zapdb/stream.go

/*
 * Copyright 2018 Dgraph Labs, Inc. and Contributors
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

package badger

import (
	"bytes"
	"context"
	"sort"
	"sync"
	"sync/atomic"
	"time"

	"github.com/dgraph-io/badger/v3/pb"
	"github.com/dgraph-io/badger/v3/table"
	"github.com/dgraph-io/badger/v3/y"
	"github.com/dgraph-io/ristretto/z"
	humanize "github.com/dustin/go-humanize"
	"github.com/pkg/errors"
)

const batchSize = 16 << 20 // 16 MB

// maxStreamSize is the maximum allowed size of a stream batch. This is a soft limit
// as a single list that is still over the limit will have to be sent as is since it
// cannot be split further. This limit prevents the framework from creating batches
// so big that sending them causes issues (e.g running into the max size gRPC limit).
var maxStreamSize = uint64(100 << 20) // 100MB

// Stream provides a framework to concurrently iterate over a snapshot of Badger, pick up
// key-values, batch them up and call Send. Stream does concurrent iteration over many smaller key
// ranges. It does NOT send keys in lexicographical sorted order. To get keys in sorted
// order, use Iterator.
type Stream struct {
	// Prefix to only iterate over certain range of keys. If set to nil (default), Stream would
	// iterate over the entire DB.
	Prefix []byte

	// Number of goroutines to use for iterating over key ranges. Defaults to 8.
	NumGo int

	// Badger would produce log entries in Infof to indicate the progress of Stream. LogPrefix can
	// be used to help differentiate them from other activities. Default is "Badger.Stream".
	LogPrefix string

	// ChooseKey is invoked each time a new key is encountered. Note that this is not called
	// on every version of the value, only the first encountered version (i.e. the highest version
	// of the value a key has). ChooseKey can be left nil to select all keys.
	//
	// Note: Calls to ChooseKey are concurrent.
	ChooseKey func(item *Item) bool

	// KeyToList, similar to ChooseKey, is only invoked on the highest version of the value. It
	// is upto the caller to iterate over the versions and generate zero, one or more KVs. It
	// is expected that the user would advance the iterator to go through the versions of the
	// values. However, the user MUST immediately return from this function on the first encounter
	// with a mismatching key. See example usage in ToList function. Can be left nil to use ToList
	// function by default.
	//
	// KeyToList has access to z.Allocator accessible via stream.Allocator(itr.ThreadId). This
	// allocator can be used to allocate KVs, to decrease the memory pressure on Go GC. Stream
	// framework takes care of releasing those resources after calling Send. AllocRef does
	// NOT need to be set in the returned KVList, as Stream framework would ignore that field,
	// instead using the allocator assigned to that thread id.
	//
	// Note: Calls to KeyToList are concurrent.
	KeyToList func(key []byte, itr *Iterator) (*pb.KVList, error)

	// This is the method where Stream sends the final output. All calls to Send are done by a
	// single goroutine, i.e. logic within Send method can expect single threaded execution.
	Send func(buf *z.Buffer) error

	// Read data above the sinceTs. All keys with version =< sinceTs will be ignored.
	SinceTs uint64
	// FullCopy should be set to true only when encryption mode is same for sender and receiver.
	FullCopy     bool
	readTs       uint64
	db           *DB
	rangeCh      chan keyRange
	kvChan       chan *z.Buffer
	nextStreamId uint32
	doneMarkers  bool
	scanned      uint64 // used to estimate the ETA for data scan.
	numProducers int32
}

// SendDoneMarkers when true would send out done markers on the stream. False by default.
func (st *Stream) SendDoneMarkers(done bool) {
	st.doneMarkers = done
}

// ToList is a default implementation of KeyToList. It picks up all valid versions of the key,
// skipping over deleted or expired keys.
func (st *Stream) ToList(key []byte, itr *Iterator) (*pb.KVList, error) {
	a := itr.Alloc
	ka := a.Copy(key)

	list := &pb.KVList{}
	for ; itr.Valid(); itr.Next() {
		item := itr.Item()
		if !bytes.Equal(key, item.Key()) {
			// Break out on the first encounter with another key.
			break
		}

		kv := y.NewKV(a)
		kv.Key = ka

		if err := item.Value(func(val []byte) error {
			kv.Value = a.Copy(val)
			return nil

		}); err != nil {
			return nil, err
		}
		kv.Version = item.Version()
		kv.ExpiresAt = item.ExpiresAt()
		// As we do full copy, we need to transmit only if it is a delete key or not.
		kv.Meta = []byte{item.meta & bitDelete}
		kv.UserMeta = a.Copy([]byte{item.UserMeta()})

		list.Kv = append(list.Kv, kv)
		if st.db.opt.NumVersionsToKeep == 1 {
			break
		}

		if item.DiscardEarlierVersions() {
			break
		}
		if item.IsDeletedOrExpired() {
			// We do a FullCopy in stream. It might happen that tables from L6 contain K(version=1),
			// while the table at L4 that was not copied contains K(version=2) with delete mark.
			// Hence, we need to send the deleted or expired item too.
			break
		}
	}
	return list, nil
}

// keyRange is [start, end), including start, excluding end. Do ensure that the start,
// end byte slices are owned by keyRange struct.
func (st *Stream) produceRanges(ctx context.Context) {
	ranges := st.db.Ranges(st.Prefix, 16)
	y.AssertTrue(len(ranges) > 0)
	y.AssertTrue(ranges[0].left == nil)
	y.AssertTrue(ranges[len(ranges)-1].right == nil)
	st.db.opt.Infof("Number of ranges found: %d\n", len(ranges))

	// Sort in descending order of size.
	sort.Slice(ranges, func(i, j int) bool {
		return ranges[i].size > ranges[j].size
	})
	for i, r := range ranges {
		st.rangeCh <- *r
		st.db.opt.Infof("Sent range %d for iteration: [%x, %x) of size: %s\n",
			i, r.left, r.right, humanize.IBytes(uint64(r.size)))
	}
	close(st.rangeCh)
}

// produceKVs picks up ranges from rangeCh, generates KV lists and sends them to kvChan.
func (st *Stream) produceKVs(ctx context.Context, itr *Iterator) error {
	atomic.AddInt32(&st.numProducers, 1)
	defer atomic.AddInt32(&st.numProducers, -1)

	// produceKVs is running iterate serially. So, we can define the outList here.
	outList := z.NewBuffer(2*batchSize, "Stream.ProduceKVs")
	defer func() {
		// The outList variable changes. So, we need to evaluate the variable in the defer. DO NOT
		// call `defer outList.Release()`.
		outList.Release()
	}()

	iterate := func(kr keyRange) error {
		itr.Alloc = z.NewAllocator(1<<20, "Stream.Iterate")
		defer itr.Alloc.Release()

		// This unique stream id is used to identify all the keys from this iteration.
		streamId := atomic.AddUint32(&st.nextStreamId, 1)
		var scanned int

		sendIt := func() error {
			select {
			case st.kvChan <- outList:
				outList = z.NewBuffer(2*batchSize, "Stream.ProduceKVs")
				atomic.AddUint64(&st.scanned, uint64(itr.scanned-scanned))
				scanned = itr.scanned
			case <-ctx.Done():
				return ctx.Err()
			}
			return nil
		}

		var prevKey []byte
		for itr.Seek(kr.left); itr.Valid(); {
			// it.Valid would only return true for keys with the provided Prefix in iterOpts.
			item := itr.Item()
			if bytes.Equal(item.Key(), prevKey) {
				itr.Next()
				continue
			}
			prevKey = append(prevKey[:0], item.Key()...)

			// Check if we reached the end of the key range.
			if len(kr.right) > 0 && bytes.Compare(item.Key(), kr.right) >= 0 {
				break
			}

			// Check if we should pick this key.
			if st.ChooseKey != nil && !st.ChooseKey(item) {
				continue
			}

			// Now convert to key value.
			itr.Alloc.Reset()
			list, err := st.KeyToList(item.KeyCopy(nil), itr)
			if err != nil {
				st.db.opt.Warningf("While reading key: %x, got error: %v", item.Key(), err)
				continue
			}
			if list == nil || len(list.Kv) == 0 {
				continue
			}
			for _, kv := range list.Kv {
				kv.StreamId = streamId
				KVToBuffer(kv, outList)
				if outList.LenNoPadding() < batchSize {
					continue
				}
				if err := sendIt(); err != nil {
					return err
				}
			}
		}
		// Mark the stream as done.
		if st.doneMarkers {
			kv := &pb.KV{
				StreamId:   streamId,
				StreamDone: true,
			}
			KVToBuffer(kv, outList)
		}
		return sendIt()
	}

	for {
		select {
		case kr, ok := <-st.rangeCh:
			if !ok {
				// Done with the keys.
				return nil
			}
			if err := iterate(kr); err != nil {
				return err
			}
		case <-ctx.Done():
			return ctx.Err()
		}
	}
}

func (st *Stream) streamKVs(ctx context.Context) error {
	onDiskSize, uncompressedSize := st.db.EstimateSize(st.Prefix)
	// Manish has seen uncompressed size to be in 20% error margin.
	uncompressedSize = uint64(float64(uncompressedSize) * 1.2)
	st.db.opt.Infof("%s Streaming about %s of uncompressed data (%s on disk)\n",
		st.LogPrefix, humanize.IBytes(uncompressedSize), humanize.IBytes(onDiskSize))

	tickerDur := 5 * time.Second
	var bytesSent uint64
	t := time.NewTicker(tickerDur)
	defer t.Stop()
	now := time.Now()

	sendBatch := func(batch *z.Buffer) error {
		defer batch.Release()
		sz := uint64(batch.LenNoPadding())
		if sz == 0 {
			return nil
		}
		bytesSent += sz
		// st.db.opt.Infof("%s Sending batch of size: %s.\n", st.LogPrefix, humanize.IBytes(sz))
		if err := st.Send(batch); err != nil {
			st.db.opt.Warningf("Error while sending: %v\n", err)
			return err
		}
		return nil
	}

	slurp := func(batch *z.Buffer) error {
	loop:
		for {
			// Send the batch immediately if it already exceeds the maximum allowed size.
			// If the size of the batch exceeds maxStreamSize, break from the loop to
			// avoid creating a batch that is so big that certain limits are reached.
			if batch.LenNoPadding() > int(maxStreamSize) {
				break loop
			}
			select {
			case kvs, ok := <-st.kvChan:
				if !ok {
					break loop
				}
				y.AssertTrue(kvs != nil)
				y.Check2(batch.Write(kvs.Bytes()))
				y.Check(kvs.Release())

			default:
				break loop
			}
		}
		return sendBatch(batch)
	} // end of slurp.

	writeRate := y.NewRateMonitor(20)
	scanRate := y.NewRateMonitor(20)
outer:
	for {
		var batch *z.Buffer
		select {
		case <-ctx.Done():
			return ctx.Err()

		case <-t.C:
			// Instead of calculating speed over the entire lifetime, we average the speed over
			// ticker duration.
			writeRate.Capture(bytesSent)
			scanned := atomic.LoadUint64(&st.scanned)
			scanRate.Capture(scanned)
			numProducers := atomic.LoadInt32(&st.numProducers)

			st.db.opt.Infof("%s [%s] Scan (%d): ~%s/%s at %s/sec. Sent: %s at %s/sec."+
				" jemalloc: %s\n",
				st.LogPrefix, y.FixedDuration(time.Since(now)), numProducers,
				y.IBytesToString(scanned, 1), humanize.IBytes(uncompressedSize),
				humanize.IBytes(scanRate.Rate()),
				y.IBytesToString(bytesSent, 1), humanize.IBytes(writeRate.Rate()),
				humanize.IBytes(uint64(z.NumAllocBytes())))

		case kvs, ok := <-st.kvChan:
			if !ok {
				break outer
			}
			y.AssertTrue(kvs != nil)
			batch = kvs

			// Otherwise, slurp more keys into this batch.
			if err := slurp(batch); err != nil {
				return err
			}
		}
	}

	st.db.opt.Infof("%s Sent data of size %s\n", st.LogPrefix, humanize.IBytes(bytesSent))
	return nil
}

func (st *Stream) copyTablesOver(ctx context.Context, tableMatrix [][]*table.Table) error {
	// TODO: See if making this concurrent would be helpful. Most likely it won't.
	// But, if it does work, then most like <3 goroutines might be sufficient.
	infof := st.db.opt.Infof
	// Make a copy of the manifest so that we don't have race condition.
	manifest := st.db.manifest.manifest.clone()
	dataKeys := make(map[uint64]struct{})
	// Iterate in reverse order so that the receiver gets the bottommost level first.
	for i := len(tableMatrix) - 1; i >= 0; i-- {
		level := i
		tables := tableMatrix[i]
		for _, t := range tables {
			// This table can be picked for copying directly.
			out := z.NewBuffer(int(t.Size())+1024, "Stream.Table")
			if dk := t.DataKey(); dk != nil {
				y.AssertTrue(dk.KeyId != 0)
				// If we have a legit data key, send it over so the table can be decrypted. The same
				// data key could have been used to encrypt many tables. Avoid sending it
				// repeatedly.
				if _, sent := dataKeys[dk.KeyId]; !sent {
					infof("Sending data key with ID: %d\n", dk.KeyId)
					val, err := dk.Marshal()
					y.Check(err)

					// This would go to key registry in destination.
					kv := &pb.KV{
						Value: val,
						Kind:  pb.KV_DATA_KEY,
					}
					KVToBuffer(kv, out)
					dataKeys[dk.KeyId] = struct{}{}
				}
			}

			infof("Sending table ID: %d at level: %d. Size: %s\n",
				t.ID(), level, humanize.IBytes(uint64(t.Size())))
			tableManifest := manifest.Tables[t.ID()]
			change := pb.ManifestChange{
				Op:    pb.ManifestChange_CREATE,
				Level: uint32(level),
				KeyId: tableManifest.KeyID,
				// Hard coding it, since we're supporting only AES for now.
				EncryptionAlgo: pb.EncryptionAlgo_aes,
				Compression:    uint32(tableManifest.Compression),
			}

			buf, err := change.Marshal()
			y.Check(err)

			// We send the table along with level to the destination, so they'd know where to
			// place the tables. We'd send all the tables first, before we start streaming. So, the
			// destination DB would write streamed keys one level above.
			kv := &pb.KV{
				// Key can be used for MANIFEST.
				Key:   buf,
				Value: t.Data,
				Kind:  pb.KV_FILE,
			}
			KVToBuffer(kv, out)

			select {
			case st.kvChan <- out:
			case <-ctx.Done():
				out.Release()
				return ctx.Err()
			}
		}
	}
	return nil
}

// Orchestrate runs Stream. It picks up ranges from the SSTables, then runs NumGo number of
// goroutines to iterate over these ranges and batch up KVs in lists. It concurrently runs a single
// goroutine to pick these lists, batch them up further and send to Output.Send. Orchestrate also
// spits logs out to Infof, using provided LogPrefix. Note that all calls to Output.Send
// are serial. In case any of these steps encounter an error, Orchestrate would stop execution and
// return that error. Orchestrate can be called multiple times, but in serial order.
func (st *Stream) Orchestrate(ctx context.Context) error {
	if st.FullCopy {
		if !st.db.opt.managedTxns || st.SinceTs != 0 || st.ChooseKey != nil && st.KeyToList != nil {
			panic("Got invalid stream options when doing full copy")
		}
	}
	ctx, cancel := context.WithCancel(ctx)
	defer cancel()
	st.rangeCh = make(chan keyRange, 3) // Contains keys for posting lists.

	// kvChan should only have a small capacity to ensure that we don't buffer up too much data if
	// sending is slow. Page size is set to 4MB, which is used to lazily cap the size of each
	// KVList. To get 128MB buffer, we can set the channel size to 32.
	st.kvChan = make(chan *z.Buffer, 32)

	if st.KeyToList == nil {
		st.KeyToList = st.ToList
	}

	// Pick up key-values from kvChan and send to stream.
	kvErr := make(chan error, 1)
	go func() {
		// Picks up KV lists from kvChan, and sends them to Output.
		err := st.streamKVs(ctx)
		if err != nil {
			cancel() // Stop all the go routines.
		}
		kvErr <- err
	}()

	// Pick all relevant tables from levels. We'd use this to copy them over,
	// or generate iterators from them.
	memTables, decr := st.db.getMemTables()
	defer decr()

	opts := DefaultIteratorOptions
	opts.Prefix = st.Prefix
	opts.SinceTs = st.SinceTs
	tableMatrix := st.db.lc.getTables(&opts)
	defer func() {
		for _, tables := range tableMatrix {
			for _, t := range tables {
				t.DecrRef()
			}
		}
	}()
	y.AssertTrue(len(tableMatrix) == st.db.opt.MaxLevels)

	infof := st.db.opt.Infof
	copyTables := func() error {
		// Figure out which tables we can copy. Only choose from the last 2 levels.
		// Say last level has data of size 100. Given a 10x level multiplier and
		// assuming the tree is balanced, second last level would have 10, and the
		// third last level would have 1. The third last level would only have 1%
		// of the data of the last level. It's OK for us to stop there and just
		// stream it, instead of trying to copy over those tables too.  When we
		// copy over tables to Level i, we can't stream any data to level i, i+1,
		// and so on. The stream has to create tables at level i-1, so there can be
		// overlap between the tables at i-1 and i.

		// Let's pick the tables which can be fully copied over from last level.
		threshold := len(tableMatrix) - 2
		toCopy := make([][]*table.Table, len(tableMatrix))
		var numCopy, numStream int
		for lev, tables := range tableMatrix {
			// We stream only the data in the two bottommost levels.
			if lev < threshold {
				numStream += len(tables)
				continue
			}
			var rem []*table.Table
			cp := tables[:0]
			for _, t := range tables {
				// We can only copy over those tables that satisfy following conditions:
				// - All the keys have version less than st.readTs
				// - st.Prefix fully covers the table
				if t.MaxVersion() > st.readTs || !t.CoveredByPrefix(st.Prefix) {
					rem = append(rem, t)
					continue
				}
				cp = append(cp, t)
			}
			toCopy[lev] = cp       // Pick tables to copy.
			tableMatrix[lev] = rem // Keep remaining for streaming.
			numCopy += len(cp)
			numStream += len(rem)
		}
		infof("Num tables to copy: %d. Num to stream: %d\n", numCopy, numStream)

		return st.copyTablesOver(ctx, toCopy)
	}

	if st.FullCopy {
		// As of now, we don't handle the non-zero SinceTs.
		if err := copyTables(); err != nil {
			return errors.Wrap(err, "while copying tables")
		}
	}

	var txn *Txn
	if st.readTs > 0 {
		txn = st.db.NewTransactionAt(st.readTs, false)
	} else {
		txn = st.db.NewTransaction(false)
	}
	defer txn.Discard()

	newIterator := func(threadId int) *Iterator {
		var itrs []y.Iterator
		for _, mt := range memTables {
			itrs = append(itrs, mt.sl.NewUniIterator(false))
		}
		if tables := tableMatrix[0]; len(tables) > 0 {
			itrs = append(itrs, iteratorsReversed(tables, 0)...)
		}
		for _, tables := range tableMatrix[1:] {
			if len(tables) == 0 {
				continue
			}
			itrs = append(itrs, table.NewConcatIterator(tables, 0))
		}

		opt := DefaultIteratorOptions
		opt.AllVersions = true
		opt.Prefix = st.Prefix
		opt.PrefetchValues = false
		opt.SinceTs = st.SinceTs

		res := &Iterator{
			txn:      txn,
			iitr:     table.NewMergeIterator(itrs, false),
			opt:      opt,
			readTs:   txn.readTs,
			ThreadId: threadId,
		}
		return res
	}

	// Picks up ranges from Badger, and sends them to rangeCh.
	// Just for simplicity, we'd consider all the tables for range production.
	go st.produceRanges(ctx)

	errCh := make(chan error, st.NumGo) // Stores error by consumeKeys.
	var wg sync.WaitGroup
	for i := 0; i < st.NumGo; i++ {
		wg.Add(1)

		go func(threadId int) {
			defer wg.Done()
			// Picks up ranges from rangeCh, generates KV lists, and sends them to kvChan.
			itr := newIterator(threadId)
			defer itr.Close()
			if err := st.produceKVs(ctx, itr); err != nil {
				select {
				case errCh <- err:
				default:
				}
			}
		}(i)
	}

	wg.Wait()        // Wait for produceKVs to be over.
	close(st.kvChan) // Now we can close kvChan.
	defer func() {
		// If due to some error, we have buffers left in kvChan, we should release them.
		for buf := range st.kvChan {
			buf.Release()
		}
	}()

	select {
	case err := <-errCh: // Check error from produceKVs.
		return err
	default:
	}

	// Wait for key streaming to be over.
	err := <-kvErr
	return err
}

func (db *DB) newStream() *Stream {
	return &Stream{
		db:        db,
		NumGo:     db.opt.NumGoroutines,
		LogPrefix: "Badger.Stream",
	}
}

// NewStream creates a new Stream.
func (db *DB) NewStream() *Stream {
	if db.opt.managedTxns {
		panic("This API can not be called in managed mode.")
	}
	return db.newStream()
}

// NewStreamAt creates a new Stream at a particular timestamp. Should only be used with managed DB.
func (db *DB) NewStreamAt(readTs uint64) *Stream {
	if !db.opt.managedTxns {
		panic("This API can only be called in managed mode.")
	}
	stream := db.newStream()
	stream.readTs = readTs
	return stream
}

func BufferToKVList(buf *z.Buffer) (*pb.KVList, error) {
	var list pb.KVList
	err := buf.SliceIterate(func(s []byte) error {
		kv := new(pb.KV)
		if err := kv.Unmarshal(s); err != nil {
			return err
		}
		list.Kv = append(list.Kv, kv)
		return nil
	})
	return &list, err
}

func KVToBuffer(kv *pb.KV, buf *z.Buffer) {
	out := buf.SliceAllocate(kv.Size())
	y.Check2(kv.MarshalToSizedBuffer(out))
}