Examples with PrimaryScanMatchCandidate com.apple.foundationdb.record.query.plan.temp.PrimaryScanMatchCandidate used on opensource projects

Example 1 with PrimaryScanMatchCandidate

use of com.apple.foundationdb.record.query.plan.temp.PrimaryScanMatchCandidate in project fdb-record-layer by FoundationDB.

the class AbstractDataAccessRule method onMatch.

/**
 * Method that does the leg work to create the appropriate expression dag for data access using value indexes or
 * value index-like scans (primary scans).
 *
 * Conceptually we do the following work:
 *
 * <ul>
 * <li> This method yields a scan plan for each matching primary candidate ({@link PrimaryScanMatchCandidate}).
 *      There is only ever going to be exactly one {@link PrimaryScanMatchCandidate} for a primary key. Due to the
 *      candidate being solely based upon a primary key, the match structure is somewhat limited. In essence, there
 *      is an implicit guarantee that we can always create a primary scan for a data source.
 * </li>
 * <li> This method yields an index scan plan for each matching value index candidate
 *      ({@link ValueIndexScanMatchCandidate}).
 * </li>
 * <li> This method yields the combinatorial expansion of intersections of distinct-ed index scan plans.
 * </li>
 * </ul>
 *
 * The work described above is semantically correct in a sense that it creates a search space that can be explored
 * and pruned in suitable ways that will eventually converge into an optimal data access plan.
 *
 * We can choose to create an index scan for every index that is available regardless what the coverage
 * of an index is. The coverage of an index is a measurement that tells us how well an index can answer what a
 * filter (or by extension a query) asks for. For instance, a high number of search arguments used in the index scan
 * can be associated with high coverage (as in the index scan covers more of the query) and vice versa.
 *
 * Similarly, we can choose to create the intersection of all possible combinations of suitable scans over indexes
 * (that we have matches for). Since we create a logical intersection of these access plans we can leave it up to
 * the respective implementation rules (e.g., {@link ImplementIntersectionRule}) to do the right thing and implement
 * the physical plan for the intersection if possible (e.g. ensuring compatibly ordered legs, etc.).
 *
 * In fact, the two before-mentioned approaches are completely valid with respect to correctness of the plan and
 * the guaranteed creation of the optimal plan. However, in reality using this approach, although valid and probably
 * the conceptually better and more orthogonal approach, will result in a ballooning of the search space very quickly.
 * While that may be acceptable for group-by engines and only few index access paths, in an OLTP world where there
 * are potentially dozens of indexes, memory footprint and the sheer number of tasks that would be created for
 * subsequent exploration and implementation of all these alternatives make the purist approach to planning these
 * indexes infeasible.
 *
 * Thus we would like to eliminate unnecessary exploration by avoiding variations we know can never be successful
 * either in creating a successful executable plan (e.g. logical expression may not ever be able to produce a
 * compatible ordering) or cannot ever create an optimal plan. In a nutshell, we try to utilize additional
 * information that is available in addition to the matching partition in order to make decisions about which
 * expression variation to create and which to avoid:
 *
 * <ul>
 * <li> For a matching primary scan candidate ({@link PrimaryScanMatchCandidate})
 *      we will not create a primary scan if the scan is incompatible with an interesting order that has been
 *      communicated downwards in the graph.
 * </li>
 * <li> For a matching index scan candidate ({@link ValueIndexScanMatchCandidate})
 *      we will not create an index scan if the scan is incompatible with an interesting order that has been
 *      communicated downwards in the graph.
 * </li>
 * <li> We will only create a scan if there is no other index scan with a greater coverage (think of coverage
 *      as the assumed amount of filtering or currently the number of bound predicates) for the search arguments
 *      which are bound by the query.
 *      For instance, an index scan {@code INDEX SCAN(i1, a = [5, 5], b = [10, 10])} is still planned along
 *      {@code INDEX SCAN(i2, x = ["hello", "hello"], y = ["world", "world"], z = [10, inf])} even though
 *      the latter utilizes three search arguments while the former one only uses two. However, an index scan
 *      {@code INDEX SCAN(i1, a = [5, 5], b = [10, 10])} is not created (and yielded) if there we also
 *      have a choice to plan {@code INDEX SCAN(i2, b = [10, 10], a = [5, 5], c = ["Guten", "Morgen"])} as that
 *      index {@code i2} has a higher coverage compared to {@code i1} <em>and</em> all bound arguments in the scan
 *      over {@code i2} are also bound in the scan over {@code i1}.
 * <li>
 *      We will only create intersections of scans if we can already establish that the logical intersection
 *      can be implemented by a {@link com.apple.foundationdb.record.query.plan.plans.RecordQueryIntersectionPlan}.
 *      That requires that the legs of the intersection are compatibly ordered <em>and</em> that that ordering follows
 *      a potentially required ordering.
 * </li>
 * </ul>
 *
 * @param call the call associated with this planner rule execution
 */
@Override
@SuppressWarnings("java:S135")
public void onMatch(@Nonnull PlannerRuleCall call) {
    final PlannerBindings bindings = call.getBindings();
    final List<? extends PartialMatch> completeMatches = bindings.getAll(getCompleteMatchMatcher());
    final R expression = bindings.get(getExpressionMatcher());
    // 
    if (completeMatches.isEmpty()) {
        return;
    }
    // 
    // return if there is no pre-determined interesting ordering
    // 
    final Optional<Set<RequestedOrdering>> requestedOrderingsOptional = call.getInterestingProperty(OrderingAttribute.ORDERING);
    if (requestedOrderingsOptional.isEmpty()) {
        return;
    }
    final Set<RequestedOrdering> requestedOrderings = requestedOrderingsOptional.get();
    // 
    // group matches by candidates
    // 
    final LinkedHashMap<MatchCandidate, ? extends ImmutableList<? extends PartialMatch>> completeMatchMap = completeMatches.stream().collect(Collectors.groupingBy(PartialMatch::getMatchCandidate, LinkedHashMap::new, ImmutableList.toImmutableList()));
    // find the best match for a candidate as there may be more than one due to partial matching
    final ImmutableSet<PartialMatch> maximumCoverageMatchPerCandidate = completeMatchMap.entrySet().stream().flatMap(entry -> {
        final List<? extends PartialMatch> completeMatchesForCandidate = entry.getValue();
        final Optional<? extends PartialMatch> bestMatchForCandidateOptional = completeMatchesForCandidate.stream().max(Comparator.comparing(PartialMatch::getNumBoundParameterPrefix));
        return bestMatchForCandidateOptional.map(Stream::of).orElse(Stream.empty());
    }).collect(ImmutableSet.toImmutableSet());
    final List<PartialMatch> bestMaximumCoverageMatches = maximumCoverageMatches(maximumCoverageMatchPerCandidate, requestedOrderings);
    if (bestMaximumCoverageMatches.isEmpty()) {
        return;
    }
    // create scans for all best matches
    final Map<PartialMatch, RelationalExpression> bestMatchToExpressionMap = createScansForMatches(bestMaximumCoverageMatches);
    final ExpressionRef<RelationalExpression> toBeInjectedReference = GroupExpressionRef.empty();
    // create single scan accesses
    for (final PartialMatch bestMatch : bestMaximumCoverageMatches) {
        final RelationalExpression dataAccessAndCompensationExpression = compensateSingleDataAccess(bestMatch, bestMatchToExpressionMap.get(bestMatch));
        toBeInjectedReference.insert(dataAccessAndCompensationExpression);
    }
    final Map<PartialMatch, RelationalExpression> bestMatchToDistinctExpressionMap = distinctMatchToScanMap(bestMatchToExpressionMap);
    @Nullable final KeyExpression commonPrimaryKey = call.getContext().getCommonPrimaryKey();
    if (commonPrimaryKey != null) {
        final var commonPrimaryKeyParts = commonPrimaryKey.normalizeKeyForPositions();
        final var boundPartitions = Lists.<List<PartialMatch>>newArrayList();
        // create intersections for all n choose k partitions from k = 2 .. n
        IntStream.range(2, bestMaximumCoverageMatches.size() + 1).mapToObj(k -> ChooseK.chooseK(bestMaximumCoverageMatches, k)).flatMap(iterable -> StreamSupport.stream(iterable.spliterator(), false)).forEach(boundPartitions::add);
        boundPartitions.stream().flatMap(partition -> createIntersectionAndCompensation(commonPrimaryKeyParts, bestMatchToDistinctExpressionMap, partition, requestedOrderings).stream()).forEach(toBeInjectedReference::insert);
    }
    call.yield(inject(expression, completeMatches, toBeInjectedReference));
}