index_join.go

// Copyright 2015 The Cockroach Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
// implied. See the License for the specific language governing
// permissions and limitations under the License.
//
// Author: Peter Mattis (peter@cockroachlabs.com)

package sql

import (
	"fmt"

	"github.com/cockroachdb/cockroach/pkg/roachpb"
	"github.com/cockroachdb/cockroach/pkg/sql/parser"
	"github.com/cockroachdb/cockroach/pkg/sql/sqlbase"
	"github.com/cockroachdb/cockroach/pkg/util/log"
)

const indexJoinBatchSize = 100

// An indexJoinNode implements joining of results from an index with the rows
// of a table.
//
// There are three parameters to an index join:
// - which index is being used;
// - which table is providing the row values;
// - which filter is applied on the result.
//   From this filter, we can further distinguish:
//   - the index-specific part of the filter, which uses only columns
//     provided by the index; and
//   - the rest of the filter, which uses (at least) non-indexed columns.
//
// The basic operation is as follows:
//
// - at instantiation:
//
//   - the original table scan is replaced by two scanNodes, one for the
//     index and one for the table.
//   - the filter expression is split in a filter-specific part and
//     table-specific part, and propagated to the respective scanNodes.
//
// - during execution:
//
//   - rows from the index scanNode are fetched; this contains
//     both the indexed columns (as pk of the index itself)
//     and the PK of the indexed table.
//   - using the PK of the indexed table, rows from the indexed
//     table are fetched using the table scanNode.
//
//   The work is batched: we pull joinBatchSize rows from the index
//   and use the primary key to construct spans that are looked up in
//   the table.
//
// In addition to this basic operation, we need to support the
// optimization implemented by setNeededColumns() (needed_columns.go)
// which aims to reduce I/O by avoiding the decoding of column data
// when it is not required downstream. This optimization needs to know
// which columns are needed from the index scanNode and which are
// needed from the table scanNode. This is determined as follows:
//
// - from the index scanNode: we need at least the indexed table's PK
//   (for otherwise the table lookup would not be possible), and
//   the columns needed by the index-specific filter.
// - from the table scanNode: we need at least the columns needed by
//   the table-specific filter.
//
// Here the question remains of where to obtain the additional columns
// needed by the downstream consumer node. For any non-indexed
// columns, the table scanNode naturally provides the values. For
// indexed columns, currently the table scanNode also provides the
// values, but really this could be optimized further to re-use the
// column data from the index scanNode instead. See the comment
// for valNeededForCol in scanNode.

type indexJoinNode struct {
	index *scanNode
	table *scanNode

	// primaryKeyPrefix is the KV key prefix of the rows
	// retrieved from the table scanNode.
	primaryKeyPrefix roachpb.Key

	// colIDtoRowIndex maps column IDs in the table scanNode into column
	// IDs in the index scanNode's results. The presence of a column ID
	// in this mapping is not sufficient to cause a column's values to
	// be produced; which columns are effectively loaded are decided by
	// the scanNodes' own valNeededForCol, which is updated by
	// setNeededColumns(). So there may be more columns in
	// colIDtoRowIndex than effectively accessed.
	colIDtoRowIndex map[sqlbase.ColumnID]int

	// primaryKeyColumns is the set of PK columns for which the
	// indexJoinNode requires a value from the index scanNode, to use as
	// lookup keys in the table scanNode. Note that the index scanNode
	// may produce more values than this, e.g. when its filter expression
	// uses more columns than the PK.
	primaryKeyColumns []bool

	explain   explainMode
	debugVals debugValues
}

// makeIndexJoin build an index join node.
// This destroys the original table scan node argument and reuses its
// storage to construct a new index scan node. A new table scan node
// is created separately as a member of the resulting index join node.
// The new index scan node is also returned alongside the new index join
// node.
func (p *planner) makeIndexJoin(
	origScan *scanNode, exactPrefix int,
) (resultPlan *indexJoinNode, indexScan *scanNode) {
	// Reuse the input argument's scanNode and its initialized parameters
	// at a starting point to build the new indexScan node.
	indexScan = origScan

	// Create a new scanNode that will be used with the primary index.
	table := p.Scan()
	table.desc = origScan.desc
	// Note: initDescDefaults can only error out if its 2nd argument is not nil.
	_ = table.initDescDefaults(origScan.scanVisibility, nil)
	table.initOrdering(0)
	table.disableBatchLimit()

	colIDtoRowIndex := map[sqlbase.ColumnID]int{}

	// primaryKeyColumns defined here will serve both as the primaryKeyColumns
	// field in the indexJoinNode, and to determine which columns are
	// provided by this index for the purpose of splitting the WHERE
	// filter into an index-specific part and a "rest" part.
	primaryKeyColumns := make([]bool, len(origScan.valNeededForCol))
	for _, colID := range table.desc.PrimaryIndex.ColumnIDs {
		// All the PK columns from the table scanNode must
		// be fetched in the index scanNode.
		idx, ok := indexScan.colIdxMap[colID]
		if !ok {
			panic(fmt.Sprintf("Unknown column %d in PrimaryIndex!", colID))
		}
		primaryKeyColumns[idx] = true
		colIDtoRowIndex[colID] = idx
	}

	// To split the WHERE filter into an index-specific part and a
	// "rest" part below the splitFilter() code must know which columns
	// are provided by the index scanNode. Since primaryKeyColumns only
	// contains the PK columns of the indexed table, we also need to
	// gather here which additional columns are indexed. This is done in
	// valProvidedIndex.
	valProvidedIndex := make([]bool, len(origScan.valNeededForCol))

	// Then, in case the index-specific part, post-split, actually
	// refers to any additional column, we also need to prepare the
	// mapping for these columns in colIDtoRowIndex.
	for _, colID := range indexScan.index.ColumnIDs {
		idx, ok := indexScan.colIdxMap[colID]
		if !ok {
			panic(fmt.Sprintf("Unknown column %d in index!", colID))
		}
		valProvidedIndex[idx] = true
		colIDtoRowIndex[colID] = idx
	}

	if origScan.filter != nil {
		// Now we split the filter by extracting the part that can be
		// evaluated using just the index columns.
		splitFunc := func(expr parser.VariableExpr) (ok bool, newExpr parser.Expr) {
			colIdx := expr.(*parser.IndexedVar).Idx
			if !(primaryKeyColumns[colIdx] || valProvidedIndex[colIdx]) {
				return false, nil
			}
			return true, indexScan.filterVars.IndexedVar(colIdx)
		}
		indexScan.filter, table.filter = splitFilter(origScan.filter, splitFunc)
	}

	// splitFilter above may have simplified the filter expression and
	// eliminated all remaining references to some of the
	// IndexedVars. Rebind the indexed vars here so that these stale
	// references are eliminated from the filterVars helper and the set
	// of needed columns is properly determined later by
	// setNeededColumns().
	indexScan.filter = indexScan.filterVars.Rebind(indexScan.filter, true, false)

	// Ensure that the remaining indexed vars are transferred to the
	// table scanNode fully.
	table.filter = table.filterVars.Rebind(table.filter, true, false)

	indexScan.initOrdering(exactPrefix)

	primaryKeyPrefix := roachpb.Key(sqlbase.MakeIndexKeyPrefix(&table.desc, table.index.ID))

	node := &indexJoinNode{
		index:             indexScan,
		table:             table,
		primaryKeyPrefix:  primaryKeyPrefix,
		colIDtoRowIndex:   colIDtoRowIndex,
		primaryKeyColumns: primaryKeyColumns,
	}

	return node, indexScan
}

func (n *indexJoinNode) Columns() ResultColumns {
	return n.table.Columns()
}

func (n *indexJoinNode) Ordering() orderingInfo {
	return n.index.Ordering()
}

func (n *indexJoinNode) Values() parser.Datums {
	return n.table.Values()
}

func (n *indexJoinNode) MarkDebug(mode explainMode) {
	if mode != explainDebug {
		panic(fmt.Sprintf("unknown debug mode %d", mode))
	}
	n.explain = mode
	// Mark both the index and the table scan nodes as debug.
	n.index.MarkDebug(mode)
	n.table.MarkDebug(mode)
}

func (n *indexJoinNode) DebugValues() debugValues {
	if n.explain != explainDebug {
		panic(fmt.Sprintf("node not in debug mode (mode %d)", n.explain))
	}
	return n.debugVals
}

func (n *indexJoinNode) Start() error {
	if err := n.table.Start(); err != nil {
		return err
	}
	return n.index.Start()
}

func (n *indexJoinNode) Next() (bool, error) {
	// Loop looking up the next row. We either are going to pull a row from the
	// table or a batch of rows from the index. If we pull a batch of rows from
	// the index we perform another iteration of the loop looking for rows in the
	// table. This outer loop is necessary because a batch of rows from the index
	// might all be filtered when the resulting rows are read from the table.
	for tableLookup := (len(n.table.spans) > 0); true; tableLookup = true {
		// First, try to pull a row from the table.
		if tableLookup {
			next, err := n.table.Next()
			if err != nil {
				return false, err
			}
			if next {
				if n.explain == explainDebug {
					n.debugVals = n.table.DebugValues()
				}
				return true, nil
			}
		}

		// The table is out of rows. Pull primary keys from the index.
		n.table.scanInitialized = false
		n.table.spans = n.table.spans[:0]

		for len(n.table.spans) < indexJoinBatchSize {
			if next, err := n.index.Next(); !next {
				// The index is out of rows or an error occurred.
				if err != nil {
					return false, err
				}
				if len(n.table.spans) == 0 {
					// The index is out of rows.
					return false, nil
				}
				break
			}

			if n.explain == explainDebug {
				n.debugVals = n.index.DebugValues()
				if n.debugVals.output != debugValueRow {
					return true, nil
				}
			}

			vals := n.index.Values()
			primaryIndexKey, _, err := sqlbase.EncodeIndexKey(
				&n.table.desc, n.table.index, n.colIDtoRowIndex, vals, n.primaryKeyPrefix)
			if err != nil {
				return false, err
			}
			key := roachpb.Key(primaryIndexKey)
			n.table.spans = append(n.table.spans, roachpb.Span{
				Key:    key,
				EndKey: key.PrefixEnd(),
			})

			if n.explain == explainDebug {
				// In debug mode, return the index information as a "partial" row.
				n.debugVals.output = debugValuePartial
				return true, nil
			}
		}

		if log.V(3) {
			log.Infof(n.index.p.ctx(), "table scan: %s", sqlbase.PrettySpans(n.table.spans, 0))
		}
	}
	return false, nil
}

func (n *indexJoinNode) Close() {
	n.index.Close()
	n.table.Close()
}