// process synchronously invokes admin split for each proposed split key.
func (sq *splitQueue) process(now roachpb.Timestamp, rng *Replica,
sysCfg *config.SystemConfig) error {
// First handle case of splitting due to zone config maps.
desc := rng.Desc()
splitKeys := sysCfg.ComputeSplitKeys(desc.StartKey, desc.EndKey)
if len(splitKeys) > 0 {
log.Infof("splitting %s at keys %v", rng, splitKeys)
for _, splitKey := range splitKeys {
if err := sq.db.AdminSplit(splitKey.AsRawKey()); err != nil {
return util.Errorf("unable to split %s at key %q: %s", rng, splitKey, err)
}
}
return nil
}
// Next handle case of splitting due to size.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return err
}
// FIXME: why is this implementation not the same as the one above?
if float64(rng.stats.GetSize())/float64(zone.RangeMaxBytes) > 1 {
log.Infof("splitting %s size=%d max=%d", rng, rng.stats.GetSize(), zone.RangeMaxBytes)
if _, pErr := client.SendWrapped(rng, rng.context(), &roachpb.AdminSplitRequest{
Span: roachpb.Span{Key: desc.StartKey.AsRawKey()},
}); pErr != nil {
return pErr.GoError()
}
}
return nil
}
// shouldQueue determines whether a replica should be queued for garbage
// collection, and if so, at what priority. Returns true for shouldQ
// in the event that the cumulative ages of GC'able bytes or extant
// intents exceed thresholds.
func (gcq *gcQueue) shouldQueue(now roachpb.Timestamp, repl *Replica,
sysCfg *config.SystemConfig) (shouldQ bool, priority float64) {
desc := repl.Desc()
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
log.Errorf("could not find GC policy for range %s: %s", repl, err)
return
}
policy := zone.GC
// GC score is the total GC'able bytes age normalized by 1 MB * the replica's TTL in seconds.
gcScore := float64(repl.stats.GetGCBytesAge(now.WallTime)) / float64(policy.TTLSeconds) / float64(gcByteCountNormalization)
// Intent score. This computes the average age of outstanding intents
// and normalizes.
intentScore := repl.stats.GetAvgIntentAge(now.WallTime) / float64(intentAgeNormalization.Nanoseconds()/1E9)
// Compute priority.
if gcScore > 1 {
priority += gcScore
}
if intentScore > 1 {
priority += intentScore
}
shouldQ = priority > 0
return
}
func (rq *replicateQueue) shouldQueue(now roachpb.Timestamp, repl *Replica,
sysCfg config.SystemConfig) (shouldQ bool, priority float64) {
if repl.needsSplitBySize() {
// If the range exceeds the split threshold, let that finish
// first. Ranges must fit in memory on both sender and receiver
// nodes while being replicated. This supplements the check
// provided by acceptsUnsplitRanges, which looks at zone config
// boundaries rather than data size.
return
}
// Find the zone config for this range.
desc := repl.Desc()
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
log.Error(err)
return
}
action, priority := rq.allocator.ComputeAction(*zone, desc)
if action != AllocatorNoop {
return true, priority
}
// See if there is a rebalancing opportunity present.
shouldRebalance := rq.allocator.ShouldRebalance(repl.store.StoreID())
return shouldRebalance, 0
}
// shouldQueue determines whether a replica should be queued for garbage
// collection, and if so, at what priority. Returns true for shouldQ
// in the event that the cumulative ages of GC'able bytes or extant
// intents exceed thresholds.
func (*gcQueue) shouldQueue(now hlc.Timestamp, repl *Replica,
sysCfg config.SystemConfig) (shouldQ bool, priority float64) {
desc := repl.Desc()
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
log.Errorf(context.TODO(), "could not find zone config for range %s: %s", repl, err)
return
}
ms := repl.GetMVCCStats()
// GC score is the total GC'able bytes age normalized by 1 MB * the replica's TTL in seconds.
gcScore := float64(ms.GCByteAge(now.WallTime)) / float64(zone.GC.TTLSeconds) / float64(gcByteCountNormalization)
// Intent score. This computes the average age of outstanding intents
// and normalizes.
intentScore := ms.AvgIntentAge(now.WallTime) / float64(intentAgeNormalization.Nanoseconds()/1E9)
// Compute priority.
if gcScore >= considerThreshold {
priority += gcScore
}
if intentScore >= considerThreshold {
priority += intentScore
}
shouldQ = priority > 0
return
}
func (rq *replicateQueue) shouldQueue(
now hlc.Timestamp,
repl *Replica,
sysCfg config.SystemConfig,
) (shouldQ bool, priority float64) {
if !repl.store.splitQueue.Disabled() && repl.needsSplitBySize() {
// If the range exceeds the split threshold, let that finish first.
// Ranges must fit in memory on both sender and receiver nodes while
// being replicated. This supplements the check provided by
// acceptsUnsplitRanges, which looks at zone config boundaries rather
// than data size.
//
// This check is ignored if the split queue is disabled, since in that
// case, the split will never come.
return
}
// Find the zone config for this range.
desc := repl.Desc()
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
log.Error(err)
return
}
action, priority := rq.allocator.ComputeAction(*zone, desc)
if action != AllocatorNoop {
return true, priority
}
// See if there is a rebalancing opportunity present.
shouldRebalance := rq.allocator.ShouldRebalance(repl.store.StoreID())
return shouldRebalance, 0
}
// process iterates through all keys in a replica's range, calling the garbage
// collector for each key and associated set of values. GC'd keys are batched
// into GC calls. Extant intents are resolved if intents are older than
// intentAgeThreshold. The transaction and abort cache records are also
// scanned and old entries evicted. During normal operation, both of these
// records are cleaned up when their respective transaction finishes, so the
// amount of work done here is expected to be small.
//
// Some care needs to be taken to avoid cyclic recreation of entries during GC:
// * a Push initiated due to an intent may recreate a transaction entry
// * resolving an intent may write a new abort cache entry
// * obtaining the transaction for a abort cache entry requires a Push
//
// The following order is taken below:
// 1) collect all intents with sufficiently old txn record
// 2) collect these intents' transactions
// 3) scan the transaction table, collecting abandoned or completed txns
// 4) push all of these transactions (possibly recreating entries)
// 5) resolve all intents (unless the txn is still PENDING), which will recreate
// abort cache entries (but with the txn timestamp; i.e. likely gc'able)
// 6) scan the abort cache table for old entries
// 7) push these transactions (again, recreating txn entries).
// 8) send a GCRequest.
func (gcq *gcQueue) process(
ctx context.Context,
now hlc.Timestamp,
repl *Replica,
sysCfg config.SystemConfig,
) error {
snap := repl.store.Engine().NewSnapshot()
desc := repl.Desc()
defer snap.Close()
// Lookup the GC policy for the zone containing this key range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return errors.Errorf("could not find zone config for range %s: %s", repl, err)
}
gcKeys, info, err := RunGC(ctx, desc, snap, now, zone.GC,
func(now hlc.Timestamp, txn *roachpb.Transaction, typ roachpb.PushTxnType) {
pushTxn(gcq.store.DB(), now, txn, typ)
},
func(intents []roachpb.Intent, poison bool, wait bool) error {
return repl.store.intentResolver.resolveIntents(ctx, intents, poison, wait)
})
if err != nil {
return err
}
gcq.eventLog.VInfof(true, "completed with stats %+v", info)
var ba roachpb.BatchRequest
var gcArgs roachpb.GCRequest
// TODO(tschottdorf): This is one of these instances in which we want
// to be more careful that the request ends up on the correct Replica,
// and we might have to worry about mixing range-local and global keys
// in a batch which might end up spanning Ranges by the time it executes.
gcArgs.Key = desc.StartKey.AsRawKey()
gcArgs.EndKey = desc.EndKey.AsRawKey()
gcArgs.Keys = gcKeys
gcArgs.Threshold = info.Threshold
// Technically not needed since we're talking directly to the Range.
ba.RangeID = desc.RangeID
ba.Timestamp = now
ba.Add(&gcArgs)
if _, pErr := repl.Send(ctx, ba); pErr != nil {
return pErr.GoError()
}
return nil
}
// process synchronously invokes admin split for each proposed split key.
func (sq *splitQueue) process(
ctx context.Context,
now hlc.Timestamp,
rng *Replica,
sysCfg config.SystemConfig,
) error {
// First handle case of splitting due to zone config maps.
desc := rng.Desc()
splitKeys := sysCfg.ComputeSplitKeys(desc.StartKey, desc.EndKey)
if len(splitKeys) > 0 {
log.Infof("splitting %s at keys %v", rng, splitKeys)
log.Trace(ctx, fmt.Sprintf("splitting at keys %v", splitKeys))
for _, splitKey := range splitKeys {
if err := sq.db.AdminSplit(splitKey.AsRawKey()); err != nil {
return errors.Errorf("unable to split %s at key %q: %s", rng, splitKey, err)
}
}
return nil
}
// Next handle case of splitting due to size.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return err
}
size := rng.GetMVCCStats().Total()
// FIXME: why is this implementation not the same as the one above?
if float64(size)/float64(zone.RangeMaxBytes) > 1 {
log.Infof("splitting %s size=%d max=%d", rng, size, zone.RangeMaxBytes)
log.Trace(ctx, fmt.Sprintf("splitting size=%d max=%d", size, zone.RangeMaxBytes))
if _, pErr := client.SendWrappedWith(rng, ctx, roachpb.Header{
Timestamp: now,
}, &roachpb.AdminSplitRequest{
Span: roachpb.Span{Key: desc.StartKey.AsRawKey()},
}); pErr != nil {
return pErr.GoError()
}
}
return nil
}
// shouldQueue determines whether a range should be queued for
// splitting. This is true if the range is intersected by a zone config
// prefix or if the range's size in bytes exceeds the limit for the zone.
func (*splitQueue) shouldQueue(now roachpb.Timestamp, rng *Replica,
sysCfg *config.SystemConfig) (shouldQ bool, priority float64) {
desc := rng.Desc()
if len(sysCfg.ComputeSplitKeys(desc.StartKey, desc.EndKey)) > 0 {
// Set priority to 1 in the event the range is split by zone configs.
priority = 1
shouldQ = true
}
// Add priority based on the size of range compared to the max
// size for the zone it's in.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
log.Error(err)
return
}
if ratio := float64(rng.stats.GetSize()) / float64(zone.RangeMaxBytes); ratio > 1 {
priority += ratio
shouldQ = true
}
return
}
func (rq *replicateQueue) shouldQueue(now roachpb.Timestamp, repl *Replica,
sysCfg config.SystemConfig) (shouldQ bool, priority float64) {
desc := repl.Desc()
if len(sysCfg.ComputeSplitKeys(desc.StartKey, desc.EndKey)) > 0 {
// If the replica's range needs splitting, wait until done.
return
}
// Find the zone config for this range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
log.Error(err)
return
}
action, priority := rq.allocator.ComputeAction(*zone, desc)
if action != AllocatorNoop {
return true, priority
}
// See if there is a rebalancing opportunity present.
shouldRebalance := rq.allocator.ShouldRebalance(repl.store.StoreID())
return shouldRebalance, 0
}
// process iterates through all keys in a replica's range, calling the garbage
// collector for each key and associated set of values. GC'd keys are batched
// into GC calls. Extant intents are resolved if intents are older than
// intentAgeThreshold.
func (gcq *gcQueue) process(now roachpb.Timestamp, repl *Replica,
sysCfg *config.SystemConfig) error {
snap := repl.rm.Engine().NewSnapshot()
desc := repl.Desc()
iter := newRangeDataIterator(desc, snap)
defer iter.Close()
defer snap.Close()
// Lookup the GC policy for the zone containing this key range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return fmt.Errorf("could not find GC policy for range %s: %s", repl, err)
}
policy := zone.GC
gcMeta := roachpb.NewGCMetadata(now.WallTime)
gc := engine.NewGarbageCollector(now, *policy)
// Compute intent expiration (intent age at which we attempt to resolve).
intentExp := now
intentExp.WallTime -= intentAgeThreshold.Nanoseconds()
// TODO(tschottdorf): execution will use a leader-assigned local
// timestamp to compute intent age. While this should be fine, could
// consider adding a Now timestamp to GCRequest which would be used
// instead.
gcArgs := &roachpb.GCRequest{
RequestHeader: roachpb.RequestHeader{
RangeID: desc.RangeID,
},
}
var mu sync.Mutex
var oldestIntentNanos int64 = math.MaxInt64
var expBaseKey roachpb.Key
var keys []roachpb.EncodedKey
var vals [][]byte
// Maps from txn ID to txn and intent key slice.
txnMap := map[string]*roachpb.Transaction{}
intentMap := map[string][]roachpb.Intent{}
// updateOldestIntent atomically updates the oldest intent.
updateOldestIntent := func(intentNanos int64) {
mu.Lock()
defer mu.Unlock()
if intentNanos < oldestIntentNanos {
oldestIntentNanos = intentNanos
}
}
// processKeysAndValues is invoked with each key and its set of
// values. Intents older than the intent age threshold are sent for
// resolution and values after the MVCC metadata, and possible
// intent, are sent for garbage collection.
processKeysAndValues := func() {
// If there's more than a single value for the key, possibly send for GC.
if len(keys) > 1 {
meta := &engine.MVCCMetadata{}
if err := proto.Unmarshal(vals[0], meta); err != nil {
log.Errorf("unable to unmarshal MVCC metadata for key %q: %s", keys[0], err)
} else {
// In the event that there's an active intent, send for
// intent resolution if older than the threshold.
startIdx := 1
if meta.Txn != nil {
// Keep track of intent to resolve if older than the intent
// expiration threshold.
if meta.Timestamp.Less(intentExp) {
id := string(meta.Txn.ID)
txnMap[id] = meta.Txn
intentMap[id] = append(intentMap[id], roachpb.Intent{Key: expBaseKey})
} else {
updateOldestIntent(meta.Txn.OrigTimestamp.WallTime)
}
// With an active intent, GC ignores MVCC metadata & intent value.
startIdx = 2
}
// See if any values may be GC'd.
if gcTS := gc.Filter(keys[startIdx:], vals[startIdx:]); !gcTS.Equal(roachpb.ZeroTimestamp) {
// TODO(spencer): need to split the requests up into
// multiple requests in the event that more than X keys
// are added to the request.
gcArgs.Keys = append(gcArgs.Keys, roachpb.GCRequest_GCKey{Key: expBaseKey, Timestamp: gcTS})
}
}
}
}
// Iterate through the keys and values of this replica's range.
for ; iter.Valid(); iter.Next() {
baseKey, ts, isValue, err := engine.MVCCDecodeKey(iter.Key())
if err != nil {
log.Errorf("unable to decode MVCC key: %q: %v", iter.Key(), err)
continue
}
if !isValue {
// Moving to the next key (& values).
processKeysAndValues()
expBaseKey = baseKey
keys = []roachpb.EncodedKey{iter.Key()}
vals = [][]byte{iter.Value()}
} else {
if !baseKey.Equal(expBaseKey) {
log.Errorf("unexpectedly found a value for %q with ts=%s; expected key %q", baseKey, ts, expBaseKey)
continue
}
keys = append(keys, iter.Key())
vals = append(vals, iter.Value())
}
}
if iter.Error() != nil {
return iter.Error()
}
// Handle last collected set of keys/vals.
processKeysAndValues()
// Process push transactions in parallel.
var wg sync.WaitGroup
for _, txn := range txnMap {
wg.Add(1)
go gcq.pushTxn(repl, now, txn, updateOldestIntent, &wg)
}
wg.Wait()
// Resolve all intents.
var intents []roachpb.Intent
for id, txn := range txnMap {
if txn.Status != roachpb.PENDING {
for _, intent := range intentMap[id] {
intent.Txn = *txn
intents = append(intents, intent)
}
}
}
done := true
if len(intents) > 0 {
done = false
repl.resolveIntents(repl.context(), intents)
}
// Set start and end keys.
if len(gcArgs.Keys) > 0 {
done = false
gcArgs.Key = gcArgs.Keys[0].Key
gcArgs.EndKey = gcArgs.Keys[len(gcArgs.Keys)-1].Key.Next()
}
if done {
return nil
}
// Send GC request through range.
gcMeta.OldestIntentNanos = proto.Int64(oldestIntentNanos)
gcArgs.GCMeta = *gcMeta
if _, err := client.SendWrapped(repl, repl.context(), gcArgs); err != nil {
return err
}
// Store current timestamp as last verification for this replica, as
// we've just successfully scanned.
if err := repl.SetLastVerificationTimestamp(now); err != nil {
log.Errorf("failed to set last verification timestamp for replica %s: %s", repl, err)
}
return nil
}
// process iterates through all keys in a replica's range, calling the garbage
// collector for each key and associated set of values. GC'd keys are batched
// into GC calls. Extant intents are resolved if intents are older than
// intentAgeThreshold. The transaction and sequence cache records are also
// scanned and old entries evicted. During normal operation, both of these
// records are cleaned up when their respective transaction finishes, so the
// amount of work done here is expected to be small.
//
// Some care needs to be taken to avoid cyclic recreation of entries during GC:
// * a Push initiated due to an intent may recreate a transaction entry
// * resolving an intent may write a new sequence cache entry
// * obtaining the transaction for a sequence cache entry requires a Push
//
// The following order is taken below:
// 1) collect all intents with sufficiently old txn record
// 2) collect these intents' transactions
// 3) scan the transaction table, collecting abandoned or completed txns
// 4) push all of these transactions (possibly recreating entries)
// 5) resolve all intents (unless the txn is still PENDING), which will recreate
// sequence cache entries (but with the txn timestamp; i.e. likely gc'able)
// 6) scan the sequence table for old entries
// 7) push these transactions (again, recreating txn entries).
// 8) send a GCRequest.
func (gcq *gcQueue) process(now roachpb.Timestamp, repl *Replica,
sysCfg config.SystemConfig) error {
snap := repl.store.Engine().NewSnapshot()
desc := repl.Desc()
iter := newReplicaDataIterator(desc, snap, true /* replicatedOnly */)
defer iter.Close()
defer snap.Close()
// Lookup the GC policy for the zone containing this key range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return util.Errorf("could not find zone config for range %s: %s", repl, err)
}
gc := engine.NewGarbageCollector(now, zone.GC)
// Compute intent expiration (intent age at which we attempt to resolve).
intentExp := now
intentExp.WallTime -= intentAgeThreshold.Nanoseconds()
txnExp := now
txnExp.WallTime -= txnCleanupThreshold.Nanoseconds()
gcArgs := &roachpb.GCRequest{}
// TODO(tschottdorf): This is one of these instances in which we want
// to be more careful that the request ends up on the correct Replica,
// and we might have to worry about mixing range-local and global keys
// in a batch which might end up spanning Ranges by the time it executes.
gcArgs.Key = desc.StartKey.AsRawKey()
gcArgs.EndKey = desc.EndKey.AsRawKey()
var expBaseKey roachpb.Key
var keys []engine.MVCCKey
var vals [][]byte
// Maps from txn ID to txn and intent key slice.
txnMap := map[uuid.UUID]*roachpb.Transaction{}
intentSpanMap := map[uuid.UUID][]roachpb.Span{}
// processKeysAndValues is invoked with each key and its set of
// values. Intents older than the intent age threshold are sent for
// resolution and values after the MVCC metadata, and possible
// intent, are sent for garbage collection.
var intentCount int
processKeysAndValues := func() {
// If there's more than a single value for the key, possibly send for GC.
if len(keys) > 1 {
meta := &engine.MVCCMetadata{}
if err := proto.Unmarshal(vals[0], meta); err != nil {
log.Errorf("unable to unmarshal MVCC metadata for key %q: %s", keys[0], err)
} else {
// In the event that there's an active intent, send for
// intent resolution if older than the threshold.
startIdx := 1
if meta.Txn != nil {
// Keep track of intent to resolve if older than the intent
// expiration threshold.
if meta.Timestamp.Less(intentExp) {
txnID := *meta.Txn.ID
txn := &roachpb.Transaction{
TxnMeta: *meta.Txn,
}
txnMap[txnID] = txn
intentCount++
intentSpanMap[txnID] = append(intentSpanMap[txnID], roachpb.Span{Key: expBaseKey})
}
// With an active intent, GC ignores MVCC metadata & intent value.
startIdx = 2
}
// See if any values may be GC'd.
if gcTS := gc.Filter(keys[startIdx:], vals[startIdx:]); !gcTS.Equal(roachpb.ZeroTimestamp) {
// TODO(spencer): need to split the requests up into
// multiple requests in the event that more than X keys
// are added to the request.
gcArgs.Keys = append(gcArgs.Keys, roachpb.GCRequest_GCKey{Key: expBaseKey, Timestamp: gcTS})
}
}
}
}
// Iterate through the keys and values of this replica's range.
for ; iter.Valid(); iter.Next() {
iterKey := iter.Key()
if !iterKey.IsValue() || !iterKey.Key.Equal(expBaseKey) {
// Moving to the next key (& values).
processKeysAndValues()
expBaseKey = iterKey.Key
if !iterKey.IsValue() {
keys = []engine.MVCCKey{iter.Key()}
vals = [][]byte{iter.Value()}
continue
}
// An implicit metadata.
keys = []engine.MVCCKey{engine.MakeMVCCMetadataKey(iterKey.Key)}
// A nil value for the encoded MVCCMetadata. This will unmarshal to an
// empty MVCCMetadata which is sufficient for processKeysAndValues to
// determine that there is no intent.
vals = [][]byte{nil}
}
keys = append(keys, iter.Key())
vals = append(vals, iter.Value())
}
if iter.Error() != nil {
return iter.Error()
}
// Handle last collected set of keys/vals.
processKeysAndValues()
gcq.eventLog.Infof(true, "assembled %d transactions from %d old intents; found %d gc'able keys", len(txnMap), intentCount, len(gcArgs.Keys))
txnKeys, err := gcq.processTransactionTable(repl, txnMap, txnExp)
if err != nil {
return err
}
// From now on, all newly added keys are range-local.
// TODO(tschottdorf): Might need to use two requests at some point since we
// hard-coded the full non-local key range in the header, but that does
// not take into account the range-local keys. It will be OK as long as
// we send directly to the Replica, though.
gcArgs.Keys = append(gcArgs.Keys, txnKeys...)
// Process push transactions in parallel.
var wg sync.WaitGroup
gcq.eventLog.Infof(true, "pushing %d txns", len(txnMap))
for _, txn := range txnMap {
if txn.Status != roachpb.PENDING {
continue
}
wg.Add(1)
go gcq.pushTxn(repl, now, txn, roachpb.PUSH_ABORT, &wg)
}
wg.Wait()
// Resolve all intents.
var intents []roachpb.Intent
for txnID, txn := range txnMap {
if txn.Status != roachpb.PENDING {
for _, intent := range intentSpanMap[txnID] {
intents = append(intents, roachpb.Intent{Span: intent, Status: txn.Status, Txn: txn.TxnMeta})
}
}
}
gcq.eventLog.Infof(true, "resolving %d intents", len(intents))
if pErr := repl.store.intentResolver.resolveIntents(repl.context(), repl, intents,
true /* wait */, false /* !poison */); pErr != nil {
return pErr.GoError()
}
// Deal with any leftover sequence cache keys. There shouldn't be many of
// them.
leftoverSeqCacheKeys := gcq.processSequenceCache(repl, now, txnExp, txnMap)
gcq.eventLog.Infof(true, "collected %d leftover sequence cache keys", len(leftoverSeqCacheKeys))
gcArgs.Keys = append(gcArgs.Keys, leftoverSeqCacheKeys...)
gcq.eventLog.Infof(true, "sending gc request for %d keys", len(gcArgs.Keys))
var ba roachpb.BatchRequest
// Technically not needed since we're talking directly to the Range.
ba.RangeID = desc.RangeID
ba.Timestamp = now
ba.Add(gcArgs)
if _, pErr := repl.Send(repl.context(), ba); pErr != nil {
return pErr.GoError()
}
return nil
}
// process iterates through all keys in a replica's range, calling the garbage
// collector for each key and associated set of values. GC'd keys are batched
// into GC calls. Extant intents are resolved if intents are older than
// intentAgeThreshold.
func (gcq *gcQueue) process(now roachpb.Timestamp, repl *Replica,
sysCfg *config.SystemConfig) error {
snap := repl.store.Engine().NewSnapshot()
desc := repl.Desc()
iter := newReplicaDataIterator(desc, snap)
defer iter.Close()
defer snap.Close()
// Lookup the GC policy for the zone containing this key range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return fmt.Errorf("could not find GC policy for range %s: %s", repl, err)
}
policy := zone.GC
gcMeta := roachpb.NewGCMetadata(now.WallTime)
gc := engine.NewGarbageCollector(now, *policy)
// Compute intent expiration (intent age at which we attempt to resolve).
intentExp := now
intentExp.WallTime -= intentAgeThreshold.Nanoseconds()
txnExp := now
txnExp.WallTime -= txnCleanupThreshold.Nanoseconds()
gcArgs := &roachpb.GCRequest{}
// TODO(tschottdorf): This is one of these instances in which we want
// to be more careful that the request ends up on the correct Replica,
// and we might have to worry about mixing range-local and global keys
// in a batch which might end up spanning Ranges by the time it executes.
gcArgs.Key = desc.StartKey.AsRawKey()
gcArgs.EndKey = desc.EndKey.AsRawKey()
var expBaseKey roachpb.Key
var keys []engine.MVCCKey
var vals [][]byte
// Maps from txn ID to txn and intent key slice.
txnMap := map[string]*roachpb.Transaction{}
intentSpanMap := map[string][]roachpb.Span{}
// processKeysAndValues is invoked with each key and its set of
// values. Intents older than the intent age threshold are sent for
// resolution and values after the MVCC metadata, and possible
// intent, are sent for garbage collection.
processKeysAndValues := func() {
// If there's more than a single value for the key, possibly send for GC.
if len(keys) > 1 {
meta := &engine.MVCCMetadata{}
if err := proto.Unmarshal(vals[0], meta); err != nil {
log.Errorf("unable to unmarshal MVCC metadata for key %q: %s", keys[0], err)
} else {
// In the event that there's an active intent, send for
// intent resolution if older than the threshold.
startIdx := 1
if meta.Txn != nil {
// Keep track of intent to resolve if older than the intent
// expiration threshold.
if meta.Timestamp.Less(intentExp) {
id := string(meta.Txn.ID)
txnMap[id] = meta.Txn
intentSpanMap[id] = append(intentSpanMap[id], roachpb.Span{Key: expBaseKey})
}
// With an active intent, GC ignores MVCC metadata & intent value.
startIdx = 2
}
// See if any values may be GC'd.
if gcTS := gc.Filter(keys[startIdx:], vals[startIdx:]); !gcTS.Equal(roachpb.ZeroTimestamp) {
// TODO(spencer): need to split the requests up into
// multiple requests in the event that more than X keys
// are added to the request.
gcArgs.Keys = append(gcArgs.Keys, roachpb.GCRequest_GCKey{Key: expBaseKey, Timestamp: gcTS})
}
}
}
}
// Iterate through the keys and values of this replica's range.
for ; iter.Valid(); iter.Next() {
baseKey, ts, isValue, err := engine.MVCCDecodeKey(iter.Key())
if err != nil {
log.Errorf("unable to decode MVCC key: %q: %v", iter.Key(), err)
continue
}
if !isValue {
// Moving to the next key (& values).
processKeysAndValues()
expBaseKey = baseKey
keys = []engine.MVCCKey{iter.Key()}
vals = [][]byte{iter.Value()}
} else {
if !baseKey.Equal(expBaseKey) {
log.Errorf("unexpectedly found a value for %q with ts=%s; expected key %q", baseKey, ts, expBaseKey)
continue
}
keys = append(keys, iter.Key())
vals = append(vals, iter.Value())
}
}
if iter.Error() != nil {
return iter.Error()
}
// Handle last collected set of keys/vals.
processKeysAndValues()
txnKeys, err := processTransactionTable(repl, txnMap, txnExp)
if err != nil {
return err
}
// From now on, all newly added keys are range-local.
// TODO(tschottdorf): Might need to use two requests at some point since we
// hard-coded the full non-local key range in the header, but that does
// not take into account the range-local keys. It will be OK as long as
// we send directly to the Replica, though.
gcArgs.Keys = append(gcArgs.Keys, txnKeys...)
// Process push transactions in parallel.
var wg sync.WaitGroup
for _, txn := range txnMap {
if txn.Status != roachpb.PENDING {
continue
}
wg.Add(1)
go pushTxn(repl, now, txn, roachpb.ABORT_TXN, &wg)
}
wg.Wait()
// Resolve all intents.
var intents []roachpb.Intent
for id, txn := range txnMap {
if txn.Status != roachpb.PENDING {
for _, intent := range intentSpanMap[id] {
intents = append(intents, roachpb.Intent{Span: intent, Txn: *txn})
}
}
}
if err := repl.resolveIntents(repl.context(), intents, true /* wait */, false /* !poison */); err != nil {
return err
}
// Deal with any leftover sequence cache keys. There shouldn't be many of
// them.
gcArgs.Keys = append(gcArgs.Keys, processSequenceCache(repl, now, txnExp, txnMap)...)
// Send GC request through range.
gcArgs.GCMeta = *gcMeta
var ba roachpb.BatchRequest
// Technically not needed since we're talking directly to the Range.
ba.RangeID = desc.RangeID
ba.Timestamp = now
ba.Add(gcArgs)
if _, pErr := repl.Send(repl.context(), ba); pErr != nil {
return pErr.GoError()
}
// Store current timestamp as last verification for this replica, as
// we've just successfully scanned.
if err := repl.SetLastVerificationTimestamp(now); err != nil {
log.Errorf("failed to set last verification timestamp for replica %s: %s", repl, err)
}
return nil
}
func (rq *replicateQueue) process(now roachpb.Timestamp, repl *Replica, sysCfg config.SystemConfig) error {
desc := repl.Desc()
// Find the zone config for this range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return err
}
action, _ := rq.allocator.ComputeAction(*zone, desc)
// Avoid taking action if the range has too many dead replicas to make
// quorum.
deadReplicas := rq.allocator.storePool.deadReplicas(desc.Replicas)
quorum := computeQuorum(len(desc.Replicas))
liveReplicaCount := len(desc.Replicas) - len(deadReplicas)
if liveReplicaCount < quorum {
return util.Errorf("range requires a replication change, but lacks a quorum of live nodes.")
}
switch action {
case AllocatorAdd:
newStore, err := rq.allocator.AllocateTarget(zone.ReplicaAttrs[0], desc.Replicas, true, nil)
if err != nil {
return err
}
newReplica := roachpb.ReplicaDescriptor{
NodeID: newStore.Node.NodeID,
StoreID: newStore.StoreID,
}
if err = repl.ChangeReplicas(roachpb.ADD_REPLICA, newReplica, desc); err != nil {
return err
}
case AllocatorRemove:
removeReplica, err := rq.allocator.RemoveTarget(desc.Replicas)
if err != nil {
return err
}
if err = repl.ChangeReplicas(roachpb.REMOVE_REPLICA, removeReplica, desc); err != nil {
return err
}
// Do not requeue if we removed ourselves.
if removeReplica.StoreID == repl.store.StoreID() {
return nil
}
case AllocatorRemoveDead:
if len(deadReplicas) == 0 {
if log.V(1) {
log.Warningf("Range of replica %s was identified as having dead replicas, but no dead replicas were found.", repl)
}
break
}
if err = repl.ChangeReplicas(roachpb.REMOVE_REPLICA, deadReplicas[0], desc); err != nil {
return err
}
case AllocatorNoop:
// The Noop case will result if this replica was queued in order to
// rebalance. Attempt to find a rebalancing target.
rebalanceStore := rq.allocator.RebalanceTarget(repl.store.StoreID(), zone.ReplicaAttrs[0], desc.Replicas)
if rebalanceStore == nil {
// No action was necessary and no rebalance target was found. Return
// without re-queuing this replica.
return nil
}
rebalanceReplica := roachpb.ReplicaDescriptor{
NodeID: rebalanceStore.Node.NodeID,
StoreID: rebalanceStore.StoreID,
}
if err = repl.ChangeReplicas(roachpb.ADD_REPLICA, rebalanceReplica, desc); err != nil {
return err
}
}
// Enqueue this replica again to see if there are more changes to be made.
rq.MaybeAdd(repl, rq.clock.Now())
return nil
}
func (rq *replicateQueue) process(
ctx context.Context,
now hlc.Timestamp,
repl *Replica,
sysCfg config.SystemConfig,
) error {
desc := repl.Desc()
// Find the zone config for this range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return err
}
action, _ := rq.allocator.ComputeAction(zone, desc)
// Avoid taking action if the range has too many dead replicas to make
// quorum.
deadReplicas := rq.allocator.storePool.deadReplicas(repl.RangeID, desc.Replicas)
quorum := computeQuorum(len(desc.Replicas))
liveReplicaCount := len(desc.Replicas) - len(deadReplicas)
if liveReplicaCount < quorum {
return errors.Errorf("range requires a replication change, but lacks a quorum of live nodes.")
}
switch action {
case AllocatorAdd:
log.Trace(ctx, "adding a new replica")
newStore, err := rq.allocator.AllocateTarget(zone.ReplicaAttrs[0], desc.Replicas, true)
if err != nil {
return err
}
newReplica := roachpb.ReplicaDescriptor{
NodeID: newStore.Node.NodeID,
StoreID: newStore.StoreID,
}
log.VTracef(1, ctx, "%s: adding replica to %+v due to under-replication", repl, newReplica)
if err = repl.ChangeReplicas(ctx, roachpb.ADD_REPLICA, newReplica, desc); err != nil {
return err
}
case AllocatorRemove:
log.Trace(ctx, "removing a replica")
// We require the lease in order to process replicas, so
// repl.store.StoreID() corresponds to the lease-holder's store ID.
removeReplica, err := rq.allocator.RemoveTarget(desc.Replicas, repl.store.StoreID())
if err != nil {
return err
}
log.VTracef(1, ctx, "%s: removing replica %+v due to over-replication", repl, removeReplica)
if err = repl.ChangeReplicas(ctx, roachpb.REMOVE_REPLICA, removeReplica, desc); err != nil {
return err
}
// Do not requeue if we removed ourselves.
if removeReplica.StoreID == repl.store.StoreID() {
return nil
}
case AllocatorRemoveDead:
log.Trace(ctx, "removing a dead replica")
if len(deadReplicas) == 0 {
if log.V(1) {
log.Warningf(ctx, "Range of replica %s was identified as having dead replicas, but no dead replicas were found.", repl)
}
break
}
deadReplica := deadReplicas[0]
log.VTracef(1, ctx, "%s: removing dead replica %+v from store", repl, deadReplica)
if err = repl.ChangeReplicas(ctx, roachpb.REMOVE_REPLICA, deadReplica, desc); err != nil {
return err
}
case AllocatorNoop:
log.Trace(ctx, "considering a rebalance")
// The Noop case will result if this replica was queued in order to
// rebalance. Attempt to find a rebalancing target.
//
// We require the lease in order to process replicas, so
// repl.store.StoreID() corresponds to the lease-holder's store ID.
rebalanceStore := rq.allocator.RebalanceTarget(
zone.ReplicaAttrs[0], desc.Replicas, repl.store.StoreID())
if rebalanceStore == nil {
log.VTracef(1, ctx, "%s: no suitable rebalance target", repl)
// No action was necessary and no rebalance target was found. Return
// without re-queuing this replica.
return nil
}
rebalanceReplica := roachpb.ReplicaDescriptor{
NodeID: rebalanceStore.Node.NodeID,
StoreID: rebalanceStore.StoreID,
}
log.VTracef(1, ctx, "%s: rebalancing to %+v", repl, rebalanceReplica)
if err = repl.ChangeReplicas(ctx, roachpb.ADD_REPLICA, rebalanceReplica, desc); err != nil {
return err
}
}
// Enqueue this replica again to see if there are more changes to be made.
rq.MaybeAdd(repl, rq.clock.Now())
return nil
}