2pc.go

// Copyright 2016 PingCAP, Inc.
//
// 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,
// See the License for the specific language governing permissions and
// limitations under the License.

package tikv

import (
	"bytes"
	"math"
	"sync"
	"sync/atomic"
	"time"

	"github.com/juju/errors"
	"github.com/opentracing/opentracing-go"
	pb "github.com/pingcap/kvproto/pkg/kvrpcpb"
	"github.com/pingcap/tidb/kv"
	"github.com/pingcap/tidb/metrics"
	"github.com/pingcap/tidb/sessionctx/binloginfo"
	"github.com/pingcap/tidb/store/tikv/tikvrpc"
	"github.com/pingcap/tidb/tablecodec"
	"github.com/pingcap/tidb/terror"
	"github.com/pingcap/tidb/util/goroutine_pool"
	binlog "github.com/pingcap/tipb/go-binlog"
	log "github.com/sirupsen/logrus"
	"golang.org/x/net/context"
)

type twoPhaseCommitAction int

const (
	actionPrewrite twoPhaseCommitAction = 1
	actionCommit   twoPhaseCommitAction = 2
	actionCleanup  twoPhaseCommitAction = 3
)

var twoPhaseCommitGP = gp.New(3 * time.Minute)

func (ca twoPhaseCommitAction) String() string {
	switch ca {
	case actionPrewrite:
		return "prewrite"
	case actionCommit:
		return "commit"
	case actionCleanup:
		return "cleanup"
	}
	return "unknown"
}

// MetricsTag returns detail tag for metrics.
func (ca twoPhaseCommitAction) MetricsTag() string {
	return "2pc_" + ca.String()
}

// twoPhaseCommitter executes a two-phase commit protocol.
type twoPhaseCommitter struct {
	store     *tikvStore
	txn       *tikvTxn
	startTS   uint64
	keys      [][]byte
	mutations map[string]*pb.Mutation
	lockTTL   uint64
	commitTS  uint64
	mu        struct {
		sync.RWMutex
		committed       bool
		undeterminedErr error // undeterminedErr saves the rpc error we encounter when commit primary key.
	}
	priority pb.CommandPri
	syncLog  bool
	connID   uint64 // connID is used for log.
	cleanWg  sync.WaitGroup
}

// newTwoPhaseCommitter creates a twoPhaseCommitter.
func newTwoPhaseCommitter(txn *tikvTxn, connID uint64) (*twoPhaseCommitter, error) {
	var (
		keys    [][]byte
		size    int
		putCnt  int
		delCnt  int
		lockCnt int
	)
	mutations := make(map[string]*pb.Mutation)
	err := txn.us.WalkBuffer(func(k kv.Key, v []byte) error {
		if len(v) > 0 {
			mutations[string(k)] = &pb.Mutation{
				Op:    pb.Op_Put,
				Key:   k,
				Value: v,
			}
			putCnt++
		} else {
			mutations[string(k)] = &pb.Mutation{
				Op:  pb.Op_Del,
				Key: k,
			}
			delCnt++
		}
		keys = append(keys, k)
		entrySize := len(k) + len(v)
		if entrySize > kv.TxnEntrySizeLimit {
			return kv.ErrEntryTooLarge
		}
		size += entrySize
		return nil
	})
	if err != nil {
		return nil, errors.Trace(err)
	}
	for _, lockKey := range txn.lockKeys {
		if _, ok := mutations[string(lockKey)]; !ok {
			mutations[string(lockKey)] = &pb.Mutation{
				Op:  pb.Op_Lock,
				Key: lockKey,
			}
			lockCnt++
			keys = append(keys, lockKey)
			size += len(lockKey)
		}
	}
	if len(keys) == 0 {
		return nil, nil
	}
	entrylimit := atomic.LoadUint64(&kv.TxnEntryCountLimit)
	if len(keys) > int(entrylimit) || size > kv.TxnTotalSizeLimit {
		return nil, kv.ErrTxnTooLarge
	}
	const logEntryCount = 10000
	const logSize = 4 * 1024 * 1024 // 4MB
	if len(keys) > logEntryCount || size > logSize {
		tableID := tablecodec.DecodeTableID(keys[0])
		log.Infof("[BIG_TXN] con:%d table id:%d size:%d, keys:%d, puts:%d, dels:%d, locks:%d, startTS:%d",
			connID, tableID, size, len(keys), putCnt, delCnt, lockCnt, txn.startTS)
	}

	metrics.TiKVTxnWriteKVCountHistogram.Observe(float64(len(keys)))
	metrics.TiKVTxnWriteSizeHistogram.Observe(float64(size))
	return &twoPhaseCommitter{
		store:     txn.store,
		txn:       txn,
		startTS:   txn.StartTS(),
		keys:      keys,
		mutations: mutations,
		lockTTL:   txnLockTTL(txn.startTime, size),
		priority:  getTxnPriority(txn),
		syncLog:   getTxnSyncLog(txn),
		connID:    connID,
	}, nil
}

func (c *twoPhaseCommitter) primary() []byte {
	return c.keys[0]
}

const bytesPerMiB = 1024 * 1024

func txnLockTTL(startTime time.Time, txnSize int) uint64 {
	// Increase lockTTL for large transactions.
	// The formula is `ttl = ttlFactor * sqrt(sizeInMiB)`.
	// When writeSize is less than 256KB, the base ttl is defaultTTL (3s);
	// When writeSize is 1MiB, 100MiB, or 400MiB, ttl is 6s, 60s, 120s correspondingly;
	lockTTL := defaultLockTTL
	if txnSize >= txnCommitBatchSize {
		sizeMiB := float64(txnSize) / bytesPerMiB
		lockTTL = uint64(float64(ttlFactor) * math.Sqrt(sizeMiB))
		if lockTTL < defaultLockTTL {
			lockTTL = defaultLockTTL
		}
		if lockTTL > maxLockTTL {
			lockTTL = maxLockTTL
		}
	}

	// Increase lockTTL by the transaction's read time.
	// When resolving a lock, we compare current ts and startTS+lockTTL to decide whether to clean up. If a txn
	// takes a long time to read, increasing its TTL will help to prevent it from been aborted soon after prewrite.
	elapsed := time.Since(startTime) / time.Millisecond
	return lockTTL + uint64(elapsed)
}

// doActionOnKeys groups keys into primary batch and secondary batches, if primary batch exists in the key,
// it does action on primary batch first, then on secondary batches. If action is commit, secondary batches
// is done in background goroutine.
func (c *twoPhaseCommitter) doActionOnKeys(bo *Backoffer, action twoPhaseCommitAction, keys [][]byte) error {
	if len(keys) == 0 {
		return nil
	}
	groups, firstRegion, err := c.store.regionCache.GroupKeysByRegion(bo, keys)
	if err != nil {
		return errors.Trace(err)
	}

	metrics.TiKVTxnRegionsNumHistogram.WithLabelValues(action.MetricsTag()).Observe(float64(len(groups)))

	var batches []batchKeys
	var sizeFunc = c.keySize
	if action == actionPrewrite {
		sizeFunc = c.keyValueSize
	}
	// Make sure the group that contains primary key goes first.
	batches = appendBatchBySize(batches, firstRegion, groups[firstRegion], sizeFunc, txnCommitBatchSize)
	delete(groups, firstRegion)
	for id, g := range groups {
		batches = appendBatchBySize(batches, id, g, sizeFunc, txnCommitBatchSize)
	}

	firstIsPrimary := bytes.Equal(keys[0], c.primary())
	if firstIsPrimary && (action == actionCommit || action == actionCleanup) {
		// primary should be committed/cleanup first
		err = c.doActionOnBatches(bo, action, batches[:1])
		if err != nil {
			return errors.Trace(err)
		}
		batches = batches[1:]
	}
	if action == actionCommit {
		// Commit secondary batches in background goroutine to reduce latency.
		twoPhaseCommitGP.Go(func() {
			e := c.doActionOnBatches(bo, action, batches)
			if e != nil {
				log.Debugf("con:%d 2PC async doActionOnBatches %s err: %v", c.connID, action, e)
				metrics.TiKVSecondaryLockCleanupFailureCounter.WithLabelValues("commit").Inc()
			}
		})
	} else {
		err = c.doActionOnBatches(bo, action, batches)
	}
	return errors.Trace(err)
}

// doActionOnBatches does action to batches in parallel.
func (c *twoPhaseCommitter) doActionOnBatches(bo *Backoffer, action twoPhaseCommitAction, batches []batchKeys) error {
	if len(batches) == 0 {
		return nil
	}
	var singleBatchActionFunc func(bo *Backoffer, batch batchKeys) error
	switch action {
	case actionPrewrite:
		singleBatchActionFunc = c.prewriteSingleBatch
	case actionCommit:
		singleBatchActionFunc = c.commitSingleBatch
	case actionCleanup:
		singleBatchActionFunc = c.cleanupSingleBatch
	}
	if len(batches) == 1 {
		e := singleBatchActionFunc(bo, batches[0])
		if e != nil {
			log.Debugf("con:%d 2PC doActionOnBatches %s failed: %v, tid: %d", c.connID, action, e, c.startTS)
		}
		return errors.Trace(e)
	}

	// For prewrite, stop sending other requests after receiving first error.
	backoffer := bo
	var cancel context.CancelFunc
	if action == actionPrewrite {
		backoffer, cancel = bo.Fork()
	}

	// Concurrently do the work for each batch.
	ch := make(chan error, len(batches))
	for _, batch1 := range batches {

		batch := batch1
		twoPhaseCommitGP.Go(func() {
			if action == actionCommit {
				// Because the secondary batches of the commit actions are implemented to be
				// committed asynchronously in background goroutines, we should not
				// fork a child context and call cancel() while the foreground goroutine exits.
				// Otherwise the background goroutines will be canceled execeptionally.
				// Here we makes a new clone of the original backoffer for this goroutine
				// exclusively to avoid the data race when using the same backoffer
				// in concurrent goroutines.
				singleBatchBackoffer := backoffer.Clone()
				ch <- singleBatchActionFunc(singleBatchBackoffer, batch)
			} else {
				singleBatchBackoffer, singleBatchCancel := backoffer.Fork()
				defer singleBatchCancel()
				ch <- singleBatchActionFunc(singleBatchBackoffer, batch)
			}
		})
	}
	var err error
	for i := 0; i < len(batches); i++ {
		if e := <-ch; e != nil {
			log.Debugf("con:%d 2PC doActionOnBatches %s failed: %v, tid: %d", c.connID, action, e, c.startTS)
			// Cancel other requests and return the first error.
			if cancel != nil {
				log.Debugf("con:%d 2PC doActionOnBatches %s to cancel other actions, tid: %d", c.connID, action, c.startTS)
				cancel()
			}
			if err == nil {
				err = e
			}
		}
	}
	return errors.Trace(err)
}

func (c *twoPhaseCommitter) keyValueSize(key []byte) int {
	size := len(key)
	if mutation := c.mutations[string(key)]; mutation != nil {
		size += len(mutation.Value)
	}
	return size
}

func (c *twoPhaseCommitter) keySize(key []byte) int {
	return len(key)
}

func (c *twoPhaseCommitter) prewriteSingleBatch(bo *Backoffer, batch batchKeys) error {
	mutations := make([]*pb.Mutation, len(batch.keys))
	for i, k := range batch.keys {
		mutations[i] = c.mutations[string(k)]
	}

	req := &tikvrpc.Request{
		Type: tikvrpc.CmdPrewrite,
		Prewrite: &pb.PrewriteRequest{
			Mutations:    mutations,
			PrimaryLock:  c.primary(),
			StartVersion: c.startTS,
			LockTtl:      c.lockTTL,
		},
		Context: pb.Context{
			Priority: c.priority,
			SyncLog:  c.syncLog,
		},
	}
	for {
		resp, err := c.store.SendReq(bo, req, batch.region, readTimeoutShort)
		if err != nil {
			return errors.Trace(err)
		}
		regionErr, err := resp.GetRegionError()
		if err != nil {
			return errors.Trace(err)
		}
		if regionErr != nil {
			err = bo.Backoff(BoRegionMiss, errors.New(regionErr.String()))
			if err != nil {
				return errors.Trace(err)
			}
			err = c.prewriteKeys(bo, batch.keys)
			return errors.Trace(err)
		}
		prewriteResp := resp.Prewrite
		if prewriteResp == nil {
			return errors.Trace(ErrBodyMissing)
		}
		keyErrs := prewriteResp.GetErrors()
		if len(keyErrs) == 0 {
			return nil
		}
		var locks []*Lock
		for _, keyErr := range keyErrs {
			lock, err1 := extractLockFromKeyErr(keyErr)
			if err1 != nil {
				return errors.Trace(err1)
			}
			log.Debugf("con:%d 2PC prewrite encounters lock: %v", c.connID, lock)
			locks = append(locks, lock)
		}
		ok, err := c.store.lockResolver.ResolveLocks(bo, locks)
		if err != nil {
			return errors.Trace(err)
		}
		if !ok {
			err = bo.Backoff(BoTxnLock, errors.Errorf("2PC prewrite lockedKeys: %d", len(locks)))
			if err != nil {
				return errors.Trace(err)
			}
		}
	}
}

func getTxnPriority(txn *tikvTxn) pb.CommandPri {
	if pri := txn.us.GetOption(kv.Priority); pri != nil {
		return kvPriorityToCommandPri(pri.(int))
	}
	return pb.CommandPri_Normal
}

func getTxnSyncLog(txn *tikvTxn) bool {
	if syncOption := txn.us.GetOption(kv.SyncLog); syncOption != nil {
		return syncOption.(bool)
	}
	return false
}

func kvPriorityToCommandPri(pri int) pb.CommandPri {
	switch pri {
	case kv.PriorityLow:
		return pb.CommandPri_Low
	case kv.PriorityHigh:
		return pb.CommandPri_High
	}
	return pb.CommandPri_Normal
}

func (c *twoPhaseCommitter) setUndeterminedErr(err error) {
	c.mu.Lock()
	defer c.mu.Unlock()
	c.mu.undeterminedErr = err
}

func (c *twoPhaseCommitter) getUndeterminedErr() error {
	c.mu.RLock()
	defer c.mu.RUnlock()
	return c.mu.undeterminedErr
}

func (c *twoPhaseCommitter) commitSingleBatch(bo *Backoffer, batch batchKeys) error {
	req := &tikvrpc.Request{
		Type: tikvrpc.CmdCommit,
		Commit: &pb.CommitRequest{
			StartVersion:  c.startTS,
			Keys:          batch.keys,
			CommitVersion: c.commitTS,
		},
		Context: pb.Context{
			Priority: c.priority,
			SyncLog:  c.syncLog,
		},
	}
	req.Context.Priority = c.priority

	sender := NewRegionRequestSender(c.store.regionCache, c.store.client)
	resp, err := sender.SendReq(bo, req, batch.region, readTimeoutShort)

	// If we fail to receive response for the request that commits primary key, it will be undetermined whether this
	// transaction has been successfully committed.
	// Under this circumstance,  we can not declare the commit is complete (may lead to data lost), nor can we throw
	// an error (may lead to the duplicated key error when upper level restarts the transaction). Currently the best
	// solution is to populate this error and let upper layer drop the connection to the corresponding mysql client.
	isPrimary := bytes.Equal(batch.keys[0], c.primary())
	if isPrimary && sender.rpcError != nil {
		c.setUndeterminedErr(errors.Trace(sender.rpcError))
	}

	if err != nil {
		return errors.Trace(err)
	}
	regionErr, err := resp.GetRegionError()
	if err != nil {
		return errors.Trace(err)
	}
	if regionErr != nil {
		err = bo.Backoff(BoRegionMiss, errors.New(regionErr.String()))
		if err != nil {
			return errors.Trace(err)
		}
		// re-split keys and commit again.
		err = c.commitKeys(bo, batch.keys)
		return errors.Trace(err)
	}
	commitResp := resp.Commit
	if commitResp == nil {
		return errors.Trace(ErrBodyMissing)
	}
	// Here we can make sure tikv has processed the commit primary key request. So
	// we can clean undetermined error.
	if isPrimary {
		c.setUndeterminedErr(nil)
	}
	if keyErr := commitResp.GetError(); keyErr != nil {
		c.mu.RLock()
		defer c.mu.RUnlock()
		err = errors.Errorf("con:%d 2PC commit failed: %v", c.connID, keyErr.String())
		if c.mu.committed {
			// No secondary key could be rolled back after it's primary key is committed.
			// There must be a serious bug somewhere.
			log.Errorf("2PC failed commit key after primary key committed: %v, tid: %d", err, c.startTS)
			return errors.Trace(err)
		}
		// The transaction maybe rolled back by concurrent transactions.
		log.Debugf("2PC failed commit primary key: %v, retry later, tid: %d", err, c.startTS)
		return errors.Annotate(err, txnRetryableMark)
	}

	c.mu.Lock()
	defer c.mu.Unlock()
	// Group that contains primary key is always the first.
	// We mark transaction's status committed when we receive the first success response.
	c.mu.committed = true
	return nil
}

func (c *twoPhaseCommitter) cleanupSingleBatch(bo *Backoffer, batch batchKeys) error {
	req := &tikvrpc.Request{
		Type: tikvrpc.CmdBatchRollback,
		BatchRollback: &pb.BatchRollbackRequest{
			Keys:         batch.keys,
			StartVersion: c.startTS,
		},
		Context: pb.Context{
			Priority: c.priority,
			SyncLog:  c.syncLog,
		},
	}
	resp, err := c.store.SendReq(bo, req, batch.region, readTimeoutShort)
	if err != nil {
		return errors.Trace(err)
	}
	regionErr, err := resp.GetRegionError()
	if err != nil {
		return errors.Trace(err)
	}
	if regionErr != nil {
		err = bo.Backoff(BoRegionMiss, errors.New(regionErr.String()))
		if err != nil {
			return errors.Trace(err)
		}
		err = c.cleanupKeys(bo, batch.keys)
		return errors.Trace(err)
	}
	if keyErr := resp.BatchRollback.GetError(); keyErr != nil {
		err = errors.Errorf("con:%d 2PC cleanup failed: %s", c.connID, keyErr)
		log.Debugf("2PC failed cleanup key: %v, tid: %d", err, c.startTS)
		return errors.Trace(err)
	}
	return nil
}

func (c *twoPhaseCommitter) prewriteKeys(bo *Backoffer, keys [][]byte) error {
	return c.doActionOnKeys(bo, actionPrewrite, keys)
}

func (c *twoPhaseCommitter) commitKeys(bo *Backoffer, keys [][]byte) error {
	return c.doActionOnKeys(bo, actionCommit, keys)
}

func (c *twoPhaseCommitter) cleanupKeys(bo *Backoffer, keys [][]byte) error {
	return c.doActionOnKeys(bo, actionCleanup, keys)
}

// The max time a Txn may use (in ms) from its startTS to commitTS.
// We use it to guarantee GC worker will not influence any active txn. The value
// should be less than `gcRunInterval`.
const maxTxnTimeUse = 590000

// execute executes the two-phase commit protocol.
func (c *twoPhaseCommitter) execute(ctx context.Context) error {
	defer func() {
		// Always clean up all written keys if the txn does not commit.
		c.mu.RLock()
		committed := c.mu.committed
		undetermined := c.mu.undeterminedErr != nil
		c.mu.RUnlock()
		if !committed && !undetermined {
			c.cleanWg.Add(1)
			twoPhaseCommitGP.Go(func() {
				err := c.cleanupKeys(NewBackoffer(context.Background(), cleanupMaxBackoff), c.keys)
				if err != nil {
					metrics.TiKVSecondaryLockCleanupFailureCounter.WithLabelValues("rollback").Inc()
					log.Infof("con:%d 2PC cleanup err: %v, tid: %d", c.connID, err, c.startTS)
				} else {
					log.Infof("con:%d 2PC clean up done, tid: %d", c.connID, c.startTS)
				}
				c.cleanWg.Done()
			})
		}
	}()

	span := opentracing.SpanFromContext(ctx)
	if span != nil {
		span = opentracing.StartSpan("twoPhaseCommit.execute", opentracing.ChildOf(span.Context()))
	} else {
		// If we lost the trace information, make a new one for 2PC commit.
		span = opentracing.StartSpan("twoPhaseCommit.execute")
	}
	defer span.Finish()

	// I'm not sure is it safe to cancel 2pc commit process at any time,
	// So use a new Background() context instead of inherit the ctx, this is by design,
	// to avoid the cancel signal from parent context.
	ctx = opentracing.ContextWithSpan(context.Background(), span)

	binlogChan := c.prewriteBinlog()
	err := c.prewriteKeys(NewBackoffer(ctx, prewriteMaxBackoff), c.keys)
	if binlogChan != nil {
		binlogErr := <-binlogChan
		if binlogErr != nil {
			return errors.Trace(binlogErr)
		}
	}
	if err != nil {
		log.Debugf("con:%d 2PC failed on prewrite: %v, tid: %d", c.connID, err, c.startTS)
		return errors.Trace(err)
	}

	commitTS, err := c.store.getTimestampWithRetry(NewBackoffer(ctx, tsoMaxBackoff))
	if err != nil {
		log.Warnf("con:%d 2PC get commitTS failed: %v, tid: %d", c.connID, err, c.startTS)
		return errors.Trace(err)
	}

	// check commitTS
	if commitTS <= c.startTS {
		err = errors.Errorf("con:%d Invalid transaction tso with start_ts=%v while commit_ts=%v",
			c.connID, c.startTS, commitTS)
		log.Error(err)
		return errors.Trace(err)
	}
	c.commitTS = commitTS
	if err = c.checkSchemaValid(); err != nil {
		return errors.Trace(err)
	}

	if c.store.oracle.IsExpired(c.startTS, maxTxnTimeUse) {
		err = errors.Errorf("con:%d txn takes too much time, start: %d, commit: %d", c.connID, c.startTS, c.commitTS)
		return errors.Annotate(err, txnRetryableMark)
	}

	err = c.commitKeys(NewBackoffer(ctx, CommitMaxBackoff), c.keys)
	if err != nil {
		if undeterminedErr := c.getUndeterminedErr(); undeterminedErr != nil {
			log.Warnf("con:%d 2PC commit result undetermined, err: %v, rpcErr: %v, tid: %v", c.connID, err, undeterminedErr, c.startTS)
			err = errors.Wrap(err, terror.ErrResultUndetermined)
		}
		if !c.mu.committed {
			log.Debugf("con:%d 2PC failed on commit: %v, tid: %d", c.connID, err, c.startTS)
			return errors.Trace(err)
		}
		log.Debugf("con:%d 2PC succeed with error: %v, tid: %d", c.connID, err, c.startTS)
	}
	return nil
}

type schemaLeaseChecker interface {
	Check(txnTS uint64) error
}

func (c *twoPhaseCommitter) checkSchemaValid() error {
	checker, ok := c.txn.us.GetOption(kv.SchemaLeaseChecker).(schemaLeaseChecker)
	if ok {
		err := checker.Check(c.commitTS)
		if err != nil {
			return errors.Trace(err)
		}
	}
	return nil
}

func (c *twoPhaseCommitter) prewriteBinlog() chan error {
	if !c.shouldWriteBinlog() {
		return nil
	}
	ch := make(chan error, 1)
	go func() {
		binInfo := c.txn.us.GetOption(kv.BinlogInfo).(*binloginfo.BinlogInfo)
		bin := binInfo.Data
		bin.StartTs = int64(c.startTS)
		if bin.Tp == binlog.BinlogType_Prewrite {
			bin.PrewriteKey = c.keys[0]
		}
		err := binInfo.WriteBinlog(c.store.clusterID)
		ch <- errors.Trace(err)
	}()
	return ch
}

func (c *twoPhaseCommitter) writeFinishBinlog(tp binlog.BinlogType, commitTS int64) {
	if !c.shouldWriteBinlog() {
		return
	}
	binInfo := c.txn.us.GetOption(kv.BinlogInfo).(*binloginfo.BinlogInfo)
	binInfo.Data.Tp = tp
	binInfo.Data.CommitTs = commitTS
	binInfo.Data.PrewriteValue = nil
	go func() {
		err := binInfo.WriteBinlog(c.store.clusterID)
		if err != nil {
			log.Errorf("failed to write binlog: %v", err)
		}
	}()
}

func (c *twoPhaseCommitter) shouldWriteBinlog() bool {
	return c.txn.us.GetOption(kv.BinlogInfo) != nil
}

// TiKV recommends each RPC packet should be less than ~1MB. We keep each packet's
// Key+Value size below 16KB.
const txnCommitBatchSize = 16 * 1024

// batchKeys is a batch of keys in the same region.
type batchKeys struct {
	region RegionVerID
	keys   [][]byte
}

// appendBatchBySize appends keys to []batchKeys. It may split the keys to make
// sure each batch's size does not exceed the limit.
func appendBatchBySize(b []batchKeys, region RegionVerID, keys [][]byte, sizeFn func([]byte) int, limit int) []batchKeys {
	var start, end int
	for start = 0; start < len(keys); start = end {
		var size int
		for end = start; end < len(keys) && size < limit; end++ {
			size += sizeFn(keys[end])
		}
		b = append(b, batchKeys{
			region: region,
			keys:   keys[start:end],
		})
	}
	return b
}