Examples with RelOptCost org.apache.calcite.plan.RelOptCost used on opensource projects

Example 6 with RelOptCost

use of org.apache.calcite.plan.RelOptCost in project hive by apache.

the class HiveRelMdDistinctRowCount method getCumulativeCost.

/*
   * Favor Broad Plans over Deep Plans.
   */
public RelOptCost getCumulativeCost(HiveJoin rel, RelMetadataQuery mq) {
    RelOptCost cost = mq.getNonCumulativeCost(rel);
    List<RelNode> inputs = rel.getInputs();
    RelOptCost maxICost = HiveCost.ZERO;
    for (RelNode input : inputs) {
        RelOptCost iCost = mq.getCumulativeCost(input);
        if (maxICost.isLt(iCost)) {
            maxICost = iCost;
        }
    }
    return cost.plus(maxICost);
}

Example 7 with RelOptCost

use of org.apache.calcite.plan.RelOptCost in project hive by apache.

the class HiveRelMdCumulativeCost method getCumulativeCost.

/*
   * Favor Broad Plans over Deep Plans.
   */
public RelOptCost getCumulativeCost(HiveJoin rel, RelMetadataQuery mq) {
    RelOptCost cost = mq.getNonCumulativeCost(rel);
    List<RelNode> inputs = rel.getInputs();
    RelOptCost maxICost = HiveCost.ZERO;
    for (RelNode input : inputs) {
        RelOptCost iCost = mq.getCumulativeCost(input);
        if (maxICost.isLt(iCost)) {
            maxICost = iCost;
        }
    }
    return cost.plus(maxICost);
}

Example 8 with RelOptCost

use of org.apache.calcite.plan.RelOptCost in project hive by apache.

the class HiveCardinalityPreservingJoinRule method choosePlan.

private RelNode choosePlan(RelNode node, RelNode optimized) {
    JaninoRelMetadataProvider original = RelMetadataQuery.THREAD_PROVIDERS.get();
    try {
        RelMetadataQuery.THREAD_PROVIDERS.set(HiveTezModelRelMetadataProvider.DEFAULT);
        node.getCluster().invalidateMetadataQuery();
        RelMetadataQuery metadataQuery = RelMetadataQuery.instance();
        RelOptCost optimizedCost = metadataQuery.getCumulativeCost(optimized);
        RelOptCost originalCost = metadataQuery.getCumulativeCost(node);
        originalCost = originalCost.multiplyBy(factor);
        LOG.debug("Original plan cost {} vs Optimized plan cost {}", originalCost, optimizedCost);
        if (optimizedCost.isLt(originalCost)) {
            if (LOG.isDebugEnabled()) {
                LOG.debug("Plan after:\n" + RelOptUtil.toString(optimized));
            }
            return optimized;
        }
        return node;
    } finally {
        node.getCluster().invalidateMetadataQuery();
        RelMetadataQuery.THREAD_PROVIDERS.set(original);
    }
}

Example 9 with RelOptCost

use of org.apache.calcite.plan.RelOptCost in project calcite by apache.

the class RelMdPercentageOriginalRows method getCumulativeCost.

// Ditto for getNonCumulativeCost
public RelOptCost getCumulativeCost(RelNode rel, RelMetadataQuery mq) {
    RelOptCost cost = mq.getNonCumulativeCost(rel);
    List<RelNode> inputs = rel.getInputs();
    for (RelNode input : inputs) {
        cost = cost.plus(mq.getCumulativeCost(input));
    }
    return cost;
}

Example 10 with RelOptCost

use of org.apache.calcite.plan.RelOptCost in project calcite by apache.

the class LoptSemiJoinOptimizer method computeScore.

/**
 * Computes a score relevant to applying a set of semijoins on a fact table.
 * The higher the score, the better.
 *
 * @param factRel fact table being filtered
 * @param dimRel dimension table that participates in semijoin
 * @param semiJoin semijoin between fact and dimension tables
 *
 * @return computed score of applying the dimension table filters on the
 * fact table
 */
private double computeScore(RelNode factRel, RelNode dimRel, SemiJoin semiJoin) {
    // Estimate savings as a result of applying semijoin filter on fact
    // table.  As a heuristic, the selectivity of the semijoin needs to
    // be less than half.  There may be instances where an even smaller
    // selectivity value is required because of the overhead of
    // index lookups on a very large fact table.  Half was chosen as
    // a middle ground based on testing that was done with a large
    // data set.
    final ImmutableBitSet dimCols = ImmutableBitSet.of(semiJoin.getRightKeys());
    final double selectivity = RelMdUtil.computeSemiJoinSelectivity(mq, factRel, dimRel, semiJoin);
    if (selectivity > .5) {
        return 0;
    }
    final RelOptCost factCost = mq.getCumulativeCost(factRel);
    // if not enough information, return a low score
    if (factCost == null) {
        return 0;
    }
    double savings = (1.0 - Math.sqrt(selectivity)) * Math.max(1.0, factCost.getRows());
    // Additional savings if the dimension columns are unique.  We can
    // ignore nulls since they will be filtered out by the semijoin.
    boolean uniq = RelMdUtil.areColumnsDefinitelyUniqueWhenNullsFiltered(mq, dimRel, dimCols);
    if (uniq) {
        savings *= 2.0;
    }
    // compute the cost of doing an extra scan on the dimension table,
    // including the distinct sort on top of the scan; if the dimension
    // columns are already unique, no need to add on the dup removal cost
    final Double dimSortCost = mq.getRowCount(dimRel);
    final Double dupRemCost = uniq ? 0 : dimSortCost;
    final RelOptCost dimCost = mq.getCumulativeCost(dimRel);
    if ((dimSortCost == null) || (dupRemCost == null) || (dimCost == null)) {
        return 0;
    }
    Double dimRows = dimCost.getRows();
    if (dimRows < 1.0) {
        dimRows = 1.0;
    }
    return savings / dimRows;
}