Esempio n. 1
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
}
Esempio n. 2
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 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
}
Esempio n. 3
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 proto.Timestamp, repl *Replica) error {
	snap := repl.rm.Engine().NewSnapshot()
	iter := newRangeDataIterator(repl.Desc(), snap)
	defer iter.Close()
	defer snap.Close()

	// Lookup the GC policy for the zone containing this key range.
	policy, err := gcq.lookupGCPolicy(repl)
	if err != nil {
		return err
	}

	gcMeta := proto.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()

	gcArgs := &proto.GCRequest{
		RequestHeader: proto.RequestHeader{
			Timestamp: now,
			RangeID:   repl.Desc().RangeID,
		},
	}
	var mu sync.Mutex
	var oldestIntentNanos int64 = math.MaxInt64
	var expBaseKey proto.Key
	var keys []proto.EncodedKey
	var vals [][]byte

	// Maps from txn ID to txn and intent key slice.
	txnMap := map[string]*proto.Transaction{}
	intentMap := map[string][]proto.Key{}

	// 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 := gogoproto.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], 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(proto.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, proto.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 := engine.MVCCDecodeKey(iter.Key())
		if !isValue {
			// Moving to the next key (& values).
			processKeysAndValues()
			expBaseKey = baseKey
			keys = []proto.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()

	// Set start and end keys.
	switch len(gcArgs.Keys) {
	case 0:
		return nil
	case 1:
		gcArgs.Key = gcArgs.Keys[0].Key
		gcArgs.EndKey = gcArgs.Key.Next()
	default:
		gcArgs.Key = gcArgs.Keys[0].Key
		gcArgs.EndKey = gcArgs.Keys[len(gcArgs.Keys)-1].Key
	}

	// 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.
	// TODO(spencer): use a batch here when available.
	for id, txn := range txnMap {
		if txn.Status != proto.PENDING {
			// The transaction was successfully pushed, so resolve the intents.
			for _, key := range intentMap[id] {
				resolveArgs := &proto.ResolveIntentRequest{
					RequestHeader: proto.RequestHeader{
						Timestamp: now,
						Key:       key,
						User:      security.RootUser,
						Txn:       txn,
					},
				}
				if _, err := repl.AddCmd(repl.context(), resolveArgs); err != nil {
					log.Warningf("resolve of key %q failed: %s", key, err)
					updateOldestIntent(txn.OrigTimestamp.WallTime)
				}
			}
		}
	}

	// Send GC request through range.
	gcMeta.OldestIntentNanos = gogoproto.Int64(oldestIntentNanos)
	gcArgs.GCMeta = *gcMeta
	if _, err := repl.AddCmd(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
}
Esempio n. 4
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.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
}
Esempio n. 5
0
// process iterates through all keys in a range, calling the garbage
// collector for each key and associated set of values. GC'd keys are
// batched into InternalGC calls. Extant intents are resolved if
// intents are older than intentAgeThreshold.
func (gcq *gcQueue) process(now proto.Timestamp, rng *Range) error {
	snap := rng.rm.Engine().NewSnapshot()
	iter := newRangeDataIterator(rng.Desc(), snap)
	defer iter.Close()
	defer snap.Close()

	// Lookup the GC policy for the zone containing this key range.
	policy, err := gcq.lookupGCPolicy(rng)
	if err != nil {
		return err
	}

	gcMeta := proto.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()

	gcArgs := &proto.InternalGCRequest{
		RequestHeader: proto.RequestHeader{
			Timestamp: now,
			RaftID:    rng.Desc().RaftID,
		},
	}
	var mu sync.Mutex
	var oldestIntentNanos int64 = math.MaxInt64
	var wg sync.WaitGroup
	var expBaseKey proto.Key
	var keys []proto.EncodedKey
	var vals [][]byte

	// 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 := gogoproto.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 {
					// Resolve intent asynchronously in a goroutine if the intent
					// is older than the intent expiration threshold.
					if meta.Timestamp.Less(intentExp) {
						wg.Add(1)
						go gcq.resolveIntent(rng, expBaseKey, meta, updateOldestIntent, &wg)
					} else {
						updateOldestIntent(meta.Timestamp.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(proto.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, proto.InternalGCRequest_GCKey{Key: expBaseKey, Timestamp: gcTS})
				}
			}
		}
	}

	// Iterate through this range's keys and values.
	for ; iter.Valid(); iter.Next() {
		baseKey, ts, isValue := engine.MVCCDecodeKey(iter.Key())
		if !isValue {
			// Moving to the next key (& values).
			processKeysAndValues()
			expBaseKey = baseKey
			keys = []proto.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()

	// Set start and end keys.
	switch len(gcArgs.Keys) {
	case 0:
		return nil
	case 1:
		gcArgs.Key = gcArgs.Keys[0].Key
		gcArgs.EndKey = gcArgs.Key.Next()
	default:
		gcArgs.Key = gcArgs.Keys[0].Key
		gcArgs.EndKey = gcArgs.Keys[len(gcArgs.Keys)-1].Key
	}

	// Wait for any outstanding intent resolves and set oldest extant intent.
	wg.Wait()
	gcMeta.OldestIntentNanos = gogoproto.Int64(oldestIntentNanos)

	// Send GC request through range.
	gcArgs.GCMeta = *gcMeta
	if err := rng.AddCmd(rng.context(), proto.Call{Args: gcArgs, Reply: &proto.InternalGCResponse{}}); err != nil {
		return err
	}

	// Store current timestamp as last verification for this range, as
	// we've just successfully scanned.
	if err := rng.SetLastVerificationTimestamp(now); err != nil {
		log.Errorf("failed to set last verification timestamp for range %s: %s", rng, err)
	}

	return nil
}