {
t.Error(err)
}
/*
if err := l.CommitIfDirty(); err != nil {
t.Error(err)
}
*/
uids = []uint64{9, 69, 81}
if err := checkUids(t, l, uids...); err != nil {
t.Error(err)
}
l.Get(&p, 0)
if string(p.Source()) != "anti-testing" {
t.Errorf("Expected: anti-testing. Got: %v", string(p.Source()))
}
/*
if err := l.CommitIfDirty(); err != nil {
t.Error(err)
}
*/
// Try reading the same data in another PostingList.
dl := NewList()
dl.init(key, ps, clog)
if err := checkUids(t, dl, uids...); err != nil {
t.Error(err)
}
if _, err := dl.MergeIfDirty(); err != nil {
t.Error(err)
}
if err := checkUids(t, dl, uids...); err != nil {
t.Error(err)
}
}
func TestAddMutation_Value(t *testing.T) {
// logrus.SetLevel(logrus.DebugLevel)
glog.Debug("Running init...")
ol := NewList()
key := Key(10, "value")
dir, err := ioutil.TempDir("", "storetest_")
if err != nil {
t.Error(err)
return
}
defer os.RemoveAll(dir)
ps := new(store.Store)
ps.Init(dir)
clog := commit.NewLogger(dir, "mutations", 50<<20)
clog.Init()
defer clog.Close()
ol.init(key, ps, clog)
glog.Debug("Init successful.")
edge := x.DirectedEdge{
Value: "oh hey there",
Source: "new-testing",
Timestamp: time.Now(),
}
if err := ol.AddMutation(edge, Set); err != nil {
t.Error(err)
}
var p types.Posting
ol.Get(&p, 0)
if p.Uid() != math.MaxUint64 {
t.Errorf("All value uids should go to MaxUint64. Got: %v", p.Uid())
}
var iout interface{}
if err := ParseValue(&iout, p.ValueBytes()); err != nil {
t.Error(err)
}
out := iout.(string)
if out != "oh hey there" {
t.Errorf("Expected a value. Got: [%q]", out)
}
// Run the same check after committing.
if _, err := ol.MergeIfDirty(); err != nil {
t.Error(err)
}
{
var tp types.Posting
if ok := ol.Get(&tp, 0); !ok {
t.Error("While retrieving posting")
}
if err := ParseValue(&iout, tp.ValueBytes()); err != nil {
t.Error(err)
}
out := iout.(string)
if out != "oh hey there" {
t.Errorf("Expected a value. Got: [%q]", out)
}
}
// The value made it to the posting list. Changing it now.
edge.Value = 119
if err := ol.AddMutation(edge, Set); err != nil {
t.Error(err)
}
if ol.Length() != 1 {
t.Errorf("Length should be one. Got: %v", ol.Length())
}
if ok := ol.Get(&p, 0); !ok {
t.Error("While retrieving posting")
}
if err := ParseValue(&iout, p.ValueBytes()); err != nil {
t.Error(err)
}
intout := iout.(float64)
if intout != 119 {
t.Errorf("Expected 119. Got: %v", intout)
}
}