Examples with HiveTableFunctionScan org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveTableFunctionScan used on opensource projects

Example 1 with HiveTableFunctionScan

use of org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveTableFunctionScan in project hive by apache.

the class HiveUnionSimpleSelectsToInlineTableRule method onMatch.

@Override
public void onMatch(RelOptRuleCall call) {
    RexBuilder rexBuilder = call.builder().getRexBuilder();
    final HiveUnion union = call.rel(0);
    if (!union.all) {
        return;
    }
    List<RelNode> inputs = new ArrayList<RelNode>();
    List<Project> projects = new ArrayList<>();
    List<HiveTableFunctionScan> inlineTables = new ArrayList<>();
    for (RelNode input : union.getInputs()) {
        input = HiveRelDecorrelator.stripHep(input);
        if (isPlainProject(input)) {
            projects.add((Project) input);
            continue;
        }
        if (isInlineTableOperand(input)) {
            inlineTables.add((HiveTableFunctionScan) input);
            continue;
        }
        inputs.add(input);
    }
    if (projects.size() + inlineTables.size() <= 1) {
        // nothing to do
        return;
    }
    RowStorage newRows = new RowStorage();
    for (HiveTableFunctionScan rel : inlineTables) {
        // inline(array(row1,row2,...))
        RexCall rex = (RexCall) ((RexCall) rel.getCall()).operands.get(0);
        for (RexNode row : rex.operands) {
            if (!(row.getType() instanceof RelRecordType)) {
                return;
            }
            newRows.addRow(row);
        }
    }
    for (Project proj : projects) {
        RexNode row = rexBuilder.makeCall(SqlStdOperatorTable.ROW, proj.getProjects());
        if (!(row.getType() instanceof RelRecordType)) {
            return;
        }
        newRows.addRow(row);
    }
    if (newRows.keySet().size() + inputs.size() == union.getInputs().size()) {
        // nothing to do
        return;
    }
    if (dummyTable == null) {
        LOG.warn("Unexpected; rule would match - but dummyTable is not available");
        return;
    }
    for (RelRecordType type : newRows.keySet()) {
        List<RexNode> rows = newRows.get(type);
        RelDataType arrayType = rexBuilder.getTypeFactory().createArrayType(type, -1);
        try {
            SqlOperator inlineFn = SqlFunctionConverter.getCalciteFn("inline", Collections.singletonList(arrayType), type, true, false);
            SqlOperator arrayFn = SqlFunctionConverter.getCalciteFn("array", Collections.nCopies(rows.size(), type), arrayType, true, false);
            RexNode expr = rexBuilder.makeCall(arrayFn, rows);
            expr = rexBuilder.makeCall(inlineFn, expr);
            RelNode newInlineTable = buildTableFunctionScan(expr, union.getCluster());
            inputs.add(newInlineTable);
        } catch (CalciteSemanticException e) {
            LOG.debug("Conversion failed with exception", e);
            return;
        }
    }
    if (inputs.size() > 1) {
        HiveUnion newUnion = (HiveUnion) union.copy(union.getTraitSet(), inputs, true);
        call.transformTo(newUnion);
    } else {
        call.transformTo(inputs.get(0));
    }
}

Example 2 with HiveTableFunctionScan

use of org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveTableFunctionScan in project hive by apache.

the class HiveIntersectRewriteRule method onMatch.

// ~ Methods ----------------------------------------------------------------
public void onMatch(RelOptRuleCall call) {
    final HiveIntersect hiveIntersect = call.rel(0);
    final RelOptCluster cluster = hiveIntersect.getCluster();
    final RexBuilder rexBuilder = cluster.getRexBuilder();
    int numOfBranch = hiveIntersect.getInputs().size();
    Builder<RelNode> bldr = new ImmutableList.Builder<RelNode>();
    // 1st level GB: create a GB (col0, col1, count(1) as c) for each branch
    for (int index = 0; index < numOfBranch; index++) {
        RelNode input = hiveIntersect.getInputs().get(index);
        final List<RexNode> gbChildProjLst = Lists.newArrayList();
        final List<Integer> groupSetPositions = Lists.newArrayList();
        for (int cInd = 0; cInd < input.getRowType().getFieldList().size(); cInd++) {
            gbChildProjLst.add(rexBuilder.makeInputRef(input, cInd));
            groupSetPositions.add(cInd);
        }
        gbChildProjLst.add(rexBuilder.makeBigintLiteral(new BigDecimal(1)));
        // create the project before GB because we need a new project with extra column '1'.
        RelNode gbInputRel = null;
        try {
            gbInputRel = HiveProject.create(input, gbChildProjLst, null);
        } catch (CalciteSemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
        // groupSetPosition includes all the positions
        final ImmutableBitSet groupSet = ImmutableBitSet.of(groupSetPositions);
        List<AggregateCall> aggregateCalls = Lists.newArrayList();
        RelDataType aggFnRetType = TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory());
        // count(1), 1's position is input.getRowType().getFieldList().size()
        AggregateCall aggregateCall = HiveCalciteUtil.createSingleArgAggCall("count", cluster, TypeInfoFactory.longTypeInfo, input.getRowType().getFieldList().size(), aggFnRetType);
        aggregateCalls.add(aggregateCall);
        HiveRelNode aggregateRel = new HiveAggregate(cluster, cluster.traitSetOf(HiveRelNode.CONVENTION), gbInputRel, groupSet, null, aggregateCalls);
        bldr.add(aggregateRel);
    }
    // create a union above all the branches
    HiveRelNode union = new HiveUnion(cluster, TraitsUtil.getDefaultTraitSet(cluster), bldr.build());
    // 2nd level GB: create a GB (col0, col1, count(c)) for each branch
    final List<Integer> groupSetPositions = Lists.newArrayList();
    // the index of c
    int cInd = union.getRowType().getFieldList().size() - 1;
    for (int index = 0; index < union.getRowType().getFieldList().size(); index++) {
        if (index != cInd) {
            groupSetPositions.add(index);
        }
    }
    List<AggregateCall> aggregateCalls = Lists.newArrayList();
    RelDataType aggFnRetType = TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory());
    AggregateCall aggregateCall = HiveCalciteUtil.createSingleArgAggCall("count", cluster, TypeInfoFactory.longTypeInfo, cInd, aggFnRetType);
    aggregateCalls.add(aggregateCall);
    if (hiveIntersect.all) {
        aggregateCall = HiveCalciteUtil.createSingleArgAggCall("min", cluster, TypeInfoFactory.longTypeInfo, cInd, aggFnRetType);
        aggregateCalls.add(aggregateCall);
    }
    final ImmutableBitSet groupSet = ImmutableBitSet.of(groupSetPositions);
    HiveRelNode aggregateRel = new HiveAggregate(cluster, cluster.traitSetOf(HiveRelNode.CONVENTION), union, groupSet, null, aggregateCalls);
    // add a filter count(c) = #branches
    int countInd = cInd;
    List<RexNode> childRexNodeLst = new ArrayList<RexNode>();
    RexInputRef ref = rexBuilder.makeInputRef(aggregateRel, countInd);
    RexLiteral literal = rexBuilder.makeBigintLiteral(new BigDecimal(numOfBranch));
    childRexNodeLst.add(ref);
    childRexNodeLst.add(literal);
    ImmutableList.Builder<RelDataType> calciteArgTypesBldr = new ImmutableList.Builder<RelDataType>();
    calciteArgTypesBldr.add(TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory()));
    calciteArgTypesBldr.add(TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory()));
    RexNode factoredFilterExpr = null;
    try {
        factoredFilterExpr = rexBuilder.makeCall(SqlFunctionConverter.getCalciteFn("=", calciteArgTypesBldr.build(), TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory()), true, false), childRexNodeLst);
    } catch (CalciteSemanticException e) {
        LOG.debug(e.toString());
        throw new RuntimeException(e);
    }
    RelNode filterRel = new HiveFilter(cluster, cluster.traitSetOf(HiveRelNode.CONVENTION), aggregateRel, factoredFilterExpr);
    if (!hiveIntersect.all) {
        // the schema for intersect distinct is like this
        // R3 on all attributes + count(c) as cnt
        // finally add a project to project out the last column
        Set<Integer> projectOutColumnPositions = new HashSet<>();
        projectOutColumnPositions.add(filterRel.getRowType().getFieldList().size() - 1);
        try {
            call.transformTo(HiveCalciteUtil.createProjectWithoutColumn(filterRel, projectOutColumnPositions));
        } catch (CalciteSemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
    } else {
        // the schema for intersect all is like this
        // R3 + count(c) as cnt + min(c) as m
        // we create a input project for udtf whose schema is like this
        // min(c) as m + R3
        List<RexNode> originalInputRefs = Lists.transform(filterRel.getRowType().getFieldList(), new Function<RelDataTypeField, RexNode>() {

            @Override
            public RexNode apply(RelDataTypeField input) {
                return new RexInputRef(input.getIndex(), input.getType());
            }
        });
        List<RexNode> copyInputRefs = new ArrayList<>();
        copyInputRefs.add(originalInputRefs.get(originalInputRefs.size() - 1));
        for (int i = 0; i < originalInputRefs.size() - 2; i++) {
            copyInputRefs.add(originalInputRefs.get(i));
        }
        RelNode srcRel = null;
        try {
            srcRel = HiveProject.create(filterRel, copyInputRefs, null);
            HiveTableFunctionScan udtf = HiveCalciteUtil.createUDTFForSetOp(cluster, srcRel);
            // finally add a project to project out the 1st column
            Set<Integer> projectOutColumnPositions = new HashSet<>();
            projectOutColumnPositions.add(0);
            call.transformTo(HiveCalciteUtil.createProjectWithoutColumn(udtf, projectOutColumnPositions));
        } catch (SemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
    }
}

Example 3 with HiveTableFunctionScan

use of org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveTableFunctionScan in project hive by apache.

the class ASTConverter method convert.

private ASTNode convert() throws CalciteSemanticException {
    /*
     * 1. Walk RelNode Graph; note from, where, gBy.. nodes.
     */
    new QBVisitor().go(root);
    /*
     * 2. convert from node.
     */
    QueryBlockInfo qb = convertSource(from);
    schema = qb.schema;
    hiveAST.from = ASTBuilder.construct(HiveParser.TOK_FROM, "TOK_FROM").add(qb.ast).node();
    /*
     * 3. convert filterNode
     */
    if (where != null) {
        ASTNode cond = where.getCondition().accept(new RexVisitor(schema, false, root.getCluster().getRexBuilder()));
        hiveAST.where = ASTBuilder.where(cond);
        planMapper.link(cond, where);
        planMapper.link(cond, RelTreeSignature.of(where));
    }
    /*
     * 4. GBy
     */
    if (groupBy != null) {
        ASTBuilder b;
        boolean groupingSetsExpression = false;
        Group aggregateType = groupBy.getGroupType();
        switch(aggregateType) {
            case SIMPLE:
                b = ASTBuilder.construct(HiveParser.TOK_GROUPBY, "TOK_GROUPBY");
                break;
            case ROLLUP:
            case CUBE:
            case OTHER:
                b = ASTBuilder.construct(HiveParser.TOK_GROUPING_SETS, "TOK_GROUPING_SETS");
                groupingSetsExpression = true;
                break;
            default:
                throw new CalciteSemanticException("Group type not recognized");
        }
        HiveAggregate hiveAgg = (HiveAggregate) groupBy;
        if (hiveAgg.getAggregateColumnsOrder() != null) {
            // Aggregation columns may have been sorted in specific order
            for (int pos : hiveAgg.getAggregateColumnsOrder()) {
                addRefToBuilder(b, groupBy.getGroupSet().nth(pos));
            }
            for (int pos = 0; pos < groupBy.getGroupCount(); pos++) {
                if (!hiveAgg.getAggregateColumnsOrder().contains(pos)) {
                    addRefToBuilder(b, groupBy.getGroupSet().nth(pos));
                }
            }
        } else {
            // Aggregation columns have not been reordered
            for (int i : groupBy.getGroupSet()) {
                addRefToBuilder(b, i);
            }
        }
        // Grouping sets expressions
        if (groupingSetsExpression) {
            for (ImmutableBitSet groupSet : groupBy.getGroupSets()) {
                ASTBuilder expression = ASTBuilder.construct(HiveParser.TOK_GROUPING_SETS_EXPRESSION, "TOK_GROUPING_SETS_EXPRESSION");
                for (int i : groupSet) {
                    addRefToBuilder(expression, i);
                }
                b.add(expression);
            }
        }
        if (!groupBy.getGroupSet().isEmpty()) {
            hiveAST.groupBy = b.node();
        }
        schema = new Schema(schema, groupBy);
    }
    /*
     * 5. Having
     */
    if (having != null) {
        ASTNode cond = having.getCondition().accept(new RexVisitor(schema, false, root.getCluster().getRexBuilder()));
        hiveAST.having = ASTBuilder.having(cond);
    }
    /*
     * 6. Project
     */
    ASTBuilder b = ASTBuilder.construct(HiveParser.TOK_SELECT, "TOK_SELECT");
    if (select instanceof Project) {
        List<RexNode> childExps = ((Project) select).getProjects();
        if (childExps.isEmpty()) {
            RexLiteral r = select.getCluster().getRexBuilder().makeExactLiteral(new BigDecimal(1));
            ASTNode selectExpr = ASTBuilder.selectExpr(ASTBuilder.literal(r), "1");
            b.add(selectExpr);
        } else {
            int i = 0;
            for (RexNode r : childExps) {
                ASTNode expr = r.accept(new RexVisitor(schema, r instanceof RexLiteral, select.getCluster().getRexBuilder()));
                String alias = select.getRowType().getFieldNames().get(i++);
                ASTNode selectExpr = ASTBuilder.selectExpr(expr, alias);
                b.add(selectExpr);
            }
        }
        hiveAST.select = b.node();
    } else {
        // select is UDTF
        HiveTableFunctionScan udtf = (HiveTableFunctionScan) select;
        List<ASTNode> children = new ArrayList<>();
        RexCall call = (RexCall) udtf.getCall();
        for (RexNode r : call.getOperands()) {
            ASTNode expr = r.accept(new RexVisitor(schema, r instanceof RexLiteral, select.getCluster().getRexBuilder()));
            children.add(expr);
        }
        ASTBuilder sel = ASTBuilder.construct(HiveParser.TOK_SELEXPR, "TOK_SELEXPR");
        ASTNode function = buildUDTFAST(call.getOperator().getName(), children);
        sel.add(function);
        for (String alias : udtf.getRowType().getFieldNames()) {
            sel.add(HiveParser.Identifier, alias);
        }
        b.add(sel);
        hiveAST.select = b.node();
    }
    /*
     * 7. Order Use in Order By from the block above. RelNode has no pointer to
     * parent hence we need to go top down; but OB at each block really belong
     * to its src/from. Hence the need to pass in sort for each block from
     * its parent.
     * 8. Limit
     */
    convertOrderToASTNode(orderLimit);
    return hiveAST.getAST();
}

Example 4 with HiveTableFunctionScan

use of org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveTableFunctionScan in project hive by apache.

the class HiveExceptRewriteRule method onMatch.

// ~ Methods ----------------------------------------------------------------
public void onMatch(RelOptRuleCall call) {
    final HiveExcept hiveExcept = call.rel(0);
    final RelOptCluster cluster = hiveExcept.getCluster();
    final RexBuilder rexBuilder = cluster.getRexBuilder();
    Builder<RelNode> bldr = new ImmutableList.Builder<RelNode>();
    // branch
    try {
        bldr.add(createFirstGB(hiveExcept.getInputs().get(0), true, cluster, rexBuilder));
        bldr.add(createFirstGB(hiveExcept.getInputs().get(1), false, cluster, rexBuilder));
    } catch (CalciteSemanticException e) {
        LOG.debug(e.toString());
        throw new RuntimeException(e);
    }
    // create a union above all the branches
    // the schema of union looks like this
    // all keys + VCol + c
    HiveRelNode union = new HiveUnion(cluster, TraitsUtil.getDefaultTraitSet(cluster), bldr.build());
    // 2nd level GB: create a GB (all keys + sum(c) as a + sum(VCol*c) as b) for
    // each branch
    final List<RexNode> gbChildProjLst = Lists.newArrayList();
    final List<Integer> groupSetPositions = Lists.newArrayList();
    int unionColumnSize = union.getRowType().getFieldList().size();
    for (int cInd = 0; cInd < unionColumnSize; cInd++) {
        gbChildProjLst.add(rexBuilder.makeInputRef(union, cInd));
        // the last 2 columns are VCol and c
        if (cInd < unionColumnSize - 2) {
            groupSetPositions.add(cInd);
        }
    }
    try {
        gbChildProjLst.add(multiply(rexBuilder.makeInputRef(union, unionColumnSize - 2), rexBuilder.makeInputRef(union, unionColumnSize - 1), cluster, rexBuilder));
    } catch (CalciteSemanticException e) {
        LOG.debug(e.toString());
        throw new RuntimeException(e);
    }
    RelNode gbInputRel = null;
    try {
        // Here we create a project for the following reasons:
        // (1) GBy only accepts arg as a position of the input, however, we need to sum on VCol*c
        // (2) This can better reuse the function createSingleArgAggCall.
        gbInputRel = HiveProject.create(union, gbChildProjLst, null);
    } catch (CalciteSemanticException e) {
        LOG.debug(e.toString());
        throw new RuntimeException(e);
    }
    // gbInputRel's schema is like this
    // all keys + VCol + c + VCol*c
    List<AggregateCall> aggregateCalls = Lists.newArrayList();
    RelDataType aggFnRetType = TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory());
    // sum(c)
    AggregateCall aggregateCall = HiveCalciteUtil.createSingleArgAggCall("sum", cluster, TypeInfoFactory.longTypeInfo, unionColumnSize - 1, aggFnRetType);
    aggregateCalls.add(aggregateCall);
    // sum(VCol*c)
    aggregateCall = HiveCalciteUtil.createSingleArgAggCall("sum", cluster, TypeInfoFactory.longTypeInfo, unionColumnSize, aggFnRetType);
    aggregateCalls.add(aggregateCall);
    final ImmutableBitSet groupSet = ImmutableBitSet.of(groupSetPositions);
    HiveRelNode aggregateRel = new HiveAggregate(cluster, cluster.traitSetOf(HiveRelNode.CONVENTION), gbInputRel, groupSet, null, aggregateCalls);
    if (!hiveExcept.all) {
        RelNode filterRel = null;
        try {
            filterRel = new HiveFilter(cluster, cluster.traitSetOf(HiveRelNode.CONVENTION), aggregateRel, makeFilterExprForExceptDistinct(aggregateRel, unionColumnSize, cluster, rexBuilder));
        } catch (CalciteSemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
        // finally add a project to project out the last 2 columns
        Set<Integer> projectOutColumnPositions = new HashSet<>();
        projectOutColumnPositions.add(filterRel.getRowType().getFieldList().size() - 2);
        projectOutColumnPositions.add(filterRel.getRowType().getFieldList().size() - 1);
        try {
            call.transformTo(HiveCalciteUtil.createProjectWithoutColumn(filterRel, projectOutColumnPositions));
        } catch (CalciteSemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
    } else {
        List<RexNode> originalInputRefs = Lists.transform(aggregateRel.getRowType().getFieldList(), new Function<RelDataTypeField, RexNode>() {

            @Override
            public RexNode apply(RelDataTypeField input) {
                return new RexInputRef(input.getIndex(), input.getType());
            }
        });
        List<RexNode> copyInputRefs = new ArrayList<>();
        try {
            copyInputRefs.add(makeExprForExceptAll(aggregateRel, unionColumnSize, cluster, rexBuilder));
        } catch (CalciteSemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
        for (int i = 0; i < originalInputRefs.size() - 2; i++) {
            copyInputRefs.add(originalInputRefs.get(i));
        }
        RelNode srcRel = null;
        try {
            srcRel = HiveProject.create(aggregateRel, copyInputRefs, null);
            HiveTableFunctionScan udtf = HiveCalciteUtil.createUDTFForSetOp(cluster, srcRel);
            // finally add a project to project out the 1st columns
            Set<Integer> projectOutColumnPositions = new HashSet<>();
            projectOutColumnPositions.add(0);
            call.transformTo(HiveCalciteUtil.createProjectWithoutColumn(udtf, projectOutColumnPositions));
        } catch (SemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
    }
}

Example 5 with HiveTableFunctionScan

use of org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveTableFunctionScan in project hive by apache.

the class HiveRelFieldTrimmer method trimFields.

/**
 * Variant of {@link #trimFields(RelNode, ImmutableBitSet, Set)} for {@link HiveTableFunctionScan}.
 * Copied {@link org.apache.calcite.sql2rel.RelFieldTrimmer#trimFields(
 * org.apache.calcite.rel.logical.LogicalTableFunctionScan, ImmutableBitSet, Set)}
 * and replaced <code>tabFun</code> to {@link HiveTableFunctionScan}.
 * Proper fix would be implement this in calcite.
 */
public TrimResult trimFields(HiveTableFunctionScan tabFun, ImmutableBitSet fieldsUsed, Set<RelDataTypeField> extraFields) {
    final RelDataType rowType = tabFun.getRowType();
    final int fieldCount = rowType.getFieldCount();
    final List<RelNode> newInputs = new ArrayList<>();
    for (RelNode input : tabFun.getInputs()) {
        final int inputFieldCount = input.getRowType().getFieldCount();
        ImmutableBitSet inputFieldsUsed = ImmutableBitSet.range(inputFieldCount);
        // Create input with trimmed columns.
        final Set<RelDataTypeField> inputExtraFields = Collections.emptySet();
        TrimResult trimResult = trimChildRestore(tabFun, input, inputFieldsUsed, inputExtraFields);
        assert trimResult.right.isIdentity();
        newInputs.add(trimResult.left);
    }
    TableFunctionScan newTabFun = tabFun;
    if (!tabFun.getInputs().equals(newInputs)) {
        newTabFun = tabFun.copy(tabFun.getTraitSet(), newInputs, tabFun.getCall(), tabFun.getElementType(), tabFun.getRowType(), tabFun.getColumnMappings());
    }
    assert newTabFun.getClass() == tabFun.getClass();
    // Always project all fields.
    Mapping mapping = Mappings.createIdentity(fieldCount);
    return result(newTabFun, mapping);
}