Incremental Rollup and Tablet Size Calculation

go.mod

@@ -18,6 +18,7 @@ require (
 	github.com/codahale/hdrhistogram v0.0.0-20161010025455-3a0bb77429bd
 	github.com/dgraph-io/badger/v2 v2.0.1-rc1.0.20200316175624-91c31ebe8c22
 	github.com/dgraph-io/dgo/v2 v2.2.0
+	github.com/dgraph-io/ristretto v0.0.2-0.20200115201040-8f368f2f2ab3
 	github.com/dgrijalva/jwt-go v3.2.0+incompatible
 	github.com/dgryski/go-farm v0.0.0-20191112170834-c2139c5d712b
 	github.com/dgryski/go-groupvarint v0.0.0-20190318181831-5ce5df8ca4e1

posting/mvcc.go

@@ -21,23 +21,121 @@ import (
 	"encoding/hex"
 	"math"
 	"strconv"
+	"sync"
 	"sync/atomic"
+	"time"
 
 	"github.com/dgraph-io/badger/v2"
+	bpb "github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/y"
 	"github.com/dgraph-io/dgo/v2/protos/api"
 	"github.com/dgraph-io/dgraph/protos/pb"
 	"github.com/dgraph-io/dgraph/x"
+	"github.com/dgraph-io/ristretto/z"
+	"github.com/golang/glog"
 	"github.com/pkg/errors"
 )
 
+// incrRollupi is used to batch keys for rollup incrementally.
+type incrRollupi struct {
+	// keysCh is populated with batch of 64 keys that needs to be rolled up during reads
+	keysCh chan *[][]byte
+	// keysPool is sync.Pool to share the batched keys to rollup.
+	keysPool *sync.Pool
+	count    uint64
+}
+
 var (
 	// ErrTsTooOld is returned when a transaction is too old to be applied.
 	ErrTsTooOld = errors.Errorf("Transaction is too old")
 	// ErrInvalidKey is returned when trying to read a posting list using
 	// an invalid key (e.g the key to a single part of a larger multi-part list).
 	ErrInvalidKey = errors.Errorf("cannot read posting list using multi-part list key")
+
+	// IncrRollup is used to batch keys for rollup incrementally.
+	IncrRollup = &incrRollupi{
+		keysCh: make(chan *[][]byte),
+		keysPool: &sync.Pool{
+			New: func() interface{} {
+				return new([][]byte)
+			},
+		},
+	}
 )
 
+// rollUpKey takes the given key's posting lists, rolls it up and writes back to badger
+func (ir *incrRollupi) rollUpKey(writer *TxnWriter, key []byte) error {
+	l, err := GetNoStore(key)
+	if err != nil {
+		return err
+	}
+
+	kvs, err := l.Rollup()
+	if err != nil {
+		return err
+	}
+	const N = uint64(1000)
+	if glog.V(2) {
+		if count := atomic.AddUint64(&ir.count, 1); count%N == 0 {
+			glog.V(2).Infof("Rolled up %d keys", count)
+		}
+	}
+	return writer.Write(&bpb.KVList{Kv: kvs})
+}
+
+func (ir *incrRollupi) addKeyToBatch(key []byte) {
+	batch := ir.keysPool.Get().(*[][]byte)
+	*batch = append(*batch, key)
+	if len(*batch) < 64 {
+		ir.keysPool.Put(batch)
+		return
+	}
+
+	select {
+	case ir.keysCh <- batch:
+	default:
+		// Drop keys and build the batch again. Lossy behavior.
+		*batch = (*batch)[:0]
+		ir.keysPool.Put(batch)
+	}
+}
+
+// Process will rollup batches of 64 keys in a go routine.
+func (ir *incrRollupi) Process(closer *y.Closer) {
+	defer closer.Done()
+
+	writer := NewTxnWriter(pstore)
+	defer writer.Flush()
+
+	m := make(map[uint64]int64) // map hash(key) to ts. hash(key) to limit the size of the map.
+	limiter := time.NewTicker(100 * time.Millisecond)
+	for {
+		select {
+		case <-closer.HasBeenClosed():
+			return
+		case batch := <-ir.keysCh:
+			currTs := time.Now().Unix()
+			for _, key := range *batch {
+				hash := z.MemHash(key)
+				if elem := m[hash]; currTs-elem >= 10 {
+					// Key not present or Key present but last roll up was more than 10 sec ago.
+					// Add/Update map and rollup.
+					m[hash] = currTs
+					if err := ir.rollUpKey(writer, key); err != nil {
+						glog.Warningf("Error %v rolling up key %v\n", err, key)
+					}
+				}
+			}
+			// clear the batch and put it back in Sync keysPool
+			*batch = (*batch)[:0]
+			ir.keysPool.Put(batch)
+
+			// throttle to 1 batch = 64 rollups per 100 ms.
+			<-limiter.C
+		}
+	}
+}
+
 // ShouldAbort returns whether the transaction should be aborted.
 func (txn *Txn) ShouldAbort() bool {
 	if txn == nil {
@@ -166,6 +264,14 @@ func ReadPostingList(key []byte, it *badger.Iterator) (*List, error) {
 	l.key = key
 	l.plist = new(pb.PostingList)
 
+	// We use the following block of code to trigger incremental rollup on this key.
+	deltaCount := 0
+	defer func() {
+		if deltaCount > 0 {
+			IncrRollup.addKeyToBatch(key)
+		}
+	}()
+
 	// Iterates from highest Ts to lowest Ts
 	for it.Valid() {
 		item := it.Item()
@@ -210,6 +316,7 @@ func ReadPostingList(key []byte, it *badger.Iterator) (*List, error) {
 			if err != nil {
 				return nil, err
 			}
+			deltaCount++
 		case BitSchemaPosting:
 			return nil, errors.Errorf(
 				"Trying to read schema in ReadPostingList for key: %s", hex.Dump(key))