// findDestinationMasters finds for each destination shard the current master.
func (scw *SplitCloneWorker) findDestinationMasters(ctx context.Context) error {
scw.setState(WorkerStateFindTargets)
// Make sure we find a master for each destination shard and log it.
scw.wr.Logger().Infof("Finding a MASTER tablet for each destination shard...")
for _, si := range scw.destinationShards {
waitCtx, waitCancel := context.WithTimeout(ctx, *waitForHealthyTabletsTimeout)
defer waitCancel()
if err := scw.tsc.WaitForTablets(waitCtx, scw.cell, si.Keyspace(), si.ShardName(), []topodatapb.TabletType{topodatapb.TabletType_MASTER}); err != nil {
return fmt.Errorf("cannot find MASTER tablet for destination shard for %v/%v (in cell: %v): %v", si.Keyspace(), si.ShardName(), scw.cell, err)
}
masters := scw.tsc.GetHealthyTabletStats(si.Keyspace(), si.ShardName(), topodatapb.TabletType_MASTER)
if len(masters) == 0 {
return fmt.Errorf("cannot find MASTER tablet for destination shard for %v/%v (in cell: %v) in HealthCheck: empty TabletStats list", si.Keyspace(), si.ShardName(), scw.cell)
}
master := masters[0]
// Get the MySQL database name of the tablet.
keyspaceAndShard := topoproto.KeyspaceShardString(si.Keyspace(), si.ShardName())
scw.destinationDbNames[keyspaceAndShard] = topoproto.TabletDbName(master.Tablet)
// TODO(mberlin): Verify on the destination master that the
// _vt.blp_checkpoint table has the latest schema.
scw.wr.Logger().Infof("Using tablet %v as destination master for %v/%v", topoproto.TabletAliasString(master.Tablet.Alias), si.Keyspace(), si.ShardName())
}
scw.wr.Logger().Infof("NOTE: The used master of a destination shard might change over the course of the copy e.g. due to a reparent. The HealthCheck module will track and log master changes and any error message will always refer the actually used master address.")
return nil
}
// Run is part of the Task interface.
func (t *SplitCloneTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
// TODO(mberlin): Add parameters for the following options?
// '--source_reader_count', '1',
// '--destination_writer_count', '1',
args := []string{"SplitClone"}
if online := parameters["online"]; online != "" {
args = append(args, "--online="+online)
}
if offline := parameters["offline"]; offline != "" {
args = append(args, "--offline="+offline)
}
if excludeTables := parameters["exclude_tables"]; excludeTables != "" {
args = append(args, "--exclude_tables="+excludeTables)
}
if writeQueryMaxRows := parameters["write_query_max_rows"]; writeQueryMaxRows != "" {
args = append(args, "--write_query_max_rows="+writeQueryMaxRows)
}
if writeQueryMaxSize := parameters["write_query_max_size"]; writeQueryMaxSize != "" {
args = append(args, "--write_query_max_size="+writeQueryMaxSize)
}
if minHealthyRdonlyTablets := parameters["min_healthy_rdonly_tablets"]; minHealthyRdonlyTablets != "" {
args = append(args, "--min_healthy_rdonly_tablets="+minHealthyRdonlyTablets)
}
args = append(args, topoproto.KeyspaceShardString(parameters["keyspace"], parameters["source_shard"]))
output, err := ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], args)
// TODO(mberlin): Remove explicit reset when vtworker supports it implicility.
if err == nil {
// Ignore output and error of the Reset.
ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], []string{"Reset"})
}
return nil, output, err
}
// shardsWithTablesSources returns all the shards that have SourceShards set
// to one value, with an array of Tables.
func shardsWithTablesSources(ctx context.Context, wr *wrangler.Wrangler) ([]map[string]string, error) {
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
keyspaces, err := wr.TopoServer().GetKeyspaces(shortCtx)
cancel()
if err != nil {
return nil, fmt.Errorf("failed to get list of keyspaces: %v", err)
}
wg := sync.WaitGroup{}
mu := sync.Mutex{} // protects result
result := make([]map[string]string, 0, len(keyspaces))
rec := concurrency.AllErrorRecorder{}
for _, keyspace := range keyspaces {
wg.Add(1)
go func(keyspace string) {
defer wg.Done()
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
shards, err := wr.TopoServer().GetShardNames(shortCtx, keyspace)
cancel()
if err != nil {
rec.RecordError(fmt.Errorf("failed to get list of shards for keyspace '%v': %v", keyspace, err))
return
}
for _, shard := range shards {
wg.Add(1)
go func(keyspace, shard string) {
defer wg.Done()
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
si, err := wr.TopoServer().GetShard(shortCtx, keyspace, shard)
cancel()
if err != nil {
rec.RecordError(fmt.Errorf("failed to get details for shard '%v': %v", topoproto.KeyspaceShardString(keyspace, shard), err))
return
}
if len(si.SourceShards) == 1 && len(si.SourceShards[0].Tables) > 0 {
mu.Lock()
result = append(result, map[string]string{
"Keyspace": keyspace,
"Shard": shard,
})
mu.Unlock()
}
}(keyspace, shard)
}
}(keyspace)
}
wg.Wait()
if rec.HasErrors() {
return nil, rec.Error()
}
if len(result) == 0 {
return nil, fmt.Errorf("there are no shards with SourceShards")
}
return result, nil
}
// Run is part of the Task interface.
func (t *VerticalSplitDiffTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
args := []string{"VerticalSplitDiff"}
args = append(args, topoproto.KeyspaceShardString(parameters["dest_keyspace"], parameters["shard"]))
output, err := ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], args)
// TODO(mberlin): Remove explicit reset when vtworker supports it implicility.
if err == nil {
// Ignore output and error of the Reset.
ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], []string{"Reset"})
}
return nil, output, err
}
// Run is part of the Task interface.
func (t *MigrateServedTypesTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
args := []string{"MigrateServedTypes"}
if cells := parameters["cells"]; cells != "" {
args = append(args, "--cells="+cells)
}
if reverse := parameters["reverse"]; reverse != "" {
args = append(args, "--reverse="+reverse)
}
args = append(args,
topoproto.KeyspaceShardString(parameters["keyspace"], parameters["source_shard"]),
parameters["type"])
output, err := ExecuteVtctl(context.TODO(), parameters["vtctld_endpoint"], args)
return nil, output, err
}
// Run is part of the Task interface.
func (t *VerticalSplitDiffTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
// Run a "Reset" first to clear the state of a previous finished command.
// This reset is best effort. We ignore the output and error of it.
// TODO(mberlin): Remove explicit reset when vtworker supports it implicility.
ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], []string{"Reset"})
args := []string{"VerticalSplitDiff"}
if minHealthyRdonlyTablets := parameters["min_healthy_rdonly_tablets"]; minHealthyRdonlyTablets != "" {
args = append(args, "--min_healthy_rdonly_tablets="+minHealthyRdonlyTablets)
}
args = append(args, topoproto.KeyspaceShardString(parameters["dest_keyspace"], parameters["shard"]))
output, err := ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], args)
return nil, output, err
}
func (scw *SplitCloneWorker) createThrottlers() error {
scw.throttlersMu.Lock()
defer scw.throttlersMu.Unlock()
for _, si := range scw.destinationShards {
// Set up the throttler for each destination shard.
keyspaceAndShard := topoproto.KeyspaceShardString(si.Keyspace(), si.ShardName())
t, err := throttler.NewThrottler(keyspaceAndShard, "transactions", scw.destinationWriterCount, scw.maxTPS, scw.maxReplicationLag)
if err != nil {
return fmt.Errorf("cannot instantiate throttler: %v", err)
}
scw.throttlers[keyspaceAndShard] = t
}
return nil
}
// Run is part of the Task interface.
func (t *VerticalSplitCloneTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
// TODO(mberlin): Add parameters for the following options?
// '--source_reader_count', '1',
// '--destination_pack_count', '1',
// '--destination_writer_count', '1',
args := []string{"VerticalSplitClone"}
args = append(args, "--tables="+parameters["tables"])
args = append(args, topoproto.KeyspaceShardString(parameters["dest_keyspace"], parameters["shard"]))
output, err := ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], args)
// TODO(mberlin): Remove explicit reset when vtworker supports it implicility.
if err == nil {
// Ignore output and error of the Reset.
ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], []string{"Reset"})
}
return nil, output, err
}
// Run is part of the Task interface.
func (t *SplitCloneTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
// Run a "Reset" first to clear the state of a previous finished command.
// This reset is best effort. We ignore the output and error of it.
// TODO(mberlin): Remove explicit reset when vtworker supports it implicility.
ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], []string{"Reset"})
// TODO(mberlin): Add parameters for the following options?
// '--source_reader_count', '1',
// '--destination_writer_count', '1',
args := []string{"SplitClone"}
if online := parameters["online"]; online != "" {
args = append(args, "--online="+online)
}
if offline := parameters["offline"]; offline != "" {
args = append(args, "--offline="+offline)
}
if excludeTables := parameters["exclude_tables"]; excludeTables != "" {
args = append(args, "--exclude_tables="+excludeTables)
}
if chunkCount := parameters["chunk_count"]; chunkCount != "" {
args = append(args, "--chunk_count="+chunkCount)
}
if minRowsPerChunk := parameters["min_rows_per_chunk"]; minRowsPerChunk != "" {
args = append(args, "--min_rows_per_chunk="+minRowsPerChunk)
}
if writeQueryMaxRows := parameters["write_query_max_rows"]; writeQueryMaxRows != "" {
args = append(args, "--write_query_max_rows="+writeQueryMaxRows)
}
if writeQueryMaxSize := parameters["write_query_max_size"]; writeQueryMaxSize != "" {
args = append(args, "--write_query_max_size="+writeQueryMaxSize)
}
if minHealthyRdonlyTablets := parameters["min_healthy_rdonly_tablets"]; minHealthyRdonlyTablets != "" {
args = append(args, "--min_healthy_rdonly_tablets="+minHealthyRdonlyTablets)
}
if maxTPS := parameters["max_tps"]; maxTPS != "" {
args = append(args, "--max_tps="+maxTPS)
}
if maxReplicationLag := parameters["max_replication_lag"]; maxReplicationLag != "" {
args = append(args, "--max_replication_lag="+maxReplicationLag)
}
args = append(args, topoproto.KeyspaceShardString(parameters["keyspace"], parameters["source_shard"]))
output, err := ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], args)
return nil, output, err
}
// Run is part of the Task interface.
func (t *SplitDiffTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
args := []string{"SplitDiff"}
if excludeTables := parameters["exclude_tables"]; excludeTables != "" {
args = append(args, "--exclude_tables="+excludeTables)
}
if minHealthyRdonlyTablets := parameters["min_healthy_rdonly_tablets"]; minHealthyRdonlyTablets != "" {
args = append(args, "--min_healthy_rdonly_tablets="+minHealthyRdonlyTablets)
}
args = append(args, topoproto.KeyspaceShardString(parameters["keyspace"], parameters["dest_shard"]))
output, err := ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], args)
// TODO(mberlin): Remove explicit reset when vtworker supports it implicility.
if err == nil {
// Ignore output and error of the Reset.
ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], []string{"Reset"})
}
return nil, output, err
}
// Run is part of the Task interface.
func (t *VerticalSplitCloneTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
// TODO(mberlin): Add parameters for the following options?
// '--source_reader_count', '1',
// '--destination_writer_count', '1',
args := []string{"VerticalSplitClone"}
args = append(args, "--tables="+parameters["tables"])
if online := parameters["online"]; online != "" {
args = append(args, "--online="+online)
}
if offline := parameters["offline"]; offline != "" {
args = append(args, "--offline="+offline)
}
if chunkCount := parameters["chunk_count"]; chunkCount != "" {
args = append(args, "--chunk_count="+chunkCount)
}
if minRowsPerChunk := parameters["min_rows_per_chunk"]; minRowsPerChunk != "" {
args = append(args, "--min_rows_per_chunk="+minRowsPerChunk)
}
if writeQueryMaxRows := parameters["write_query_max_rows"]; writeQueryMaxRows != "" {
args = append(args, "--write_query_max_rows="+writeQueryMaxRows)
}
if writeQueryMaxSize := parameters["write_query_max_size"]; writeQueryMaxSize != "" {
args = append(args, "--write_query_max_size="+writeQueryMaxSize)
}
if minHealthyRdonlyTablets := parameters["min_healthy_rdonly_tablets"]; minHealthyRdonlyTablets != "" {
args = append(args, "--min_healthy_rdonly_tablets="+minHealthyRdonlyTablets)
}
if maxTPS := parameters["max_tps"]; maxTPS != "" {
args = append(args, "--max_tps="+maxTPS)
}
if maxReplicationLag := parameters["max_replication_lag"]; maxReplicationLag != "" {
args = append(args, "--max_replication_lag="+maxReplicationLag)
}
args = append(args, topoproto.KeyspaceShardString(parameters["dest_keyspace"], parameters["shard"]))
output, err := ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], args)
// TODO(mberlin): Remove explicit reset when vtworker supports it implicility.
if err == nil {
// Ignore output and error of the Reset.
ExecuteVtworker(context.TODO(), parameters["vtworker_endpoint"], []string{"Reset"})
}
return nil, output, err
}
// findTargets phase:
// - find one rdonly in the source shard
// - mark it as 'worker' pointing back to us
// - get the aliases of all the targets
func (vscw *VerticalSplitCloneWorker) findTargets(ctx context.Context) error {
vscw.setState(WorkerStateFindTargets)
// find an appropriate tablet in the source shard
var err error
vscw.sourceAlias, err = FindWorkerTablet(ctx, vscw.wr, vscw.cleaner, nil /* tsc */, vscw.cell, vscw.sourceKeyspace, "0", vscw.minHealthyRdonlyTablets)
if err != nil {
return fmt.Errorf("FindWorkerTablet() failed for %v/%v/0: %v", vscw.cell, vscw.sourceKeyspace, err)
}
vscw.wr.Logger().Infof("Using tablet %v as the source", topoproto.TabletAliasString(vscw.sourceAlias))
// get the tablet info for it
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
ti, err := vscw.wr.TopoServer().GetTablet(shortCtx, vscw.sourceAlias)
cancel()
if err != nil {
return fmt.Errorf("cannot read tablet %v: %v", topoproto.TabletAliasString(vscw.sourceAlias), err)
}
vscw.sourceTablet = ti.Tablet
// stop replication on it
shortCtx, cancel = context.WithTimeout(ctx, *remoteActionsTimeout)
err = vscw.wr.TabletManagerClient().StopSlave(shortCtx, vscw.sourceTablet)
cancel()
if err != nil {
return fmt.Errorf("cannot stop replication on tablet %v", topoproto.TabletAliasString(vscw.sourceAlias))
}
wrangler.RecordStartSlaveAction(vscw.cleaner, vscw.sourceTablet)
// Initialize healthcheck and add destination shards to it.
vscw.healthCheck = discovery.NewHealthCheck(*remoteActionsTimeout, *healthcheckRetryDelay, *healthCheckTimeout)
vscw.tsc = discovery.NewTabletStatsCache(vscw.healthCheck, vscw.cell)
watcher := discovery.NewShardReplicationWatcher(vscw.wr.TopoServer(), vscw.healthCheck,
vscw.cell, vscw.destinationKeyspace, vscw.destinationShard,
*healthCheckTopologyRefresh, discovery.DefaultTopoReadConcurrency)
vscw.destinationShardWatchers = append(vscw.destinationShardWatchers, watcher)
// Make sure we find a master for each destination shard and log it.
vscw.wr.Logger().Infof("Finding a MASTER tablet for each destination shard...")
waitCtx, waitCancel := context.WithTimeout(ctx, *waitForHealthyTabletsTimeout)
defer waitCancel()
if err := vscw.tsc.WaitForTablets(waitCtx, vscw.cell, vscw.destinationKeyspace, vscw.destinationShard, []topodatapb.TabletType{topodatapb.TabletType_MASTER}); err != nil {
return fmt.Errorf("cannot find MASTER tablet for destination shard for %v/%v (in cell: %v): %v", vscw.destinationKeyspace, vscw.destinationShard, vscw.cell, err)
}
masters := vscw.tsc.GetHealthyTabletStats(vscw.destinationKeyspace, vscw.destinationShard, topodatapb.TabletType_MASTER)
if len(masters) == 0 {
return fmt.Errorf("cannot find MASTER tablet for destination shard for %v/%v (in cell: %v) in HealthCheck: empty TabletStats list", vscw.destinationKeyspace, vscw.destinationShard, vscw.cell)
}
master := masters[0]
// Get the MySQL database name of the tablet.
keyspaceAndShard := topoproto.KeyspaceShardString(vscw.destinationKeyspace, vscw.destinationShard)
vscw.destinationDbNames[keyspaceAndShard] = topoproto.TabletDbName(master.Tablet)
// TODO(mberlin): Verify on the destination master that the
// _vt.blp_checkpoint table has the latest schema.
vscw.wr.Logger().Infof("Using tablet %v as destination master for %v/%v", topoproto.TabletAliasString(master.Tablet.Alias), vscw.destinationKeyspace, vscw.destinationShard)
vscw.wr.Logger().Infof("NOTE: The used master of a destination shard might change over the course of the copy e.g. due to a reparent. The HealthCheck module will track and log master changes and any error message will always refer the actually used master address.")
return nil
}
func (p *shardTabletProvider) description() string {
return topoproto.KeyspaceShardString(p.keyspace, p.shard)
}
// copy phase:
// - copy the data from source tablets to destination masters (with replication on)
// Assumes that the schema has already been created on each destination tablet
// (probably from vtctl's CopySchemaShard)
func (scw *LegacySplitCloneWorker) copy(ctx context.Context) error {
scw.setState(WorkerStateCloneOffline)
start := time.Now()
defer func() {
statsStateDurationsNs.Set(string(WorkerStateCloneOffline), time.Now().Sub(start).Nanoseconds())
}()
// 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)
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
sourceSchemaDefinition, err := scw.wr.GetSchema(shortCtx, scw.sourceAliases[0], nil, scw.excludeTables, false /* includeViews */)
cancel()
if err != nil {
return fmt.Errorf("cannot get schema from source %v: %v", topoproto.TabletAliasString(scw.sourceAliases[0]), err)
}
if len(sourceSchemaDefinition.TableDefinitions) == 0 {
return fmt.Errorf("no tables matching the table filter in tablet %v", topoproto.TabletAliasString(scw.sourceAliases[0]))
}
for _, td := range sourceSchemaDefinition.TableDefinitions {
if len(td.Columns) == 0 {
return fmt.Errorf("schema for table %v has no columns", td.Name)
}
}
scw.wr.Logger().Infof("Source tablet 0 has %v tables to copy", len(sourceSchemaDefinition.TableDefinitions))
scw.tableStatusList.initialize(sourceSchemaDefinition)
// In parallel, setup the channels to send SQL data chunks to for each destination tablet:
//
// mu protects the context for cancelation, and firstError
mu := sync.Mutex{}
var firstError error
ctx, cancelCopy := context.WithCancel(ctx)
processError := func(format string, args ...interface{}) {
scw.wr.Logger().Errorf(format, args...)
mu.Lock()
if firstError == nil {
firstError = fmt.Errorf(format, args...)
cancelCopy()
}
mu.Unlock()
}
insertChannels := make([]chan string, len(scw.destinationShards))
destinationWaitGroup := sync.WaitGroup{}
for shardIndex, si := range scw.destinationShards {
// 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] = make(chan string, scw.destinationWriterCount*2)
go func(keyspace, shard string, insertChannel chan string) {
for j := 0; j < scw.destinationWriterCount; j++ {
destinationWaitGroup.Add(1)
go func(threadID int) {
defer destinationWaitGroup.Done()
keyspaceAndShard := topoproto.KeyspaceShardString(keyspace, shard)
throttler := scw.destinationThrottlers[keyspaceAndShard]
defer throttler.ThreadFinished(threadID)
executor := newExecutor(scw.wr, scw.tsc, throttler, keyspace, shard, threadID)
if err := executor.fetchLoop(ctx, insertChannel); err != nil {
processError("executer.FetchLoop failed: %v", err)
}
}(j)
}
}(si.Keyspace(), si.ShardName(), insertChannels[shardIndex])
}
// read the vschema if needed
var keyspaceSchema *vindexes.KeyspaceSchema
if *useV3ReshardingMode {
kschema, err := scw.wr.TopoServer().GetVSchema(ctx, scw.keyspace)
if err != nil {
return fmt.Errorf("cannot load VSchema for keyspace %v: %v", scw.keyspace, err)
}
if kschema == nil {
return fmt.Errorf("no VSchema for keyspace %v", scw.keyspace)
}
keyspaceSchema, err = vindexes.BuildKeyspaceSchema(kschema, scw.keyspace)
if err != nil {
return fmt.Errorf("cannot build vschema for keyspace %v: %v", scw.keyspace, err)
}
}
// 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 {
var keyResolver keyspaceIDResolver
if *useV3ReshardingMode {
keyResolver, err = newV3ResolverFromTableDefinition(keyspaceSchema, td)
if err != nil {
return fmt.Errorf("cannot resolve v3 sharding keys for keyspace %v: %v", scw.keyspace, err)
}
} else {
keyResolver, err = newV2Resolver(scw.keyspaceInfo, td)
if err != nil {
return fmt.Errorf("cannot resolve sharding keys for keyspace %v: %v", scw.keyspace, err)
}
}
rowSplitter := NewRowSplitter(scw.destinationShards, keyResolver)
chunks, err := generateChunks(ctx, scw.wr, scw.sourceTablets[shardIndex], td, scw.sourceReaderCount, defaultMinRowsPerChunk)
if err != nil {
return err
}
scw.tableStatusList.setThreadCount(tableIndex, len(chunks)-1)
for _, c := range chunks {
sourceWaitGroup.Add(1)
go func(td *tabletmanagerdatapb.TableDefinition, tableIndex int, chunk chunk) {
defer sourceWaitGroup.Done()
sema.Acquire()
defer sema.Release()
scw.tableStatusList.threadStarted(tableIndex)
// Start streaming from the source tablets.
tp := newSingleTabletProvider(ctx, scw.wr.TopoServer(), scw.sourceAliases[shardIndex])
rr, err := NewRestartableResultReader(ctx, scw.wr.Logger(), tp, td, chunk, false /* allowMultipleRetries */)
if err != nil {
processError("NewRestartableResultReader failed: %v", err)
return
}
defer rr.Close(ctx)
// process the data
dbNames := make([]string, len(scw.destinationShards))
for i, si := range scw.destinationShards {
keyspaceAndShard := topoproto.KeyspaceShardString(si.Keyspace(), si.ShardName())
dbNames[i] = scw.destinationDbNames[keyspaceAndShard]
}
if err := scw.processData(ctx, dbNames, td, tableIndex, rr, rowSplitter, insertChannels, scw.destinationPackCount); err != nil {
processError("processData failed: %v", err)
}
scw.tableStatusList.threadDone(tableIndex)
}(td, tableIndex, c)
}
}
}
sourceWaitGroup.Wait()
for shardIndex := range scw.destinationShards {
close(insertChannels[shardIndex])
}
destinationWaitGroup.Wait()
if firstError != nil {
return firstError
}
// then create and populate the blp_checkpoint table
if scw.strategy.skipPopulateBlpCheckpoint {
scw.wr.Logger().Infof("Skipping populating the blp_checkpoint table")
} else {
queries := make([]string, 0, 4)
queries = append(queries, binlogplayer.CreateBlpCheckpoint()...)
flags := ""
if scw.strategy.dontStartBinlogPlayer {
flags = binlogplayer.BlpFlagDontStart
}
// get the current position from the sources
for shardIndex := range scw.sourceShards {
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
status, err := scw.wr.TabletManagerClient().SlaveStatus(shortCtx, scw.sourceTablets[shardIndex])
cancel()
if err != nil {
return err
}
queries = append(queries, binlogplayer.PopulateBlpCheckpoint(uint32(shardIndex), status.Position, scw.maxTPS, throttler.ReplicationLagModuleDisabled, time.Now().Unix(), flags))
}
for _, si := range scw.destinationShards {
destinationWaitGroup.Add(1)
go func(keyspace, shard string) {
defer destinationWaitGroup.Done()
scw.wr.Logger().Infof("Making and populating blp_checkpoint table")
keyspaceAndShard := topoproto.KeyspaceShardString(keyspace, shard)
if err := runSQLCommands(ctx, scw.wr, scw.tsc, keyspace, shard, scw.destinationDbNames[keyspaceAndShard], queries); err != nil {
processError("blp_checkpoint queries failed: %v", err)
}
}(si.Keyspace(), si.ShardName())
}
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())
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
err := scw.wr.SetSourceShards(shortCtx, si.Keyspace(), si.ShardName(), scw.sourceAliases, nil)
cancel()
if err != nil {
return fmt.Errorf("failed to set source shards: %v", err)
}
}
}
err = scw.findRefreshTargets(ctx)
if err != nil {
return fmt.Errorf("failed before refreshing state on destination tablets: %v", err)
}
// And force a state refresh (re-read topo) 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.refreshAliases[shardIndex] {
destinationWaitGroup.Add(1)
go func(ti *topo.TabletInfo) {
defer destinationWaitGroup.Done()
scw.wr.Logger().Infof("Refreshing state on tablet %v", ti.AliasString())
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
err := scw.wr.TabletManagerClient().RefreshState(shortCtx, ti.Tablet)
cancel()
if err != nil {
processError("RefreshState failed on tablet %v: %v", ti.AliasString(), err)
}
}(scw.refreshTablets[shardIndex][*tabletAlias])
}
}
destinationWaitGroup.Wait()
return firstError
}
// findTargets phase:
// - find one rdonly in the source shard
// - mark it as 'worker' pointing back to us
// - get the aliases of all the targets
func (scw *LegacySplitCloneWorker) findTargets(ctx context.Context) error {
scw.setState(WorkerStateFindTargets)
var err error
// find an appropriate tablet in the source shards
scw.sourceAliases = make([]*topodatapb.TabletAlias, len(scw.sourceShards))
for i, si := range scw.sourceShards {
scw.sourceAliases[i], err = FindWorkerTablet(ctx, scw.wr, scw.cleaner, scw.tsc, scw.cell, si.Keyspace(), si.ShardName(), scw.minHealthyRdonlyTablets)
if err != nil {
return fmt.Errorf("FindWorkerTablet() failed for %v/%v/%v: %v", scw.cell, si.Keyspace(), si.ShardName(), err)
}
scw.wr.Logger().Infof("Using tablet %v as source for %v/%v", topoproto.TabletAliasString(scw.sourceAliases[i]), si.Keyspace(), si.ShardName())
}
// get the tablet info for them, and stop their replication
scw.sourceTablets = make([]*topodatapb.Tablet, len(scw.sourceAliases))
for i, alias := range scw.sourceAliases {
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
ti, err := scw.wr.TopoServer().GetTablet(shortCtx, alias)
cancel()
if err != nil {
return fmt.Errorf("cannot read tablet %v: %v", topoproto.TabletAliasString(alias), err)
}
scw.sourceTablets[i] = ti.Tablet
shortCtx, cancel = context.WithTimeout(ctx, *remoteActionsTimeout)
err = scw.wr.TabletManagerClient().StopSlave(shortCtx, scw.sourceTablets[i])
cancel()
if err != nil {
return fmt.Errorf("cannot stop replication on tablet %v", topoproto.TabletAliasString(alias))
}
wrangler.RecordStartSlaveAction(scw.cleaner, scw.sourceTablets[i])
}
// Initialize healthcheck and add destination shards to it.
scw.healthCheck = discovery.NewHealthCheck(*remoteActionsTimeout, *healthcheckRetryDelay, *healthCheckTimeout)
scw.tsc = discovery.NewTabletStatsCache(scw.healthCheck, scw.cell)
for _, si := range scw.destinationShards {
watcher := discovery.NewShardReplicationWatcher(scw.wr.TopoServer(), scw.healthCheck,
scw.cell, si.Keyspace(), si.ShardName(),
*healthCheckTopologyRefresh, discovery.DefaultTopoReadConcurrency)
scw.destinationShardWatchers = append(scw.destinationShardWatchers, watcher)
}
// Make sure we find a master for each destination shard and log it.
scw.wr.Logger().Infof("Finding a MASTER tablet for each destination shard...")
for _, si := range scw.destinationShards {
waitCtx, waitCancel := context.WithTimeout(ctx, 10*time.Second)
defer waitCancel()
if err := scw.tsc.WaitForTablets(waitCtx, scw.cell, si.Keyspace(), si.ShardName(), []topodatapb.TabletType{topodatapb.TabletType_MASTER}); err != nil {
return fmt.Errorf("cannot find MASTER tablet for destination shard for %v/%v: %v", si.Keyspace(), si.ShardName(), err)
}
masters := scw.tsc.GetHealthyTabletStats(si.Keyspace(), si.ShardName(), topodatapb.TabletType_MASTER)
if len(masters) == 0 {
return fmt.Errorf("cannot find MASTER tablet for destination shard for %v/%v in HealthCheck: empty TabletStats list", si.Keyspace(), si.ShardName())
}
master := masters[0]
// Get the MySQL database name of the tablet.
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
ti, err := scw.wr.TopoServer().GetTablet(shortCtx, master.Tablet.Alias)
cancel()
if err != nil {
return fmt.Errorf("cannot get the TabletInfo for destination master (%v) to find out its db name: %v", topoproto.TabletAliasString(master.Tablet.Alias), err)
}
keyspaceAndShard := topoproto.KeyspaceShardString(si.Keyspace(), si.ShardName())
scw.destinationDbNames[keyspaceAndShard] = ti.DbName()
// TODO(mberlin): Verify on the destination master that the
// _vt.blp_checkpoint table has the latest schema.
scw.wr.Logger().Infof("Using tablet %v as destination master for %v/%v", topoproto.TabletAliasString(master.Tablet.Alias), si.Keyspace(), si.ShardName())
}
scw.wr.Logger().Infof("NOTE: The used master of a destination shard might change over the course of the copy e.g. due to a reparent. The HealthCheck module will track and log master changes and any error message will always refer the actually used master address.")
// Set up the throttler for each destination shard.
for _, si := range scw.destinationShards {
keyspaceAndShard := topoproto.KeyspaceShardString(si.Keyspace(), si.ShardName())
t, err := throttler.NewThrottler(
keyspaceAndShard, "transactions", scw.destinationWriterCount, scw.maxTPS, throttler.ReplicationLagModuleDisabled)
if err != nil {
return fmt.Errorf("cannot instantiate throttler: %v", err)
}
scw.destinationThrottlers[keyspaceAndShard] = t
}
return nil
}
// copy phase:
// - copy the data from source tablets to destination masters (with replication on)
// Assumes that the schema has already been created on each destination tablet
// (probably from vtctl's CopySchemaShard)
func (scw *SplitCloneWorker) clone(ctx context.Context, state StatusWorkerState) error {
if state != WorkerStateCloneOnline && state != WorkerStateCloneOffline {
panic(fmt.Sprintf("invalid state passed to clone(): %v", state))
}
scw.setState(state)
start := time.Now()
defer func() {
statsStateDurationsNs.Set(string(state), time.Now().Sub(start).Nanoseconds())
}()
var firstSourceTablet *topodatapb.Tablet
if state == WorkerStateCloneOffline {
// Use the first source tablet which we took offline.
firstSourceTablet = scw.sourceTablets[0]
} else {
// Pick any healthy serving source tablet.
si := scw.sourceShards[0]
tablets := scw.tsc.GetTabletStats(si.Keyspace(), si.ShardName(), topodatapb.TabletType_RDONLY)
if len(tablets) == 0 {
// We fail fast on this problem and don't retry because at the start all tablets should be healthy.
return fmt.Errorf("no healthy RDONLY tablet in source shard (%v) available (required to find out the schema)", topoproto.KeyspaceShardString(si.Keyspace(), si.ShardName()))
}
firstSourceTablet = tablets[0].Tablet
}
var statsCounters []*stats.Counters
var tableStatusList *tableStatusList
switch state {
case WorkerStateCloneOnline:
statsCounters = []*stats.Counters{statsOnlineInsertsCounters, statsOnlineUpdatesCounters, statsOnlineDeletesCounters, statsOnlineEqualRowsCounters}
tableStatusList = scw.tableStatusListOnline
case WorkerStateCloneOffline:
statsCounters = []*stats.Counters{statsOfflineInsertsCounters, statsOfflineUpdatesCounters, statsOfflineDeletesCounters, statsOfflineEqualRowsCounters}
tableStatusList = scw.tableStatusListOffline
}
// The throttlers exist only for the duration of this clone() call.
// That means a SplitClone invocation with both online and offline phases
// will create throttlers for each phase.
if err := scw.createThrottlers(); err != nil {
return err
}
defer scw.closeThrottlers()
sourceSchemaDefinition, err := scw.getSourceSchema(ctx, firstSourceTablet)
if err != nil {
return err
}
scw.wr.Logger().Infof("Source tablet 0 has %v tables to copy", len(sourceSchemaDefinition.TableDefinitions))
tableStatusList.initialize(sourceSchemaDefinition)
// In parallel, setup the channels to send SQL data chunks to for each destination tablet:
//
// mu protects the context for cancelation, and firstError
mu := sync.Mutex{}
var firstError error
ctx, cancelCopy := context.WithCancel(ctx)
processError := func(format string, args ...interface{}) {
scw.wr.Logger().Errorf(format, args...)
mu.Lock()
if firstError == nil {
firstError = fmt.Errorf(format, args...)
cancelCopy()
}
mu.Unlock()
}
insertChannels := make([]chan string, len(scw.destinationShards))
destinationWaitGroup := sync.WaitGroup{}
for shardIndex, si := range scw.destinationShards {
// 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] = make(chan string, scw.destinationWriterCount*2)
for j := 0; j < scw.destinationWriterCount; j++ {
destinationWaitGroup.Add(1)
go func(keyspace, shard string, insertChannel chan string, throttler *throttler.Throttler, threadID int) {
defer destinationWaitGroup.Done()
defer throttler.ThreadFinished(threadID)
executor := newExecutor(scw.wr, scw.tsc, throttler, keyspace, shard, threadID)
if err := executor.fetchLoop(ctx, insertChannel); err != nil {
processError("executer.FetchLoop failed: %v", err)
}
}(si.Keyspace(), si.ShardName(), insertChannels[shardIndex], scw.getThrottler(si.Keyspace(), si.ShardName()), j)
}
}
// Now for each table, read data chunks and send them to all
// insertChannels
sourceWaitGroup := sync.WaitGroup{}
sema := sync2.NewSemaphore(scw.sourceReaderCount, 0)
for tableIndex, td := range sourceSchemaDefinition.TableDefinitions {
td = reorderColumnsPrimaryKeyFirst(td)
keyResolver, err := scw.createKeyResolver(td)
if err != nil {
return fmt.Errorf("cannot resolve sharding keys for keyspace %v: %v", scw.destinationKeyspace, err)
}
// TODO(mberlin): We're going to chunk *all* source shards based on the MIN
// and MAX values of the *first* source shard. Is this going to be a problem?
chunks, err := generateChunks(ctx, scw.wr, firstSourceTablet, td, scw.chunkCount, scw.minRowsPerChunk)
if err != nil {
return err
}
tableStatusList.setThreadCount(tableIndex, len(chunks))
for _, c := range chunks {
sourceWaitGroup.Add(1)
go func(td *tabletmanagerdatapb.TableDefinition, tableIndex int, chunk chunk) {
defer sourceWaitGroup.Done()
errPrefix := fmt.Sprintf("table=%v chunk=%v", td.Name, chunk)
// We need our own error per Go routine to avoid races.
var err error
sema.Acquire()
defer sema.Release()
tableStatusList.threadStarted(tableIndex)
if state == WorkerStateCloneOnline {
// Wait for enough healthy tablets (they might have become unhealthy
// and their replication lag might have increased since we started.)
if err := scw.waitForTablets(ctx, scw.sourceShards, *retryDuration); err != nil {
processError("%v: No healthy source tablets found (gave up after %v): ", errPrefix, *retryDuration, err)
return
}
}
// Set up readers for the diff. There will be one reader for every
// source and destination shard.
sourceReaders := make([]ResultReader, len(scw.sourceShards))
destReaders := make([]ResultReader, len(scw.destinationShards))
for shardIndex, si := range scw.sourceShards {
var tp tabletProvider
allowMultipleRetries := true
if state == WorkerStateCloneOffline {
tp = newSingleTabletProvider(ctx, scw.wr.TopoServer(), scw.offlineSourceAliases[shardIndex])
// allowMultipleRetries is false to avoid that we'll keep retrying
// on the same tablet alias for hours. This guards us against the
// situation that an offline tablet gets restarted and serves again.
// In that case we cannot use it because its replication is no
// longer stopped at the same point as we took it offline initially.
allowMultipleRetries = false
} else {
tp = newShardTabletProvider(scw.tsc, scw.tabletTracker, si.Keyspace(), si.ShardName())
}
sourceResultReader, err := NewRestartableResultReader(ctx, scw.wr.Logger(), tp, td, chunk, allowMultipleRetries)
if err != nil {
processError("%v: NewRestartableResultReader for source: %v failed", errPrefix, tp.description())
return
}
defer sourceResultReader.Close()
sourceReaders[shardIndex] = sourceResultReader
}
// Wait for enough healthy tablets (they might have become unhealthy
// and their replication lag might have increased due to a previous
// chunk pipeline.)
if err := scw.waitForTablets(ctx, scw.destinationShards, *retryDuration); err != nil {
processError("%v: No healthy destination tablets found (gave up after %v): ", errPrefix, *retryDuration, err)
return
}
for shardIndex, si := range scw.destinationShards {
tp := newShardTabletProvider(scw.tsc, scw.tabletTracker, si.Keyspace(), si.ShardName())
destResultReader, err := NewRestartableResultReader(ctx, scw.wr.Logger(), tp, td, chunk, true /* allowMultipleRetries */)
if err != nil {
processError("%v: NewRestartableResultReader for destination: %v failed: %v", errPrefix, tp.description(), err)
return
}
defer destResultReader.Close()
destReaders[shardIndex] = destResultReader
}
var sourceReader ResultReader
var destReader ResultReader
if len(sourceReaders) >= 2 {
sourceReader, err = NewResultMerger(sourceReaders, len(td.PrimaryKeyColumns))
if err != nil {
processError("%v: NewResultMerger for source tablets failed: %v", errPrefix, err)
return
}
} else {
sourceReader = sourceReaders[0]
}
if len(destReaders) >= 2 {
destReader, err = NewResultMerger(destReaders, len(td.PrimaryKeyColumns))
if err != nil {
processError("%v: NewResultMerger for destination tablets failed: %v", errPrefix, err)
return
}
} else {
destReader = destReaders[0]
}
dbNames := make([]string, len(scw.destinationShards))
for i, si := range scw.destinationShards {
keyspaceAndShard := topoproto.KeyspaceShardString(si.Keyspace(), si.ShardName())
dbNames[i] = scw.destinationDbNames[keyspaceAndShard]
}
// Compare the data and reconcile any differences.
differ, err := NewRowDiffer2(ctx, sourceReader, destReader, td, tableStatusList, tableIndex,
scw.destinationShards, keyResolver,
insertChannels, ctx.Done(), dbNames, scw.writeQueryMaxRows, scw.writeQueryMaxSize, scw.writeQueryMaxRowsDelete, statsCounters)
if err != nil {
processError("%v: NewRowDiffer2 failed: %v", errPrefix, err)
return
}
// Ignore the diff report because all diffs should get reconciled.
_ /* DiffReport */, err = differ.Diff()
if err != nil {
processError("%v: RowDiffer2 failed: %v", errPrefix, err)
return
}
tableStatusList.threadDone(tableIndex)
}(td, tableIndex, c)
}
}
sourceWaitGroup.Wait()
for shardIndex := range scw.destinationShards {
close(insertChannels[shardIndex])
}
destinationWaitGroup.Wait()
if firstError != nil {
return firstError
}
if state == WorkerStateCloneOffline {
// Create and populate the blp_checkpoint table to give filtered replication
// a starting point.
if scw.strategy.skipPopulateBlpCheckpoint {
scw.wr.Logger().Infof("Skipping populating the blp_checkpoint table")
} else {
queries := make([]string, 0, 4)
queries = append(queries, binlogplayer.CreateBlpCheckpoint()...)
flags := ""
if scw.strategy.dontStartBinlogPlayer {
flags = binlogplayer.BlpFlagDontStart
}
// get the current position from the sources
for shardIndex := range scw.sourceShards {
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
status, err := scw.wr.TabletManagerClient().SlaveStatus(shortCtx, scw.sourceTablets[shardIndex])
cancel()
if err != nil {
return err
}
// TODO(mberlin): Fill in scw.maxReplicationLag once the adapative
// throttler is enabled by default.
queries = append(queries, binlogplayer.PopulateBlpCheckpoint(uint32(shardIndex), status.Position, scw.maxTPS, throttler.ReplicationLagModuleDisabled, time.Now().Unix(), flags))
}
for _, si := range scw.destinationShards {
destinationWaitGroup.Add(1)
go func(keyspace, shard string) {
defer destinationWaitGroup.Done()
scw.wr.Logger().Infof("Making and populating blp_checkpoint table")
keyspaceAndShard := topoproto.KeyspaceShardString(keyspace, shard)
if err := runSQLCommands(ctx, scw.wr, scw.tsc, keyspace, shard, scw.destinationDbNames[keyspaceAndShard], queries); err != nil {
processError("blp_checkpoint queries failed: %v", err)
}
}(si.Keyspace(), si.ShardName())
}
destinationWaitGroup.Wait()
if firstError != nil {
return firstError
}
}
// Configure filtered replication by setting the SourceShard info.
// The master tablets won't enable filtered replication (the binlog player)
// until they re-read the topology due to a restart or a reload.
// 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 (tables: %v)", si.Keyspace(), si.ShardName(), scw.tables)
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
err := scw.wr.SetSourceShards(shortCtx, si.Keyspace(), si.ShardName(), scw.offlineSourceAliases, scw.tables)
cancel()
if err != nil {
return fmt.Errorf("failed to set source shards: %v", err)
}
}
}
// Force a state refresh (re-read topo) on all destination tablets.
// The master tablet will end up starting filtered replication at this point.
//
// Find all tablets first, then refresh the state on each in parallel.
err = scw.findRefreshTargets(ctx)
if err != nil {
return fmt.Errorf("failed before refreshing state on destination tablets: %v", err)
}
for shardIndex := range scw.destinationShards {
for _, tabletAlias := range scw.refreshAliases[shardIndex] {
destinationWaitGroup.Add(1)
go func(ti *topo.TabletInfo) {
defer destinationWaitGroup.Done()
scw.wr.Logger().Infof("Refreshing state on tablet %v", ti.AliasString())
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
err := scw.wr.TabletManagerClient().RefreshState(shortCtx, ti.Tablet)
cancel()
if err != nil {
processError("RefreshState failed on tablet %v: %v", ti.AliasString(), err)
}
}(scw.refreshTablets[shardIndex][*tabletAlias])
}
}
} // clonePhase == offline
destinationWaitGroup.Wait()
return firstError
}
func (scw *SplitCloneWorker) getThrottlerLocked(keyspace, shard string) *throttler.Throttler {
keyspaceAndShard := topoproto.KeyspaceShardString(keyspace, shard)
return scw.throttlers[keyspaceAndShard]
}
// Run is part of the Task interface.
func (t *WaitForFilteredReplicationTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
keyspaceAndShard := topoproto.KeyspaceShardString(parameters["keyspace"], parameters["shard"])
output, err := ExecuteVtctl(context.TODO(), parameters["vtctld_endpoint"],
[]string{"WaitForFilteredReplication", "-max_delay", parameters["max_delay"], keyspaceAndShard})
return nil, output, err
}
// clone phase:
// - copy the data from source tablets to destination masters (with replication on)
// Assumes that the schema has already been created on each destination tablet
// (probably from vtctl's CopySchemaShard)
func (vscw *VerticalSplitCloneWorker) clone(ctx context.Context) error {
vscw.setState(WorkerStateCloneOffline)
start := time.Now()
defer func() {
statsStateDurationsNs.Set(string(WorkerStateCloneOffline), time.Now().Sub(start).Nanoseconds())
}()
// get source schema
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
sourceSchemaDefinition, err := vscw.wr.GetSchema(shortCtx, vscw.sourceAlias, vscw.tables, nil, true)
cancel()
if err != nil {
return fmt.Errorf("cannot get schema from source %v: %v", topoproto.TabletAliasString(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.tableStatusList.initialize(sourceSchemaDefinition)
// In parallel, setup the channels to send SQL data chunks to
// for each destination tablet.
//
// mu protects firstError
mu := sync.Mutex{}
var firstError error
ctx, cancelCopy := context.WithCancel(ctx)
processError := func(format string, args ...interface{}) {
vscw.wr.Logger().Errorf(format, args...)
mu.Lock()
if firstError == nil {
firstError = fmt.Errorf(format, args...)
cancelCopy()
}
mu.Unlock()
}
destinationWaitGroup := sync.WaitGroup{}
// we create one channel for the 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.
insertChannel := make(chan string, vscw.destinationWriterCount*2)
// Set up the throttler for the destination shard.
keyspaceAndShard := topoproto.KeyspaceShardString(vscw.destinationKeyspace, vscw.destinationShard)
destinationThrottler, err := throttler.NewThrottler(
keyspaceAndShard, "transactions", vscw.destinationWriterCount, vscw.maxTPS, throttler.ReplicationLagModuleDisabled)
if err != nil {
return fmt.Errorf("cannot instantiate throttler: %v", err)
}
for j := 0; j < vscw.destinationWriterCount; j++ {
destinationWaitGroup.Add(1)
go func(threadID int) {
defer destinationWaitGroup.Done()
defer destinationThrottler.ThreadFinished(threadID)
executor := newExecutor(vscw.wr, vscw.tsc, destinationThrottler, vscw.destinationKeyspace, vscw.destinationShard, threadID)
if err := executor.fetchLoop(ctx, insertChannel); err != nil {
processError("executer.FetchLoop failed: %v", err)
}
}(j)
}
// Now for each table, read data chunks and send them to insertChannel
sourceWaitGroup := sync.WaitGroup{}
sema := sync2.NewSemaphore(vscw.sourceReaderCount, 0)
dbName := vscw.destinationDbNames[topoproto.KeyspaceShardString(vscw.destinationKeyspace, vscw.destinationShard)]
for tableIndex, td := range sourceSchemaDefinition.TableDefinitions {
if td.Type == tmutils.TableView {
continue
}
chunks, err := generateChunks(ctx, vscw.wr, vscw.sourceTablet, td, vscw.minTableSizeForSplit, vscw.sourceReaderCount)
if err != nil {
return err
}
vscw.tableStatusList.setThreadCount(tableIndex, len(chunks)-1)
for _, c := range chunks {
sourceWaitGroup.Add(1)
go func(td *tabletmanagerdatapb.TableDefinition, tableIndex int, chunk chunk) {
defer sourceWaitGroup.Done()
sema.Acquire()
defer sema.Release()
vscw.tableStatusList.threadStarted(tableIndex)
// Start streaming from the source tablet.
rr, err := NewRestartableResultReader(ctx, vscw.wr.Logger(), vscw.wr.TopoServer(), vscw.sourceAlias, td, chunk)
if err != nil {
processError("NewRestartableResultReader failed: %v", err)
return
}
defer rr.Close()
// process the data
if err := vscw.processData(ctx, dbName, td, tableIndex, rr, insertChannel, vscw.destinationPackCount); err != nil {
processError("ResultReader failed: %v", err)
}
vscw.tableStatusList.threadDone(tableIndex)
}(td, tableIndex, c)
}
}
sourceWaitGroup.Wait()
close(insertChannel)
destinationWaitGroup.Wait()
// Stop Throttler.
destinationThrottler.Close()
if firstError != nil {
return firstError
}
// then create and populate the blp_checkpoint table
if vscw.strategy.skipPopulateBlpCheckpoint {
vscw.wr.Logger().Infof("Skipping populating the blp_checkpoint table")
} else {
// get the current position from the source
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
status, err := vscw.wr.TabletManagerClient().SlaveStatus(shortCtx, vscw.sourceTablet)
cancel()
if err != nil {
return err
}
queries := make([]string, 0, 4)
queries = append(queries, binlogplayer.CreateBlpCheckpoint()...)
flags := ""
if vscw.strategy.dontStartBinlogPlayer {
flags = binlogplayer.BlpFlagDontStart
}
queries = append(queries, binlogplayer.PopulateBlpCheckpoint(0, status.Position, vscw.maxTPS, throttler.ReplicationLagModuleDisabled, time.Now().Unix(), flags))
vscw.wr.Logger().Infof("Making and populating blp_checkpoint table")
if err := runSQLCommands(ctx, vscw.wr, vscw.tsc, vscw.destinationKeyspace, vscw.destinationShard, dbName, queries); err != nil {
processError("blp_checkpoint queries failed: %v", err)
}
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)
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
err := vscw.wr.SetSourceShards(shortCtx, vscw.destinationKeyspace, vscw.destinationShard, []*topodatapb.TabletAlias{vscw.sourceAlias}, vscw.tables)
cancel()
if err != nil {
return fmt.Errorf("Failed to set source shards: %v", err)
}
}
err = vscw.findRefreshTargets(ctx)
if err != nil {
return fmt.Errorf("failed before refreshing state on destination tablets: %v", err)
}
// And force a state refresh (re-read topo) on all destination tablets.
// The master tablet will end up starting filtered replication
// at this point.
for _, tabletAlias := range vscw.refreshAliases {
destinationWaitGroup.Add(1)
go func(ti *topo.TabletInfo) {
defer destinationWaitGroup.Done()
vscw.wr.Logger().Infof("Refreshing state on tablet %v", ti.AliasString())
shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
err := vscw.wr.TabletManagerClient().RefreshState(shortCtx, ti.Tablet)
cancel()
if err != nil {
processError("RefreshState failed on tablet %v: %v", ti.AliasString(), err)
}
}(vscw.refreshTablets[*tabletAlias])
}
destinationWaitGroup.Wait()
return firstError
}
// Run is part of the Task interface.
func (t *HorizontalReshardingTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
// Required parameters.
// Example: test_keyspace
keyspace := parameters["keyspace"]
// Example: 10-20
sourceShards := strings.Split(parameters["source_shard_list"], ",")
// Example: 10-18,18-20
destShards := strings.Split(parameters["dest_shard_list"], ",")
// Example: localhost:15000
vtctldEndpoint := parameters["vtctld_endpoint"]
// Example: localhost:15001
vtworkerEndpoint := parameters["vtworker_endpoint"]
// Optional parameters.
// Example: unrelated1,unrelated2
excludeTables := parameters["exclude_tables"]
// Example: 1
minHealthyRdonlyTablets := parameters["min_healthy_rdonly_tablets"]
var newTasks []*automationpb.TaskContainer
copySchemaTasks := NewTaskContainer()
for _, destShard := range destShards {
AddTask(copySchemaTasks, "CopySchemaShardTask", map[string]string{
"source_keyspace_and_shard": topoproto.KeyspaceShardString(keyspace, sourceShards[0]),
"dest_keyspace_and_shard": topoproto.KeyspaceShardString(keyspace, destShard),
"exclude_tables": excludeTables,
"vtctld_endpoint": vtctldEndpoint,
})
}
newTasks = append(newTasks, copySchemaTasks)
splitCloneTasks := NewTaskContainer()
for _, sourceShard := range sourceShards {
// TODO(mberlin): Add a semaphore as argument to limit the parallism.
AddTask(splitCloneTasks, "SplitCloneTask", map[string]string{
"keyspace": keyspace,
"source_shard": sourceShard,
"vtworker_endpoint": vtworkerEndpoint,
"exclude_tables": excludeTables,
"min_healthy_rdonly_tablets": minHealthyRdonlyTablets,
})
}
newTasks = append(newTasks, splitCloneTasks)
// TODO(mberlin): When the framework supports nesting tasks, these wait tasks should be run before each SplitDiff.
waitTasks := NewTaskContainer()
for _, destShard := range destShards {
AddTask(waitTasks, "WaitForFilteredReplicationTask", map[string]string{
"keyspace": keyspace,
"shard": destShard,
"max_delay": "30s",
"vtctld_endpoint": vtctldEndpoint,
})
}
newTasks = append(newTasks, waitTasks)
// TODO(mberlin): Run all SplitDiffTasks in parallel which do not use overlapping source shards for the comparison.
for _, destShard := range destShards {
splitDiffTask := NewTaskContainer()
AddTask(splitDiffTask, "SplitDiffTask", map[string]string{
"keyspace": keyspace,
"dest_shard": destShard,
"vtworker_endpoint": vtworkerEndpoint,
"exclude_tables": excludeTables,
"min_healthy_rdonly_tablets": minHealthyRdonlyTablets,
})
newTasks = append(newTasks, splitDiffTask)
}
for _, servedType := range []string{"rdonly", "replica", "master"} {
migrateServedTypesTasks := NewTaskContainer()
for _, sourceShard := range sourceShards {
AddTask(migrateServedTypesTasks, "MigrateServedTypesTask", map[string]string{
"keyspace": keyspace,
"source_shard": sourceShard,
"type": servedType,
"vtctld_endpoint": vtctldEndpoint,
})
}
newTasks = append(newTasks, migrateServedTypesTasks)
}
return newTasks, "", nil
}
// Run is part of the Task interface.
func (t *VerticalSplitTask) Run(parameters map[string]string) ([]*automationpb.TaskContainer, string, error) {
// Example: source_keyspace
sourceKeyspace := parameters["source_keyspace"]
// Example: destination_keyspace
destKeyspace := parameters["dest_keyspace"]
// Example: 10-20
shards := strings.Split(parameters["shard_list"], ",")
// Example: table1,table2
tables := parameters["tables"]
// Example: localhost:15000
vtctldEndpoint := parameters["vtctld_endpoint"]
// Example: localhost:15001
vtworkerEndpoint := parameters["vtworker_endpoint"]
var newTasks []*automationpb.TaskContainer
copySchemaTasks := NewTaskContainer()
for _, shard := range shards {
AddTask(copySchemaTasks, "CopySchemaShardTask", map[string]string{
"source_keyspace_and_shard": topoproto.KeyspaceShardString(sourceKeyspace, shard),
"dest_keyspace_and_shard": topoproto.KeyspaceShardString(destKeyspace, shard),
"tables": tables,
"vtctld_endpoint": vtctldEndpoint,
})
}
newTasks = append(newTasks, copySchemaTasks)
vSplitCloneTasks := NewTaskContainer()
for _, shard := range shards {
// TODO(mberlin): Add a semaphore as argument to limit the parallism.
AddTask(vSplitCloneTasks, "VerticalSplitCloneTask", map[string]string{
"dest_keyspace": destKeyspace,
"shard": shard,
"tables": tables,
"vtworker_endpoint": vtworkerEndpoint,
})
}
newTasks = append(newTasks, vSplitCloneTasks)
// TODO(mberlin): When the framework supports nesting tasks, these wait tasks should be run before each SplitDiff.
waitTasks := NewTaskContainer()
for _, shard := range shards {
AddTask(waitTasks, "WaitForFilteredReplicationTask", map[string]string{
"keyspace": destKeyspace,
"shard": shard,
"max_delay": "30s",
"vtctld_endpoint": vtctldEndpoint,
})
}
newTasks = append(newTasks, waitTasks)
// TODO(mberlin): Run all SplitDiffTasks in parallel which do not use overlapping source shards for the comparison.
for _, shard := range shards {
vSplitDiffTask := NewTaskContainer()
AddTask(vSplitDiffTask, "VerticalSplitDiffTask", map[string]string{
"dest_keyspace": destKeyspace,
"shard": shard,
"vtworker_endpoint": vtworkerEndpoint,
})
newTasks = append(newTasks, vSplitDiffTask)
}
for _, servedType := range []string{"rdonly", "replica", "master"} {
migrateServedTypesTasks := NewTaskContainer()
for _, shard := range shards {
AddTask(migrateServedTypesTasks, "MigrateServedFromTask", map[string]string{
"dest_keyspace": destKeyspace,
"shard": shard,
"type": servedType,
"vtctld_endpoint": vtctldEndpoint,
})
}
newTasks = append(newTasks, migrateServedTypesTasks)
}
return newTasks, "", nil
}