vdiff.go

/*
Copyright 2019 The Vitess Authors.

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 wrangler

import (
	"context"
	"encoding/json"
	"fmt"
	"math"
	"reflect"
	"sort"
	"strings"
	"sync"
	"time"

	"google.golang.org/protobuf/encoding/prototext"

	"vitess.io/vitess/go/mysql/replication"
	"vitess.io/vitess/go/mysql/sqlerror"
	"vitess.io/vitess/go/ptr"
	"vitess.io/vitess/go/vt/vtenv"

	"vitess.io/vitess/go/mysql/collations"
	"vitess.io/vitess/go/sqltypes"
	"vitess.io/vitess/go/vt/binlog/binlogplayer"
	"vitess.io/vitess/go/vt/concurrency"
	"vitess.io/vitess/go/vt/discovery"
	"vitess.io/vitess/go/vt/grpcclient"
	"vitess.io/vitess/go/vt/key"
	"vitess.io/vitess/go/vt/log"
	"vitess.io/vitess/go/vt/logutil"
	"vitess.io/vitess/go/vt/schema"
	"vitess.io/vitess/go/vt/schemadiff"
	"vitess.io/vitess/go/vt/sqlparser"
	"vitess.io/vitess/go/vt/topo"
	"vitess.io/vitess/go/vt/topo/topoproto"
	"vitess.io/vitess/go/vt/vtctl/schematools"
	"vitess.io/vitess/go/vt/vtctl/workflow"
	"vitess.io/vitess/go/vt/vterrors"
	"vitess.io/vitess/go/vt/vtgate/engine"
	"vitess.io/vitess/go/vt/vtgate/engine/opcode"
	"vitess.io/vitess/go/vt/vtgate/evalengine"
	"vitess.io/vitess/go/vt/vtgate/vindexes"
	"vitess.io/vitess/go/vt/vttablet/tabletconn"
	"vitess.io/vitess/go/vt/vttablet/tabletmanager/vreplication"

	binlogdatapb "vitess.io/vitess/go/vt/proto/binlogdata"
	querypb "vitess.io/vitess/go/vt/proto/query"
	tabletmanagerdatapb "vitess.io/vitess/go/vt/proto/tabletmanagerdata"
	topodatapb "vitess.io/vitess/go/vt/proto/topodata"
)

// At most how many samples we should show for row differences in the final report
const maxVDiffReportSampleRows = 10

// DiffReport is the summary of differences for one table.
type DiffReport struct {
	ProcessedRows        int
	MatchingRows         int
	MismatchedRows       int
	ExtraRowsSource      int
	ExtraRowsSourceDiffs []*RowDiff
	ExtraRowsTarget      int
	ExtraRowsTargetDiffs []*RowDiff
	MismatchedRowsSample []*DiffMismatch
	TableName            string
}

// DiffMismatch is a sample of row diffs between source and target.
type DiffMismatch struct {
	Source *RowDiff
	Target *RowDiff
}

// RowDiff is a row that didn't match as part of the comparison.
type RowDiff struct {
	Row   map[string]sqltypes.Value
	Query string
}

// vdiff contains the metadata for performing vdiff for one workflow.
type vdiff struct {
	env            *vtenv.Environment
	ts             *trafficSwitcher
	sourceCell     string
	targetCell     string
	tabletTypesStr string

	// differs uses the target table name for its key.
	differs map[string]*tableDiffer

	// The key for sources and targets is the shard name.
	// The source and target keyspaces are pulled from ts.
	sources map[string]*shardStreamer
	targets map[string]*shardStreamer

	workflow       string
	targetKeyspace string
	tables         []string
	sourceTimeZone string
	targetTimeZone string
}

// compareColInfo contains the metadata for a column of the table being diffed
type compareColInfo struct {
	colIndex  int                       // index of the column in the filter's select
	collation collations.ID             // is the collation of the column, if any
	values    *evalengine.EnumSetValues // is the list of enum or set values for the column, if any
	isPK      bool                      // is this column part of the primary key
}

// tableDiffer performs a diff for one table in the workflow.
type tableDiffer struct {
	targetTable string
	// sourceExpression and targetExpression are select queries.
	sourceExpression string
	targetExpression string

	// compareCols is the list of non-pk columns to compare.
	// If the value is -1, it's a pk column and should not be
	// compared.
	compareCols []compareColInfo
	// comparePKs is the list of pk columns to compare. The logic
	// for comparing pk columns is different from compareCols
	comparePKs []compareColInfo
	// pkCols has the indices of PK cols in the select list
	pkCols []int

	// selectPks is the list of pk columns as they appear in the select clause for the diff.
	selectPks []int

	// source Primitive and targetPrimitive are used for streaming
	sourcePrimitive engine.Primitive
	targetPrimitive engine.Primitive

	collationEnv *collations.Environment
	parser       *sqlparser.Parser
}

// shardStreamer streams rows from one shard. This works for
// the source as well as the target.
// shardStreamer satisfies engine.StreamExecutor, and can be
// added to Primitives of engine.MergeSort.
// shardStreamer is a member of vdiff, and gets reused by
// every tableDiffer. A new result channel gets instantiated
// for every tableDiffer iteration.
type shardStreamer struct {
	primary          *topo.TabletInfo
	tablet           *topodatapb.Tablet
	position         replication.Position
	snapshotPosition string
	result           chan *sqltypes.Result
	err              error
}

// VDiff reports differences between the sources and targets of a vreplication workflow.
func (wr *Wrangler) VDiff(ctx context.Context, targetKeyspace, workflowName, sourceCell, targetCell, tabletTypesStr string,
	filteredReplicationWaitTime time.Duration, format string, maxRows int64, tables string, debug, onlyPks bool,
	maxExtraRowsToCompare int) (map[string]*DiffReport, error) {
	log.Infof("Starting VDiff for %s.%s, sourceCell %s, targetCell %s, tabletTypes %s, timeout %s",
		targetKeyspace, workflowName, sourceCell, targetCell, tabletTypesStr, filteredReplicationWaitTime.String())
	// Assign defaults to sourceCell and targetCell if not specified.
	if sourceCell == "" && targetCell == "" {
		cells, err := wr.ts.GetCellInfoNames(ctx)
		if err != nil {
			return nil, err
		}
		if len(cells) == 0 {
			// Unreachable
			return nil, fmt.Errorf("there are no cells in the topo")
		}
		sourceCell = cells[0]
		targetCell = sourceCell
	}
	if sourceCell == "" {
		sourceCell = targetCell
	}
	if targetCell == "" {
		targetCell = sourceCell
	}

	// Reuse migrater code to fetch and validate initial metadata about the workflow.
	ts, err := wr.buildTrafficSwitcher(ctx, targetKeyspace, workflowName)
	if err != nil {
		wr.Logger().Errorf("buildTrafficSwitcher: %v", err)
		return nil, err
	}
	if ts.frozen {
		return nil, fmt.Errorf("invalid VDiff run: writes have been already been switched for workflow %s.%s",
			targetKeyspace, workflowName)
	}
	if err := ts.validate(ctx); err != nil {
		ts.Logger().Errorf("validate: %v", err)
		return nil, err
	}
	tables = strings.TrimSpace(tables)
	var includeTables []string
	if tables != "" {
		includeTables = strings.Split(tables, ",")
	}
	// Initialize vdiff
	df := &vdiff{
		env:            wr.env,
		ts:             ts,
		sourceCell:     sourceCell,
		targetCell:     targetCell,
		tabletTypesStr: tabletTypesStr,
		sources:        make(map[string]*shardStreamer),
		targets:        make(map[string]*shardStreamer),
		workflow:       workflowName,
		targetKeyspace: targetKeyspace,
		tables:         includeTables,
		sourceTimeZone: ts.sourceTimeZone,
		targetTimeZone: ts.targetTimeZone,
	}
	for shard, source := range ts.Sources() {
		df.sources[shard] = &shardStreamer{
			primary: source.GetPrimary(),
		}
	}
	var oneTarget *workflow.MigrationTarget
	for shard, target := range ts.Targets() {
		df.targets[shard] = &shardStreamer{
			primary: target.GetPrimary(),
		}
		oneTarget = target
	}
	var oneFilter *binlogdatapb.Filter
	for _, bls := range oneTarget.Sources {
		oneFilter = bls.Filter
		break
	}
	req := &tabletmanagerdatapb.GetSchemaRequest{}
	schm, err := schematools.GetSchema(ctx, wr.ts, wr.tmc, oneTarget.GetPrimary().Alias, req)
	if err != nil {
		return nil, vterrors.Wrap(err, "GetSchema")
	}
	if err = df.buildVDiffPlan(oneFilter, schm, df.tables); err != nil {
		return nil, vterrors.Wrap(err, "buildVDiffPlan")
	}

	if err := df.selectTablets(ctx, ts); err != nil {
		return nil, vterrors.Wrap(err, "selectTablets")
	}
	defer func() {
		// We use a new context as we want to reset the state even
		// when the parent context has timed out or been canceled.
		log.Infof("Restarting the %q VReplication workflow on target tablets in keyspace %q", df.workflow, df.targetKeyspace)
		restartCtx, restartCancel := context.WithTimeout(context.Background(), DefaultActionTimeout)
		defer restartCancel()
		if err := df.restartTargets(restartCtx); err != nil {
			wr.Logger().Errorf("Could not restart workflow %q on target tablets in keyspace %q: %v, please restart it manually",
				df.workflow, df.targetKeyspace, err)
		}
	}()

	// Perform the diffs.
	// We need a cancelable context to abort all running streams
	// if one stream returns an error.
	ctx, cancel := context.WithCancel(ctx)
	defer cancel()

	// Get the VSchema on the target and source keyspaces. We can
	// then use this for handling edge cases, such as adjusting
	// results for reference tables when the shard count is
	// different between the source and target as then there will
	// be a extra rows reported on the side with more shards.
	srcTopo := wr.ts
	if ts.ExternalTopo() != nil {
		srcTopo = ts.ExternalTopo()
	}
	srcvschema, err := srcTopo.GetVSchema(ctx, ts.SourceKeyspaceName())
	if err != nil {
		return nil, err
	}
	tgtvschema, err := wr.ts.GetVSchema(ctx, ts.TargetKeyspaceName())
	if err != nil {
		return nil, err
	}
	numSourceShards := len(ts.SourceShards())
	numTargetShards := len(ts.TargetShards())
	numShardDiff := int(math.Abs(float64(numSourceShards - numTargetShards)))

	// TODO(sougou): parallelize
	rowsToCompare := maxRows
	diffReports := make(map[string]*DiffReport)
	jsonOutput := ""
	for table, td := range df.differs {
		// Skip internal operation tables for vdiff
		if schema.IsInternalOperationTableName(table) {
			continue
		}
		if err := df.diffTable(ctx, wr, table, td, filteredReplicationWaitTime); err != nil {
			return nil, err
		}
		// Perform the diff of source and target streams.
		dr, err := td.diff(ctx, &rowsToCompare, debug, onlyPks, maxExtraRowsToCompare)
		if err != nil {
			return nil, vterrors.Wrap(err, "diff")
		}
		dr.TableName = table
		// This could be a reference table with a different number of shards on
		// the source and target, so let's check and adjust if needed. In that
		// case we should have no mismatched rows and the number of extra rows
		// should be a multiple of the extra shards.
		if numShardDiff > 0 && (dr.ExtraRowsSource > 0 || dr.ExtraRowsTarget > 0) && dr.MismatchedRows == 0 {
			// Each shard should have the same number of rows for a reference table.
			perShardRows := (dr.ProcessedRows + dr.MatchingRows) / (numSourceShards + numTargetShards)
			if perShardRows == dr.MatchingRows { // If not then there's a legitimate mismatch
				svt, sok := srcvschema.Tables[table]
				tvt, tok := tgtvschema.Tables[table]
				if numSourceShards > numTargetShards && sok && svt.Type == vindexes.TypeReference &&
					dr.ExtraRowsSource/numShardDiff == perShardRows {
					dr.ExtraRowsSource = 0
					dr.ExtraRowsSourceDiffs = nil
					dr.ProcessedRows = dr.MatchingRows
				}
				if numTargetShards > numSourceShards && tok && tvt.Type == vindexes.TypeReference &&
					dr.ExtraRowsTarget/numShardDiff == perShardRows {
					dr.ExtraRowsTarget = 0
					dr.ExtraRowsTargetDiffs = nil
					dr.ProcessedRows = dr.MatchingRows
				}
			}
		}
		// If the only difference is the order in which the rows were returned
		// by MySQL on each side then we'll have the same number of extras on
		// both sides. If that's the case, then let's see if the extra rows on
		// both sides are actually different.
		if (dr.ExtraRowsSource == dr.ExtraRowsTarget) && (dr.ExtraRowsSource <= maxExtraRowsToCompare) {
			for i := 0; i < len(dr.ExtraRowsSourceDiffs); i++ {
				foundMatch := false
				for j := 0; j < len(dr.ExtraRowsTargetDiffs); j++ {
					if reflect.DeepEqual(dr.ExtraRowsSourceDiffs[i], dr.ExtraRowsTargetDiffs[j]) {
						dr.ExtraRowsSourceDiffs = append(dr.ExtraRowsSourceDiffs[:i], dr.ExtraRowsSourceDiffs[i+1:]...)
						dr.ExtraRowsTargetDiffs = append(dr.ExtraRowsTargetDiffs[:j], dr.ExtraRowsTargetDiffs[j+1:]...)
						dr.ExtraRowsSource--
						dr.ExtraRowsTarget--
						dr.ProcessedRows--
						dr.MatchingRows++
						// We've removed an element from both slices at the current index
						// so we need to shift the counters back as well to process the
						// new elements at the index and avoid using an index out of range.
						i--
						j--
						foundMatch = true
						break
					}
				}
				// If we didn't find a match then the tables are in fact different and we can short circuit the second pass
				if !foundMatch {
					break
				}
			}
		}
		// We can now trim the extra rows diffs on both sides to the maxVDiffReportSampleRows value
		if len(dr.ExtraRowsSourceDiffs) > maxVDiffReportSampleRows {
			dr.ExtraRowsSourceDiffs = dr.ExtraRowsSourceDiffs[:maxVDiffReportSampleRows-1]
		}
		if len(dr.ExtraRowsTargetDiffs) > maxVDiffReportSampleRows {
			dr.ExtraRowsTargetDiffs = dr.ExtraRowsTargetDiffs[:maxVDiffReportSampleRows-1]
		}
		diffReports[table] = dr
	}
	if format == "json" {
		j, err := json.MarshalIndent(diffReports, "", "")
		if err != nil {
			wr.Logger().Printf("Error converting report to json: %v", err.Error())
		}
		jsonOutput += string(j)
		wr.logger.Printf("%s", jsonOutput)
	} else {
		for table, dr := range diffReports {
			wr.Logger().Printf("Summary for table %v:\n", table)
			wr.Logger().Printf("\tProcessedRows: %v\n", dr.ProcessedRows)
			wr.Logger().Printf("\tMatchingRows: %v\n", dr.MatchingRows)
			wr.Logger().Printf("\tMismatchedRows: %v\n", dr.MismatchedRows)
			wr.Logger().Printf("\tExtraRowsSource: %v\n", dr.ExtraRowsSource)
			wr.Logger().Printf("\tExtraRowsTarget: %v\n", dr.ExtraRowsTarget)
			for i, rs := range dr.ExtraRowsSourceDiffs {
				wr.Logger().Printf("\tSample extra row in source %v:\n", i)
				formatSampleRow(wr.Logger(), rs, debug)
			}
			for i, rs := range dr.ExtraRowsTargetDiffs {
				wr.Logger().Printf("\tSample extra row in target %v:\n", i)
				formatSampleRow(wr.Logger(), rs, debug)
			}
			for i, rs := range dr.MismatchedRowsSample {
				wr.Logger().Printf("\tSample rows with mismatch %v:\n", i)
				wr.Logger().Printf("\t\tSource row:\n")
				formatSampleRow(wr.Logger(), rs.Source, debug)
				wr.Logger().Printf("\t\tTarget row:\n")
				formatSampleRow(wr.Logger(), rs.Target, debug)
			}
		}
	}
	return diffReports, nil
}

func (df *vdiff) diffTable(ctx context.Context, wr *Wrangler, table string, td *tableDiffer, filteredReplicationWaitTime time.Duration) error {
	log.Infof("Starting vdiff for table %s", table)

	log.Infof("Locking target keyspace %s", df.targetKeyspace)
	ctx, unlock, lockErr := wr.ts.LockKeyspace(ctx, df.targetKeyspace, "vdiff")
	if lockErr != nil {
		log.Errorf("LockKeyspace failed: %v", lockErr)
		wr.Logger().Errorf("LockKeyspace %s failed: %v", df.targetKeyspace)
		return lockErr
	}

	var err error
	defer func() {
		unlock(&err)
		if err != nil {
			log.Errorf("UnlockKeyspace %s failed: %v", df.targetKeyspace, lockErr)
		}
	}()

	// Stop the targets and record their source positions.
	if err := df.stopTargets(ctx); err != nil {
		return vterrors.Wrap(err, "stopTargets")
	}
	// Make sure all sources are past the target's positions and start a query stream that records the current source positions.
	if err := df.startQueryStreams(ctx, df.ts.SourceKeyspaceName(), df.sources, td.sourceExpression, filteredReplicationWaitTime); err != nil {
		return vterrors.Wrap(err, "startQueryStreams(sources)")
	}
	// Fast forward the targets to the newly recorded source positions.
	if err := df.syncTargets(ctx, filteredReplicationWaitTime); err != nil {
		return vterrors.Wrap(err, "syncTargets")
	}
	// Sources and targets are in sync. Start query streams on the targets.
	if err := df.startQueryStreams(ctx, df.ts.TargetKeyspaceName(), df.targets, td.targetExpression, filteredReplicationWaitTime); err != nil {
		return vterrors.Wrap(err, "startQueryStreams(targets)")
	}
	// Now that queries are running, target vreplication streams can be restarted.
	return nil
}

// buildVDiffPlan builds all the differs.
func (df *vdiff) buildVDiffPlan(filter *binlogdatapb.Filter, schm *tabletmanagerdatapb.SchemaDefinition, tablesToInclude []string) error {
	df.differs = make(map[string]*tableDiffer)
	for _, table := range schm.TableDefinitions {
		rule, err := vreplication.MatchTable(table.Name, filter)
		if err != nil {
			return err
		}
		if rule == nil || rule.Filter == "exclude" {
			continue
		}
		query := rule.Filter
		if rule.Filter == "" || key.IsValidKeyRange(rule.Filter) {
			buf := sqlparser.NewTrackedBuffer(nil)
			buf.Myprintf("select * from %v", sqlparser.NewIdentifierCS(table.Name))
			query = buf.String()
		}
		include := true
		if len(tablesToInclude) > 0 {
			include = false
			for _, t := range tablesToInclude {
				if t == table.Name {
					include = true
					break
				}
			}
		}
		if include {
			df.differs[table.Name], err = df.buildTablePlan(table, query)
			if err != nil {
				return err
			}
		}
	}
	if len(tablesToInclude) > 0 && len(tablesToInclude) != len(df.differs) {
		log.Errorf("one or more tables provided are not present in the workflow: %v, %+v", tablesToInclude, df.differs)
		return fmt.Errorf("one or more tables provided are not present in the workflow: %v, %+v", tablesToInclude, df.differs)
	}
	return nil
}

// findPKs identifies PKs, determines any collations to be used for
// them, and removes them from the columns used for data comparison.
func findPKs(env *vtenv.Environment, table *tabletmanagerdatapb.TableDefinition, targetSelect *sqlparser.Select, td *tableDiffer) (sqlparser.OrderBy, error) {
	columnCollations, columnValues, err := getColumnCollations(env, table)
	if err != nil {
		return nil, err
	}
	var orderby sqlparser.OrderBy
	for _, pk := range table.PrimaryKeyColumns {
		found := false
		for i, selExpr := range targetSelect.SelectExprs {
			expr := selExpr.(*sqlparser.AliasedExpr).Expr
			colname := ""
			switch ct := expr.(type) {
			case *sqlparser.ColName:
				colname = ct.Name.String()
			case *sqlparser.FuncExpr: // eg. weight_string()
				// no-op
			default:
				log.Warningf("Not considering column %v for PK, type %v not handled", selExpr, ct)
			}
			if strings.EqualFold(pk, colname) {
				td.compareCols[i].isPK = true
				td.compareCols[i].collation = columnCollations[strings.ToLower(colname)]
				td.compareCols[i].values = columnValues[strings.ToLower(colname)]
				td.comparePKs = append(td.comparePKs, td.compareCols[i])
				td.selectPks = append(td.selectPks, i)
				// We'll be comparing pks separately. So, remove them from compareCols.
				td.pkCols = append(td.pkCols, i)
				found = true
				break
			}
		}
		if !found {
			// Unreachable.
			return nil, fmt.Errorf("column %v not found in table %v", pk, table.Name)
		}
		orderby = append(orderby, &sqlparser.Order{
			Expr:      &sqlparser.ColName{Name: sqlparser.NewIdentifierCI(pk)},
			Direction: sqlparser.AscOrder,
		})
	}
	return orderby, nil
}

// getColumnCollations determines the proper collation to use for each
// column in the table definition leveraging MySQL's collation inheritance
// rules.
func getColumnCollations(venv *vtenv.Environment, table *tabletmanagerdatapb.TableDefinition) (map[string]collations.ID, map[string]*evalengine.EnumSetValues, error) {
	createstmt, err := venv.Parser().Parse(table.Schema)
	if err != nil {
		return nil, nil, err
	}
	createtable, ok := createstmt.(*sqlparser.CreateTable)
	if !ok {
		return nil, nil, vterrors.Wrapf(err, "invalid table schema %s for table %s", table.Schema, table.Name)
	}
	env := schemadiff.NewEnv(venv, venv.CollationEnv().DefaultConnectionCharset())
	tableschema, err := schemadiff.NewCreateTableEntity(env, createtable)
	if err != nil {
		return nil, nil, vterrors.Wrapf(err, "invalid table schema %s for table %s", table.Schema, table.Name)
	}
	tableCharset := tableschema.GetCharset()
	tableCollation := tableschema.GetCollation()
	// If no explicit collation is specified for the column then we need
	// to walk the inheritance tree.
	getColumnCollation := func(column *sqlparser.ColumnDefinition) collations.ID {
		// If there's an explicit collation listed then use that.
		if column.Type.Options.Collate != "" {
			return env.CollationEnv().LookupByName(strings.ToLower(column.Type.Options.Collate))
		}
		// If the column has a charset listed then the default collation
		// for that charset is used.
		if column.Type.Charset.Name != "" {
			return env.CollationEnv().DefaultCollationForCharset(strings.ToLower(column.Type.Charset.Name))
		}
		// If the table has an explicit collation listed then use that.
		if tableCollation != "" {
			return env.CollationEnv().LookupByName(strings.ToLower(tableCollation))
		}
		// If the table has a charset listed then use the default collation
		// for that charset.
		if tableCharset != "" {
			return env.CollationEnv().DefaultCollationForCharset(strings.ToLower(tableCharset))
		}
		// The table is using the global default charset and collation and
		// we inherit that.
		return env.CollationEnv().DefaultConnectionCharset()
	}

	columnCollations := make(map[string]collations.ID)
	columnValues := make(map[string]*evalengine.EnumSetValues)
	for _, column := range tableschema.TableSpec.Columns {
		// If it's not a character based type then no collation is used.
		if !sqltypes.IsQuoted(column.Type.SQLType()) {
			columnCollations[column.Name.Lowered()] = collations.Unknown
			continue
		}
		columnCollations[column.Name.Lowered()] = getColumnCollation(column)
		if len(column.Type.EnumValues) == 0 {
			continue
		}
		columnValues[column.Name.Lowered()] = ptr.Of(evalengine.EnumSetValues(column.Type.EnumValues))
	}
	return columnCollations, columnValues, nil
}

// If SourceTimeZone is defined in the BinlogSource, the VReplication workflow would have converted the datetime
// columns expecting the source to have been in the SourceTimeZone and target in TargetTimeZone. We need to do the reverse
// conversion in VDiff before comparing to the source
func (df *vdiff) adjustForSourceTimeZone(targetSelectExprs sqlparser.SelectExprs, fields map[string]querypb.Type) sqlparser.SelectExprs {
	if df.sourceTimeZone == "" {
		return targetSelectExprs
	}
	log.Infof("source time zone specified: %s", df.sourceTimeZone)
	var newSelectExprs sqlparser.SelectExprs
	var modified bool
	for _, expr := range targetSelectExprs {
		converted := false
		switch selExpr := expr.(type) {
		case *sqlparser.AliasedExpr:
			if colAs, ok := selExpr.Expr.(*sqlparser.ColName); ok {
				var convertTZFuncExpr *sqlparser.FuncExpr
				colName := colAs.Name.Lowered()
				fieldType := fields[colName]
				if fieldType == querypb.Type_DATETIME {
					convertTZFuncExpr = &sqlparser.FuncExpr{
						Name: sqlparser.NewIdentifierCI("convert_tz"),
						Exprs: sqlparser.Exprs{
							colAs,
							sqlparser.NewStrLiteral(df.targetTimeZone),
							sqlparser.NewStrLiteral(df.sourceTimeZone),
						},
					}
					log.Infof("converting datetime column %s using convert_tz()", colName)
					newSelectExprs = append(newSelectExprs, &sqlparser.AliasedExpr{Expr: convertTZFuncExpr, As: colAs.Name})
					converted = true
					modified = true
				}
			}
		}
		if !converted { // not datetime
			newSelectExprs = append(newSelectExprs, expr)
		}
	}
	if modified { // at least one datetime was found
		log.Infof("Found datetime columns when SourceTimeZone was set, resetting target SelectExprs after convert_tz()")
		return newSelectExprs
	}
	return targetSelectExprs
}

func getColumnNameForSelectExpr(selectExpression sqlparser.SelectExpr) (string, error) {
	aliasedExpr := selectExpression.(*sqlparser.AliasedExpr)
	expr := aliasedExpr.Expr
	var colname string
	switch t := expr.(type) {
	case *sqlparser.ColName:
		colname = t.Name.Lowered()
	case *sqlparser.FuncExpr: // only in case datetime was converted using convert_tz()
		colname = aliasedExpr.As.Lowered()
	default:
		return "", fmt.Errorf("found target SelectExpr which was neither ColName or FuncExpr: %+v", aliasedExpr)
	}
	return colname, nil
}

// buildTablePlan builds one tableDiffer.
func (df *vdiff) buildTablePlan(table *tabletmanagerdatapb.TableDefinition, query string) (*tableDiffer, error) {
	statement, err := df.env.Parser().Parse(query)
	if err != nil {
		return nil, err
	}
	sel, ok := statement.(*sqlparser.Select)
	if !ok {
		return nil, fmt.Errorf("unexpected: %v", sqlparser.String(statement))
	}
	td := &tableDiffer{
		targetTable:  table.Name,
		collationEnv: df.env.CollationEnv(),
		parser:       df.env.Parser(),
	}
	sourceSelect := &sqlparser.Select{}
	targetSelect := &sqlparser.Select{}
	// aggregates contains the list if Aggregate functions, if any.
	var aggregates []*engine.AggregateParams
	for _, selExpr := range sel.SelectExprs {
		switch selExpr := selExpr.(type) {
		case *sqlparser.StarExpr:
			// If it's a '*' expression, expand column list from the schema.
			for _, fld := range table.Fields {
				aliased := &sqlparser.AliasedExpr{Expr: &sqlparser.ColName{Name: sqlparser.NewIdentifierCI(fld.Name)}}
				sourceSelect.SelectExprs = append(sourceSelect.SelectExprs, aliased)
				targetSelect.SelectExprs = append(targetSelect.SelectExprs, aliased)
			}
		case *sqlparser.AliasedExpr:
			var targetCol *sqlparser.ColName
			if selExpr.As.IsEmpty() {
				if colAs, ok := selExpr.Expr.(*sqlparser.ColName); ok {
					targetCol = colAs
				} else {
					return nil, fmt.Errorf("expression needs an alias: %v", sqlparser.String(selExpr))
				}
			} else {
				targetCol = &sqlparser.ColName{Name: selExpr.As}
			}
			// If the input was "select a as b", then source will use "a" and target will use "b".
			sourceSelect.SelectExprs = append(sourceSelect.SelectExprs, selExpr)
			targetSelect.SelectExprs = append(targetSelect.SelectExprs, &sqlparser.AliasedExpr{Expr: targetCol})

			// Check if it's an aggregate expression
			if expr, ok := selExpr.Expr.(sqlparser.AggrFunc); ok {
				switch fname := expr.AggrName(); fname {
				case "count", "sum":
					// this will only work as long as aggregates can be pushed down to tablets
					// this won't work: "select count(*) from (select id from t limit 1)"
					// since vreplication only handles simple tables (no joins/derived tables) this is fine for now
					// but will need to be revisited when we add such support to vreplication
					aggregates = append(aggregates, engine.NewAggregateParam(
						/*opcode*/ opcode.AggregateSum,
						/*offset*/ len(sourceSelect.SelectExprs)-1,
						/*alias*/ "", df.env.CollationEnv()))
				}
			}
		default:
			return nil, fmt.Errorf("unexpected: %v", sqlparser.String(statement))
		}
	}

	fields := make(map[string]querypb.Type)
	for _, field := range table.Fields {
		fields[strings.ToLower(field.Name)] = field.Type
	}

	targetSelect.SelectExprs = df.adjustForSourceTimeZone(targetSelect.SelectExprs, fields)
	// Start with adding all columns for comparison.
	td.compareCols = make([]compareColInfo, len(sourceSelect.SelectExprs))
	for i := range td.compareCols {
		td.compareCols[i].colIndex = i
		colname, err := getColumnNameForSelectExpr(targetSelect.SelectExprs[i])
		if err != nil {
			return nil, err
		}
		_, ok = fields[colname]
		if !ok {
			return nil, fmt.Errorf("column %v not found in table %v", colname, table.Name)
		}
	}

	sourceSelect.From = sel.From
	// The target table name should the one that matched the rule.
	// It can be different from the source table.
	targetSelect.From = sqlparser.TableExprs{
		&sqlparser.AliasedTableExpr{
			Expr: &sqlparser.TableName{
				Name: sqlparser.NewIdentifierCS(table.Name),
			},
		},
	}

	orderby, err := findPKs(df.env, table, targetSelect, td)
	if err != nil {
		return nil, err
	}
	// Remove in_keyrange. It's not understood by mysql.
	sourceSelect.Where = removeKeyrange(sel.Where)
	// The source should also perform the group by.
	sourceSelect.GroupBy = sel.GroupBy
	sourceSelect.OrderBy = orderby

	// The target should perform the order by, but not the group by.
	targetSelect.OrderBy = orderby

	td.sourceExpression = sqlparser.String(sourceSelect)
	td.targetExpression = sqlparser.String(targetSelect)

	td.sourcePrimitive = newMergeSorter(df.sources, td.comparePKs, df.env.CollationEnv())
	td.targetPrimitive = newMergeSorter(df.targets, td.comparePKs, df.env.CollationEnv())
	// If there were aggregate expressions, we have to re-aggregate
	// the results, which engine.OrderedAggregate can do.
	if len(aggregates) != 0 {
		td.sourcePrimitive = &engine.OrderedAggregate{
			Aggregates:  aggregates,
			GroupByKeys: pkColsToGroupByParams(td.pkCols, td.collationEnv),
			Input:       td.sourcePrimitive,
		}
	}

	return td, nil
}

func pkColsToGroupByParams(pkCols []int, collationEnv *collations.Environment) []*engine.GroupByParams {
	var res []*engine.GroupByParams
	for _, col := range pkCols {
		res = append(res, &engine.GroupByParams{KeyCol: col, WeightStringCol: -1, Type: evalengine.Type{}, CollationEnv: collationEnv})
	}
	return res
}

// newMergeSorter creates an engine.MergeSort based on the shard streamers and pk columns.
func newMergeSorter(participants map[string]*shardStreamer, comparePKs []compareColInfo, collationEnv *collations.Environment) *engine.MergeSort {
	prims := make([]engine.StreamExecutor, 0, len(participants))
	for _, participant := range participants {
		prims = append(prims, participant)
	}
	ob := make([]evalengine.OrderByParams, 0, len(comparePKs))
	for _, cpk := range comparePKs {
		weightStringCol := -1
		// if the collation is nil or unknown, use binary collation to compare as bytes
		var collation collations.ID = collations.CollationBinaryID
		if cpk.collation != collations.Unknown {
			collation = cpk.collation
		}
		ob = append(ob, evalengine.OrderByParams{Col: cpk.colIndex, WeightStringCol: weightStringCol, Type: evalengine.NewType(sqltypes.Unknown, collation), CollationEnv: collationEnv})
	}
	return &engine.MergeSort{
		Primitives: prims,
		OrderBy:    ob,
	}
}

// selectTablets selects the tablets that will be used for the diff.
func (df *vdiff) selectTablets(ctx context.Context, ts *trafficSwitcher) error {
	var wg sync.WaitGroup
	var err1, err2 error

	// Parallelize all discovery.
	wg.Add(1)
	go func() {
		defer wg.Done()
		err1 = df.forAll(df.sources, func(shard string, source *shardStreamer) error {
			sourceTopo := df.ts.TopoServer()
			if ts.ExternalTopo() != nil {
				sourceTopo = ts.ExternalTopo()
			}
			tp, err := discovery.NewTabletPicker(ctx, sourceTopo, []string{df.sourceCell}, df.sourceCell,
				df.ts.SourceKeyspaceName(), shard, df.tabletTypesStr, discovery.TabletPickerOptions{})
			if err != nil {
				return err
			}

			tablet, err := tp.PickForStreaming(ctx)
			if err != nil {
				return err
			}
			source.tablet = tablet
			return nil
		})
	}()

	wg.Add(1)
	go func() {
		defer wg.Done()
		includeNonServingTablets := false
		if df.ts.workflowType == binlogdatapb.VReplicationWorkflowType_Reshard {
			// For resharding, the target shards could be non-serving if traffic has already been switched once.
			// When shards are created their IsPrimaryServing attribute is set to true. However, when the traffic is switched
			// it is set to false for the shards we are switching from. We don't have a way to know if we have
			// switched or not, so we just include non-serving tablets for all reshards.
			includeNonServingTablets = true
		}
		err2 = df.forAll(df.targets, func(shard string, target *shardStreamer) error {
			tp, err := discovery.NewTabletPicker(ctx, df.ts.TopoServer(), []string{df.targetCell}, df.targetCell,
				df.ts.TargetKeyspaceName(), shard, df.tabletTypesStr,
				discovery.TabletPickerOptions{IncludeNonServingTablets: includeNonServingTablets})
			if err != nil {
				return err
			}

			tablet, err := tp.PickForStreaming(ctx)
			if err != nil {
				return err
			}
			target.tablet = tablet
			return nil
		})
	}()

	wg.Wait()
	if err1 != nil {
		return err1
	}
	return err2
}

// stopTargets stops all the targets and records their source positions.
func (df *vdiff) stopTargets(ctx context.Context) error {
	var mu sync.Mutex

	err := df.forAll(df.targets, func(shard string, target *shardStreamer) error {
		query := fmt.Sprintf("update _vt.vreplication set state='Stopped', message='for vdiff' where db_name=%s and workflow=%s", encodeString(target.primary.DbName()), encodeString(df.ts.WorkflowName()))
		_, err := df.ts.TabletManagerClient().VReplicationExec(ctx, target.primary.Tablet, query)
		if err != nil {
			return err
		}
		query = fmt.Sprintf("select source, pos from _vt.vreplication where db_name=%s and workflow=%s", encodeString(target.primary.DbName()), encodeString(df.ts.WorkflowName()))
		p3qr, err := df.ts.TabletManagerClient().VReplicationExec(ctx, target.primary.Tablet, query)
		if err != nil {
			return err
		}
		qr := sqltypes.Proto3ToResult(p3qr)

		for _, row := range qr.Rows {
			var bls binlogdatapb.BinlogSource
			rowBytes, err := row[0].ToBytes()
			if err != nil {
				return err
			}
			if err := prototext.Unmarshal(rowBytes, &bls); err != nil {
				return err
			}
			pos, err := binlogplayer.DecodePosition(row[1].ToString())
			if err != nil {
				return err
			}
			func() {
				mu.Lock()
				defer mu.Unlock()

				source, ok := df.sources[bls.Shard]
				if !ok {
					// Unreachable.
					return
				}
				if !source.position.IsZero() && source.position.AtLeast(pos) {
					return
				}
				source.position = pos
			}()
		}
		return nil
	})
	if err != nil {
		return err
	}
	return nil
}

// starQueryStreams makes sure the sources are past the target's positions, starts the query streams,
// and records the snapshot position of the query. It creates a result channel which StreamExecute
// will use to serve rows.
func (df *vdiff) startQueryStreams(ctx context.Context, keyspace string, participants map[string]*shardStreamer, query string, filteredReplicationWaitTime time.Duration) error {
	waitCtx, cancel := context.WithTimeout(ctx, filteredReplicationWaitTime)
	defer cancel()
	return df.forAll(participants, func(shard string, participant *shardStreamer) error {
		// Iteration for each participant.
		if participant.position.IsZero() {
			return fmt.Errorf("workflow %s.%s: stream has not started on tablet %s", df.targetKeyspace, df.workflow, participant.primary.Alias.String())
		}
		log.Infof("WaitForPosition: tablet %s should reach position %s", participant.tablet.Alias.String(), replication.EncodePosition(participant.position))
		if err := df.ts.TabletManagerClient().WaitForPosition(waitCtx, participant.tablet, replication.EncodePosition(participant.position)); err != nil {
			log.Errorf("WaitForPosition error: %s", err)
			return vterrors.Wrapf(err, "WaitForPosition for tablet %v", topoproto.TabletAliasString(participant.tablet.Alias))
		}
		participant.result = make(chan *sqltypes.Result, 1)
		gtidch := make(chan string, 1)

		// Start the stream in a separate goroutine.
		go df.streamOne(ctx, keyspace, shard, participant, query, gtidch)

		// Wait for the gtid to be sent. If it's not received, there was an error
		// which would be stored in participant.err.
		gtid, ok := <-gtidch
		if !ok {
			return participant.err
		}
		// Save the new position, as of when the query executed.
		participant.snapshotPosition = gtid
		return nil
	})
}

// streamOne is called as a goroutine, and communicates its results through channels.
// It first sends the snapshot gtid to gtidch.
// Then it streams results to participant.result.
// Before returning, it sets participant.err, and closes all channels.
// If any channel is closed, then participant.err can be checked if there was an error.
// The shardStreamer's StreamExecute consumes the result channel.
func (df *vdiff) streamOne(ctx context.Context, keyspace, shard string, participant *shardStreamer, query string, gtidch chan string) {
	defer close(participant.result)
	defer close(gtidch)

	// Wrap the streaming in a separate function so we can capture the error.
	// This shows that the error will be set before the channels are closed.
	participant.err = func() error {
		conn, err := tabletconn.GetDialer()(ctx, participant.tablet, grpcclient.FailFast(false))
		if err != nil {
			return err
		}
		defer conn.Close(ctx)

		target := &querypb.Target{
			Keyspace:   keyspace,
			Shard:      shard,
			TabletType: participant.tablet.Type,
		}
		var fields []*querypb.Field
		return conn.VStreamResults(ctx, target, query, func(vrs *binlogdatapb.VStreamResultsResponse) error {
			if vrs.Fields != nil {
				fields = vrs.Fields
				gtidch <- vrs.Gtid
			}
			p3qr := &querypb.QueryResult{
				Fields: fields,
				Rows:   vrs.Rows,
			}
			result := sqltypes.Proto3ToResult(p3qr)
			// Fields should be received only once, and sent only once.
			if vrs.Fields == nil {
				result.Fields = nil
			}
			select {
			case participant.result <- result:
			case <-ctx.Done():
				return vterrors.Wrap(ctx.Err(), "VStreamResults")
			}
			return nil
		})
	}()
}

// syncTargets fast-forwards the vreplication to the source snapshot positions
// and waits for the selected tablets to catch up to that point.
func (df *vdiff) syncTargets(ctx context.Context, filteredReplicationWaitTime time.Duration) error {
	waitCtx, cancel := context.WithTimeout(ctx, filteredReplicationWaitTime)
	defer cancel()
	err := df.ts.ForAllUIDs(func(target *workflow.MigrationTarget, uid int32) error {
		bls := target.Sources[uid]
		pos := df.sources[bls.Shard].snapshotPosition
		query := fmt.Sprintf("update _vt.vreplication set state='Running', stop_pos='%s', message='synchronizing for vdiff' where id=%d", pos, uid)
		if _, err := df.ts.TabletManagerClient().VReplicationExec(ctx, target.GetPrimary().Tablet, query); err != nil {
			return err
		}
		if err := df.ts.TabletManagerClient().VReplicationWaitForPos(waitCtx, target.GetPrimary().Tablet, uid, pos); err != nil {
			return vterrors.Wrapf(err, "VReplicationWaitForPos for tablet %v", topoproto.TabletAliasString(target.GetPrimary().Tablet.Alias))
		}
		return nil
	})
	if err != nil {
		return err
	}

	err = df.forAll(df.targets, func(shard string, target *shardStreamer) error {
		pos, err := df.ts.TabletManagerClient().PrimaryPosition(ctx, target.primary.Tablet)
		if err != nil {
			return err
		}
		mpos, err := binlogplayer.DecodePosition(pos)
		if err != nil {
			return err
		}
		target.position = mpos
		return nil
	})
	return err
}

// restartTargets restarts the stopped target vreplication streams.
func (df *vdiff) restartTargets(ctx context.Context) error {
	return df.forAll(df.targets, func(shard string, target *shardStreamer) error {
		query := fmt.Sprintf("update _vt.vreplication set state='Running', message='', stop_pos='' where db_name=%s and workflow=%s",
			encodeString(target.primary.DbName()), encodeString(df.ts.WorkflowName()))
		log.Infof("Restarting the %q VReplication workflow on %q using %q", df.ts.WorkflowName(), target.primary.Alias, query)
		var err error
		// Let's retry a few times if we get a retryable error.
		for i := 1; i <= 3; i++ {
			_, err = df.ts.TabletManagerClient().VReplicationExec(ctx, target.primary.Tablet, query)
			if err == nil || !sqlerror.IsEphemeralError(err) {
				break
			}
			log.Warningf("Encountered the following error while restarting the %q VReplication workflow on %q, will retry (attempt #%d): %v",
				df.ts.WorkflowName(), target.primary.Alias, i, err)
		}
		return err
	})
}

func (df *vdiff) forAll(participants map[string]*shardStreamer, f func(string, *shardStreamer) error) error {
	var wg sync.WaitGroup
	allErrors := &concurrency.AllErrorRecorder{}
	for shard, participant := range participants {
		wg.Add(1)
		go func(shard string, participant *shardStreamer) {
			defer wg.Done()

			if err := f(shard, participant); err != nil {
				allErrors.RecordError(err)
			}
		}(shard, participant)
	}
	wg.Wait()
	return allErrors.AggrError(vterrors.Aggregate)
}

// -----------------------------------------------------------------
// primitiveExecutor

// primitiveExecutor starts execution on the top level primitive
// and provides convenience functions for row-by-row iteration.
type primitiveExecutor struct {
	prim     engine.Primitive
	rows     [][]sqltypes.Value
	resultch chan *sqltypes.Result
	err      error
}

func newPrimitiveExecutor(ctx context.Context, prim engine.Primitive) *primitiveExecutor {
	pe := &primitiveExecutor{
		prim:     prim,
		resultch: make(chan *sqltypes.Result, 1),
	}
	vcursor := &contextVCursor{}
	go func() {
		defer close(pe.resultch)
		pe.err = vcursor.StreamExecutePrimitive(ctx, pe.prim, make(map[string]*querypb.BindVariable), true, func(qr *sqltypes.Result) error {
			select {
			case pe.resultch <- qr:
			case <-ctx.Done():
				return vterrors.Wrap(ctx.Err(), "LHS Stream")
			}
			return nil
		})
	}()
	return pe
}

func (pe *primitiveExecutor) next() ([]sqltypes.Value, error) {
	for len(pe.rows) == 0 {
		qr, ok := <-pe.resultch
		if !ok {
			return nil, pe.err
		}
		pe.rows = qr.Rows
	}

	row := pe.rows[0]
	pe.rows = pe.rows[1:]
	return row, nil
}

func (pe *primitiveExecutor) drain(ctx context.Context) (int, error) {
	count := 0
	for {
		row, err := pe.next()
		if err != nil {
			return 0, err
		}
		if row == nil {
			return count, nil
		}
		count++
	}
}

// -----------------------------------------------------------------
// shardStreamer

func (sm *shardStreamer) StreamExecute(ctx context.Context, vcursor engine.VCursor, bindVars map[string]*querypb.BindVariable, wantfields bool, callback func(*sqltypes.Result) error) error {
	for result := range sm.result {
		if err := callback(result); err != nil {
			return err
		}
	}
	return sm.err
}

// humanInt formats large integers to a value easier to the eye: 100000=100k 1e12=1b 234000000=234m ...
func humanInt(n int64) string { // nolint
	var val float64
	var unit string
	switch true {
	case n < 1000:
		val = float64(n)
	case n < 1e6:
		val = float64(n) / 1000
		unit = "k"
	case n < 1e9:
		val = float64(n) / 1e6
		unit = "m"
	default:
		val = float64(n) / 1e9
		unit = "b"
	}
	s := fmt.Sprintf("%0.3f", val)
	s = strings.Replace(s, ".000", "", -1)

	return fmt.Sprintf("%s%s", s, unit)
}

// -----------------------------------------------------------------
// tableDiffer

func (td *tableDiffer) diff(ctx context.Context, rowsToCompare *int64, debug, onlyPks bool, maxExtraRowsToCompare int) (*DiffReport, error) {
	sourceExecutor := newPrimitiveExecutor(ctx, td.sourcePrimitive)
	targetExecutor := newPrimitiveExecutor(ctx, td.targetPrimitive)
	dr := &DiffReport{}
	var sourceRow, targetRow []sqltypes.Value
	var err error
	advanceSource := true
	advanceTarget := true
	for {
		if dr.ProcessedRows%1e7 == 0 { // log progress every 10 million rows
			log.Infof("VDiff progress:: table %s: %s rows", td.targetTable, humanInt(int64(dr.ProcessedRows)))
		}
		*rowsToCompare--
		if *rowsToCompare < 0 {
			log.Infof("Stopping vdiff, specified limit reached")
			return dr, nil
		}
		if advanceSource {
			sourceRow, err = sourceExecutor.next()
			if err != nil {
				return nil, err
			}
		}
		if advanceTarget {
			targetRow, err = targetExecutor.next()
			if err != nil {
				return nil, err
			}
		}

		if sourceRow == nil && targetRow == nil {
			return dr, nil
		}

		advanceSource = true
		advanceTarget = true

		if sourceRow == nil {
			diffRow, err := td.genRowDiff(td.sourceExpression, targetRow, debug, onlyPks)
			if err != nil {
				return nil, vterrors.Wrap(err, "unexpected error generating diff")
			}
			dr.ExtraRowsTargetDiffs = append(dr.ExtraRowsTargetDiffs, diffRow)

			// drain target, update count
			count, err := targetExecutor.drain(ctx)
			if err != nil {
				return nil, err
			}
			dr.ExtraRowsTarget += 1 + count
			dr.ProcessedRows += 1 + count
			return dr, nil
		}
		if targetRow == nil {
			// no more rows from the target
			// we know we have rows from source, drain, update count
			diffRow, err := td.genRowDiff(td.sourceExpression, sourceRow, debug, onlyPks)
			if err != nil {
				return nil, vterrors.Wrap(err, "unexpected error generating diff")
			}
			dr.ExtraRowsSourceDiffs = append(dr.ExtraRowsSourceDiffs, diffRow)

			count, err := sourceExecutor.drain(ctx)
			if err != nil {
				return nil, err
			}
			dr.ExtraRowsSource += 1 + count
			dr.ProcessedRows += 1 + count
			return dr, nil
		}

		dr.ProcessedRows++

		// Compare pk values.
		c, err := td.compare(sourceRow, targetRow, td.comparePKs, false)
		switch {
		case err != nil:
			return nil, err
		case c < 0:
			if dr.ExtraRowsSource < maxExtraRowsToCompare {
				diffRow, err := td.genRowDiff(td.sourceExpression, sourceRow, debug, onlyPks)
				if err != nil {
					return nil, vterrors.Wrap(err, "unexpected error generating diff")
				}
				dr.ExtraRowsSourceDiffs = append(dr.ExtraRowsSourceDiffs, diffRow)
			}
			dr.ExtraRowsSource++
			advanceTarget = false
			continue
		case c > 0:
			if dr.ExtraRowsTarget < maxExtraRowsToCompare {
				diffRow, err := td.genRowDiff(td.targetExpression, targetRow, debug, onlyPks)
				if err != nil {
					return nil, vterrors.Wrap(err, "unexpected error generating diff")
				}
				dr.ExtraRowsTargetDiffs = append(dr.ExtraRowsTargetDiffs, diffRow)
			}
			dr.ExtraRowsTarget++
			advanceSource = false
			continue
		}

		// c == 0
		// Compare the non-pk values.
		c, err = td.compare(sourceRow, targetRow, td.compareCols, true)
		switch {
		case err != nil:
			return nil, err
		case c != 0:
			// We don't do a second pass to compare mismatched rows so we can cap the slice here
			if dr.MismatchedRows < maxVDiffReportSampleRows {
				sourceDiffRow, err := td.genRowDiff(td.targetExpression, sourceRow, debug, onlyPks)
				if err != nil {
					return nil, vterrors.Wrap(err, "unexpected error generating diff")
				}
				targetDiffRow, err := td.genRowDiff(td.targetExpression, targetRow, debug, onlyPks)
				if err != nil {
					return nil, vterrors.Wrap(err, "unexpected error generating diff")
				}
				dr.MismatchedRowsSample = append(dr.MismatchedRowsSample, &DiffMismatch{Source: sourceDiffRow, Target: targetDiffRow})
			}
			dr.MismatchedRows++
		default:
			dr.MatchingRows++
		}
	}
}

func (td *tableDiffer) compare(sourceRow, targetRow []sqltypes.Value, cols []compareColInfo, compareOnlyNonPKs bool) (int, error) {
	for _, col := range cols {
		if col.isPK && compareOnlyNonPKs {
			continue
		}
		compareIndex := col.colIndex
		var c int
		var err error
		var collationID collations.ID
		// if the collation is nil or unknown, use binary collation to compare as bytes
		collationID = col.collation
		if col.collation == collations.Unknown {
			collationID = collations.CollationBinaryID
		}
		c, err = evalengine.NullsafeCompare(sourceRow[compareIndex], targetRow[compareIndex], td.collationEnv, collationID, col.values)
		if err != nil {
			return 0, err
		}
		if c != 0 {
			return c, nil
		}
	}
	return 0, nil
}

func (td *tableDiffer) genRowDiff(queryStmt string, row []sqltypes.Value, debug, onlyPks bool) (*RowDiff, error) {
	drp := &RowDiff{}
	drp.Row = make(map[string]sqltypes.Value)
	statement, err := td.parser.Parse(queryStmt)
	if err != nil {
		return nil, err
	}
	sel, ok := statement.(*sqlparser.Select)
	if !ok {
		return nil, fmt.Errorf("unexpected: %v", sqlparser.String(statement))
	}

	if debug {
		drp.Query = td.genDebugQueryDiff(sel, row, onlyPks)
	}

	if onlyPks {
		for _, pkI := range td.selectPks {
			buf := sqlparser.NewTrackedBuffer(nil)
			sel.SelectExprs[pkI].Format(buf)
			col := buf.String()
			drp.Row[col] = row[pkI]
		}
		return drp, nil
	}

	for i := range sel.SelectExprs {
		buf := sqlparser.NewTrackedBuffer(nil)
		sel.SelectExprs[i].Format(buf)
		col := buf.String()
		drp.Row[col] = row[i]
	}

	return drp, nil
}

func (td *tableDiffer) genDebugQueryDiff(sel *sqlparser.Select, row []sqltypes.Value, onlyPks bool) string {
	buf := sqlparser.NewTrackedBuffer(nil)
	buf.Myprintf("select ")

	if onlyPks {
		for i, pkI := range td.selectPks {
			pk := sel.SelectExprs[pkI]
			pk.Format(buf)
			if i != len(td.selectPks)-1 {
				buf.Myprintf(", ")
			}
		}
	} else {
		sel.SelectExprs.Format(buf)
	}
	buf.Myprintf(" from ")
	buf.Myprintf(sqlparser.ToString(sel.From))
	buf.Myprintf(" where ")
	for i, pkI := range td.selectPks {
		sel.SelectExprs[pkI].Format(buf)
		buf.Myprintf("=")
		row[pkI].EncodeSQL(buf)
		if i != len(td.selectPks)-1 {
			buf.Myprintf(" AND ")
		}
	}
	buf.Myprintf(";")
	return buf.String()
}

// -----------------------------------------------------------------
// contextVCursor

// contextVCursor satisfies VCursor interface
type contextVCursor struct {
	engine.VCursor
}

func (vc *contextVCursor) ConnCollation() collations.ID {
	return collations.CollationBinaryID
}

func (vc *contextVCursor) ExecutePrimitive(ctx context.Context, primitive engine.Primitive, bindVars map[string]*querypb.BindVariable, wantfields bool) (*sqltypes.Result, error) {
	return primitive.TryExecute(ctx, vc, bindVars, wantfields)
}

func (vc *contextVCursor) StreamExecutePrimitive(ctx context.Context, primitive engine.Primitive, bindVars map[string]*querypb.BindVariable, wantfields bool, callback func(*sqltypes.Result) error) error {
	return primitive.TryStreamExecute(ctx, vc, bindVars, wantfields, callback)
}

// -----------------------------------------------------------------
// Utility functions

func removeKeyrange(where *sqlparser.Where) *sqlparser.Where {
	if where == nil {
		return nil
	}
	if isFuncKeyrange(where.Expr) {
		return nil
	}
	where.Expr = removeExprKeyrange(where.Expr)
	return where
}

func removeExprKeyrange(node sqlparser.Expr) sqlparser.Expr {
	switch node := node.(type) {
	case *sqlparser.AndExpr:
		if isFuncKeyrange(node.Left) {
			return removeExprKeyrange(node.Right)
		}
		if isFuncKeyrange(node.Right) {
			return removeExprKeyrange(node.Left)
		}
		return &sqlparser.AndExpr{
			Left:  removeExprKeyrange(node.Left),
			Right: removeExprKeyrange(node.Right),
		}
	}
	return node
}

func isFuncKeyrange(expr sqlparser.Expr) bool {
	funcExpr, ok := expr.(*sqlparser.FuncExpr)
	return ok && funcExpr.Name.EqualString("in_keyrange")
}

func formatSampleRow(logger logutil.Logger, rd *RowDiff, debug bool) {
	keys := make([]string, 0, len(rd.Row))
	for k := range rd.Row {
		keys = append(keys, k)
	}

	sort.Strings(keys)

	for _, k := range keys {
		logger.Printf("\t\t\t %s: %s\n", k, formatValue(rd.Row[k]))
	}

	if debug {
		logger.Printf("\t\tDebugQuery: %v\n", rd.Query)
	}
}

func formatValue(val sqltypes.Value) string {
	if val.Type() == sqltypes.Null {
		return "null (NULL_TYPE)"
	}
	if val.IsQuoted() || val.Type() == sqltypes.Bit {
		if len(val.Raw()) >= 20 {
			rawBytes := val.Raw()[:20]
			rawBytes = append(rawBytes, []byte("...[TRUNCATED]")...)
			return fmt.Sprintf("%q (%v)", rawBytes, val.Type())
		}
		return fmt.Sprintf("%q (%v)", val.Raw(), val.Type())
	}
	return fmt.Sprintf("%s (%v)", val.Raw(), val.Type())
}