func (s *store) Restore() error {
s.mu.Lock()
defer s.mu.Unlock()
min, max := newRevBytes(), newRevBytes()
revToBytes(revision{}, min)
revToBytes(revision{main: math.MaxInt64, sub: math.MaxInt64}, max)
// restore index
tx := s.b.BatchTx()
tx.Lock()
_, finishedCompactBytes := tx.UnsafeRange(metaBucketName, finishedCompactKeyName, nil, 0)
if len(finishedCompactBytes) != 0 {
s.compactMainRev = bytesToRev(finishedCompactBytes[0]).main
log.Printf("storage: restore compact to %d", s.compactMainRev)
}
// TODO: limit N to reduce max memory usage
keys, vals := tx.UnsafeRange(keyBucketName, min, max, 0)
for i, key := range keys {
var kv storagepb.KeyValue
if err := kv.Unmarshal(vals[i]); err != nil {
log.Fatalf("storage: cannot unmarshal event: %v", err)
}
rev := bytesToRev(key[:revBytesLen])
// restore index
switch {
case isTombstone(key):
s.kvindex.Tombstone(kv.Key, rev)
default:
s.kvindex.Restore(kv.Key, revision{kv.CreateRevision, 0}, rev, kv.Version)
}
// update revision
s.currentRev = rev
}
_, scheduledCompactBytes := tx.UnsafeRange(metaBucketName, scheduledCompactKeyName, nil, 0)
if len(scheduledCompactBytes) != 0 {
scheduledCompact := bytesToRev(scheduledCompactBytes[0]).main
if scheduledCompact > s.compactMainRev {
log.Printf("storage: resume scheduled compaction at %d", scheduledCompact)
go s.Compact(scheduledCompact)
}
}
tx.Unlock()
return nil
}
// range is a keyword in Go, add Keys suffix.
func (s *store) rangeKeys(key, end []byte, limit, rangeRev int64) (kvs []storagepb.KeyValue, curRev int64, err error) {
curRev = int64(s.currentRev.main)
if s.currentRev.sub > 0 {
curRev += 1
}
if rangeRev > curRev {
return nil, s.currentRev.main, ErrFutureRev
}
var rev int64
if rangeRev <= 0 {
rev = curRev
} else {
rev = rangeRev
}
if rev <= s.compactMainRev {
return nil, 0, ErrCompacted
}
_, revpairs := s.kvindex.Range(key, end, int64(rev))
if len(revpairs) == 0 {
return nil, curRev, nil
}
for _, revpair := range revpairs {
start, end := revBytesRange(revpair)
_, vs := s.tx.UnsafeRange(keyBucketName, start, end, 0)
if len(vs) != 1 {
log.Fatalf("storage: range cannot find rev (%d,%d)", revpair.main, revpair.sub)
}
var kv storagepb.KeyValue
if err := kv.Unmarshal(vs[0]); err != nil {
log.Fatalf("storage: cannot unmarshal event: %v", err)
}
kvs = append(kvs, kv)
if limit > 0 && len(kvs) >= int(limit) {
break
}
}
return kvs, curRev, nil
}
// kvsToEvents gets all events for the watchers from all key-value pairs
func kvsToEvents(wg *watcherGroup, revs, vals [][]byte) (evs []storagepb.Event) {
for i, v := range vals {
var kv storagepb.KeyValue
if err := kv.Unmarshal(v); err != nil {
log.Panicf("storage: cannot unmarshal event: %v", err)
}
if !wg.contains(string(kv.Key)) {
continue
}
ty := storagepb.PUT
if isTombstone(revs[i]) {
ty = storagepb.DELETE
// patch in mod revision so watchers won't skip
kv.ModRevision = bytesToRev(revs[i]).main
}
evs = append(evs, storagepb.Event{Kv: &kv, Type: ty})
}
return evs
}
func (s *store) delete(key []byte, rev revision) {
mainrev := s.currentRev.main + 1
ibytes := newRevBytes()
revToBytes(revision{main: mainrev, sub: s.currentRev.sub}, ibytes)
ibytes = appendMarkTombstone(ibytes)
kv := storagepb.KeyValue{
Key: key,
}
d, err := kv.Marshal()
if err != nil {
log.Fatalf("storage: cannot marshal event: %v", err)
}
s.tx.UnsafePut(keyBucketName, ibytes, d)
err = s.kvindex.Tombstone(key, revision{main: mainrev, sub: s.currentRev.sub})
if err != nil {
log.Fatalf("storage: cannot tombstone an existing key (%s): %v", string(key), err)
}
s.changes = append(s.changes, kv)
s.currentRev.sub += 1
ibytes = newRevBytes()
revToBytes(rev, ibytes)
_, vs := s.tx.UnsafeRange(keyBucketName, ibytes, nil, 0)
kv.Reset()
if err = kv.Unmarshal(vs[0]); err != nil {
log.Fatalf("storage: cannot unmarshal value: %v", err)
}
if lease.LeaseID(kv.Lease) != lease.NoLease {
err = s.le.Detach(lease.LeaseID(kv.Lease), []lease.LeaseItem{{Key: string(kv.Key)}})
if err != nil {
log.Fatalf("storage: cannot detach %v", err)
}
}
}
// RangeHistory ranges the history from key to end starting from startRev.
// If `end` is nil, the request only observes the events on key.
// If `end` is not nil, it observes the events on key range [key, range_end).
// Limit limits the number of events returned.
// If startRev <=0, rangeEvents returns events from the beginning of uncompacted history.
//
// If the required start rev is compacted, ErrCompacted will be returned.
// If the required start rev has not happened, ErrFutureRev will be returned.
//
// RangeHistory returns revision bytes slice and key-values that satisfy the requirement (0 <= n <= limit).
// If history in the revision range has not all happened, it returns immeidately
// what is available.
// It also returns nextRev which indicates the start revision used for the following
// RangeEvents call. The nextRev could be smaller than the given endRev if the store
// has not progressed so far or it hits the event limit.
//
// TODO: return byte slices instead of keyValues to avoid meaningless encode and decode.
// This also helps to return raw (key, val) pair directly to make API consistent.
func (s *store) RangeHistory(key, end []byte, limit, startRev int64) (revbs [][]byte, kvs []storagepb.KeyValue, nextRev int64, err error) {
s.mu.Lock()
defer s.mu.Unlock()
if startRev > 0 && startRev <= s.compactMainRev {
return nil, nil, 0, ErrCompacted
}
if startRev > s.currentRev.main {
return nil, nil, 0, ErrFutureRev
}
revs := s.kvindex.RangeSince(key, end, startRev)
if len(revs) == 0 {
return nil, nil, s.currentRev.main + 1, nil
}
tx := s.b.BatchTx()
tx.Lock()
defer tx.Unlock()
// fetch events from the backend using revisions
for _, rev := range revs {
start, end := revBytesRange(rev)
ks, vs := tx.UnsafeRange(keyBucketName, start, end, 0)
if len(vs) != 1 {
log.Fatalf("storage: range cannot find rev (%d,%d)", rev.main, rev.sub)
}
var kv storagepb.KeyValue
if err := kv.Unmarshal(vs[0]); err != nil {
log.Fatalf("storage: cannot unmarshal event: %v", err)
}
revbs = append(revbs, ks[0])
kvs = append(kvs, kv)
if limit > 0 && len(kvs) >= int(limit) {
return revbs, kvs, rev.main + 1, nil
}
}
return revbs, kvs, s.currentRev.main + 1, nil
}
// kvsToEvents gets all events for the watchers from all key-value pairs
func kvsToEvents(revs, vals [][]byte, wsk watcherSetByKey, pfxs map[string]struct{}) (evs []storagepb.Event) {
for i, v := range vals {
var kv storagepb.KeyValue
if err := kv.Unmarshal(v); err != nil {
log.Panicf("storage: cannot unmarshal event: %v", err)
}
k := string(kv.Key)
if _, ok := wsk.getSetByKey(k); !ok && !matchPrefix(k, pfxs) {
continue
}
ty := storagepb.PUT
if isTombstone(revs[i]) {
ty = storagepb.DELETE
// patch in mod revision so watchers won't skip
kv.ModRevision = bytesToRev(revs[i]).main
}
evs = append(evs, storagepb.Event{Kv: &kv, Type: ty})
}
return evs
}
// syncWatchers periodically syncs unsynced watchers by: Iterate all unsynced
// watchers to get the minimum revision within its range, skipping the
// watcher if its current revision is behind the compact revision of the
// store. And use this minimum revision to get all key-value pairs. Then send
// those events to watchers.
func (s *watchableStore) syncWatchers() {
s.store.mu.Lock()
defer s.store.mu.Unlock()
if len(s.unsynced) == 0 {
return
}
// in order to find key-value pairs from unsynced watchers, we need to
// find min revision index, and these revisions can be used to
// query the backend store of key-value pairs
minRev := int64(math.MaxInt64)
curRev := s.store.currentRev.main
compactionRev := s.store.compactMainRev
prefixes := make(map[string]struct{})
for _, set := range s.unsynced {
for w := range set {
k := string(w.key)
if w.cur > curRev {
panic("watcher current revision should not exceed current revision")
}
if w.cur < compactionRev {
select {
case w.ch <- WatchResponse{WatchID: w.id, CompactRevision: compactionRev}:
s.unsynced.delete(w)
default:
// retry next time
}
continue
}
if minRev >= w.cur {
minRev = w.cur
}
if w.prefix {
prefixes[k] = struct{}{}
}
}
}
minBytes, maxBytes := newRevBytes(), newRevBytes()
revToBytes(revision{main: minRev}, minBytes)
revToBytes(revision{main: curRev + 1}, maxBytes)
// UnsafeRange returns keys and values. And in boltdb, keys are revisions.
// values are actual key-value pairs in backend.
tx := s.store.b.BatchTx()
tx.Lock()
ks, vs := tx.UnsafeRange(keyBucketName, minBytes, maxBytes, 0)
evs := []storagepb.Event{}
// get the list of all events from all key-value pairs
for i, v := range vs {
var kv storagepb.KeyValue
if err := kv.Unmarshal(v); err != nil {
log.Panicf("storage: cannot unmarshal event: %v", err)
}
k := string(kv.Key)
if _, ok := s.unsynced.getSetByKey(k); !ok && !matchPrefix(k, prefixes) {
continue
}
var ev storagepb.Event
switch {
case isTombstone(ks[i]):
ev.Type = storagepb.DELETE
default:
ev.Type = storagepb.PUT
}
ev.Kv = &kv
evs = append(evs, ev)
}
tx.Unlock()
for w, es := range newWatcherToEventMap(s.unsynced, evs) {
select {
// s.store.Rev also uses Lock, so just return directly
case w.ch <- WatchResponse{WatchID: w.id, Events: es, Revision: s.store.currentRev.main}:
pendingEventsGauge.Add(float64(len(es)))
default:
// TODO: handle the full unsynced watchers.
// continue to process other watchers for now, the full ones
// will be processed next time and hopefully it will not be full.
continue
}
w.cur = curRev
s.synced.add(w)
s.unsynced.delete(w)
}
slowWatcherGauge.Set(float64(len(s.unsynced)))
}
func (s *store) put(key, value []byte, leaseID lease.LeaseID) {
rev := s.currentRev.main + 1
c := rev
oldLease := lease.NoLease
// if the key exists before, use its previous created and
// get its previous leaseID
grev, created, ver, err := s.kvindex.Get(key, rev)
if err == nil {
c = created.main
ibytes := newRevBytes()
revToBytes(grev, ibytes)
_, vs := s.tx.UnsafeRange(keyBucketName, ibytes, nil, 0)
var kv storagepb.KeyValue
if err = kv.Unmarshal(vs[0]); err != nil {
log.Fatalf("storage: cannot unmarshal value: %v", err)
}
oldLease = lease.LeaseID(kv.Lease)
}
ibytes := newRevBytes()
revToBytes(revision{main: rev, sub: s.currentRev.sub}, ibytes)
ver = ver + 1
kv := storagepb.KeyValue{
Key: key,
Value: value,
CreateRevision: c,
ModRevision: rev,
Version: ver,
Lease: int64(leaseID),
}
d, err := kv.Marshal()
if err != nil {
log.Fatalf("storage: cannot marshal event: %v", err)
}
s.tx.UnsafePut(keyBucketName, ibytes, d)
s.kvindex.Put(key, revision{main: rev, sub: s.currentRev.sub})
s.changes = append(s.changes, kv)
s.currentRev.sub += 1
if oldLease != lease.NoLease {
if s.le == nil {
panic("no lessor to detach lease")
}
err = s.le.Detach(oldLease, []lease.LeaseItem{{Key: string(key)}})
if err != nil {
panic("unexpected error from lease detach")
}
}
if leaseID != lease.NoLease {
if s.le == nil {
panic("no lessor to attach lease")
}
err = s.le.Attach(leaseID, []lease.LeaseItem{{Key: string(key)}})
if err != nil {
panic("unexpected error from lease Attach")
}
}
}
func (s *store) restore() error {
min, max := newRevBytes(), newRevBytes()
revToBytes(revision{main: 1}, min)
revToBytes(revision{main: math.MaxInt64, sub: math.MaxInt64}, max)
// restore index
tx := s.b.BatchTx()
tx.Lock()
_, finishedCompactBytes := tx.UnsafeRange(metaBucketName, finishedCompactKeyName, nil, 0)
if len(finishedCompactBytes) != 0 {
s.compactMainRev = bytesToRev(finishedCompactBytes[0]).main
log.Printf("storage: restore compact to %d", s.compactMainRev)
}
// TODO: limit N to reduce max memory usage
keys, vals := tx.UnsafeRange(keyBucketName, min, max, 0)
for i, key := range keys {
var kv storagepb.KeyValue
if err := kv.Unmarshal(vals[i]); err != nil {
log.Fatalf("storage: cannot unmarshal event: %v", err)
}
rev := bytesToRev(key[:revBytesLen])
// restore index
switch {
case isTombstone(key):
s.kvindex.Tombstone(kv.Key, rev)
if lease.LeaseID(kv.Lease) != lease.NoLease {
err := s.le.Detach(lease.LeaseID(kv.Lease), []lease.LeaseItem{{Key: string(kv.Key)}})
if err != nil && err != lease.ErrLeaseNotFound {
log.Fatalf("storage: unexpected Detach error %v", err)
}
}
default:
s.kvindex.Restore(kv.Key, revision{kv.CreateRevision, 0}, rev, kv.Version)
if lease.LeaseID(kv.Lease) != lease.NoLease {
if s.le == nil {
panic("no lessor to attach lease")
}
err := s.le.Attach(lease.LeaseID(kv.Lease), []lease.LeaseItem{{Key: string(kv.Key)}})
// We are walking through the kv history here. It is possible that we attached a key to
// the lease and the lease was revoked later.
// Thus attaching an old version of key to a none existing lease is possible here, and
// we should just ignore the error.
if err != nil && err != lease.ErrLeaseNotFound {
panic("unexpected Attach error")
}
}
}
// update revision
s.currentRev = rev
}
_, scheduledCompactBytes := tx.UnsafeRange(metaBucketName, scheduledCompactKeyName, nil, 0)
if len(scheduledCompactBytes) != 0 {
scheduledCompact := bytesToRev(scheduledCompactBytes[0]).main
if scheduledCompact > s.compactMainRev {
log.Printf("storage: resume scheduled compaction at %d", scheduledCompact)
go s.Compact(scheduledCompact)
}
}
tx.Unlock()
return nil
}
// syncWatchers periodically syncs unsynced watchers by: Iterate all unsynced
// watchers to get the minimum revision within its range, skipping the
// watcher if its current revision is behind the compact revision of the
// store. And use this minimum revision to get all key-value pairs. Then send
// those events to watchers.
func (s *watchableStore) syncWatchers() {
s.store.mu.Lock()
defer s.store.mu.Unlock()
if len(s.unsynced) == 0 {
return
}
// in order to find key-value pairs from unsynced watchers, we need to
// find min revision index, and these revisions can be used to
// query the backend store of key-value pairs
minRev := int64(math.MaxInt64)
curRev := s.store.currentRev.main
compactionRev := s.store.compactMainRev
// TODO: change unsynced struct type same to this
keyToUnsynced := make(map[string]map[*watcher]struct{})
for w := range s.unsynced {
k := string(w.key)
if w.cur > curRev {
panic("watcher current revision should not exceed current revision")
}
if w.cur < compactionRev {
// TODO: return error compacted to that watcher instead of
// just removing it sliently from unsynced.
delete(s.unsynced, w)
continue
}
if minRev >= w.cur {
minRev = w.cur
}
if _, ok := keyToUnsynced[k]; !ok {
keyToUnsynced[k] = make(map[*watcher]struct{})
}
keyToUnsynced[k][w] = struct{}{}
}
minBytes, maxBytes := newRevBytes(), newRevBytes()
revToBytes(revision{main: minRev}, minBytes)
revToBytes(revision{main: curRev + 1}, maxBytes)
// UnsafeRange returns keys and values. And in boltdb, keys are revisions.
// values are actual key-value pairs in backend.
tx := s.store.b.BatchTx()
tx.Lock()
ks, vs := tx.UnsafeRange(keyBucketName, minBytes, maxBytes, 0)
tx.Unlock()
evs := []storagepb.Event{}
// get the list of all events from all key-value pairs
for i, v := range vs {
var kv storagepb.KeyValue
if err := kv.Unmarshal(v); err != nil {
log.Panicf("storage: cannot unmarshal event: %v", err)
}
k := string(kv.Key)
if _, ok := keyToUnsynced[k]; !ok {
continue
}
var ev storagepb.Event
switch {
case isTombstone(ks[i]):
ev.Type = storagepb.DELETE
default:
ev.Type = storagepb.PUT
}
ev.Kv = &kv
evs = append(evs, ev)
}
for w, es := range newWatcherToEventMap(keyToUnsynced, evs) {
wr := WatchResponse{WatchID: w.id, Events: es}
select {
case w.ch <- wr:
pendingEventsGauge.Add(float64(len(es)))
default:
// TODO: handle the full unsynced watchers.
// continue to process other watchers for now, the full ones
// will be processed next time and hopefully it will not be full.
continue
}
k := string(w.key)
if err := unsafeAddWatcher(&s.synced, k, w); err != nil {
log.Panicf("error unsafeAddWatcher (%v) for key %s", err, k)
}
delete(s.unsynced, w)
}
slowWatcherGauge.Set(float64(len(s.unsynced)))
}