示例#1
0
// 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)()
	defer config.TestingDisableTableSplits()()
	store, stopper, _ := createTestStore(t)
	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.MakeColumnKey(append([]byte(nil), rowKey...), 1)
	col2Key := keys.MakeColumnKey(append([]byte(nil), rowKey...), 2)

	// We don't care about the value, so just store any old thing.
	if pErr := store.DB().Put(col1Key, "column 1"); pErr != nil {
		t.Fatal(pErr)
	}
	if pErr := store.DB().Put(col2Key, "column 2"); pErr != nil {
		t.Fatal(pErr)
	}

	// Split between col1Key and col2Key by splitting before col2Key.
	args := adminSplitArgs(col2Key, col2Key)
	_, pErr := client.SendWrapped(rg1(store), nil, &args)
	if pErr != nil {
		t.Fatalf("%s: split unexpected error: %s", col1Key, pErr)
	}

	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)
	}
}
示例#2
0
// 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.MakeColumnKey(k, uint32(namespaceTable.Columns[2].ID))
	}
	return k
}
示例#3
0
// 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.MakeColumnKey(k, uint32(descriptorTable.Columns[1].ID))
}
示例#4
0
// 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.MakeColumnKey(k, uint32(zonesTable.Columns[1].ID))
}
示例#5
0
// 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) error {
	if len(values) != len(ri.insertCols) {
		return util.Errorf("got %d values but expected %d", len(values), len(ri.insertCols))
	}

	// 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
		}
	}

	primaryIndexKey, secondaryIndexEntries, err := ri.helper.encodeIndexes(ri.insertColIDtoRowIndex, values)
	if err != nil {
		return err
	}

	// Write the row sentinel. We want to write the sentinel first in case
	// we are trying to insert a duplicate primary key: if we write the
	// secondary indexes first, we may get an error that looks like a
	// uniqueness violation on a non-unique index.
	ri.key = keys.MakeNonColumnKey(primaryIndexKey)
	if log.V(2) {
		log.Infof("CPut %s -> NULL", ri.key)
	}
	// Each sentinel value needs a distinct RawBytes field as the computed
	// checksum includes the key the value is associated with.
	ri.sentinelValue.SetBytes([]byte{})
	b.CPut(&ri.key, &ri.sentinelValue, nil)
	ri.key = nil

	for _, secondaryIndexEntry := range secondaryIndexEntries {
		if log.V(2) {
			log.Infof("CPut %s -> %v", secondaryIndexEntry.Key, secondaryIndexEntry.Value)
		}
		ri.key = secondaryIndexEntry.Key
		b.CPut(&ri.key, secondaryIndexEntry.Value, nil)
	}
	ri.key = nil

	// Write the row columns.
	for i, val := range values {
		col := ri.insertCols[i]

		if ri.helper.columnInPK(col.ID) {
			// Skip primary key columns as their values are encoded in the row
			// sentinel key which is guaranteed to exist for as long as the row
			// exists.
			continue
		}

		if ri.marshalled[i].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.MakeColumnKey(primaryIndexKey, uint32(col.ID))
			if log.V(2) {
				log.Infof("CPut %s -> %v", ri.key, val)
			}

			b.CPut(&ri.key, &ri.marshalled[i], nil)
			ri.key = nil
		}
	}

	return nil
}
示例#6
0
// 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, util.Errorf("got %d values but expected %d", len(oldValues), len(ru.fetchCols))
	}
	if len(updateValues) != len(ru.updateCols) {
		return nil, util.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
	}

	// 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]
	}

	newPrimaryIndexKey := primaryIndexKey
	rowPrimaryKeyChanged := false
	var newSecondaryIndexEntries []sqlbase.IndexEntry
	if ru.primaryKeyColChange {
		newPrimaryIndexKey, newSecondaryIndexEntries, err = ru.helper.encodeIndexes(ru.fetchColIDtoRowIndex, ru.newValues)
		if err != nil {
			return nil, err
		}
		rowPrimaryKeyChanged = !bytes.Equal(primaryIndexKey, newPrimaryIndexKey)
	} else {
		newSecondaryIndexEntries, err = sqlbase.EncodeSecondaryIndexes(
			ru.helper.tableDesc.ID, ru.helper.indexes, ru.fetchColIDtoRowIndex, ru.newValues)
		if err != nil {
			return nil, err
		}
	}

	if rowPrimaryKeyChanged {
		err := ru.rd.deleteRow(b, oldValues)
		if err != nil {
			return nil, err
		}
		err = ru.ri.insertRow(b, ru.newValues)
		return ru.newValues, err
	}

	// Update secondary indexes.
	for i, newSecondaryIndexEntry := range newSecondaryIndexEntries {
		secondaryIndexEntry := secondaryIndexEntries[i]
		secondaryKeyChanged := !bytes.Equal(newSecondaryIndexEntry.Key, secondaryIndexEntry.Key)
		if secondaryKeyChanged {
			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)
				}
				b.CPut(newSecondaryIndexEntry.Key, newSecondaryIndexEntry.Value, nil)
			}
		}
	}

	// Add the new values.
	for i, val := range updateValues {
		col := ru.updateCols[i]

		if ru.helper.columnInPK(col.ID) {
			// Skip primary key columns as their values are encoded in the row
			// sentinel key which is guaranteed to exist for as long as the row
			// exists.
			continue
		}

		ru.key = keys.MakeColumnKey(newPrimaryIndexKey, uint32(col.ID))
		if ru.marshalled[i].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.
			if log.V(2) {
				log.Infof("Put %s -> %v", ru.key, val)
			}

			b.Put(&ru.key, &ru.marshalled[i])
		} else {
			// The column might have already existed but is being set to NULL, so
			// delete it.
			if log.V(2) {
				log.Infof("Del %s", ru.key)
			}

			b.Del(&ru.key)
		}
		ru.key = nil
	}

	return ru.newValues, nil
}
示例#7
0
func (sc *SchemaChanger) truncateAndBackfillColumnsChunk(
	added []sqlbase.ColumnDescriptor,
	dropped []sqlbase.ColumnDescriptor,
	nonNullableColumn string,
	defaultExprs []parser.TypedExpr,
	evalCtx parser.EvalContext,
	sp sqlbase.Span,
) (roachpb.Key, bool, error) {
	var curSentinel roachpb.Key
	done := false
	err := sc.db.Txn(func(txn *client.Txn) error {
		tableDesc, err := getTableDescFromID(txn, sc.tableID)
		if err != nil {
			return err
		}
		// Short circuit the backfill if the table has been deleted.
		if tableDesc.Deleted {
			done = true
			return nil
		}

		// Run a scan across the table using the primary key. Running
		// the scan and applying the changes in many transactions is
		// fine because the schema change is in the correct state to
		// handle intermediate OLTP commands which delete and add
		// values during the scan.
		b := &client.Batch{}
		b.Scan(sp.Start, sp.End, ColumnTruncateAndBackfillChunkSize)
		if err := txn.Run(b); err != nil {
			return err
		}

		// Use a different batch to truncate/backfill columns.
		writeBatch := &client.Batch{}
		marshalled := make([]roachpb.Value, len(defaultExprs))
		done = true
		for _, result := range b.Results {
			var sentinelKey roachpb.Key
			for _, kv := range result.Rows {
				// Still processing table.
				done = false
				if nonNullableColumn != "" {
					return fmt.Errorf("column %s contains null values", nonNullableColumn)
				}

				if sentinelKey == nil || !bytes.HasPrefix(kv.Key, sentinelKey) {
					// Sentinel keys have a 0 suffix indicating 0 bytes of
					// column ID. Strip off that suffix to determine the
					// prefix shared with the other keys for the row.
					sentinelKey = sqlbase.StripColumnIDLength(kv.Key)
					// Store away key for the next table row as the point from
					// which to start from.
					curSentinel = sentinelKey

					// Delete the entire dropped columns. This used to use SQL
					// UPDATE in the past to update the dropped column to
					// NULL; but a column in the process of being dropped is
					// placed in the table descriptor mutations, and a SQL
					// UPDATE of a column in mutations will fail.
					for _, columnDesc := range dropped {
						// Delete the dropped column.
						colKey := keys.MakeColumnKey(sentinelKey, uint32(columnDesc.ID))
						if log.V(2) {
							log.Infof("Del %s", colKey)
						}
						writeBatch.Del(colKey)
					}

					// Add the new columns and backfill the values.
					for i, expr := range defaultExprs {
						if expr == nil {
							continue
						}
						col := added[i]
						colKey := keys.MakeColumnKey(sentinelKey, uint32(col.ID))
						d, err := expr.Eval(evalCtx)
						if err != nil {
							return err
						}
						marshalled[i], err = sqlbase.MarshalColumnValue(col, d)
						if err != nil {
							return err
						}

						if log.V(2) {
							log.Infof("Put %s -> %v", colKey, d)
						}
						// Insert default value into the column. If this row
						// was recently added the default value might have
						// already been populated, because the
						// ColumnDescriptor is in the WRITE_ONLY state.
						// Reinserting the default value is not a big deal.
						//
						// Note: a column in the WRITE_ONLY state cannot be
						// populated directly through SQL. A SQL INSERT cannot
						// directly reference the column, and the INSERT
						// populates the column with the default value.
						writeBatch.Put(colKey, &marshalled[i])
					}
				}
			}
		}
		if err := txn.Run(writeBatch); err != nil {
			for _, r := range writeBatch.Results {
				if r.PErr != nil {
					return convertBackfillError(tableDesc, writeBatch, r.PErr)
				}
			}
			return err
		}
		return nil
	})
	return curSentinel.PrefixEnd(), done, err
}
示例#8
0
func (p *planner) backfillBatch(b *client.Batch, oldTableDesc *TableDescriptor, mutationID MutationID) *roachpb.Error {
	var droppedColumnDescs []ColumnDescriptor
	var droppedIndexDescs []IndexDescriptor
	var newIndexDescs []IndexDescriptor
	// Collect the elements that are part of the mutation.
	for _, m := range oldTableDesc.Mutations {
		if m.MutationID != mutationID {
			// Mutations are applied in a FIFO order. Only apply the first set of
			// mutations if they have the mutation ID we're looking for.
			break
		}
		switch m.Direction {
		case DescriptorMutation_ADD:
			switch t := m.Descriptor_.(type) {
			case *DescriptorMutation_Column:
				// TODO(vivek): Add column to new columns and use it
				// to fill in default values.

			case *DescriptorMutation_Index:
				newIndexDescs = append(newIndexDescs, *t.Index)
			}

		case DescriptorMutation_DROP:
			switch t := m.Descriptor_.(type) {
			case *DescriptorMutation_Column:
				droppedColumnDescs = append(droppedColumnDescs, *t.Column)

			case *DescriptorMutation_Index:
				droppedIndexDescs = append(droppedIndexDescs, *t.Index)
			}
		}
	}

	// TODO(vivek): Break these backfill operations into chunks. All of them
	// will fail on big tables (see #3274).

	// Delete the entire dropped columns.
	// This used to use SQL UPDATE in the past to update the dropped
	// column to NULL; but a column in the process of being
	// dropped is placed in the table descriptor mutations, and
	// a SQL UPDATE of a column in mutations will fail.
	if len(droppedColumnDescs) > 0 {
		// Run a scan across the table using the primary key.
		start := roachpb.Key(MakeIndexKeyPrefix(oldTableDesc.ID, oldTableDesc.PrimaryIndex.ID))
		// Use a different batch to perform the scan.
		batch := &client.Batch{}
		batch.Scan(start, start.PrefixEnd(), 0)
		if pErr := p.txn.Run(batch); pErr != nil {
			return pErr
		}
		for _, result := range batch.Results {
			var sentinelKey roachpb.Key
			for _, kv := range result.Rows {
				if sentinelKey == nil || !bytes.HasPrefix(kv.Key, sentinelKey) {
					// Sentinel keys have a 0 suffix indicating 0 bytes of column
					// ID. Strip off that suffix to determine the prefix shared with the
					// other keys for the row.
					sentinelKey = stripColumnIDLength(kv.Key)
					for _, columnDesc := range droppedColumnDescs {
						// Delete the dropped column.
						colKey := keys.MakeColumnKey(sentinelKey, uint32(columnDesc.ID))
						if log.V(2) {
							log.Infof("Del %s", colKey)
						}
						b.Del(colKey)
					}
				}
			}
		}
	}

	for _, indexDescriptor := range droppedIndexDescs {
		indexPrefix := MakeIndexKeyPrefix(oldTableDesc.ID, indexDescriptor.ID)

		// Delete the index.
		indexStartKey := roachpb.Key(indexPrefix)
		indexEndKey := indexStartKey.PrefixEnd()
		if log.V(2) {
			log.Infof("DelRange %s - %s", indexStartKey, indexEndKey)
		}
		b.DelRange(indexStartKey, indexEndKey)
	}

	if len(newIndexDescs) > 0 {
		// Get all the rows affected.
		// TODO(vivek): Avoid going through Select.
		// TODO(tamird): Support partial indexes?
		// Use a scanNode with SELECT to pass in a TableDescriptor
		// to the SELECT without needing to use a parser.QualifiedName,
		// because we want to run schema changes from a gossip feed of
		// table IDs.
		scan := &scanNode{
			planner: p,
			txn:     p.txn,
			desc:    oldTableDesc,
		}
		scan.initDescDefaults()
		rows, pErr := p.selectIndex(&selectNode{}, scan, nil, false)
		if pErr != nil {
			return pErr
		}

		// Construct a map from column ID to the index the value appears at within a
		// row.
		colIDtoRowIndex, pErr := makeColIDtoRowIndex(rows, oldTableDesc)
		if pErr != nil {
			return pErr
		}

		for rows.Next() {
			rowVals := rows.Values()

			for _, newIndexDesc := range newIndexDescs {
				secondaryIndexEntries, pErr := encodeSecondaryIndexes(
					oldTableDesc.ID, []IndexDescriptor{newIndexDesc}, colIDtoRowIndex, rowVals)
				if pErr != nil {
					return pErr
				}

				for _, secondaryIndexEntry := range secondaryIndexEntries {
					if log.V(2) {
						log.Infof("CPut %s -> %v", secondaryIndexEntry.key,
							secondaryIndexEntry.value)
					}
					b.CPut(secondaryIndexEntry.key, secondaryIndexEntry.value, nil)
				}
			}
		}

		return rows.PErr()
	}

	return nil
}
示例#9
0
func (p *planner) backfillBatch(b *client.Batch, oldTableDesc, newTableDesc *TableDescriptor) error {
	var droppedColumnDescs []ColumnDescriptor
	var droppedIndexDescs []IndexDescriptor
	var newIndexDescs []IndexDescriptor
	for _, m := range oldTableDesc.Mutations {
		switch m.Direction {
		case DescriptorMutation_ADD:
			switch t := m.Descriptor_.(type) {
			case *DescriptorMutation_Column:
				// TODO(vivek): Add column to new columns and use it
				// to fill in default values.

			case *DescriptorMutation_Index:
				newIndexDescs = append(newIndexDescs, *t.Index)
			}

		case DescriptorMutation_DROP:
			switch t := m.Descriptor_.(type) {
			case *DescriptorMutation_Column:
				droppedColumnDescs = append(droppedColumnDescs, *t.Column)

			case *DescriptorMutation_Index:
				droppedIndexDescs = append(droppedIndexDescs, *t.Index)
			}
		}
	}

	// TODO(vivek): Break these backfill operations into chunks. All of them
	// will fail on big tables.

	// Delete the entire dropped columns.
	// This used to use SQL UPDATE in the past to update the dropped
	// column to NULL; but a column in the process of being
	// dropped is placed in the table descriptor mutations, and
	// a SQL UPDATE of a column in mutations will fail.
	if len(droppedColumnDescs) > 0 {
		// Run a scan across the table using the primary key.
		start := roachpb.Key(MakeIndexKeyPrefix(newTableDesc.ID, newTableDesc.PrimaryIndex.ID))
		// Use a different batch to perform the scan.
		batch := &client.Batch{}
		batch.Scan(start, start.PrefixEnd(), 0)
		if err := p.txn.Run(batch); err != nil {
			return err
		}
		for _, result := range batch.Results {
			var sentinelKey roachpb.Key
			for _, kv := range result.Rows {
				if sentinelKey == nil || !bytes.HasPrefix(kv.Key, sentinelKey) {
					// Sentinel keys have a 0 suffix indicating 0 bytes of column
					// ID. Strip off that suffix to determine the prefix shared with the
					// other keys for the row.
					sentinelKey = stripColumnIDLength(kv.Key)
					for _, columnDesc := range droppedColumnDescs {
						// Delete the dropped column.
						colKey := keys.MakeColumnKey(sentinelKey, uint32(columnDesc.ID))
						if log.V(2) {
							log.Infof("Del %s", colKey)
						}
						b.Del(colKey)
					}
				}
			}
		}
	}

	for _, indexDescriptor := range droppedIndexDescs {
		indexPrefix := MakeIndexKeyPrefix(newTableDesc.ID, indexDescriptor.ID)

		// Delete the index.
		indexStartKey := roachpb.Key(indexPrefix)
		indexEndKey := indexStartKey.PrefixEnd()
		if log.V(2) {
			log.Infof("DelRange %s - %s", indexStartKey, indexEndKey)
		}
		b.DelRange(indexStartKey, indexEndKey)
	}

	if len(newIndexDescs) > 0 {
		// Get all the rows affected.
		// TODO(vivek): Avoid going through Select.
		// TODO(tamird): Support partial indexes?
		// Use a scanNode with SELECT to pass in a TableDescriptor
		// to the SELECT without needing to use a parser.QualifiedName,
		// because we want to run schema changes from a gossip feed of
		// table IDs.
		scan := &scanNode{
			planner: p,
			txn:     p.txn,
			desc:    oldTableDesc,
		}
		scan.initDescDefaults()
		rows, err := p.selectWithScan(scan, &parser.Select{Exprs: oldTableDesc.allColumnsSelector()})
		if err != nil {
			return err
		}

		// Construct a map from column ID to the index the value appears at within a
		// row.
		colIDtoRowIndex, err := makeColIDtoRowIndex(rows, oldTableDesc)
		if err != nil {
			return err
		}

		// TODO(tamird): This will fall down in production use. We need to do
		// something better (see #2036). In particular, this implementation
		// has the following problems:
		// - Very large tables will generate an enormous batch here. This
		// isn't really a problem in itself except that it will exacerbate
		// the other issue:
		// - Any non-quiescent table that this runs against will end up with
		// an inconsistent index. This is because as inserts/updates continue
		// to roll in behind this operation's read front, the written index
		// will become incomplete/stale before it's written.

		for rows.Next() {
			rowVals := rows.Values()

			for _, newIndexDesc := range newIndexDescs {
				secondaryIndexEntries, err := encodeSecondaryIndexes(
					oldTableDesc.ID, []IndexDescriptor{newIndexDesc}, colIDtoRowIndex, rowVals)
				if err != nil {
					return err
				}

				for _, secondaryIndexEntry := range secondaryIndexEntries {
					if log.V(2) {
						log.Infof("CPut %s -> %v", secondaryIndexEntry.key,
							secondaryIndexEntry.value)
					}
					b.CPut(secondaryIndexEntry.key, secondaryIndexEntry.value, nil)
				}
			}
		}

		return rows.Err()
	}

	return nil
}
示例#10
0
// Insert inserts rows into the database.
// Privileges: INSERT on table
//   Notes: postgres requires INSERT. No "on duplicate key update" option.
//          mysql requires INSERT. Also requires UPDATE on "ON DUPLICATE KEY UPDATE".
func (p *planner) Insert(n *parser.Insert, autoCommit bool) (planNode, *roachpb.Error) {
	// TODO(marcb): We can't use the cached descriptor here because a recent
	// update of the schema (e.g. the addition of an index) might not be
	// reflected in the cached version (yet). Perhaps schema modification
	// routines such as CREATE INDEX should not return until the schema change
	// has been pushed everywhere.
	tableDesc, pErr := p.getTableLease(n.Table)
	if pErr != nil {
		return nil, pErr
	}

	if err := p.checkPrivilege(&tableDesc, privilege.INSERT); err != nil {
		return nil, roachpb.NewError(err)
	}

	var cols []ColumnDescriptor
	// Determine which columns we're inserting into.
	if n.DefaultValues() {
		cols = tableDesc.Columns
	} else {
		var err error
		if cols, err = p.processColumns(&tableDesc, n.Columns); err != nil {
			return nil, roachpb.NewError(err)
		}
	}
	// Number of columns expecting an input. This doesn't include the
	// columns receiving a default value.
	numInputColumns := len(cols)

	// Construct a map from column ID to the index the value appears at within a
	// row.
	colIDtoRowIndex := map[ColumnID]int{}
	for i, c := range cols {
		colIDtoRowIndex[c.ID] = i
	}

	// Add the column if it has a DEFAULT expression.
	addIfDefault := func(col ColumnDescriptor) {
		if col.DefaultExpr != nil {
			if _, ok := colIDtoRowIndex[col.ID]; !ok {
				colIDtoRowIndex[col.ID] = len(cols)
				cols = append(cols, col)
			}
		}
	}

	// Add any column that has a DEFAULT expression.
	for _, col := range tableDesc.Columns {
		addIfDefault(col)
	}
	// Also add any column in a mutation that is WRITE_ONLY and has
	// a DEFAULT expression.
	for _, m := range tableDesc.Mutations {
		if m.State != DescriptorMutation_WRITE_ONLY {
			continue
		}
		if col := m.GetColumn(); col != nil {
			addIfDefault(*col)
		}
	}

	// Verify we have at least the columns that are part of the primary key.
	primaryKeyCols := map[ColumnID]struct{}{}
	for i, id := range tableDesc.PrimaryIndex.ColumnIDs {
		if _, ok := colIDtoRowIndex[id]; !ok {
			return nil, roachpb.NewUErrorf("missing %q primary key column", tableDesc.PrimaryIndex.ColumnNames[i])
		}
		primaryKeyCols[id] = struct{}{}
	}

	// Construct the default expressions. The returned slice will be nil if no
	// column in the table has a default expression.
	defaultExprs, err := p.makeDefaultExprs(cols)
	if err != nil {
		return nil, roachpb.NewError(err)
	}

	// Replace any DEFAULT markers with the corresponding default expressions.
	insertRows := p.fillDefaults(defaultExprs, cols, n)

	// Transform the values into a rows object. This expands SELECT statements or
	// generates rows from the values contained within the query.
	rows, pErr := p.makePlan(insertRows, false)
	if pErr != nil {
		return nil, pErr
	}

	if expressions := len(rows.Columns()); expressions > numInputColumns {
		return nil, roachpb.NewUErrorf("INSERT has more expressions than target columns: %d/%d", expressions, numInputColumns)
	}

	primaryIndex := tableDesc.PrimaryIndex
	primaryIndexKeyPrefix := MakeIndexKeyPrefix(tableDesc.ID, primaryIndex.ID)

	marshalled := make([]interface{}, len(cols))

	b := p.txn.NewBatch()
	rh, err := makeReturningHelper(p, n.Returning, tableDesc.Name, cols)
	if err != nil {
		return nil, roachpb.NewError(err)
	}
	for rows.Next() {
		rowVals := rows.Values()

		// The values for the row may be shorter than the number of columns being
		// inserted into. Generate default values for those columns using the
		// default expressions.
		for i := len(rowVals); i < len(cols); i++ {
			if defaultExprs == nil {
				rowVals = append(rowVals, parser.DNull)
				continue
			}
			d, err := defaultExprs[i].Eval(p.evalCtx)
			if err != nil {
				return nil, roachpb.NewError(err)
			}
			rowVals = append(rowVals, d)
		}

		// Check to see if NULL is being inserted into any non-nullable column.
		for _, col := range tableDesc.Columns {
			if !col.Nullable {
				if i, ok := colIDtoRowIndex[col.ID]; !ok || rowVals[i] == parser.DNull {
					return nil, roachpb.NewUErrorf("null value in column %q violates not-null constraint", col.Name)
				}
			}
		}

		// Check that the row 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 rowVals {
			// Make sure the value can be written to the column before proceeding.
			var mErr error
			if marshalled[i], mErr = marshalColumnValue(cols[i], val, p.evalCtx.Args); mErr != nil {
				return nil, roachpb.NewError(mErr)
			}
		}

		if p.prepareOnly {
			continue
		}

		primaryIndexKey, _, eErr := encodeIndexKey(
			&primaryIndex, colIDtoRowIndex, rowVals, primaryIndexKeyPrefix)
		if eErr != nil {
			return nil, roachpb.NewError(eErr)
		}

		// Write the secondary indexes.
		indexes := tableDesc.Indexes
		// Also include the secondary indexes in mutation state WRITE_ONLY.
		for _, m := range tableDesc.Mutations {
			if m.State == DescriptorMutation_WRITE_ONLY {
				if index := m.GetIndex(); index != nil {
					indexes = append(indexes, *index)
				}
			}
		}
		secondaryIndexEntries, eErr := encodeSecondaryIndexes(
			tableDesc.ID, indexes, colIDtoRowIndex, rowVals)
		if eErr != nil {
			return nil, roachpb.NewError(eErr)
		}

		for _, secondaryIndexEntry := range secondaryIndexEntries {
			if log.V(2) {
				log.Infof("CPut %s -> %v", secondaryIndexEntry.key,
					secondaryIndexEntry.value)
			}
			b.CPut(secondaryIndexEntry.key, secondaryIndexEntry.value, nil)
		}

		// Write the row sentinel.
		sentinelKey := keys.MakeNonColumnKey(primaryIndexKey)
		if log.V(2) {
			log.Infof("CPut %s -> NULL", roachpb.Key(sentinelKey))
		}
		// This is subtle: An interface{}(nil) deletes the value, so we pass in
		// []byte{} as a non-nil value.
		b.CPut(sentinelKey, []byte{}, nil)

		// Write the row columns.
		for i, val := range rowVals {
			col := cols[i]
			if _, ok := primaryKeyCols[col.ID]; ok {
				// Skip primary key columns as their values are encoded in the row
				// sentinel key which is guaranteed to exist for as long as the row
				// exists.
				continue
			}

			if marshalled[i] != 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.

				key := keys.MakeColumnKey(primaryIndexKey, uint32(col.ID))
				if log.V(2) {
					log.Infof("CPut %s -> %v", roachpb.Key(key), val)
				}

				b.CPut(key, marshalled[i], nil)
			}
		}

		if err := rh.append(rowVals); err != nil {
			return nil, roachpb.NewError(err)
		}
	}
	if pErr := rows.PErr(); pErr != nil {
		return nil, pErr
	}

	if p.prepareOnly {
		// Return the result column types.
		return rh.getResults(), nil
	}

	if isSystemConfigID(tableDesc.GetID()) {
		// Mark transaction as operating on the system DB.
		p.txn.SetSystemConfigTrigger()
	}

	if autoCommit {
		// An auto-txn can commit the transaction with the batch. This is an
		// optimization to avoid an extra round-trip to the transaction
		// coordinator.
		pErr = p.txn.CommitInBatch(b)
	} else {
		pErr = p.txn.Run(b)
	}
	if pErr != nil {
		return nil, convertBatchError(&tableDesc, *b, pErr)
	}
	return rh.getResults(), nil
}
示例#11
0
func (sc *SchemaChanger) truncateAndBackfillColumns(
	lease *TableDescriptor_SchemaChangeLease,
	added []ColumnDescriptor,
	dropped []ColumnDescriptor,
	version DescriptorVersion,
) *roachpb.Error {
	evalCtx := parser.EvalContext{}
	// Set the eval context timestamps.
	pTime := timeutil.Now()
	evalCtx.SetTxnTimestamp(pTime)
	evalCtx.SetStmtTimestamp(pTime)
	defaultExprs, err := makeDefaultExprs(added, &parser.Parser{}, evalCtx)
	if err != nil {
		return roachpb.NewError(err)
	}

	// Remember any new non nullable column with no default value.
	nonNullableColumn := ""
	for _, columnDesc := range added {
		if columnDesc.DefaultExpr == nil && !columnDesc.Nullable {
			nonNullableColumn = columnDesc.Name
		}
	}

	// Add or Drop a column.
	if len(dropped) > 0 || nonNullableColumn != "" || len(defaultExprs) > 0 {
		// First extend the schema change lease.
		l, pErr := sc.ExtendLease(*lease)
		if pErr != nil {
			return pErr
		}
		*lease = l

		pErr = sc.db.Txn(func(txn *client.Txn) *roachpb.Error {
			tableDesc, pErr := getTableDescAtVersion(txn, sc.tableID, version)
			if pErr != nil {
				return pErr
			}

			// Run a scan across the table using the primary key.
			start := roachpb.Key(MakeIndexKeyPrefix(tableDesc.ID, tableDesc.PrimaryIndex.ID))
			b := &client.Batch{}
			b.Scan(start, start.PrefixEnd(), 0)
			if pErr := txn.Run(b); pErr != nil {
				return pErr
			}

			if nonNullableColumn != "" {
				for _, result := range b.Results {
					if len(result.Rows) > 0 {
						return roachpb.NewErrorf("column %s contains null values", nonNullableColumn)
					}
				}
			}

			// Use a different batch to truncate/backfill columns.
			writeBatch := &client.Batch{}
			for _, result := range b.Results {
				var sentinelKey roachpb.Key
				for _, kv := range result.Rows {
					if sentinelKey == nil || !bytes.HasPrefix(kv.Key, sentinelKey) {
						// Sentinel keys have a 0 suffix indicating 0 bytes of column
						// ID. Strip off that suffix to determine the prefix shared with the
						// other keys for the row.
						sentinelKey = stripColumnIDLength(kv.Key)

						// Delete the entire dropped columns.
						// This used to use SQL UPDATE in the past to update the dropped
						// column to NULL; but a column in the process of being
						// dropped is placed in the table descriptor mutations, and
						// a SQL UPDATE of a column in mutations will fail.
						for _, columnDesc := range dropped {
							// Delete the dropped column.
							colKey := keys.MakeColumnKey(sentinelKey, uint32(columnDesc.ID))
							if log.V(2) {
								log.Infof("Del %s", colKey)
							}
							writeBatch.Del(colKey)
						}

						// Add the new columns and backfill the values.
						for i, expr := range defaultExprs {
							if expr == nil {
								continue
							}
							col := added[i]
							colKey := keys.MakeColumnKey(sentinelKey, uint32(col.ID))
							d, err := expr.Eval(evalCtx)
							if err != nil {
								return roachpb.NewError(err)
							}
							val, err := marshalColumnValue(col, d, evalCtx.Args)
							if err != nil {
								return roachpb.NewError(err)
							}

							if log.V(2) {
								log.Infof("Put %s -> %v", colKey, val)
							}
							// Insert default value into the column. If this row
							// was recently added the default value might have
							// already been populated, because the
							// ColumnDescriptor is in the WRITE_ONLY state.
							// Reinserting the default value is not a big deal.
							//
							// Note: a column in the WRITE_ONLY state cannot be
							// populated directly through SQL. A SQL INSERT cannot
							// directly reference the column, and the INSERT
							// populates the column with the default value.
							writeBatch.Put(colKey, val)
						}
					}
				}
			}
			if pErr := txn.Run(writeBatch); pErr != nil {
				return convertBackfillError(tableDesc, writeBatch, pErr)
			}
			return nil
		})
		return pErr
	}
	return nil
}
示例#12
0
// Update updates columns for a selection of rows from a table.
// Privileges: UPDATE and SELECT on table. We currently always use a select statement.
//   Notes: postgres requires UPDATE. Requires SELECT with WHERE clause with table.
//          mysql requires UPDATE. Also requires SELECT with WHERE clause with table.
func (p *planner) Update(n *parser.Update) (planNode, *roachpb.Error) {
	tableDesc, pErr := p.getAliasedTableLease(n.Table)
	if pErr != nil {
		return nil, pErr
	}

	if err := p.checkPrivilege(tableDesc, privilege.UPDATE); err != nil {
		return nil, roachpb.NewError(err)
	}

	// Determine which columns we're inserting into.
	var names parser.QualifiedNames
	for _, expr := range n.Exprs {
		var epErr *roachpb.Error
		expr.Expr, epErr = p.expandSubqueries(expr.Expr, len(expr.Names))
		if epErr != nil {
			return nil, epErr
		}

		if expr.Tuple {
			// TODO(pmattis): The distinction between Tuple and DTuple here is
			// irritating. We'll see a DTuple if the expression was a subquery that
			// has been evaluated. We'll see a Tuple in other cases.
			n := 0
			switch t := expr.Expr.(type) {
			case parser.Tuple:
				n = len(t)
			case parser.DTuple:
				n = len(t)
			default:
				return nil, roachpb.NewErrorf("unsupported tuple assignment: %T", expr.Expr)
			}
			if len(expr.Names) != n {
				return nil, roachpb.NewUErrorf("number of columns (%d) does not match number of values (%d)",
					len(expr.Names), n)
			}
		}
		names = append(names, expr.Names...)
	}
	cols, err := p.processColumns(tableDesc, names)
	if err != nil {
		return nil, roachpb.NewError(err)
	}

	// Set of columns being updated
	colIDSet := map[ColumnID]struct{}{}
	for _, c := range cols {
		colIDSet[c.ID] = struct{}{}
	}
	// Don't allow updating any column that is part of the primary key.
	for i, id := range tableDesc.PrimaryIndex.ColumnIDs {
		if _, ok := colIDSet[id]; ok {
			return nil, roachpb.NewUErrorf("primary key column %q cannot be updated", tableDesc.PrimaryIndex.ColumnNames[i])
		}
	}

	defaultExprs, err := p.makeDefaultExprs(cols)
	if err != nil {
		return nil, roachpb.NewError(err)
	}

	// Generate the list of select targets. We need to select all of the columns
	// plus we select all of the update expressions in case those expressions
	// reference columns (e.g. "UPDATE t SET v = v + 1"). Note that we flatten
	// expressions for tuple assignments just as we flattened the column names
	// above. So "UPDATE t SET (a, b) = (1, 2)" translates into select targets of
	// "*, 1, 2", not "*, (1, 2)".
	targets := tableDesc.allColumnsSelector()
	i := 0
	for _, expr := range n.Exprs {
		if expr.Tuple {
			switch t := expr.Expr.(type) {
			case parser.Tuple:
				for _, e := range t {
					e = fillDefault(e, i, defaultExprs)
					targets = append(targets, parser.SelectExpr{Expr: e})
					i++
				}
			case parser.DTuple:
				for _, e := range t {
					targets = append(targets, parser.SelectExpr{Expr: e})
					i++
				}
			}
		} else {
			e := fillDefault(expr.Expr, i, defaultExprs)
			targets = append(targets, parser.SelectExpr{Expr: e})
			i++
		}
	}

	// Query the rows that need updating.
	rows, pErr := p.Select(&parser.Select{
		Exprs: targets,
		From:  parser.TableExprs{n.Table},
		Where: n.Where,
	})
	if pErr != nil {
		return nil, pErr
	}

	// ValArgs have their types populated in the above Select if they are part
	// of an expression ("SET a = 2 + $1") in the type check step where those
	// types are inferred. For the simpler case ("SET a = $1"), populate them
	// using marshalColumnValue. This step also verifies that the expression
	// types match the column types.
	if p.prepareOnly {
		i := 0
		f := func(expr parser.Expr) *roachpb.Error {
			idx := i
			i++
			// DefaultVal doesn't implement TypeCheck
			if _, ok := expr.(parser.DefaultVal); ok {
				return nil
			}
			d, err := expr.TypeCheck(p.evalCtx.Args)
			if err != nil {
				return roachpb.NewError(err)
			}
			if _, err := marshalColumnValue(cols[idx], d, p.evalCtx.Args); err != nil {
				return roachpb.NewError(err)
			}
			return nil
		}
		for _, expr := range n.Exprs {
			if expr.Tuple {
				switch t := expr.Expr.(type) {
				case parser.Tuple:
					for _, e := range t {
						if err := f(e); err != nil {
							return nil, err
						}
					}
				case parser.DTuple:
					for _, e := range t {
						if err := f(e); err != nil {
							return nil, err
						}
					}
				}
			} else {
				if err := f(expr.Expr); err != nil {
					return nil, err
				}
			}
		}

		return nil, nil
	}

	// Construct a map from column ID to the index the value appears at within a
	// row.
	colIDtoRowIndex := map[ColumnID]int{}
	for i, col := range tableDesc.Columns {
		colIDtoRowIndex[col.ID] = i
	}

	primaryIndex := tableDesc.PrimaryIndex
	primaryIndexKeyPrefix := MakeIndexKeyPrefix(tableDesc.ID, primaryIndex.ID)

	// Secondary indexes needing updating.
	needsUpdate := func(index IndexDescriptor) bool {
		for _, id := range index.ColumnIDs {
			if _, ok := colIDSet[id]; ok {
				return true
			}
		}
		return false
	}

	indexes := make([]IndexDescriptor, 0, len(tableDesc.Indexes)+len(tableDesc.Mutations))
	var deleteOnlyIndex map[int]struct{}

	for _, index := range tableDesc.Indexes {
		if needsUpdate(index) {
			indexes = append(indexes, index)
		}
	}
	for _, m := range tableDesc.Mutations {
		if index := m.GetIndex(); index != nil {
			if needsUpdate(*index) {
				indexes = append(indexes, *index)

				switch m.State {
				case DescriptorMutation_DELETE_ONLY:
					if deleteOnlyIndex == nil {
						// Allocate at most once.
						deleteOnlyIndex = make(map[int]struct{}, len(tableDesc.Mutations))
					}
					deleteOnlyIndex[len(indexes)-1] = struct{}{}

				case DescriptorMutation_WRITE_ONLY:
				}
			}
		}
	}

	marshalled := make([]interface{}, len(cols))

	b := client.Batch{}
	result := &valuesNode{}
	for rows.Next() {
		rowVals := rows.Values()
		result.rows = append(result.rows, parser.DTuple(nil))

		primaryIndexKey, _, err := encodeIndexKey(
			&primaryIndex, colIDtoRowIndex, rowVals, primaryIndexKeyPrefix)
		if err != nil {
			return nil, roachpb.NewError(err)
		}
		// Compute the current secondary index key:value pairs for this row.
		secondaryIndexEntries, err := encodeSecondaryIndexes(
			tableDesc.ID, indexes, colIDtoRowIndex, rowVals)
		if err != nil {
			return nil, roachpb.NewError(err)
		}

		// Our updated value expressions occur immediately after the plain
		// columns in the output.
		newVals := rowVals[len(tableDesc.Columns):]
		// Update the row values.
		for i, col := range cols {
			val := newVals[i]
			if !col.Nullable && val == parser.DNull {
				return nil, roachpb.NewUErrorf("null value in column %q violates not-null constraint", col.Name)
			}
			rowVals[colIDtoRowIndex[col.ID]] = val
		}

		// 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 newVals {
			var mErr error
			if marshalled[i], mErr = marshalColumnValue(cols[i], val, p.evalCtx.Args); mErr != nil {
				return nil, roachpb.NewError(mErr)
			}
		}

		// Compute the new secondary index key:value pairs for this row.
		newSecondaryIndexEntries, eErr := encodeSecondaryIndexes(
			tableDesc.ID, indexes, colIDtoRowIndex, rowVals)
		if eErr != nil {
			return nil, roachpb.NewError(eErr)
		}

		// Update secondary indexes.
		for i, newSecondaryIndexEntry := range newSecondaryIndexEntries {
			secondaryIndexEntry := secondaryIndexEntries[i]
			if !bytes.Equal(newSecondaryIndexEntry.key, secondaryIndexEntry.key) {
				// Do not update Indexes in the DELETE_ONLY state.
				if _, ok := deleteOnlyIndex[i]; !ok {
					if log.V(2) {
						log.Infof("CPut %s -> %v", newSecondaryIndexEntry.key,
							newSecondaryIndexEntry.value)
					}
					b.CPut(newSecondaryIndexEntry.key, newSecondaryIndexEntry.value, nil)
				}
				if log.V(2) {
					log.Infof("Del %s", secondaryIndexEntry.key)
				}
				b.Del(secondaryIndexEntry.key)
			}
		}

		// Add the new values.
		for i, val := range newVals {
			col := cols[i]

			key := keys.MakeColumnKey(primaryIndexKey, uint32(col.ID))
			if marshalled[i] != 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.
				if log.V(2) {
					log.Infof("Put %s -> %v", key, val)
				}

				b.Put(key, marshalled[i])
			} else {
				// The column might have already existed but is being set to NULL, so
				// delete it.
				if log.V(2) {
					log.Infof("Del %s", key)
				}

				b.Del(key)
			}
		}
	}

	if pErr := rows.PErr(); pErr != nil {
		return nil, pErr
	}
	if pErr := p.txn.Run(&b); pErr != nil {
		return nil, convertBatchError(tableDesc, b, pErr)
	}

	return result, nil
}
示例#13
0
// Update updates columns for a selection of rows from a table.
// Privileges: UPDATE and SELECT on table. We currently always use a select statement.
//   Notes: postgres requires UPDATE. Requires SELECT with WHERE clause with table.
//          mysql requires UPDATE. Also requires SELECT with WHERE clause with table.
// TODO(guanqun): need to support CHECK in UPDATE
func (p *planner) Update(n *parser.Update, autoCommit bool) (planNode, *roachpb.Error) {
	tracing.AnnotateTrace()
	tableDesc, pErr := p.getAliasedTableLease(n.Table)
	if pErr != nil {
		return nil, pErr
	}

	if err := p.checkPrivilege(tableDesc, privilege.UPDATE); err != nil {
		return nil, roachpb.NewError(err)
	}

	// TODO(dan): Consider caching this on the TableDescriptor.
	primaryKeyCols := map[ColumnID]struct{}{}
	for _, id := range tableDesc.PrimaryIndex.ColumnIDs {
		primaryKeyCols[id] = struct{}{}
	}

	exprs := make([]parser.UpdateExpr, len(n.Exprs))
	for i, expr := range n.Exprs {
		exprs[i] = *expr
	}

	// Determine which columns we're inserting into.
	var names parser.QualifiedNames
	for i, expr := range exprs {
		newExpr, epErr := p.expandSubqueries(expr.Expr, len(expr.Names))
		if epErr != nil {
			return nil, epErr
		}
		exprs[i].Expr = newExpr

		if expr.Tuple {
			// TODO(pmattis): The distinction between Tuple and DTuple here is
			// irritating. We'll see a DTuple if the expression was a subquery that
			// has been evaluated. We'll see a Tuple in other cases.
			n := 0
			switch t := newExpr.(type) {
			case *parser.Tuple:
				n = len(t.Exprs)
			case parser.DTuple:
				n = len(t)
			default:
				return nil, roachpb.NewErrorf("unsupported tuple assignment: %T", newExpr)
			}
			if len(expr.Names) != n {
				return nil, roachpb.NewUErrorf("number of columns (%d) does not match number of values (%d)",
					len(expr.Names), n)
			}
		}
		names = append(names, expr.Names...)
	}
	cols, err := p.processColumns(tableDesc, names)
	if err != nil {
		return nil, roachpb.NewError(err)
	}

	// Set of columns being updated
	var primaryKeyColChange bool
	colIDSet := map[ColumnID]struct{}{}
	for _, c := range cols {
		colIDSet[c.ID] = struct{}{}
		if _, ok := primaryKeyCols[c.ID]; ok {
			primaryKeyColChange = true
		}
	}

	defaultExprs, err := makeDefaultExprs(cols, &p.parser, p.evalCtx)
	if err != nil {
		return nil, roachpb.NewError(err)
	}

	// Generate the list of select targets. We need to select all of the columns
	// plus we select all of the update expressions in case those expressions
	// reference columns (e.g. "UPDATE t SET v = v + 1"). Note that we flatten
	// expressions for tuple assignments just as we flattened the column names
	// above. So "UPDATE t SET (a, b) = (1, 2)" translates into select targets of
	// "*, 1, 2", not "*, (1, 2)".
	// TODO(radu): we only need to select columns necessary to generate primary and
	// secondary indexes keys, and columns needed by returningHelper.
	targets := tableDesc.allColumnsSelector()
	i := 0
	// Remember the index where the targets for exprs start.
	exprTargetIdx := len(targets)
	for _, expr := range exprs {
		if expr.Tuple {
			switch t := expr.Expr.(type) {
			case *parser.Tuple:
				for _, e := range t.Exprs {
					e = fillDefault(e, i, defaultExprs)
					targets = append(targets, parser.SelectExpr{Expr: e})
					i++
				}
			case parser.DTuple:
				for _, e := range t {
					targets = append(targets, parser.SelectExpr{Expr: e})
					i++
				}
			}
		} else {
			e := fillDefault(expr.Expr, i, defaultExprs)
			targets = append(targets, parser.SelectExpr{Expr: e})
			i++
		}
	}

	tracing.AnnotateTrace()

	// Query the rows that need updating.
	rows, pErr := p.SelectClause(&parser.SelectClause{
		Exprs: targets,
		From:  []parser.TableExpr{n.Table},
		Where: n.Where,
	})
	if pErr != nil {
		return nil, pErr
	}

	rh, err := makeReturningHelper(p, n.Returning, tableDesc.Name, tableDesc.Columns)
	if err != nil {
		return nil, roachpb.NewError(err)
	}

	// ValArgs have their types populated in the above Select if they are part
	// of an expression ("SET a = 2 + $1") in the type check step where those
	// types are inferred. For the simpler case ("SET a = $1"), populate them
	// using marshalColumnValue. This step also verifies that the expression
	// types match the column types.
	if p.evalCtx.PrepareOnly {
		for i, target := range rows.(*selectNode).render[exprTargetIdx:] {
			// DefaultVal doesn't implement TypeCheck
			if _, ok := target.(parser.DefaultVal); ok {
				continue
			}
			d, err := target.TypeCheck(p.evalCtx.Args)
			if err != nil {
				return nil, roachpb.NewError(err)
			}
			if _, err := marshalColumnValue(cols[i], d, p.evalCtx.Args); err != nil {
				return nil, roachpb.NewError(err)
			}
		}
		// Return the result column types.
		return rh.getResults()
	}

	// Construct a map from column ID to the index the value appears at within a
	// row.
	colIDtoRowIndex := map[ColumnID]int{}
	for i, col := range tableDesc.Columns {
		colIDtoRowIndex[col.ID] = i
	}

	primaryIndex := tableDesc.PrimaryIndex
	primaryIndexKeyPrefix := MakeIndexKeyPrefix(tableDesc.ID, primaryIndex.ID)

	// Secondary indexes needing updating.
	needsUpdate := func(index IndexDescriptor) bool {
		// If the primary key changed, we need to update all of them.
		if primaryKeyColChange {
			return true
		}
		for _, id := range index.ColumnIDs {
			if _, ok := colIDSet[id]; ok {
				return true
			}
		}
		return false
	}

	indexes := make([]IndexDescriptor, 0, len(tableDesc.Indexes)+len(tableDesc.Mutations))
	var deleteOnlyIndex map[int]struct{}

	for _, index := range tableDesc.Indexes {
		if needsUpdate(index) {
			indexes = append(indexes, index)
		}
	}
	for _, m := range tableDesc.Mutations {
		if index := m.GetIndex(); index != nil {
			if needsUpdate(*index) {
				indexes = append(indexes, *index)

				switch m.State {
				case DescriptorMutation_DELETE_ONLY:
					if deleteOnlyIndex == nil {
						// Allocate at most once.
						deleteOnlyIndex = make(map[int]struct{}, len(tableDesc.Mutations))
					}
					deleteOnlyIndex[len(indexes)-1] = struct{}{}

				case DescriptorMutation_WRITE_ONLY:
				}
			}
		}
	}

	marshalled := make([]interface{}, len(cols))

	b := p.txn.NewBatch()
	tracing.AnnotateTrace()
	for rows.Next() {
		tracing.AnnotateTrace()

		rowVals := rows.Values()

		primaryIndexKey, _, err := encodeIndexKey(
			&primaryIndex, colIDtoRowIndex, rowVals, primaryIndexKeyPrefix)
		if err != nil {
			return nil, roachpb.NewError(err)
		}
		// Compute the current secondary index key:value pairs for this row.
		secondaryIndexEntries, err := encodeSecondaryIndexes(
			tableDesc.ID, indexes, colIDtoRowIndex, rowVals)
		if err != nil {
			return nil, roachpb.NewError(err)
		}

		// Our updated value expressions occur immediately after the plain
		// columns in the output.
		newVals := rowVals[len(tableDesc.Columns):]

		// Ensure that the values honor the specified column widths.
		for i := range newVals {
			if err := checkValueWidth(cols[i], newVals[i]); err != nil {
				return nil, roachpb.NewError(err)
			}
		}

		// Update the row values.
		for i, col := range cols {
			val := newVals[i]
			if !col.Nullable && val == parser.DNull {
				return nil, roachpb.NewUErrorf("null value in column %q violates not-null constraint", col.Name)
			}
			rowVals[colIDtoRowIndex[col.ID]] = val
		}

		// 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 newVals {
			var mErr error
			if marshalled[i], mErr = marshalColumnValue(cols[i], val, p.evalCtx.Args); mErr != nil {
				return nil, roachpb.NewError(mErr)
			}
		}

		// Compute the new primary index key for this row.
		newPrimaryIndexKey := primaryIndexKey
		var rowPrimaryKeyChanged bool
		if primaryKeyColChange {
			newPrimaryIndexKey, _, err = encodeIndexKey(
				&primaryIndex, colIDtoRowIndex, rowVals, primaryIndexKeyPrefix)
			if err != nil {
				return nil, roachpb.NewError(err)
			}
			// Note that even if primaryIndexColChange is true, it's possible that
			// primary key fields in this particular row didn't change.
			rowPrimaryKeyChanged = !bytes.Equal(primaryIndexKey, newPrimaryIndexKey)
		}

		// Compute the new secondary index key:value pairs for this row.
		newSecondaryIndexEntries, eErr := encodeSecondaryIndexes(
			tableDesc.ID, indexes, colIDtoRowIndex, rowVals)
		if eErr != nil {
			return nil, roachpb.NewError(eErr)
		}

		if rowPrimaryKeyChanged {
			// Delete all the data stored under the old primary key.
			rowStartKey := roachpb.Key(primaryIndexKey)
			rowEndKey := rowStartKey.PrefixEnd()
			if log.V(2) {
				log.Infof("DelRange %s - %s", rowStartKey, rowEndKey)
			}
			b.DelRange(rowStartKey, rowEndKey, false)

			// Delete all the old secondary indexes.
			for _, secondaryIndexEntry := range secondaryIndexEntries {
				if log.V(2) {
					log.Infof("Del %s", secondaryIndexEntry.key)
				}
				b.Del(secondaryIndexEntry.key)
			}

			// Write the new row sentinel. We want to write the sentinel first in case
			// we are trying to insert a duplicate primary key: if we write the
			// secondary indexes first, we may get an error that looks like a
			// uniqueness violation on a non-unique index.
			sentinelKey := keys.MakeNonColumnKey(newPrimaryIndexKey)
			if log.V(2) {
				log.Infof("CPut %s -> NULL", roachpb.Key(sentinelKey))
			}
			// This is subtle: An interface{}(nil) deletes the value, so we pass in
			// []byte{} as a non-nil value.
			b.CPut(sentinelKey, []byte{}, nil)

			// Write any fields from the old row that were not modified by the UPDATE.
			for i, col := range tableDesc.Columns {
				if _, ok := colIDSet[col.ID]; ok {
					continue
				}
				if _, ok := primaryKeyCols[col.ID]; ok {
					continue
				}
				key := keys.MakeColumnKey(newPrimaryIndexKey, uint32(col.ID))
				val := rowVals[i]
				marshalledVal, mErr := marshalColumnValue(col, val, p.evalCtx.Args)
				if mErr != nil {
					return nil, roachpb.NewError(mErr)
				}

				if log.V(2) {
					log.Infof("Put %s -> %v", roachpb.Key(key), val)
				}
				b.Put(key, marshalledVal)
			}
			// At this point, we've deleted the old row and associated index data and
			// written the sentinel keys and column keys for non-updated columns. Fall
			// through to below where the index keys and updated column keys will be
			// written.
		}

		// Update secondary indexes.
		for i, newSecondaryIndexEntry := range newSecondaryIndexEntries {
			secondaryIndexEntry := secondaryIndexEntries[i]
			secondaryKeyChanged := !bytes.Equal(newSecondaryIndexEntry.key, secondaryIndexEntry.key)
			if secondaryKeyChanged {
				if log.V(2) {
					log.Infof("Del %s", secondaryIndexEntry.key)
				}
				b.Del(secondaryIndexEntry.key)
			}
			if rowPrimaryKeyChanged || secondaryKeyChanged {
				// Do not update Indexes in the DELETE_ONLY state.
				if _, ok := deleteOnlyIndex[i]; !ok {
					if log.V(2) {
						log.Infof("CPut %s -> %v", newSecondaryIndexEntry.key,
							newSecondaryIndexEntry.value)
					}
					b.CPut(newSecondaryIndexEntry.key, newSecondaryIndexEntry.value, nil)
				}
			}
		}

		// Add the new values.
		for i, val := range newVals {
			col := cols[i]

			if _, ok := primaryKeyCols[col.ID]; ok {
				// Skip primary key columns as their values are encoded in the row
				// sentinel key which is guaranteed to exist for as long as the row
				// exists.
				continue
			}

			key := keys.MakeColumnKey(newPrimaryIndexKey, uint32(col.ID))
			if marshalled[i] != 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.
				if log.V(2) {
					log.Infof("Put %s -> %v", roachpb.Key(key), val)
				}

				b.Put(key, marshalled[i])
			} else {
				// The column might have already existed but is being set to NULL, so
				// delete it.
				if log.V(2) {
					log.Infof("Del %s", key)
				}

				b.Del(key)
			}
		}

		// rowVals[:len(tableDesc.Columns)] have been updated with the new values above.
		if err := rh.append(rowVals[:len(tableDesc.Columns)]); err != nil {
			return nil, roachpb.NewError(err)
		}
	}
	tracing.AnnotateTrace()

	if pErr := rows.PErr(); pErr != nil {
		return nil, pErr
	}

	if isSystemConfigID(tableDesc.GetID()) {
		// Mark transaction as operating on the system DB.
		p.txn.SetSystemConfigTrigger()
	}

	if autoCommit {
		// An auto-txn can commit the transaction with the batch. This is an
		// optimization to avoid an extra round-trip to the transaction
		// coordinator.
		pErr = p.txn.CommitInBatch(b)
	} else {
		pErr = p.txn.Run(b)
	}
	if pErr != nil {
		return nil, convertBatchError(tableDesc, *b, pErr)
	}

	tracing.AnnotateTrace()
	return rh.getResults()
}