// copy phase:
// - copy the data from source tablets to destination masters (wtih replication on)
// Assumes that the schema has already been created on each destination tablet
// (probably from vtctl's CopySchemaShard)
func (scw *SplitCloneWorker) copy() error {
scw.setState(stateSCCopy)
// get source schema from the first shard
// TODO(alainjobart): for now, we assume the schema is compatible
// on all source shards. Furthermore, we estimate the number of rows
// in each source shard for each table to be about the same
// (rowCount is used to estimate an ETA)
sourceSchemaDefinition, err := scw.wr.GetSchema(scw.sourceAliases[0], nil, scw.excludeTables, true)
if err != nil {
return fmt.Errorf("cannot get schema from source %v: %v", scw.sourceAliases[0], err)
}
if len(sourceSchemaDefinition.TableDefinitions) == 0 {
return fmt.Errorf("no tables matching the table filter in tablet %v", scw.sourceAliases[0])
}
scw.wr.Logger().Infof("Source tablet 0 has %v tables to copy", len(sourceSchemaDefinition.TableDefinitions))
scw.mu.Lock()
scw.tableStatus = make([]*tableStatus, len(sourceSchemaDefinition.TableDefinitions))
for i, td := range sourceSchemaDefinition.TableDefinitions {
scw.tableStatus[i] = &tableStatus{
name: td.Name,
rowCount: td.RowCount * uint64(len(scw.sourceAliases)),
}
}
scw.startTime = time.Now()
scw.mu.Unlock()
// Find the column index for the sharding columns in all the databases, and count rows
columnIndexes := make([]int, len(sourceSchemaDefinition.TableDefinitions))
for tableIndex, td := range sourceSchemaDefinition.TableDefinitions {
if td.Type == myproto.TABLE_BASE_TABLE {
// find the column to split on
columnIndexes[tableIndex] = -1
for i, name := range td.Columns {
if name == scw.keyspaceInfo.ShardingColumnName {
columnIndexes[tableIndex] = i
break
}
}
if columnIndexes[tableIndex] == -1 {
return fmt.Errorf("table %v doesn't have a column named '%v'", td.Name, scw.keyspaceInfo.ShardingColumnName)
}
scw.tableStatus[tableIndex].mu.Lock()
scw.tableStatus[tableIndex].rowCount = td.RowCount
scw.tableStatus[tableIndex].mu.Unlock()
} else {
scw.tableStatus[tableIndex].mu.Lock()
scw.tableStatus[tableIndex].isView = true
scw.tableStatus[tableIndex].mu.Unlock()
}
}
// In parallel, setup the channels to send SQL data chunks to for each destination tablet:
//
// mu protects the abort channel for closing, and firstError
mu := sync.Mutex{}
abort := make(chan struct{})
var firstError error
processError := func(format string, args ...interface{}) {
scw.wr.Logger().Errorf(format, args...)
mu.Lock()
if abort != nil {
close(abort)
abort = nil
firstError = fmt.Errorf(format, args...)
}
mu.Unlock()
}
// since we're writing only to masters, we need to enable bin logs so that replication happens
disableBinLogs := false
insertChannels := make([][]chan string, len(scw.destinationShards))
destinationWaitGroup := sync.WaitGroup{}
for shardIndex, _ := range scw.destinationShards {
insertChannels[shardIndex] = make([]chan string, len(scw.destinationAliases[shardIndex]))
for i, tabletAlias := range scw.destinationAliases[shardIndex] {
// we create one channel per destination tablet. It
// is sized to have a buffer of a maximum of
// destinationWriterCount * 2 items, to hopefully
// always have data. We then have
// destinationWriterCount go routines reading from it.
insertChannels[shardIndex][i] = make(chan string, scw.destinationWriterCount*2)
go func(ti *topo.TabletInfo, insertChannel chan string) {
for j := 0; j < scw.destinationWriterCount; j++ {
destinationWaitGroup.Add(1)
go func() {
defer destinationWaitGroup.Done()
if err := executeFetchLoop(scw.wr, ti, insertChannel, abort, disableBinLogs); err != nil {
processError("executeFetchLoop failed: %v", err)
}
}()
}
}(scw.destinationTablets[shardIndex][tabletAlias], insertChannels[shardIndex][i])
}
}
// Now for each table, read data chunks and send them to all
// insertChannels
sourceWaitGroup := sync.WaitGroup{}
for shardIndex, _ := range scw.sourceShards {
sema := sync2.NewSemaphore(scw.sourceReaderCount, 0)
for tableIndex, td := range sourceSchemaDefinition.TableDefinitions {
if td.Type == myproto.TABLE_VIEW {
continue
}
rowSplitter := NewRowSplitter(scw.destinationShards, scw.keyspaceInfo.ShardingColumnType, columnIndexes[tableIndex])
chunks, err := findChunks(scw.wr, scw.sourceTablets[shardIndex], td, scw.minTableSizeForSplit, scw.sourceReaderCount)
if err != nil {
return err
}
scw.tableStatus[tableIndex].setThreadCount(len(chunks) - 1)
for chunkIndex := 0; chunkIndex < len(chunks)-1; chunkIndex++ {
sourceWaitGroup.Add(1)
go func(td *myproto.TableDefinition, tableIndex, chunkIndex int) {
defer sourceWaitGroup.Done()
sema.Acquire()
defer sema.Release()
scw.tableStatus[tableIndex].threadStarted()
// build the query, and start the streaming
selectSQL := buildSQLFromChunks(scw.wr, td, chunks, chunkIndex, scw.sourceAliases[shardIndex].String())
qrr, err := NewQueryResultReaderForTablet(scw.wr.TopoServer(), scw.sourceAliases[shardIndex], selectSQL)
if err != nil {
processError("NewQueryResultReaderForTablet failed: %v", err)
return
}
defer qrr.Close()
// process the data
if err := scw.processData(td, tableIndex, qrr, rowSplitter, insertChannels, scw.destinationPackCount, abort); err != nil {
processError("processData failed: %v", err)
}
scw.tableStatus[tableIndex].threadDone()
}(td, tableIndex, chunkIndex)
}
}
}
sourceWaitGroup.Wait()
for shardIndex, _ := range scw.destinationShards {
for _, c := range insertChannels[shardIndex] {
close(c)
}
}
destinationWaitGroup.Wait()
if firstError != nil {
return firstError
}
// then create and populate the blp_checkpoint table
if scw.strategy.PopulateBlpCheckpoint {
queries := make([]string, 0, 4)
queries = append(queries, binlogplayer.CreateBlpCheckpoint()...)
flags := ""
if scw.strategy.DontStartBinlogPlayer {
flags = binlogplayer.BLP_FLAG_DONT_START
}
// get the current position from the sources
for shardIndex, _ := range scw.sourceShards {
ctx, cancel := context.WithTimeout(context.TODO(), 30*time.Second)
status, err := scw.wr.TabletManagerClient().SlaveStatus(ctx, scw.sourceTablets[shardIndex])
if err != nil {
return err
}
cancel()
queries = append(queries, binlogplayer.PopulateBlpCheckpoint(0, status.Position, time.Now().Unix(), flags))
}
for shardIndex, _ := range scw.destinationShards {
for _, tabletAlias := range scw.destinationAliases[shardIndex] {
destinationWaitGroup.Add(1)
go func(ti *topo.TabletInfo) {
defer destinationWaitGroup.Done()
scw.wr.Logger().Infof("Making and populating blp_checkpoint table on tablet %v", ti.Alias)
if err := runSqlCommands(scw.wr, ti, queries, abort, disableBinLogs); err != nil {
processError("blp_checkpoint queries failed on tablet %v: %v", ti.Alias, err)
}
}(scw.destinationTablets[shardIndex][tabletAlias])
}
}
destinationWaitGroup.Wait()
if firstError != nil {
return firstError
}
}
// Now we're done with data copy, update the shard's source info.
// TODO(alainjobart) this is a superset, some shards may not
// overlap, have to deal with this better (for N -> M splits
// where both N>1 and M>1)
if scw.strategy.SkipSetSourceShards {
scw.wr.Logger().Infof("Skipping setting SourceShard on destination shards.")
} else {
for _, si := range scw.destinationShards {
scw.wr.Logger().Infof("Setting SourceShard on shard %v/%v", si.Keyspace(), si.ShardName())
if err := scw.wr.SetSourceShards(si.Keyspace(), si.ShardName(), scw.sourceAliases, nil); err != nil {
return fmt.Errorf("Failed to set source shards: %v", err)
}
}
}
// And force a schema reload on all destination tablets.
// The master tablet will end up starting filtered replication
// at this point.
for shardIndex, _ := range scw.destinationShards {
for _, tabletAlias := range scw.reloadAliases[shardIndex] {
destinationWaitGroup.Add(1)
go func(ti *topo.TabletInfo) {
defer destinationWaitGroup.Done()
scw.wr.Logger().Infof("Reloading schema on tablet %v", ti.Alias)
ctx, cancel := context.WithTimeout(context.TODO(), 30*time.Second)
if err := scw.wr.TabletManagerClient().ReloadSchema(ctx, ti); err != nil {
processError("ReloadSchema failed on tablet %v: %v", ti.Alias, err)
}
cancel()
}(scw.reloadTablets[shardIndex][tabletAlias])
}
}
destinationWaitGroup.Wait()
return firstError
}
// copy phase:
// - copy the data from source tablets to destination masters (wtih replication on)
// Assumes that the schema has already been created on each destination tablet
// (probably from vtctl's CopySchemaShard)
func (vscw *VerticalSplitCloneWorker) copy() error {
vscw.setState(stateVSCCopy)
// get source schema
sourceSchemaDefinition, err := vscw.wr.GetSchema(vscw.sourceAlias, vscw.tables, nil, true)
if err != nil {
return fmt.Errorf("cannot get schema from source %v: %v", vscw.sourceAlias, err)
}
if len(sourceSchemaDefinition.TableDefinitions) == 0 {
return fmt.Errorf("no tables matching the table filter")
}
vscw.wr.Logger().Infof("Source tablet has %v tables to copy", len(sourceSchemaDefinition.TableDefinitions))
vscw.mu.Lock()
vscw.tableStatus = make([]*tableStatus, len(sourceSchemaDefinition.TableDefinitions))
for i, td := range sourceSchemaDefinition.TableDefinitions {
vscw.tableStatus[i] = &tableStatus{
name: td.Name,
rowCount: td.RowCount,
}
}
vscw.startTime = time.Now()
vscw.mu.Unlock()
// Count rows
for i, td := range sourceSchemaDefinition.TableDefinitions {
vscw.tableStatus[i].mu.Lock()
if td.Type == myproto.TABLE_BASE_TABLE {
vscw.tableStatus[i].rowCount = td.RowCount
} else {
vscw.tableStatus[i].isView = true
}
vscw.tableStatus[i].mu.Unlock()
}
// In parallel, setup the channels to send SQL data chunks to for each destination tablet.
//
// mu protects the abort channel for closing, and firstError
mu := sync.Mutex{}
abort := make(chan struct{})
var firstError error
processError := func(format string, args ...interface{}) {
vscw.wr.Logger().Errorf(format, args...)
mu.Lock()
if abort != nil {
close(abort)
abort = nil
firstError = fmt.Errorf(format, args...)
}
mu.Unlock()
}
// since we're writing only to masters, we need to enable bin logs so that replication happens
disableBinLogs := false
insertChannels := make([]chan string, len(vscw.destinationAliases))
destinationWaitGroup := sync.WaitGroup{}
for i, tabletAlias := range vscw.destinationAliases {
// we create one channel per destination tablet. It
// is sized to have a buffer of a maximum of
// destinationWriterCount * 2 items, to hopefully
// always have data. We then have
// destinationWriterCount go routines reading from it.
insertChannels[i] = make(chan string, vscw.destinationWriterCount*2)
go func(ti *topo.TabletInfo, insertChannel chan string) {
for j := 0; j < vscw.destinationWriterCount; j++ {
destinationWaitGroup.Add(1)
go func() {
defer destinationWaitGroup.Done()
if err := executeFetchLoop(vscw.wr, ti, insertChannel, abort, disableBinLogs); err != nil {
processError("executeFetchLoop failed: %v", err)
}
}()
}
}(vscw.destinationTablets[tabletAlias], insertChannels[i])
}
// Now for each table, read data chunks and send them to all
// insertChannels
sourceWaitGroup := sync.WaitGroup{}
sema := sync2.NewSemaphore(vscw.sourceReaderCount, 0)
for tableIndex, td := range sourceSchemaDefinition.TableDefinitions {
if td.Type == myproto.TABLE_VIEW {
continue
}
chunks, err := findChunks(vscw.wr, vscw.sourceTablet, td, vscw.minTableSizeForSplit, vscw.sourceReaderCount)
if err != nil {
return err
}
vscw.tableStatus[tableIndex].setThreadCount(len(chunks) - 1)
for chunkIndex := 0; chunkIndex < len(chunks)-1; chunkIndex++ {
sourceWaitGroup.Add(1)
go func(td *myproto.TableDefinition, tableIndex, chunkIndex int) {
defer sourceWaitGroup.Done()
sema.Acquire()
defer sema.Release()
vscw.tableStatus[tableIndex].threadStarted()
// build the query, and start the streaming
selectSQL := buildSQLFromChunks(vscw.wr, td, chunks, chunkIndex, vscw.sourceAlias.String())
qrr, err := NewQueryResultReaderForTablet(vscw.wr.TopoServer(), vscw.sourceAlias, selectSQL)
if err != nil {
processError("NewQueryResultReaderForTablet failed: %v", err)
return
}
defer qrr.Close()
// process the data
if err := vscw.processData(td, tableIndex, qrr, insertChannels, vscw.destinationPackCount, abort); err != nil {
processError("QueryResultReader failed: %v", err)
}
vscw.tableStatus[tableIndex].threadDone()
}(td, tableIndex, chunkIndex)
}
}
sourceWaitGroup.Wait()
for _, c := range insertChannels {
close(c)
}
destinationWaitGroup.Wait()
if firstError != nil {
return firstError
}
// then create and populate the blp_checkpoint table
if vscw.strategy.PopulateBlpCheckpoint {
// get the current position from the source
ctx, cancel := context.WithTimeout(context.TODO(), 30*time.Second)
status, err := vscw.wr.TabletManagerClient().SlaveStatus(ctx, vscw.sourceTablet)
if err != nil {
return err
}
cancel()
queries := make([]string, 0, 4)
queries = append(queries, binlogplayer.CreateBlpCheckpoint()...)
flags := ""
if vscw.strategy.DontStartBinlogPlayer {
flags = binlogplayer.BLP_FLAG_DONT_START
}
queries = append(queries, binlogplayer.PopulateBlpCheckpoint(0, status.Position, time.Now().Unix(), flags))
for _, tabletAlias := range vscw.destinationAliases {
destinationWaitGroup.Add(1)
go func(ti *topo.TabletInfo) {
defer destinationWaitGroup.Done()
vscw.wr.Logger().Infof("Making and populating blp_checkpoint table on tablet %v", ti.Alias)
if err := runSqlCommands(vscw.wr, ti, queries, abort, disableBinLogs); err != nil {
processError("blp_checkpoint queries failed on tablet %v: %v", ti.Alias, err)
}
}(vscw.destinationTablets[tabletAlias])
}
destinationWaitGroup.Wait()
if firstError != nil {
return firstError
}
}
// Now we're done with data copy, update the shard's source info.
if vscw.strategy.SkipSetSourceShards {
vscw.wr.Logger().Infof("Skipping setting SourceShard on destination shard.")
} else {
vscw.wr.Logger().Infof("Setting SourceShard on shard %v/%v", vscw.destinationKeyspace, vscw.destinationShard)
if err := vscw.wr.SetSourceShards(vscw.destinationKeyspace, vscw.destinationShard, []topo.TabletAlias{vscw.sourceAlias}, vscw.tables); err != nil {
return fmt.Errorf("Failed to set source shards: %v", err)
}
}
// And force a schema reload on all destination tablets.
// The master tablet will end up starting filtered replication
// at this point.
for _, tabletAlias := range vscw.reloadAliases {
destinationWaitGroup.Add(1)
go func(ti *topo.TabletInfo) {
defer destinationWaitGroup.Done()
vscw.wr.Logger().Infof("Reloading schema on tablet %v", ti.Alias)
ctx, cancel := context.WithTimeout(context.TODO(), 30*time.Second)
if err := vscw.wr.TabletManagerClient().ReloadSchema(ctx, ti); err != nil {
processError("ReloadSchema failed on tablet %v: %v", ti.Alias, err)
}
cancel()
}(vscw.reloadTablets[tabletAlias])
}
destinationWaitGroup.Wait()
return firstError
}