Search in sources :

Example 1 with MaterializedScanPlanNode

use of org.voltdb.plannodes.MaterializedScanPlanNode in project voltdb by VoltDB.

the class SubPlanAssembler method injectIndexedJoinWithMaterializedScan.

// Generate a plan for an IN-LIST-driven index scan
private static AbstractPlanNode injectIndexedJoinWithMaterializedScan(AbstractExpression listElements, IndexScanPlanNode scanNode) {
    MaterializedScanPlanNode matScan = new MaterializedScanPlanNode();
    assert (listElements instanceof VectorValueExpression || listElements instanceof ParameterValueExpression);
    matScan.setRowData(listElements);
    matScan.setSortDirection(scanNode.getSortDirection());
    NestLoopIndexPlanNode nlijNode = new NestLoopIndexPlanNode();
    nlijNode.setJoinType(JoinType.INNER);
    nlijNode.addInlinePlanNode(scanNode);
    nlijNode.addAndLinkChild(matScan);
    // resolve the sort direction
    nlijNode.resolveSortDirection();
    return nlijNode;
}
Also used : VectorValueExpression(org.voltdb.expressions.VectorValueExpression) MaterializedScanPlanNode(org.voltdb.plannodes.MaterializedScanPlanNode) ParameterValueExpression(org.voltdb.expressions.ParameterValueExpression) NestLoopIndexPlanNode(org.voltdb.plannodes.NestLoopIndexPlanNode)

Example 2 with MaterializedScanPlanNode

use of org.voltdb.plannodes.MaterializedScanPlanNode in project voltdb by VoltDB.

the class SelectSubPlanAssembler method getSelectSubPlanForJoin.

/**
     * Given a join node and plan-sub-graph for outer and inner sub-nodes,
     * construct the plan-sub-graph for that node.
     *
     * @param joinNode A parent join node.
     * @param outerPlan The outer node plan-sub-graph.
     * @param innerPlan The inner node plan-sub-graph.
     * @return A completed plan-sub-graph
     * or null if a valid plan can not be produced for given access paths.
     */
private IndexSortablePlanNode getSelectSubPlanForJoin(BranchNode joinNode, AbstractPlanNode outerPlan, AbstractPlanNode innerPlan) {
    // Filter (post-join) expressions
    ArrayList<AbstractExpression> whereClauses = new ArrayList<>();
    whereClauses.addAll(joinNode.m_whereInnerList);
    whereClauses.addAll(joinNode.m_whereInnerOuterList);
    if (joinNode.getJoinType() == JoinType.FULL) {
        // For all other join types, the whereOuterList expressions were pushed down to the outer node
        whereClauses.addAll(joinNode.m_whereOuterList);
    }
    assert (joinNode.getRightNode() != null);
    JoinNode innerJoinNode = joinNode.getRightNode();
    AccessPath innerAccessPath = innerJoinNode.m_currentAccessPath;
    // We may need to add a send/receive pair to the inner plan for the special case.
    // This trick only works once per plan, BUT once the partitioned data has been
    // received on the coordinator, it can be treated as replicated data in later
    // joins, which MAY help with later outer joins with replicated data.
    boolean needInnerSendReceive = m_partitioning.requiresTwoFragments() && !innerPlan.hasReplicatedResult() && outerPlan.hasReplicatedResult() && joinNode.getJoinType() != JoinType.INNER;
    // When the inner plan is an IndexScan, there MAY be a choice of whether to join using a
    // NestLoopJoin (NLJ) or a NestLoopIndexJoin (NLIJ). The NLJ will have an advantage over the
    // NLIJ in the cases where it applies, since it does a single access or iteration over the index
    // and caches the result, where the NLIJ does an index access or iteration for each outer row.
    // The NestLoopJoin applies when the inner IndexScan is driven only by parameter and constant
    // expressions determined at the start of the query. That requires that none of the IndexScan's
    // various expressions that drive the index access may reference columns from the outer row
    // -- they can only reference columns of the index's base table (the indexed expressions)
    // as well as constants and parameters. The IndexScan's "otherExprs" expressions that only
    // drive post-filtering are not an issue since the NestLoopJoin does feature per-outer-tuple
    // post-filtering on each pass over the cached index scan result.
    // The special case of an OUTER JOIN of replicated outer row data with a partitioned inner
    // table requires that the partitioned data be sent to the coordinator prior to the join.
    // This limits the join option to NLJ. The index scan must make a single index access on
    // each partition and cache the result at the coordinator for post-filtering.
    // This requires that the index access be based on parameters or constants only
    // -- the replicated outer row data will only be available later at the coordinator,
    // so it can not drive the per-partition index scan.
    // If the NLJ option is precluded for the usual reason (outer-row-based indexing) AND
    // the NLIJ is precluded by the special case (OUTER JOIN of replicated outer rows and
    // partitioned inner rows) this method returns null, effectively rejecting this indexed
    // access path for the inner node. Other access paths or join orders may prove more successful.
    boolean canHaveNLJ = true;
    boolean canHaveNLIJ = true;
    if (innerPlan instanceof IndexScanPlanNode) {
        if (hasInnerOuterIndexExpression(joinNode.getRightNode().getTableAlias(), innerAccessPath.indexExprs, innerAccessPath.initialExpr, innerAccessPath.endExprs)) {
            canHaveNLJ = false;
        }
    } else {
        canHaveNLIJ = false;
    }
    if (needInnerSendReceive) {
        canHaveNLIJ = false;
    }
    // partition columns
    if (joinNode.getJoinType() == JoinType.FULL && m_partitioning.requiresTwoFragments() && !outerPlan.hasReplicatedResult() && innerPlan.hasReplicatedResult()) {
        canHaveNLIJ = false;
        canHaveNLJ = false;
    }
    AbstractJoinPlanNode ajNode = null;
    if (canHaveNLJ) {
        NestLoopPlanNode nljNode = new NestLoopPlanNode();
        // get all the clauses that join the applicable two tables
        // Copy innerAccessPath.joinExprs to leave it unchanged,
        // avoiding accumulation of redundant expressions when
        // joinClauses gets built up for various alternative plans.
        ArrayList<AbstractExpression> joinClauses = new ArrayList<>(innerAccessPath.joinExprs);
        if ((innerPlan instanceof IndexScanPlanNode) || (innerPlan instanceof NestLoopIndexPlanNode && innerPlan.getChild(0) instanceof MaterializedScanPlanNode)) {
            // InnerPlan is an IndexScan OR an NLIJ of a MaterializedScan
            // (IN LIST) and an IndexScan. In this case, the inner and
            // inner-outer non-index join expressions (if any) are in the
            // indexScan's otherExpr. The former should stay as IndexScanPlan
            // predicates but the latter need to be pulled up into NLJ
            // predicates because the IndexScan is executed once, not once
            // per outer tuple.
            ArrayList<AbstractExpression> otherExprs = new ArrayList<>();
            // PLEASE do not update the "innerAccessPath.otherExprs", it may be reused
            // for other path evaluation on the other outer side join.
            List<AbstractExpression> innerExpr = filterSingleTVEExpressions(innerAccessPath.otherExprs, otherExprs);
            joinClauses.addAll(otherExprs);
            IndexScanPlanNode scanNode = null;
            if (innerPlan instanceof IndexScanPlanNode) {
                scanNode = (IndexScanPlanNode) innerPlan;
            } else {
                assert (innerPlan instanceof NestLoopIndexPlanNode);
                scanNode = ((NestLoopIndexPlanNode) innerPlan).getInlineIndexScan();
            }
            scanNode.setPredicate(innerExpr);
        } else if (innerJoinNode instanceof BranchNode && joinNode.getJoinType() != JoinType.INNER) {
            // If the innerJoinNode is a LEAF node OR if the join type is an INNER join,
            // the conditions that apply to the inner side
            // have been applied as predicates to the inner scan node already.
            // otherExpr of innerAccessPath comes from its parentNode's joinInnerList.
            // For Outer join (LEFT or FULL), it could mean a join predicate on the table of
            // the inner node ONLY, that can not be pushed down.
            joinClauses.addAll(innerAccessPath.otherExprs);
        }
        nljNode.setJoinPredicate(ExpressionUtil.combinePredicates(joinClauses));
        // combine the tails plan graph with the new head node
        nljNode.addAndLinkChild(outerPlan);
        // right child node.
        if (needInnerSendReceive) {
            // This trick only works once per plan.
            if (outerPlan.hasAnyNodeOfClass(AbstractReceivePlanNode.class) || innerPlan.hasAnyNodeOfClass(AbstractReceivePlanNode.class)) {
                return null;
            }
            innerPlan = addSendReceivePair(innerPlan);
        }
        nljNode.addAndLinkChild(innerPlan);
        ajNode = nljNode;
    } else if (canHaveNLIJ) {
        NestLoopIndexPlanNode nlijNode = new NestLoopIndexPlanNode();
        IndexScanPlanNode innerNode = (IndexScanPlanNode) innerPlan;
        // Set IndexScan predicate. The INNER join expressions for a FULL join come from
        // the innerAccessPath.joinExprs and need to be combined with the other join expressions
        innerNode.setPredicate(innerAccessPath.joinExprs, innerAccessPath.otherExprs);
        nlijNode.addInlinePlanNode(innerPlan);
        // combine the tails plan graph with the new head node
        nlijNode.addAndLinkChild(outerPlan);
        ajNode = nlijNode;
    } else {
        m_recentErrorMsg = "Unsupported special case of complex OUTER JOIN between replicated outer table and partitioned inner table.";
        return null;
    }
    ajNode.setJoinType(joinNode.getJoinType());
    ajNode.setPreJoinPredicate(ExpressionUtil.combinePredicates(joinNode.m_joinOuterList));
    ajNode.setWherePredicate(ExpressionUtil.combinePredicates(whereClauses));
    ajNode.resolveSortDirection();
    return ajNode;
}
Also used : AbstractReceivePlanNode(org.voltdb.plannodes.AbstractReceivePlanNode) JoinNode(org.voltdb.planner.parseinfo.JoinNode) IndexScanPlanNode(org.voltdb.plannodes.IndexScanPlanNode) AbstractJoinPlanNode(org.voltdb.plannodes.AbstractJoinPlanNode) ArrayList(java.util.ArrayList) NestLoopPlanNode(org.voltdb.plannodes.NestLoopPlanNode) BranchNode(org.voltdb.planner.parseinfo.BranchNode) AbstractExpression(org.voltdb.expressions.AbstractExpression) MaterializedScanPlanNode(org.voltdb.plannodes.MaterializedScanPlanNode) NestLoopIndexPlanNode(org.voltdb.plannodes.NestLoopIndexPlanNode)

Example 3 with MaterializedScanPlanNode

use of org.voltdb.plannodes.MaterializedScanPlanNode in project voltdb by VoltDB.

the class SubPlanAssembler method getIndexAccessPlanForTable.

/**
     * Get an index scan access plan for a table.
     *
     * @param tableAliasIndex The table to get data from.
     * @param path The access path to access the data in the table (index/scan/etc).
     * @return An index scan plan node OR,
               in one edge case, an NLIJ of a MaterializedScan and an index scan plan node.
     */
private static AbstractPlanNode getIndexAccessPlanForTable(StmtTableScan tableScan, AccessPath path) {
    // now assume this will be an index scan and get the relevant index
    Index index = path.index;
    IndexScanPlanNode scanNode = new IndexScanPlanNode(tableScan, index);
    AbstractPlanNode resultNode = scanNode;
    // set sortDirection here because it might be used for IN list
    scanNode.setSortDirection(path.sortDirection);
    // the one element of indexExprs.
    for (AbstractExpression expr : path.indexExprs) {
        if (path.lookupType == IndexLookupType.GEO_CONTAINS) {
            scanNode.addSearchKeyExpression(expr);
            scanNode.addCompareNotDistinctFlag(false);
            continue;
        }
        AbstractExpression exprRightChild = expr.getRight();
        assert (exprRightChild != null);
        if (expr.getExpressionType() == ExpressionType.COMPARE_IN) {
            // Replace this method's result with an injected NLIJ.
            resultNode = injectIndexedJoinWithMaterializedScan(exprRightChild, scanNode);
            // Extract a TVE from the LHS MaterializedScan for use by the IndexScan in its new role.
            MaterializedScanPlanNode matscan = (MaterializedScanPlanNode) resultNode.getChild(0);
            AbstractExpression elemExpr = matscan.getOutputExpression();
            assert (elemExpr != null);
            // Replace the IN LIST condition in the end expression referencing all the list elements
            // with a more efficient equality filter referencing the TVE for each element in turn.
            replaceInListFilterWithEqualityFilter(path.endExprs, exprRightChild, elemExpr);
            // Set up the similar VectorValue --> TVE replacement of the search key expression.
            exprRightChild = elemExpr;
        }
        if (exprRightChild instanceof AbstractSubqueryExpression) {
            // DEAD CODE with the guards on index: ENG-8203
            assert (false);
        }
        scanNode.addSearchKeyExpression(exprRightChild);
        // If the index expression is an "IS NOT DISTINCT FROM" comparison, let the NULL values go through. (ENG-11096)
        scanNode.addCompareNotDistinctFlag(expr.getExpressionType() == ExpressionType.COMPARE_NOTDISTINCT);
    }
    // create the IndexScanNode with all its metadata
    scanNode.setLookupType(path.lookupType);
    scanNode.setBindings(path.bindings);
    scanNode.setEndExpression(ExpressionUtil.combinePredicates(path.endExprs));
    scanNode.setPredicate(path.otherExprs);
    // Propagate the sorting information
    // into the scan node from the access path.
    // The initial expression is needed to control a (short?) forward scan to adjust the start of a reverse
    // iteration after it had to initially settle for starting at "greater than a prefix key".
    scanNode.setInitialExpression(ExpressionUtil.combinePredicates(path.initialExpr));
    scanNode.setSkipNullPredicate();
    scanNode.setEliminatedPostFilters(path.eliminatedPostExprs);
    if (scanNode instanceof IndexSortablePlanNode) {
        IndexUseForOrderBy indexUse = ((IndexSortablePlanNode) scanNode).indexUse();
        indexUse.setWindowFunctionUsesIndex(path.m_windowFunctionUsesIndex);
        indexUse.setSortOrderFromIndexScan(path.sortDirection);
        indexUse.setWindowFunctionIsCompatibleWithOrderBy(path.m_stmtOrderByIsCompatible);
        indexUse.setFinalExpressionOrderFromIndexScan(path.m_finalExpressionOrder);
    }
    return resultNode;
}
Also used : AbstractPlanNode(org.voltdb.plannodes.AbstractPlanNode) IndexSortablePlanNode(org.voltdb.plannodes.IndexSortablePlanNode) AbstractExpression(org.voltdb.expressions.AbstractExpression) IndexScanPlanNode(org.voltdb.plannodes.IndexScanPlanNode) MaterializedScanPlanNode(org.voltdb.plannodes.MaterializedScanPlanNode) AbstractSubqueryExpression(org.voltdb.expressions.AbstractSubqueryExpression) Index(org.voltdb.catalog.Index) IndexUseForOrderBy(org.voltdb.plannodes.IndexUseForOrderBy)

Aggregations

MaterializedScanPlanNode (org.voltdb.plannodes.MaterializedScanPlanNode)3 AbstractExpression (org.voltdb.expressions.AbstractExpression)2 IndexScanPlanNode (org.voltdb.plannodes.IndexScanPlanNode)2 NestLoopIndexPlanNode (org.voltdb.plannodes.NestLoopIndexPlanNode)2 ArrayList (java.util.ArrayList)1 Index (org.voltdb.catalog.Index)1 AbstractSubqueryExpression (org.voltdb.expressions.AbstractSubqueryExpression)1 ParameterValueExpression (org.voltdb.expressions.ParameterValueExpression)1 VectorValueExpression (org.voltdb.expressions.VectorValueExpression)1 BranchNode (org.voltdb.planner.parseinfo.BranchNode)1 JoinNode (org.voltdb.planner.parseinfo.JoinNode)1 AbstractJoinPlanNode (org.voltdb.plannodes.AbstractJoinPlanNode)1 AbstractPlanNode (org.voltdb.plannodes.AbstractPlanNode)1 AbstractReceivePlanNode (org.voltdb.plannodes.AbstractReceivePlanNode)1 IndexSortablePlanNode (org.voltdb.plannodes.IndexSortablePlanNode)1 IndexUseForOrderBy (org.voltdb.plannodes.IndexUseForOrderBy)1 NestLoopPlanNode (org.voltdb.plannodes.NestLoopPlanNode)1