// findCellsForRebuild will find all the cells in the given keyspace
// and create an entry if the map for them
func (wr *Wrangler) findCellsForRebuild(ki *topo.KeyspaceInfo, keyspace string, shards []string, cells []string, srvKeyspaceMap map[string]*topo.SrvKeyspace) error {
er := concurrency.FirstErrorRecorder{}
mu := sync.Mutex{}
wg := sync.WaitGroup{}
for _, shard := range shards {
wg.Add(1)
go func(shard string) {
if si, err := wr.ts.GetShard(keyspace, shard); err != nil {
er.RecordError(fmt.Errorf("GetShard(%v,%v) failed: %v", keyspace, shard, err))
} else {
mu.Lock()
for _, cell := range si.Cells {
if !topo.InCellList(cell, cells) {
continue
}
if _, ok := srvKeyspaceMap[cell]; !ok {
srvKeyspaceMap[cell] = &topo.SrvKeyspace{
Shards: make([]topo.SrvShard, 0, 16),
ShardingColumnName: ki.ShardingColumnName,
ShardingColumnType: ki.ShardingColumnType,
ServedFrom: ki.ServedFrom,
}
}
}
mu.Unlock()
}
wg.Done()
}(shard)
}
wg.Wait()
return er.Error()
}
// Clone will do all the necessary actions to copy all the data from a
// source to a set of destinations.
func (wr *Wrangler) Clone(srcTabletAlias topo.TabletAlias, dstTabletAliases []topo.TabletAlias, forceMasterSnapshot bool, snapshotConcurrency, fetchConcurrency, fetchRetryCount int, serverMode bool) error {
// make sure the destination can be restored into (otherwise
// there is no point in taking the snapshot in the first place),
// and reserve it.
reserved := make([]topo.TabletAlias, 0, len(dstTabletAliases))
for _, dstTabletAlias := range dstTabletAliases {
err := wr.ReserveForRestore(srcTabletAlias, dstTabletAlias)
if err != nil {
wr.UnreserveForRestoreMulti(reserved)
return err
}
reserved = append(reserved, dstTabletAlias)
wr.Logger().Infof("Successfully reserved %v for restore", dstTabletAlias)
}
// take the snapshot, or put the server in SnapshotSource mode
// srcFilePath, parentAlias, slaveStartRequired, readWrite
sr, originalType, err := wr.Snapshot(srcTabletAlias, forceMasterSnapshot, snapshotConcurrency, serverMode)
if err != nil {
// The snapshot failed so un-reserve the destinations and return
wr.UnreserveForRestoreMulti(reserved)
return err
}
// try to restore the snapshot
// In serverMode, and in the case where we're replicating from
// the master, we can't wait for replication, as the master is down.
wg := sync.WaitGroup{}
rec := concurrency.FirstErrorRecorder{}
for _, dstTabletAlias := range dstTabletAliases {
wg.Add(1)
go func(dstTabletAlias topo.TabletAlias) {
e := wr.Restore(srcTabletAlias, sr.ManifestPath, dstTabletAlias, sr.ParentAlias, fetchConcurrency, fetchRetryCount, true, serverMode && originalType == topo.TYPE_MASTER)
rec.RecordError(e)
wg.Done()
}(dstTabletAlias)
}
wg.Wait()
err = rec.Error()
// in any case, fix the server
if serverMode {
resetErr := wr.SnapshotSourceEnd(srcTabletAlias, sr.SlaveStartRequired, sr.ReadOnly, originalType)
if resetErr != nil {
if err == nil {
// If there is no other error, this matters.
err = resetErr
} else {
// In the context of a larger failure, just log a note to cleanup.
wr.Logger().Errorf("Failed to reset snapshot source: %v - vtctl SnapshotSourceEnd is required", resetErr)
}
}
}
return err
}
// FindAllShardsInKeyspace reads and returns all the existing shards in
// a keyspace. It doesn't take any lock.
func (ts Server) FindAllShardsInKeyspace(ctx context.Context, keyspace string) (map[string]*ShardInfo, error) {
shards, err := ts.GetShardNames(ctx, keyspace)
if err != nil {
return nil, err
}
result := make(map[string]*ShardInfo, len(shards))
wg := sync.WaitGroup{}
mu := sync.Mutex{}
rec := concurrency.FirstErrorRecorder{}
for _, shard := range shards {
wg.Add(1)
go func(shard string) {
defer wg.Done()
si, err := ts.GetShard(ctx, keyspace, shard)
if err != nil {
rec.RecordError(fmt.Errorf("GetShard(%v,%v) failed: %v", keyspace, shard, err))
return
}
mu.Lock()
result[shard] = si
mu.Unlock()
}(shard)
}
wg.Wait()
if rec.HasErrors() {
return nil, rec.Error()
}
return result, nil
}
// FindAllShardsInKeyspace reads and returns all the existing shards in
// a keyspace. It doesn't take any lock.
func (ts Server) FindAllShardsInKeyspace(ctx context.Context, keyspace string) (map[string]*ShardInfo, error) {
shards, err := ts.GetShardNames(ctx, keyspace)
if err != nil {
return nil, fmt.Errorf("failed to get list of shards for keyspace '%v': %v", keyspace, err)
}
result := make(map[string]*ShardInfo, len(shards))
wg := sync.WaitGroup{}
mu := sync.Mutex{}
rec := concurrency.FirstErrorRecorder{}
for _, shard := range shards {
wg.Add(1)
go func(shard string) {
defer wg.Done()
si, err := ts.GetShard(ctx, keyspace, shard)
if err != nil {
if err == ErrNoNode {
log.Warningf("GetShard(%v,%v) returned ErrNoNode, consider checking the topology.", keyspace, shard)
} else {
rec.RecordError(fmt.Errorf("GetShard(%v,%v) failed: %v", keyspace, shard, err))
}
return
}
mu.Lock()
result[shard] = si
mu.Unlock()
}(shard)
}
wg.Wait()
if rec.HasErrors() {
return nil, rec.Error()
}
return result, nil
}
// This function should only be used with an action lock on the keyspace
// - otherwise the consistency of the serving graph data can't be
// guaranteed.
//
// Take data from the global keyspace and rebuild the local serving
// copies in each cell.
func (wr *Wrangler) rebuildKeyspace(keyspace string, cells []string) error {
log.Infof("rebuildKeyspace %v", keyspace)
ki, err := wr.ts.GetKeyspace(keyspace)
if err != nil {
// Temporary change: we try to keep going even if node
// doesn't exist
if err != topo.ErrNoNode {
return err
}
ki = topo.NewKeyspaceInfo(keyspace, &topo.Keyspace{})
}
shards, err := wr.ts.GetShardNames(keyspace)
if err != nil {
return err
}
// Rebuild all shards in parallel.
wg := sync.WaitGroup{}
er := concurrency.FirstErrorRecorder{}
for _, shard := range shards {
wg.Add(1)
go func(shard string) {
if err := wr.RebuildShardGraph(keyspace, shard, cells); err != nil {
er.RecordError(fmt.Errorf("RebuildShardGraph failed: %v/%v %v", keyspace, shard, err))
}
wg.Done()
}(shard)
}
wg.Wait()
if er.HasErrors() {
return er.Error()
}
// Scan the first shard to discover which cells need local serving data.
aliases, err := topo.FindAllTabletAliasesInShard(wr.ts, keyspace, shards[0])
if err != nil {
return err
}
// srvKeyspaceMap is a map:
// key: local keyspace {cell,keyspace}
// value: topo.SrvKeyspace object being built
srvKeyspaceMap := make(map[cellKeyspace]*topo.SrvKeyspace)
for _, alias := range aliases {
keyspaceLocation := cellKeyspace{alias.Cell, keyspace}
if _, ok := srvKeyspaceMap[keyspaceLocation]; !ok {
// before adding keyspaceLocation to the map of
// of KeyspaceByPath, we check this is a
// serving tablet. No serving tablet in shard
// 0 means we're not rebuilding the serving
// graph in that cell. This is somewhat
// expensive, but we only do it on all the
// non-serving tablets in a shard before we
// find a serving tablet.
ti, err := wr.ts.GetTablet(alias)
if err != nil {
return err
}
if !ti.IsInServingGraph() {
continue
}
srvKeyspaceMap[keyspaceLocation] = &topo.SrvKeyspace{
Shards: make([]topo.SrvShard, 0, 16),
ShardingColumnName: ki.ShardingColumnName,
ShardingColumnType: ki.ShardingColumnType,
ServedFrom: ki.ServedFrom,
}
}
}
// for each entry in the srvKeyspaceMap map, we do the following:
// - read the ShardInfo structures for each shard
// - compute the union of the db types (replica, master, ...)
// - sort the shards in the list by range
// - check the ranges are compatible (no hole, covers everything)
for ck, srvKeyspace := range srvKeyspaceMap {
keyspaceDbTypes := make(map[topo.TabletType]bool)
srvKeyspace.Partitions = make(map[topo.TabletType]*topo.KeyspacePartition)
for _, shard := range shards {
srvShard, err := wr.ts.GetSrvShard(ck.cell, ck.keyspace, shard)
if err != nil {
return err
}
for _, tabletType := range srvShard.TabletTypes {
keyspaceDbTypes[tabletType] = true
}
// for each type this shard is supposed to serve,
// add it to srvKeyspace.Partitions
for _, tabletType := range srvShard.ServedTypes {
if _, ok := srvKeyspace.Partitions[tabletType]; !ok {
srvKeyspace.Partitions[tabletType] = &topo.KeyspacePartition{
Shards: make([]topo.SrvShard, 0)}
}
srvKeyspace.Partitions[tabletType].Shards = append(srvKeyspace.Partitions[tabletType].Shards, *srvShard)
}
}
srvKeyspace.TabletTypes = make([]topo.TabletType, 0, len(keyspaceDbTypes))
for dbType := range keyspaceDbTypes {
srvKeyspace.TabletTypes = append(srvKeyspace.TabletTypes, dbType)
}
first := true
for tabletType, partition := range srvKeyspace.Partitions {
topo.SrvShardArray(partition.Shards).Sort()
// check the first Start is MinKey, the last End is MaxKey,
// and the values in between match: End[i] == Start[i+1]
if partition.Shards[0].KeyRange.Start != key.MinKey {
return fmt.Errorf("Keyspace partition for %v does not start with %v", tabletType, key.MinKey)
}
if partition.Shards[len(partition.Shards)-1].KeyRange.End != key.MaxKey {
return fmt.Errorf("Keyspace partition for %v does not end with %v", tabletType, key.MaxKey)
}
for i := range partition.Shards[0 : len(partition.Shards)-1] {
if partition.Shards[i].KeyRange.End != partition.Shards[i+1].KeyRange.Start {
return fmt.Errorf("Non-contiguous KeyRange values for %v at shard %v to %v: %v != %v", tabletType, i, i+1, partition.Shards[i].KeyRange.End.Hex(), partition.Shards[i+1].KeyRange.Start.Hex())
}
}
// backfill Shards
if first {
first = false
srvKeyspace.Shards = partition.Shards
}
}
}
// and then finally save the keyspace objects
for ck, srvKeyspace := range srvKeyspaceMap {
if err := wr.ts.UpdateSrvKeyspace(ck.cell, ck.keyspace, srvKeyspace); err != nil {
return fmt.Errorf("writing serving data failed: %v", err)
}
}
return nil
}
// This function should only be used with an action lock on the keyspace
// - otherwise the consistency of the serving graph data can't be
// guaranteed.
//
// Take data from the global keyspace and rebuild the local serving
// copies in each cell.
func (wr *Wrangler) rebuildKeyspace(keyspace string, cells []string) error {
relog.Info("rebuildKeyspace %v", keyspace)
shards, err := wr.ts.GetShardNames(keyspace)
if err != nil {
return err
}
// Rebuild all shards in parallel.
wg := sync.WaitGroup{}
er := concurrency.FirstErrorRecorder{}
for _, shard := range shards {
wg.Add(1)
go func(shard string) {
if err := wr.RebuildShardGraph(keyspace, shard, cells); err != nil {
er.RecordError(fmt.Errorf("RebuildShardGraph failed: %v/%v %v", keyspace, shard, err))
}
wg.Done()
}(shard)
}
wg.Wait()
if er.HasErrors() {
return er.Error()
}
// Scan the first shard to discover which cells need local serving data.
aliases, err := topo.FindAllTabletAliasesInShard(wr.ts, keyspace, shards[0])
if err != nil {
return err
}
// srvKeyspaceByPath is a map:
// key: local keyspace {cell,keyspace}
// value: topo.SrvKeyspace object being built
srvKeyspaceByPath := make(map[cellKeyspace]*topo.SrvKeyspace)
for _, alias := range aliases {
keyspaceLocation := cellKeyspace{alias.Cell, keyspace}
if _, ok := srvKeyspaceByPath[keyspaceLocation]; !ok {
// before adding keyspaceLocation to the map of
// of KeyspaceByPath, we check this is a
// serving tablet. No serving tablet in shard
// 0 means we're not rebuilding the serving
// graph in that cell. This is somewhat
// expensive, but we only do it on all the
// non-serving tablets in a shard before we
// find a serving tablet.
ti, err := wr.ts.GetTablet(alias)
if err != nil {
return err
}
if !ti.IsServingType() {
continue
}
srvKeyspaceByPath[keyspaceLocation] = &topo.SrvKeyspace{Shards: make([]topo.SrvShard, 0, 16)}
}
}
// for each entry in the srvKeyspaceByPath map, we do the following:
// - read the ShardInfo structures for each shard
// - prune the AddrsByType field, result would be too big
// - compute the union of the db types (replica, master, ...)
// - sort the shards in the list by range
// - check the ranges are compatible (no hole, covers everything)
for srvPath, srvKeyspace := range srvKeyspaceByPath {
keyspaceDbTypes := make(map[topo.TabletType]bool)
for _, shard := range shards {
srvShard, err := wr.ts.GetSrvShard(srvPath.cell, srvPath.keyspace, shard)
if err != nil {
return err
}
for dbType, _ := range srvShard.AddrsByType {
keyspaceDbTypes[topo.TabletType(dbType)] = true
}
// Prune addrs, this is unnecessarily expensive right now. It is easier to
// load on-demand since we have to do that anyway on a reconnect.
srvShard.AddrsByType = nil
srvKeyspace.Shards = append(srvKeyspace.Shards, *srvShard)
}
tabletTypes := make([]topo.TabletType, 0, len(keyspaceDbTypes))
for dbType, _ := range keyspaceDbTypes {
tabletTypes = append(tabletTypes, dbType)
}
srvKeyspace.TabletTypes = tabletTypes
// FIXME(msolomon) currently this only works when the shards are range-based
topo.SrvShardArray(srvKeyspace.Shards).Sort()
// check the first Start is MinKey, the last End is MaxKey,
// and the values in between match: End[i] == Start[i+1]
if srvKeyspace.Shards[0].KeyRange.Start != key.MinKey {
return fmt.Errorf("Keyspace does not start with %v", key.MinKey)
}
if srvKeyspace.Shards[len(srvKeyspace.Shards)-1].KeyRange.End != key.MaxKey {
return fmt.Errorf("Keyspace does not end with %v", key.MaxKey)
}
for i, _ := range srvKeyspace.Shards[0 : len(srvKeyspace.Shards)-1] {
if srvKeyspace.Shards[i].KeyRange.End != srvKeyspace.Shards[i+1].KeyRange.Start {
return fmt.Errorf("Non-contiguous KeyRange values at shard %v to %v: %v != %v", i, i+1, srvKeyspace.Shards[i].KeyRange.End.Hex(), srvKeyspace.Shards[i+1].KeyRange.Start.Hex())
}
}
}
// and then finally save the keyspace objects
for srvPath, srvKeyspace := range srvKeyspaceByPath {
if err := wr.ts.UpdateSrvKeyspace(srvPath.cell, srvPath.keyspace, srvKeyspace); err != nil {
return fmt.Errorf("writing serving data failed: %v", err)
}
}
return nil
}
// This function should only be used with an action lock on the keyspace
// - otherwise the consistency of the serving graph data can't be
// guaranteed.
//
// Take data from the global keyspace and rebuild the local serving
// copies in each cell.
func (wr *Wrangler) rebuildKeyspace(keyspace string, cells []string, shardCache map[string]*topo.ShardInfo) error {
wr.logger.Infof("rebuildKeyspace %v", keyspace)
ki, err := wr.ts.GetKeyspace(keyspace)
if err != nil {
return err
}
shards, err := wr.ts.GetShardNames(keyspace)
if err != nil {
return err
}
// Rebuild all shards in parallel.
wg := sync.WaitGroup{}
er := concurrency.FirstErrorRecorder{}
for _, shard := range shards {
wg.Add(1)
go func(shard string) {
if err := wr.RebuildShardGraph(keyspace, shard, cells); err != nil {
er.RecordError(fmt.Errorf("RebuildShardGraph failed: %v/%v %v", keyspace, shard, err))
}
wg.Done()
}(shard)
}
wg.Wait()
if er.HasErrors() {
return er.Error()
}
// Build the list of cells to work on: we get the union
// of all the Cells of all the Shards, limited to the provided cells.
//
// srvKeyspaceMap is a map:
// key: cell
// value: topo.SrvKeyspace object being built
srvKeyspaceMap := make(map[string]*topo.SrvKeyspace)
if err := wr.findCellsForRebuild(ki, keyspace, shards, cells, srvKeyspaceMap); err != nil {
return err
}
// Then we add the cells from the keyspaces we might be 'ServedFrom'.
for _, servedFrom := range ki.ServedFrom {
servedFromShards, err := wr.ts.GetShardNames(servedFrom)
if err != nil {
return err
}
if err := wr.findCellsForRebuild(ki, servedFrom, servedFromShards, cells, srvKeyspaceMap); err != nil {
return err
}
}
// for each entry in the srvKeyspaceMap map, we do the following:
// - read the SrvShard structures for each shard / cell
// - if not present, build an empty one from global Shard
// - compute the union of the db types (replica, master, ...)
// - sort the shards in the list by range
// - check the ranges are compatible (no hole, covers everything)
if shardCache == nil {
shardCache = make(map[string]*topo.ShardInfo)
}
for cell, srvKeyspace := range srvKeyspaceMap {
keyspaceDbTypes := make(map[topo.TabletType]bool)
srvKeyspace.Partitions = make(map[topo.TabletType]*topo.KeyspacePartition)
for _, shard := range shards {
srvShard, err := wr.ts.GetSrvShard(cell, keyspace, shard)
switch err {
case nil:
// we keep going
case topo.ErrNoNode:
wr.logger.Infof("Cell %v for %v/%v has no SvrShard, using Shard data with no TabletTypes instead", cell, keyspace, shard)
si, ok := shardCache[shard]
if !ok {
si, err = wr.ts.GetShard(keyspace, shard)
if err != nil {
return fmt.Errorf("GetShard(%v, %v) (backup for GetSrvShard in cell %v) failed: %v", keyspace, shard, cell, err)
}
shardCache[shard] = si
}
srvShard = &topo.SrvShard{
Name: si.ShardName(),
KeyRange: si.KeyRange,
ServedTypes: si.ServedTypes,
MasterCell: si.MasterAlias.Cell,
}
default:
return err
}
for _, tabletType := range srvShard.TabletTypes {
keyspaceDbTypes[tabletType] = true
}
// for each type this shard is supposed to serve,
// add it to srvKeyspace.Partitions
for _, tabletType := range srvShard.ServedTypes {
if _, ok := srvKeyspace.Partitions[tabletType]; !ok {
srvKeyspace.Partitions[tabletType] = &topo.KeyspacePartition{
Shards: make([]topo.SrvShard, 0)}
}
srvKeyspace.Partitions[tabletType].Shards = append(srvKeyspace.Partitions[tabletType].Shards, *srvShard)
}
}
srvKeyspace.TabletTypes = make([]topo.TabletType, 0, len(keyspaceDbTypes))
for dbType := range keyspaceDbTypes {
srvKeyspace.TabletTypes = append(srvKeyspace.TabletTypes, dbType)
}
first := true
for tabletType, partition := range srvKeyspace.Partitions {
topo.SrvShardArray(partition.Shards).Sort()
// check the first Start is MinKey, the last End is MaxKey,
// and the values in between match: End[i] == Start[i+1]
if partition.Shards[0].KeyRange.Start != key.MinKey {
return fmt.Errorf("keyspace partition for %v in cell %v does not start with %v", tabletType, cell, key.MinKey)
}
if partition.Shards[len(partition.Shards)-1].KeyRange.End != key.MaxKey {
return fmt.Errorf("keyspace partition for %v in cell %v does not end with %v", tabletType, cell, key.MaxKey)
}
for i := range partition.Shards[0 : len(partition.Shards)-1] {
if partition.Shards[i].KeyRange.End != partition.Shards[i+1].KeyRange.Start {
return fmt.Errorf("non-contiguous KeyRange values for %v in cell %v at shard %v to %v: %v != %v", tabletType, cell, i, i+1, partition.Shards[i].KeyRange.End.Hex(), partition.Shards[i+1].KeyRange.Start.Hex())
}
}
// backfill Shards
if first {
first = false
srvKeyspace.Shards = partition.Shards
}
}
}
// and then finally save the keyspace objects
for cell, srvKeyspace := range srvKeyspaceMap {
wr.logger.Infof("updating keyspace serving graph in cell %v for %v", cell, keyspace)
if err := wr.ts.UpdateSrvKeyspace(cell, keyspace, srvKeyspace); err != nil {
return fmt.Errorf("writing serving data failed: %v", err)
}
}
return nil
}
// rebuildKeyspace should only be used with an action lock on the keyspace
// - otherwise the consistency of the serving graph data can't be
// guaranteed.
//
// Take data from the global keyspace and rebuild the local serving
// copies in each cell.
func rebuildKeyspace(ctx context.Context, log logutil.Logger, ts topo.Server, keyspace string, cells []string, rebuildSrvShards bool) error {
log.Infof("rebuildKeyspace %v", keyspace)
ki, err := ts.GetKeyspace(ctx, keyspace)
if err != nil {
return err
}
var shardCache map[string]*topo.ShardInfo
if rebuildSrvShards {
shards, err := ts.GetShardNames(ctx, keyspace)
if err != nil {
return nil
}
// Rebuild all shards in parallel, save the shards
shardCache = make(map[string]*topo.ShardInfo)
wg := sync.WaitGroup{}
mu := sync.Mutex{}
rec := concurrency.FirstErrorRecorder{}
for _, shard := range shards {
wg.Add(1)
go func(shard string) {
if shardInfo, err := RebuildShard(ctx, log, ts, keyspace, shard, cells); err != nil {
rec.RecordError(fmt.Errorf("RebuildShard failed: %v/%v %v", keyspace, shard, err))
} else {
mu.Lock()
shardCache[shard] = shardInfo
mu.Unlock()
}
wg.Done()
}(shard)
}
wg.Wait()
if rec.HasErrors() {
return rec.Error()
}
} else {
shardCache, err = ts.FindAllShardsInKeyspace(ctx, keyspace)
if err != nil {
return err
}
}
// Build the list of cells to work on: we get the union
// of all the Cells of all the Shards, limited to the provided cells.
//
// srvKeyspaceMap is a map:
// key: cell
// value: topo.SrvKeyspace object being built
srvKeyspaceMap := make(map[string]*topodatapb.SrvKeyspace)
findCellsForRebuild(ki, shardCache, cells, srvKeyspaceMap)
// Then we add the cells from the keyspaces we might be 'ServedFrom'.
for _, ksf := range ki.ServedFroms {
servedFromShards, err := ts.FindAllShardsInKeyspace(ctx, ksf.Keyspace)
if err != nil {
return err
}
findCellsForRebuild(ki, servedFromShards, cells, srvKeyspaceMap)
}
// for each entry in the srvKeyspaceMap map, we do the following:
// - read the SrvShard structures for each shard / cell
// - if not present, build an empty one from global Shard
// - compute the union of the db types (replica, master, ...)
// - sort the shards in the list by range
// - check the ranges are compatible (no hole, covers everything)
for cell, srvKeyspace := range srvKeyspaceMap {
for _, si := range shardCache {
servedTypes := si.GetServedTypesPerCell(cell)
// for each type this shard is supposed to serve,
// add it to srvKeyspace.Partitions
for _, tabletType := range servedTypes {
partition := topoproto.SrvKeyspaceGetPartition(srvKeyspace, tabletType)
if partition == nil {
partition = &topodatapb.SrvKeyspace_KeyspacePartition{
ServedType: tabletType,
}
srvKeyspace.Partitions = append(srvKeyspace.Partitions, partition)
}
partition.ShardReferences = append(partition.ShardReferences, &topodatapb.ShardReference{
Name: si.ShardName(),
KeyRange: si.KeyRange,
})
}
}
if err := orderAndCheckPartitions(cell, srvKeyspace); err != nil {
return err
}
}
// and then finally save the keyspace objects
for cell, srvKeyspace := range srvKeyspaceMap {
log.Infof("updating keyspace serving graph in cell %v for %v", cell, keyspace)
if err := ts.UpdateSrvKeyspace(ctx, cell, keyspace, srvKeyspace); err != nil {
return fmt.Errorf("writing serving data failed: %v", err)
}
}
return nil
}
// This function should only be used with an action lock on the keyspace
// - otherwise the consistency of the serving graph data can't be
// guaranteed.
//
// Take data from the global keyspace and rebuild the local serving
// copies in each cell.
func (wr *Wrangler) rebuildKeyspace(keyspace string, cells []string) error {
wr.logger.Infof("rebuildKeyspace %v", keyspace)
ki, err := wr.ts.GetKeyspace(keyspace)
if err != nil {
return err
}
shards, err := wr.ts.GetShardNames(keyspace)
if err != nil {
return err
}
// Rebuild all shards in parallel, save the shards
shardCache := make(map[string]*topo.ShardInfo)
wg := sync.WaitGroup{}
mu := sync.Mutex{}
rec := concurrency.FirstErrorRecorder{}
for _, shard := range shards {
wg.Add(1)
go func(shard string) {
if shardInfo, err := wr.RebuildShardGraph(keyspace, shard, cells); err != nil {
rec.RecordError(fmt.Errorf("RebuildShardGraph failed: %v/%v %v", keyspace, shard, err))
} else {
mu.Lock()
shardCache[shard] = shardInfo
mu.Unlock()
}
wg.Done()
}(shard)
}
wg.Wait()
if rec.HasErrors() {
return rec.Error()
}
// Build the list of cells to work on: we get the union
// of all the Cells of all the Shards, limited to the provided cells.
//
// srvKeyspaceMap is a map:
// key: cell
// value: topo.SrvKeyspace object being built
srvKeyspaceMap := make(map[string]*topo.SrvKeyspace)
wr.findCellsForRebuild(ki, shardCache, cells, srvKeyspaceMap)
// Then we add the cells from the keyspaces we might be 'ServedFrom'.
for _, ksf := range ki.ServedFromMap {
servedFromShards, err := topo.FindAllShardsInKeyspace(wr.ts, ksf.Keyspace)
if err != nil {
return err
}
wr.findCellsForRebuild(ki, servedFromShards, cells, srvKeyspaceMap)
}
// for each entry in the srvKeyspaceMap map, we do the following:
// - read the SrvShard structures for each shard / cell
// - if not present, build an empty one from global Shard
// - compute the union of the db types (replica, master, ...)
// - sort the shards in the list by range
// - check the ranges are compatible (no hole, covers everything)
for cell, srvKeyspace := range srvKeyspaceMap {
keyspaceDbTypes := make(map[topo.TabletType]bool)
srvKeyspace.Partitions = make(map[topo.TabletType]*topo.KeyspacePartition)
for shard, si := range shardCache {
srvShard, err := wr.ts.GetSrvShard(cell, keyspace, shard)
switch err {
case nil:
// we keep going
case topo.ErrNoNode:
wr.logger.Infof("Cell %v for %v/%v has no SvrShard, using Shard data with no TabletTypes instead", cell, keyspace, shard)
srvShard = &topo.SrvShard{
Name: si.ShardName(),
KeyRange: si.KeyRange,
ServedTypes: si.GetServedTypesPerCell(cell),
MasterCell: si.MasterAlias.Cell,
}
default:
return err
}
for _, tabletType := range srvShard.TabletTypes {
keyspaceDbTypes[tabletType] = true
}
// for each type this shard is supposed to serve,
// add it to srvKeyspace.Partitions
for _, tabletType := range srvShard.ServedTypes {
if _, ok := srvKeyspace.Partitions[tabletType]; !ok {
srvKeyspace.Partitions[tabletType] = &topo.KeyspacePartition{
Shards: make([]topo.SrvShard, 0)}
}
srvKeyspace.Partitions[tabletType].Shards = append(srvKeyspace.Partitions[tabletType].Shards, *srvShard)
}
}
srvKeyspace.TabletTypes = make([]topo.TabletType, 0, len(keyspaceDbTypes))
for dbType := range keyspaceDbTypes {
srvKeyspace.TabletTypes = append(srvKeyspace.TabletTypes, dbType)
}
if err := wr.checkPartitions(cell, srvKeyspace); err != nil {
return err
}
}
// and then finally save the keyspace objects
for cell, srvKeyspace := range srvKeyspaceMap {
wr.logger.Infof("updating keyspace serving graph in cell %v for %v", cell, keyspace)
if err := wr.ts.UpdateSrvKeyspace(cell, keyspace, srvKeyspace); err != nil {
return fmt.Errorf("writing serving data failed: %v", err)
}
}
return nil
}