// TestStoreRangeSplitInsideRow verifies an attempt to split a range inside of
// a table row will cause a split at a boundary between rows.
func TestStoreRangeSplitInsideRow(t *testing.T) {
defer leaktest.AfterTest(t)()
sCtx := storage.TestStoreContext()
sCtx.TestingKnobs.DisableSplitQueue = true
store, stopper, _ := createTestStoreWithContext(t, sCtx)
defer stopper.Stop()
// Manually create some the column keys corresponding to the table:
//
// CREATE TABLE t (id STRING PRIMARY KEY, col1 INT, col2 INT)
tableKey := keys.MakeTablePrefix(keys.MaxReservedDescID + 1)
rowKey := roachpb.Key(encoding.EncodeVarintAscending(append([]byte(nil), tableKey...), 1))
rowKey = encoding.EncodeStringAscending(encoding.EncodeVarintAscending(rowKey, 1), "a")
col1Key := keys.MakeFamilyKey(append([]byte(nil), rowKey...), 1)
col2Key := keys.MakeFamilyKey(append([]byte(nil), rowKey...), 2)
// We don't care about the value, so just store any old thing.
if err := store.DB().Put(col1Key, "column 1"); err != nil {
t.Fatal(err)
}
if err := store.DB().Put(col2Key, "column 2"); err != nil {
t.Fatal(err)
}
// Split between col1Key and col2Key by splitting before col2Key.
args := adminSplitArgs(col2Key, col2Key)
_, err := client.SendWrapped(rg1(store), nil, &args)
if err != nil {
t.Fatalf("%s: split unexpected error: %s", col1Key, err)
}
rng1 := store.LookupReplica(col1Key, nil)
rng2 := store.LookupReplica(col2Key, nil)
// Verify the two columns are still on the same range.
if !reflect.DeepEqual(rng1, rng2) {
t.Fatalf("%s: ranges differ: %+v vs %+v", roachpb.Key(col1Key), rng1, rng2)
}
// Verify we split on a row key.
if startKey := rng1.Desc().StartKey; !startKey.Equal(rowKey) {
t.Fatalf("%s: expected split on %s, but found %s",
roachpb.Key(col1Key), roachpb.Key(rowKey), startKey)
}
// Verify the previous range was split on a row key.
rng3 := store.LookupReplica(tableKey, nil)
if endKey := rng3.Desc().EndKey; !endKey.Equal(rowKey) {
t.Fatalf("%s: expected split on %s, but found %s",
roachpb.Key(col1Key), roachpb.Key(rowKey), endKey)
}
}
// MakeNameMetadataKey returns the key for the name. Pass name == "" in order
// to generate the prefix key to use to scan over all of the names for the
// specified parentID.
func MakeNameMetadataKey(parentID ID, name string) roachpb.Key {
name = NormalizeName(name)
k := keys.MakeTablePrefix(uint32(namespaceTable.ID))
k = encoding.EncodeUvarintAscending(k, uint64(namespaceTable.PrimaryIndex.ID))
k = encoding.EncodeUvarintAscending(k, uint64(parentID))
if name != "" {
k = encoding.EncodeBytesAscending(k, []byte(name))
k = keys.MakeFamilyKey(k, uint32(namespaceTable.Columns[2].ID))
}
return k
}
// updateRow adds to the batch the kv operations necessary to update a table row
// with the given values.
//
// The row corresponding to oldValues is updated with the ones in updateValues.
// Note that updateValues only contains the ones that are changing.
//
// The return value is only good until the next call to UpdateRow.
func (ru *rowUpdater) updateRow(
b *client.Batch,
oldValues []parser.Datum,
updateValues []parser.Datum,
) ([]parser.Datum, error) {
if len(oldValues) != len(ru.fetchCols) {
return nil, errors.Errorf("got %d values but expected %d", len(oldValues), len(ru.fetchCols))
}
if len(updateValues) != len(ru.updateCols) {
return nil, errors.Errorf("got %d values but expected %d", len(updateValues), len(ru.updateCols))
}
primaryIndexKey, secondaryIndexEntries, err := ru.helper.encodeIndexes(ru.fetchColIDtoRowIndex, oldValues)
if err != nil {
return nil, err
}
// The secondary index entries returned by rowHelper.encodeIndexes are only
// valid until the next call to encodeIndexes. We need to copy them so that
// we can compare against the new secondary index entries.
secondaryIndexEntries = append(ru.indexEntriesBuf[:0], secondaryIndexEntries...)
ru.indexEntriesBuf = secondaryIndexEntries
// Check that the new value types match the column types. This needs to
// happen before index encoding because certain datum types (i.e. tuple)
// cannot be used as index values.
for i, val := range updateValues {
if ru.marshalled[i], err = sqlbase.MarshalColumnValue(ru.updateCols[i], val); err != nil {
return nil, err
}
}
// Update the row values.
copy(ru.newValues, oldValues)
for i, updateCol := range ru.updateCols {
ru.newValues[ru.fetchColIDtoRowIndex[updateCol.ID]] = updateValues[i]
}
rowPrimaryKeyChanged := false
var newSecondaryIndexEntries []sqlbase.IndexEntry
if ru.primaryKeyColChange {
var newPrimaryIndexKey []byte
newPrimaryIndexKey, newSecondaryIndexEntries, err =
ru.helper.encodeIndexes(ru.fetchColIDtoRowIndex, ru.newValues)
if err != nil {
return nil, err
}
rowPrimaryKeyChanged = !bytes.Equal(primaryIndexKey, newPrimaryIndexKey)
} else {
newSecondaryIndexEntries, err =
ru.helper.encodeSecondaryIndexes(ru.fetchColIDtoRowIndex, ru.newValues)
if err != nil {
return nil, err
}
}
if rowPrimaryKeyChanged {
if err := ru.fks.checkIdx(ru.helper.tableDesc.PrimaryIndex.ID, oldValues, ru.newValues); err != nil {
return nil, err
}
for i := range newSecondaryIndexEntries {
if !bytes.Equal(newSecondaryIndexEntries[i].Key, secondaryIndexEntries[i].Key) {
if err := ru.fks.checkIdx(ru.helper.indexes[i].ID, oldValues, ru.newValues); err != nil {
return nil, err
}
}
}
if err := ru.rd.deleteRow(b, oldValues); err != nil {
return nil, err
}
if err := ru.ri.insertRow(b, ru.newValues, false); err != nil {
return nil, err
}
return ru.newValues, nil
}
// Add the new values.
// TODO(dan): This has gotten very similar to the loop in insertRow, see if
// they can be DRY'd. Ideally, this would also work for
// truncateAndBackfillColumnsChunk, which is currently abusing rowUdpdater.
for i, family := range ru.helper.tableDesc.Families {
update := false
for _, colID := range family.ColumnIDs {
if _, ok := ru.updateColIDtoRowIndex[colID]; ok {
update = true
break
}
}
if !update {
continue
}
if i > 0 {
// HACK: MakeFamilyKey appends to its argument, so on every loop iteration
// after the first, trim primaryIndexKey so nothing gets overwritten.
// TODO(dan): Instead of this, use something like engine.ChunkAllocator.
primaryIndexKey = primaryIndexKey[:len(primaryIndexKey):len(primaryIndexKey)]
}
if len(family.ColumnIDs) == 1 && family.ColumnIDs[0] == family.DefaultColumnID {
// Storage optimization to store DefaultColumnID directly as a value. Also
// backwards compatible with the original BaseFormatVersion.
idx, ok := ru.updateColIDtoRowIndex[family.DefaultColumnID]
if !ok {
continue
}
ru.key = keys.MakeFamilyKey(primaryIndexKey, uint32(family.ID))
if log.V(2) {
log.Infof("Put %s -> %v", ru.key, ru.marshalled[idx].PrettyPrint())
}
b.Put(&ru.key, &ru.marshalled[idx])
ru.key = nil
continue
}
ru.key = keys.MakeFamilyKey(primaryIndexKey, uint32(family.ID))
ru.valueBuf = ru.valueBuf[:0]
var lastColID sqlbase.ColumnID
familySortedColumnIDs, ok := ru.helper.sortedColumnFamily(family.ID)
if !ok {
panic("invalid family sorted column id map")
}
for _, colID := range familySortedColumnIDs {
if ru.helper.columnInPK(colID) {
if family.ID != 0 {
return nil, errors.Errorf("primary index column %d must be in family 0, was %d", colID, family.ID)
}
// Skip primary key columns as their values are encoded in the key of
// each family. Family 0 is guaranteed to exist and acts as a sentinel.
continue
}
idx, ok := ru.fetchColIDtoRowIndex[colID]
if !ok {
return nil, errors.Errorf("column %d was expected to be fetched, but wasn't", colID)
}
col := ru.fetchCols[idx]
if ru.newValues[idx].Compare(parser.DNull) == 0 {
continue
}
if lastColID > col.ID {
panic(fmt.Errorf("cannot write column id %d after %d", col.ID, lastColID))
}
colIDDiff := col.ID - lastColID
lastColID = col.ID
ru.valueBuf, err = sqlbase.EncodeTableValue(ru.valueBuf, colIDDiff, ru.newValues[idx])
if err != nil {
return nil, err
}
}
if family.ID != 0 && len(ru.valueBuf) == 0 {
// The family might have already existed but every column in it is being
// set to NULL, so delete it.
if log.V(2) {
log.Infof("Del %s", ru.key)
}
b.Del(&ru.key)
} else {
ru.value.SetTuple(ru.valueBuf)
if log.V(2) {
log.Infof("Put %s -> %v", ru.key, ru.value.PrettyPrint())
}
b.Put(&ru.key, &ru.value)
}
ru.key = nil
}
// Update secondary indexes.
for i, newSecondaryIndexEntry := range newSecondaryIndexEntries {
secondaryIndexEntry := secondaryIndexEntries[i]
secondaryKeyChanged := !bytes.Equal(newSecondaryIndexEntry.Key, secondaryIndexEntry.Key)
if secondaryKeyChanged {
if err := ru.fks.checkIdx(ru.helper.indexes[i].ID, oldValues, ru.newValues); err != nil {
return nil, err
}
if log.V(2) {
log.Infof("Del %s", secondaryIndexEntry.Key)
}
b.Del(secondaryIndexEntry.Key)
// Do not update Indexes in the DELETE_ONLY state.
if _, ok := ru.deleteOnlyIndex[i]; !ok {
if log.V(2) {
log.Infof("CPut %s -> %v", newSecondaryIndexEntry.Key, newSecondaryIndexEntry.Value.PrettyPrint())
}
b.CPut(newSecondaryIndexEntry.Key, &newSecondaryIndexEntry.Value, nil)
}
}
}
return ru.newValues, nil
}
// insertRow adds to the batch the kv operations necessary to insert a table row
// with the given values.
func (ri *rowInserter) insertRow(b *client.Batch, values []parser.Datum, ignoreConflicts bool) error {
if len(values) != len(ri.insertCols) {
return errors.Errorf("got %d values but expected %d", len(values), len(ri.insertCols))
}
putFn := insertCPutFn
if ignoreConflicts {
putFn = insertPutFn
}
// Encode the values to the expected column type. This needs to
// happen before index encoding because certain datum types (i.e. tuple)
// cannot be used as index values.
for i, val := range values {
// Make sure the value can be written to the column before proceeding.
var err error
if ri.marshalled[i], err = sqlbase.MarshalColumnValue(ri.insertCols[i], val); err != nil {
return err
}
}
if err := ri.fks.checkAll(values); err != nil {
return err
}
primaryIndexKey, secondaryIndexEntries, err := ri.helper.encodeIndexes(ri.insertColIDtoRowIndex, values)
if err != nil {
return err
}
// Add the new values.
// TODO(dan): This has gotten very similar to the loop in updateRow, see if
// they can be DRY'd. Ideally, this would also work for
// truncateAndBackfillColumnsChunk, which is currently abusing rowUdpdater.
for i, family := range ri.helper.tableDesc.Families {
if i > 0 {
// HACK: MakeFamilyKey appends to its argument, so on every loop iteration
// after the first, trim primaryIndexKey so nothing gets overwritten.
// TODO(dan): Instead of this, use something like engine.ChunkAllocator.
primaryIndexKey = primaryIndexKey[:len(primaryIndexKey):len(primaryIndexKey)]
}
if len(family.ColumnIDs) == 1 && family.ColumnIDs[0] == family.DefaultColumnID {
// Storage optimization to store DefaultColumnID directly as a value. Also
// backwards compatible with the original BaseFormatVersion.
idx, ok := ri.insertColIDtoRowIndex[family.DefaultColumnID]
if !ok {
continue
}
if ri.marshalled[idx].RawBytes != nil {
// We only output non-NULL values. Non-existent column keys are
// considered NULL during scanning and the row sentinel ensures we know
// the row exists.
ri.key = keys.MakeFamilyKey(primaryIndexKey, uint32(family.ID))
putFn(b, &ri.key, &ri.marshalled[idx])
ri.key = nil
}
continue
}
ri.key = keys.MakeFamilyKey(primaryIndexKey, uint32(family.ID))
ri.valueBuf = ri.valueBuf[:0]
var lastColID sqlbase.ColumnID
familySortedColumnIDs, ok := ri.helper.sortedColumnFamily(family.ID)
if !ok {
panic("invalid family sorted column id map")
}
for _, colID := range familySortedColumnIDs {
if ri.helper.columnInPK(colID) {
if family.ID != 0 {
return errors.Errorf("primary index column %d must be in family 0, was %d", colID, family.ID)
}
// Skip primary key columns as their values are encoded in the key of
// each family. Family 0 is guaranteed to exist and acts as a sentinel.
continue
}
idx, ok := ri.insertColIDtoRowIndex[colID]
if !ok {
// Column not being inserted.
continue
}
col := ri.insertCols[idx]
if values[idx].Compare(parser.DNull) == 0 {
continue
}
if lastColID > col.ID {
panic(fmt.Errorf("cannot write column id %d after %d", col.ID, lastColID))
}
colIDDiff := col.ID - lastColID
lastColID = col.ID
ri.valueBuf, err = sqlbase.EncodeTableValue(ri.valueBuf, colIDDiff, values[idx])
if err != nil {
return err
}
}
if family.ID == 0 || len(ri.valueBuf) > 0 {
ri.value.SetTuple(ri.valueBuf)
putFn(b, &ri.key, &ri.value)
}
ri.key = nil
}
for i := range secondaryIndexEntries {
e := &secondaryIndexEntries[i]
putFn(b, &e.Key, &e.Value)
}
return nil
}
// MakeDescMetadataKey returns the key for the descriptor.
func MakeDescMetadataKey(descID ID) roachpb.Key {
k := keys.MakeTablePrefix(uint32(DescriptorTable.ID))
k = encoding.EncodeUvarintAscending(k, uint64(DescriptorTable.PrimaryIndex.ID))
k = encoding.EncodeUvarintAscending(k, uint64(descID))
return keys.MakeFamilyKey(k, uint32(DescriptorTable.Columns[1].ID))
}
// MakeZoneKey returns the key for 'id's entry in the system.zones table.
func MakeZoneKey(id ID) roachpb.Key {
k := keys.MakeTablePrefix(uint32(zonesTable.ID))
k = encoding.EncodeUvarintAscending(k, uint64(zonesTable.PrimaryIndex.ID))
k = encoding.EncodeUvarintAscending(k, uint64(id))
return keys.MakeFamilyKey(k, uint32(zonesTable.Columns[1].ID))
}
// MakeDescMetadataKey returns the key for the descriptor.
func MakeDescMetadataKey(descID ID) roachpb.Key {
k := MakeAllDescsMetadataKey()
k = encoding.EncodeUvarintAscending(k, uint64(descID))
return keys.MakeFamilyKey(k, uint32(DescriptorTable.Columns[1].ID))
}