// Calculate the set of index blocks that need to be loaded.
// blocksByVb stores which blocks need to be processed for each vbucket, in ascending vbucket order. Multiple vb map
// values can point to the same IndexBlock (i.e. when those vbs share a partition).
// blocksByKey stores all IndexBlocks to be retrieved, indexed by block key - no duplicates
func (b *BitFlagStorage) calculateChangedBlocks(fromSeq base.SequenceClock, channelClock base.SequenceClock) (blocksByKey map[string]IndexBlock, blocksByVb []vbBlockSet, err error) {
blocksByKey = make(map[string]IndexBlock, 1)
blocksByVb = make([]vbBlockSet, 0)
for vbNo, clockVbSeq := range channelClock.Value() {
fromVbSeq := fromSeq.GetSequence(uint16(vbNo))
// Verify that the requested from value is less than the channel clock sequence (there are
// new entries for this vbucket in the channel)
if fromVbSeq >= clockVbSeq {
continue
}
blockSet := vbBlockSet{vbNo: uint16(vbNo)}
partition := b.partitions.VbMap[uint16(vbNo)]
for _, blockIndex := range generateBitFlagBlockIndexes(b.channelName, fromVbSeq, clockVbSeq, partition) {
blockKey := getIndexBlockKey(b.channelName, blockIndex, partition)
block, found := blocksByKey[blockKey]
if !found {
block = newBitFlagBufferBlockForKey(blockKey, b.channelName, blockIndex, partition, b.partitions.VbPositionMaps[partition])
blocksByKey[blockKey] = block
}
blockSet.blocks = append(blockSet.blocks, block)
}
blocksByVb = append(blocksByVb, blockSet)
}
return blocksByKey, blocksByVb, nil
}
func (b *BitFlagStorage) GetChanges(fromSeq base.SequenceClock, toSeq base.SequenceClock) ([]*LogEntry, error) {
// Determine which blocks have changed, and load those blocks
blocksByKey, blocksByVb, err := b.calculateChangedBlocks(fromSeq, toSeq)
if err != nil {
return nil, err
}
b.bulkLoadBlocks(blocksByKey)
// For each vbucket, create the entries from the blocks. Create in reverse sequence order, for
// deduplication once we've retrieved the full entry
entries := make([]*LogEntry, 0)
entryKeys := make([]string, 0)
for blockSetIndex := len(blocksByVb) - 1; blockSetIndex >= 0; blockSetIndex-- {
blockSet := blocksByVb[blockSetIndex]
vbNo := blockSet.vbNo
blocks := blockSet.blocks
fromVbSeq := fromSeq.GetSequence(vbNo) + 1
toVbSeq := toSeq.GetSequence(vbNo)
for blockIndex := len(blocks) - 1; blockIndex >= 0; blockIndex-- {
blockEntries, keys := blocks[blockIndex].GetEntries(vbNo, fromVbSeq, toVbSeq, true)
entries = append(entries, blockEntries...)
entryKeys = append(entryKeys, keys...)
}
}
// Bulk retrieval of individual entries. Performs deduplication, and reordering into ascending vb and sequence order
results := b.bulkLoadEntries(entryKeys, entries)
return results, nil
}
// calculateHash does a simple sum of the vector clock values, modulo s.mod. This intentionally
// doesn't generate a random distribution of clocks to hash values. Instead, since vbucket values within the
// clock are monotonically increasing, it's going to generate increasing, non-colliding hash values
// until s.mod value is reached. The goal is to tune s.mod to the update frequency (and so the expiry
// of previously used hash values). so that the previously used hash value (n) is obsolete by the time
// the clock wraps around to (mod + n).
func (s *sequenceHasher) calculateHash(clock base.SequenceClock) uint64 {
var sum uint64
sum = uint64(0)
for _, value := range clock.Value() {
sum += value & s.modMinus1
}
return sum & s.modMinus1
}
func (k *kvChannelIndex) pollForChanges(stableClock base.SequenceClock, newChannelClock base.SequenceClock) (hasChanges bool, cancelPolling bool) {
changeCacheExpvars.Add(fmt.Sprintf("pollCount-%s", k.channelName), 1)
// Increment the overall poll count since a changes request (regardless of whether there have been polled changes)
totalPollCount := atomic.AddUint32(&k.pollCount, 1)
unreadPollCount := atomic.LoadUint32(&k.unreadPollCount)
if unreadPollCount > kMaxUnreadPollCount {
// We've sent a notify, but had (kMaxUnreadPollCount) polls without anyone calling getChanges.
// Assume nobody is listening for updates - cancel polling for this channel
return false, true
}
if unreadPollCount > 0 {
// Give listeners more time to call getChanges, but increment
atomic.AddUint32(&k.unreadPollCount, 1)
}
k.lastPolledLock.Lock()
defer k.lastPolledLock.Unlock()
// First poll handling
if k.lastPolledChannelClock == nil {
k.lastPolledChannelClock = k.clock.Copy()
k.lastPolledSince = k.clock.Copy()
k.lastPolledValidTo = k.clock.Copy()
}
if !newChannelClock.AnyAfter(k.lastPolledChannelClock) {
// No changes to channel clock - update validTo based on the new stable sequence
k.lastPolledValidTo.SetTo(stableClock)
// If we've exceeded empty poll count, return hasChanges=true to trigger the "is
// anyone listening" check
if totalPollCount > kMaxEmptyPollCount {
return true, false
} else {
return false, false
}
}
// The clock has changed - load the changes and store in last polled
if err := k.updateLastPolled(stableClock, newChannelClock); err != nil {
base.Warn("Error updating last polled for channel %s: %v", k.channelName, err)
return false, false
}
// We have changes - increment unread counter if we haven't already
if unreadPollCount == 0 {
atomic.AddUint32(&k.unreadPollCount, 1)
}
return true, false
}
func (k *kvChannelIndex) getChannelClock() (base.SequenceClock, error) {
var channelClock base.SequenceClock
var err error
// If we're polling, return a copy
k.lastPolledLock.RLock()
defer k.lastPolledLock.RUnlock()
if k.lastPolledChannelClock != nil {
channelClock = base.NewSequenceClockImpl()
channelClock.SetTo(k.lastPolledChannelClock)
} else {
channelClock, err = k.loadChannelClock()
if err != nil {
return nil, err
}
}
return channelClock, nil
}
// calculateChangedPartitions identifies which vbuckets have changed after sinceClock until toClock, groups these
// by partition, and returns as a array of PartitionRanges, indexed by partition number.
func (ds *DenseStorageReader) calculateChanged(sinceClock, toClock base.SequenceClock) (changedVbs []uint16, changedPartitions []*PartitionRange) {
changedVbs = make([]uint16, 0)
changedPartitions = make([]*PartitionRange, ds.partitions.PartitionCount())
for vbNoInt, toSeq := range toClock.Value() {
vbNo := uint16(vbNoInt)
sinceSeq := sinceClock.GetSequence(vbNo)
if sinceSeq < toSeq {
changedVbs = append(changedVbs, vbNo)
partitionNo := ds.partitions.PartitionForVb(vbNo)
if changedPartitions[partitionNo] == nil {
partitionRange := NewPartitionRange()
changedPartitions[partitionNo] = &partitionRange
}
changedPartitions[partitionNo].SetRange(vbNo, sinceSeq, toSeq)
}
}
return changedVbs, changedPartitions
}
func (k *kvChannelIndex) checkLastPolled(since base.SequenceClock) (results []*LogEntry) {
k.lastPolledLock.RLock()
defer k.lastPolledLock.RUnlock()
if k.lastPolledValidTo == nil || k.lastPolledSince == nil {
return results
}
matchesLastPolledSince := true
lastPolledValue := k.lastPolledSince.Value()
validToValue := k.lastPolledValidTo.Value()
sinceValue := since.Value()
for vb, sequence := range sinceValue {
lastPolledVbValue := lastPolledValue[vb]
if sequence != lastPolledVbValue {
matchesLastPolledSince = false
if sequence < lastPolledVbValue || sequence > validToValue[vb] {
// poll results aren't sufficient for this request - return empty set
return results
}
}
}
// The last polled results are sufficient to serve this request. If there's a match on the since values,
// return the entire last polled changes. If not, filter the last polled changes and return all entries greater
// than the since value
if matchesLastPolledSince {
// TODO: come up with a solution that doesn't make as much GC work on every checkLastPolled hit,
// but doesn't break when k.lastPolledChanges gets updated mid-request.
results := make([]*LogEntry, len(k.lastPolledChanges))
copy(results, k.lastPolledChanges)
return results
} else {
for _, entry := range k.lastPolledChanges {
if entry.Sequence > sinceValue[entry.VbNo] {
results = append(results, entry)
}
}
}
return results
}
// Returns the set of index entries for the channel more recent than the
// specified since SequenceClock. Index entries with sequence values greater than
// the index stable sequence are not returned.
func (k *kvChannelIndex) getChanges(since base.SequenceClock) ([]*LogEntry, error) {
var results []*LogEntry
// Someone is still interested in this channel - reset poll counts
atomic.StoreUint32(&k.pollCount, 0)
atomic.StoreUint32(&k.unreadPollCount, 0)
chanClock, err := k.getChannelClock()
if err != nil {
// Note: gocb returns "Key not found.", go-couchbase returns "MCResponse status=KEY_ENOENT, opcode=GET, opaque=0, msg: Not found"
// Using string matching to identify key not found for now - really need a better API in go-couchbase/gocb for gets that allows us to distinguish
// between errors and key not found with something more robust than string matching.
if IsNotFoundError(err) {
// initialize chanClock as empty clock
chanClock = base.NewSequenceClockImpl()
} else {
return results, err
}
}
// If requested clock is later than the channel clock, return empty
if since.AllAfter(chanClock) {
base.LogTo("DIndex+", "requested clock is later than channel clock - no new changes to report")
return results, nil
}
// If the since value is more recent than the last polled clock, return the results from the
// last polling. Has the potential to return values earlier than since and later than
// lastPolledClock, but these duplicates will be ignored by replication. Could validate
// greater than since inside this if clause, but leaving out as a performance optimization for
// now
if lastPolledResults := k.checkLastPolled(since); len(lastPolledResults) > 0 {
indexExpvars.Add("getChanges_lastPolled_hit", 1)
return lastPolledResults, nil
}
indexExpvars.Add("getChanges_lastPolled_miss", 1)
return k.channelStorage.GetChanges(since, chanClock)
}
func (s *sequenceHasher) GetHash(clock base.SequenceClock) (string, error) {
if clock == nil {
return "", errors.New("Can't calculate hash for nil clock")
}
hashValue := s.calculateHash(clock)
// Load stored clocks for this hash, to see if it's already been defined.
// Note: getCacheValue and load are handled as separate operations to optimize locking.
// 1. getCacheValue locks the cache, and retrieves the current cache entry (or creates a new empty entry if not found)
// 2. cachedValue.load locks the entry, and loads from the DB if no previous entry is found
cachedValue := s.getCacheValue(hashValue)
cachedClocks, err := cachedValue.load(s.loadClocks)
if err != nil {
return "", err
}
// Check whether the cached clocks for the hash value match our clock
exists, index := cachedClocks.Contains(clock.Value())
if exists {
seqHash := sequenceHash{
hashValue: hashValue,
collisionIndex: uint16(index),
}
indexExpvars.Add("seqHash_getHash_hits", 1)
return seqHash.String(), nil
}
// Didn't find a match in cache - update the index and the cache. Get a write lock on the index value
// first, to ensure only one goroutine on this SG attempts to write. writeCas handling below handles
// the case where other SGs are updating the value concurrently
indexExpvars.Add("seqHash_getHash_misses", 1)
// First copy the clock value, to ensure we store a non-mutable version in the cache
clockValue := make([]uint64, len(clock.Value()))
copy(clockValue, clock.Value())
updateErr := func() error {
cachedValue.lock.Lock()
defer cachedValue.lock.Unlock()
// If the number of cached clocks has changed, check whether someone else has added this clock
// while we waited for the lock
if len(cachedValue.clocks.Sequences) > len(cachedClocks.Sequences) {
exists, index = cachedValue.clocks.Contains(clockValue)
if exists {
return nil
}
}
// Add our clock to the cached clocks for this hash
existingClocks := cachedValue.clocks
existingClocks.Sequences = append(existingClocks.Sequences, clockValue)
// Update the hash entry in the bucket
key := kHashPrefix + strconv.FormatUint(hashValue, 10)
initialValue, err := existingClocks.Marshal()
index = len(existingClocks.Sequences) - 1
if err != nil {
return err
}
_, err = base.WriteCasRaw(s.bucket, key, initialValue, existingClocks.cas, int(s.hashExpiry), func(value []byte) (updatedValue []byte, err error) {
// Note: The following is invoked upon cas failure - may be called multiple times
base.LogTo("DIndex+", "CAS fail - reapplying changes for hash storage for key: %s", key)
var sClocks storedClocks
err = sClocks.Unmarshal(value)
if err != nil {
base.Warn("Error unmarshalling hash storage during update", err)
return nil, err
}
exists, index = sClocks.Contains(clockValue)
if exists {
// return empty byte array to cancel the update
return []byte{}, nil
}
// Not found - add
sClocks.Sequences = append(sClocks.Sequences, clockValue)
base.LogTo("DIndex+", "Reattempting stored hash write for key %s:", key)
index = len(sClocks.Sequences) - 1
return sClocks.Marshal()
})
return nil
}()
if updateErr != nil {
return "", updateErr
}
indexExpvars.Add("writeCasRaw_hash", 1)
if err != nil && err.Error() != "Already Exists" {
return "", err
}
seqHash := &sequenceHash{
hashValue: hashValue,
collisionIndex: uint16(index),
}
return seqHash.String(), nil
}
func (clock PartitionClock) AddToClock(seqClock base.SequenceClock) {
for vbNo, seq := range clock {
seqClock.SetSequence(vbNo, seq)
}
}