Examples with NAryUnionPlanNode org.apache.flink.optimizer.plan.NAryUnionPlanNode used on opensource projects

Example 11 with NAryUnionPlanNode

use of org.apache.flink.optimizer.plan.NAryUnionPlanNode in project flink by apache.

the class GenericFlatTypePostPass method traverse.

@SuppressWarnings("unchecked")
protected void traverse(PlanNode node, T parentSchema, boolean createUtilities) {
    // distinguish the node types
    if (node instanceof SinkPlanNode) {
        SinkPlanNode sn = (SinkPlanNode) node;
        Channel inchannel = sn.getInput();
        T schema = createEmptySchema();
        sn.postPassHelper = schema;
        // add the sinks information to the schema
        try {
            getSinkSchema(sn, schema);
        } catch (ConflictingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Conflicting type infomation for the data sink '" + sn.getSinkNode().getOperator().getName() + "'.");
        }
        // descend to the input channel
        try {
            propagateToChannel(schema, inchannel, createUtilities);
        } catch (MissingFieldTypeInfoException ex) {
            throw new CompilerPostPassException("Missing type infomation for the channel that inputs to the data sink '" + sn.getSinkNode().getOperator().getName() + "'.");
        }
    } else if (node instanceof SourcePlanNode) {
        if (createUtilities) {
            ((SourcePlanNode) node).setSerializer(createSerializer(parentSchema, node));
        // nothing else to be done here. the source has no input and no strategy itself
        }
    } else if (node instanceof BulkIterationPlanNode) {
        BulkIterationPlanNode iterationNode = (BulkIterationPlanNode) node;
        // get the nodes current schema
        T schema;
        if (iterationNode.postPassHelper == null) {
            schema = createEmptySchema();
            iterationNode.postPassHelper = schema;
        } else {
            schema = (T) iterationNode.postPassHelper;
        }
        schema.increaseNumConnectionsThatContributed();
        // add the parent schema to the schema
        if (propagateParentSchemaDown) {
            addSchemaToSchema(parentSchema, schema, iterationNode.getProgramOperator().getName());
        }
        // check whether all outgoing channels have not yet contributed. come back later if not.
        if (schema.getNumConnectionsThatContributed() < iterationNode.getOutgoingChannels().size()) {
            return;
        }
        if (iterationNode.getRootOfStepFunction() instanceof NAryUnionPlanNode) {
            throw new CompilerException("Optimizer cannot compile an iteration step function where next partial solution is created by a Union node.");
        }
        // traverse the termination criterion for the first time. create schema only, no utilities. Needed in case of intermediate termination criterion
        if (iterationNode.getRootOfTerminationCriterion() != null) {
            SingleInputPlanNode addMapper = (SingleInputPlanNode) iterationNode.getRootOfTerminationCriterion();
            traverse(addMapper.getInput().getSource(), createEmptySchema(), false);
            try {
                addMapper.getInput().setSerializer(createSerializer(createEmptySchema()));
            } catch (MissingFieldTypeInfoException e) {
                throw new RuntimeException(e);
            }
        }
        // traverse the step function for the first time. create schema only, no utilities
        traverse(iterationNode.getRootOfStepFunction(), schema, false);
        T pss = (T) iterationNode.getPartialSolutionPlanNode().postPassHelper;
        if (pss == null) {
            throw new CompilerException("Error in Optimizer Post Pass: Partial solution schema is null after first traversal of the step function.");
        }
        // traverse the step function for the second time, taking the schema of the partial solution
        traverse(iterationNode.getRootOfStepFunction(), pss, createUtilities);
        if (iterationNode.getRootOfTerminationCriterion() != null) {
            SingleInputPlanNode addMapper = (SingleInputPlanNode) iterationNode.getRootOfTerminationCriterion();
            traverse(addMapper.getInput().getSource(), createEmptySchema(), createUtilities);
            try {
                addMapper.getInput().setSerializer(createSerializer(createEmptySchema()));
            } catch (MissingFieldTypeInfoException e) {
                throw new RuntimeException(e);
            }
        }
        // take the schema from the partial solution node and add its fields to the iteration result schema.
        // input and output schema need to be identical, so this is essentially a sanity check
        addSchemaToSchema(pss, schema, iterationNode.getProgramOperator().getName());
        // set the serializer
        if (createUtilities) {
            iterationNode.setSerializerForIterationChannel(createSerializer(pss, iterationNode.getPartialSolutionPlanNode()));
        }
        // done, we can now propagate our info down
        try {
            propagateToChannel(schema, iterationNode.getInput(), createUtilities);
        } catch (MissingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Could not set up runtime strategy for input channel to node '" + iterationNode.getProgramOperator().getName() + "'. Missing type information for key field " + e.getFieldNumber());
        }
    } else if (node instanceof WorksetIterationPlanNode) {
        WorksetIterationPlanNode iterationNode = (WorksetIterationPlanNode) node;
        // get the nodes current schema
        T schema;
        if (iterationNode.postPassHelper == null) {
            schema = createEmptySchema();
            iterationNode.postPassHelper = schema;
        } else {
            schema = (T) iterationNode.postPassHelper;
        }
        schema.increaseNumConnectionsThatContributed();
        // add the parent schema to the schema (which refers to the solution set schema)
        if (propagateParentSchemaDown) {
            addSchemaToSchema(parentSchema, schema, iterationNode.getProgramOperator().getName());
        }
        // check whether all outgoing channels have not yet contributed. come back later if not.
        if (schema.getNumConnectionsThatContributed() < iterationNode.getOutgoingChannels().size()) {
            return;
        }
        if (iterationNode.getNextWorkSetPlanNode() instanceof NAryUnionPlanNode) {
            throw new CompilerException("Optimizer cannot compile a workset iteration step function where the next workset is produced by a Union node.");
        }
        if (iterationNode.getSolutionSetDeltaPlanNode() instanceof NAryUnionPlanNode) {
            throw new CompilerException("Optimizer cannot compile a workset iteration step function where the solution set delta is produced by a Union node.");
        }
        // traverse the step function
        // pass an empty schema to the next workset and the parent schema to the solution set delta
        // these first traversals are schema only
        traverse(iterationNode.getNextWorkSetPlanNode(), createEmptySchema(), false);
        traverse(iterationNode.getSolutionSetDeltaPlanNode(), schema, false);
        T wss = (T) iterationNode.getWorksetPlanNode().postPassHelper;
        T sss = (T) iterationNode.getSolutionSetPlanNode().postPassHelper;
        if (wss == null) {
            throw new CompilerException("Error in Optimizer Post Pass: Workset schema is null after first traversal of the step function.");
        }
        if (sss == null) {
            throw new CompilerException("Error in Optimizer Post Pass: Solution set schema is null after first traversal of the step function.");
        }
        // make the second pass and instantiate the utilities
        traverse(iterationNode.getNextWorkSetPlanNode(), wss, createUtilities);
        traverse(iterationNode.getSolutionSetDeltaPlanNode(), sss, createUtilities);
        // the solution set input and the result must have the same schema, this acts as a sanity check.
        try {
            for (Map.Entry<Integer, X> entry : sss) {
                Integer pos = entry.getKey();
                schema.addType(pos, entry.getValue());
            }
        } catch (ConflictingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Conflicting type information for field " + e.getFieldNumber() + " in node '" + iterationNode.getProgramOperator().getName() + "'. Contradicting types between the " + "result of the iteration and the solution set schema: " + e.getPreviousType() + " and " + e.getNewType() + ". Most probable cause: Invalid constant field annotations.");
        }
        // set the serializers and comparators
        if (createUtilities) {
            WorksetIterationNode optNode = iterationNode.getIterationNode();
            iterationNode.setWorksetSerializer(createSerializer(wss, iterationNode.getWorksetPlanNode()));
            iterationNode.setSolutionSetSerializer(createSerializer(sss, iterationNode.getSolutionSetPlanNode()));
            try {
                iterationNode.setSolutionSetComparator(createComparator(optNode.getSolutionSetKeyFields(), null, sss));
            } catch (MissingFieldTypeInfoException ex) {
                throw new CompilerPostPassException("Could not set up the solution set for workset iteration '" + optNode.getOperator().getName() + "'. Missing type information for key field " + ex.getFieldNumber() + '.');
            }
        }
        // done, we can now propagate our info down
        try {
            propagateToChannel(schema, iterationNode.getInitialSolutionSetInput(), createUtilities);
            propagateToChannel(wss, iterationNode.getInitialWorksetInput(), createUtilities);
        } catch (MissingFieldTypeInfoException ex) {
            throw new CompilerPostPassException("Could not set up runtime strategy for input channel to node '" + iterationNode.getProgramOperator().getName() + "'. Missing type information for key field " + ex.getFieldNumber());
        }
    } else if (node instanceof SingleInputPlanNode) {
        SingleInputPlanNode sn = (SingleInputPlanNode) node;
        // get the nodes current schema
        T schema;
        if (sn.postPassHelper == null) {
            schema = createEmptySchema();
            sn.postPassHelper = schema;
        } else {
            schema = (T) sn.postPassHelper;
        }
        schema.increaseNumConnectionsThatContributed();
        SingleInputNode optNode = sn.getSingleInputNode();
        // add the parent schema to the schema
        if (propagateParentSchemaDown) {
            addSchemaToSchema(parentSchema, schema, optNode, 0);
        }
        // check whether all outgoing channels have not yet contributed. come back later if not.
        if (schema.getNumConnectionsThatContributed() < sn.getOutgoingChannels().size()) {
            return;
        }
        // add the nodes local information
        try {
            getSingleInputNodeSchema(sn, schema);
        } catch (ConflictingFieldTypeInfoException e) {
            throw new CompilerPostPassException(getConflictingTypeErrorMessage(e, optNode.getOperator().getName()));
        }
        if (createUtilities) {
            // parameterize the node's driver strategy
            for (int i = 0; i < sn.getDriverStrategy().getNumRequiredComparators(); i++) {
                try {
                    sn.setComparator(createComparator(sn.getKeys(i), sn.getSortOrders(i), schema), i);
                } catch (MissingFieldTypeInfoException e) {
                    throw new CompilerPostPassException("Could not set up runtime strategy for node '" + optNode.getOperator().getName() + "'. Missing type information for key field " + e.getFieldNumber());
                }
            }
        }
        // done, we can now propagate our info down
        try {
            propagateToChannel(schema, sn.getInput(), createUtilities);
        } catch (MissingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Could not set up runtime strategy for input channel to node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
        }
        // don't forget the broadcast inputs
        for (Channel c : sn.getBroadcastInputs()) {
            try {
                propagateToChannel(createEmptySchema(), c, createUtilities);
            } catch (MissingFieldTypeInfoException e) {
                throw new CompilerPostPassException("Could not set up runtime strategy for broadcast channel in node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
            }
        }
    } else if (node instanceof DualInputPlanNode) {
        DualInputPlanNode dn = (DualInputPlanNode) node;
        // get the nodes current schema
        T schema1;
        T schema2;
        if (dn.postPassHelper1 == null) {
            schema1 = createEmptySchema();
            schema2 = createEmptySchema();
            dn.postPassHelper1 = schema1;
            dn.postPassHelper2 = schema2;
        } else {
            schema1 = (T) dn.postPassHelper1;
            schema2 = (T) dn.postPassHelper2;
        }
        schema1.increaseNumConnectionsThatContributed();
        schema2.increaseNumConnectionsThatContributed();
        TwoInputNode optNode = dn.getTwoInputNode();
        // add the parent schema to the schema
        if (propagateParentSchemaDown) {
            addSchemaToSchema(parentSchema, schema1, optNode, 0);
            addSchemaToSchema(parentSchema, schema2, optNode, 1);
        }
        // check whether all outgoing channels have not yet contributed. come back later if not.
        if (schema1.getNumConnectionsThatContributed() < dn.getOutgoingChannels().size()) {
            return;
        }
        // add the nodes local information
        try {
            getDualInputNodeSchema(dn, schema1, schema2);
        } catch (ConflictingFieldTypeInfoException e) {
            throw new CompilerPostPassException(getConflictingTypeErrorMessage(e, optNode.getOperator().getName()));
        }
        // parameterize the node's driver strategy
        if (createUtilities) {
            if (dn.getDriverStrategy().getNumRequiredComparators() > 0) {
                // set the individual comparators
                try {
                    dn.setComparator1(createComparator(dn.getKeysForInput1(), dn.getSortOrders(), schema1));
                    dn.setComparator2(createComparator(dn.getKeysForInput2(), dn.getSortOrders(), schema2));
                } catch (MissingFieldTypeInfoException e) {
                    throw new CompilerPostPassException("Could not set up runtime strategy for node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
                }
                // set the pair comparator
                try {
                    dn.setPairComparator(createPairComparator(dn.getKeysForInput1(), dn.getKeysForInput2(), dn.getSortOrders(), schema1, schema2));
                } catch (MissingFieldTypeInfoException e) {
                    throw new CompilerPostPassException("Could not set up runtime strategy for node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
                }
            }
        }
        // done, we can now propagate our info down
        try {
            propagateToChannel(schema1, dn.getInput1(), createUtilities);
        } catch (MissingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Could not set up runtime strategy for the first input channel to node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
        }
        try {
            propagateToChannel(schema2, dn.getInput2(), createUtilities);
        } catch (MissingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Could not set up runtime strategy for the second input channel to node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
        }
        // don't forget the broadcast inputs
        for (Channel c : dn.getBroadcastInputs()) {
            try {
                propagateToChannel(createEmptySchema(), c, createUtilities);
            } catch (MissingFieldTypeInfoException e) {
                throw new CompilerPostPassException("Could not set up runtime strategy for broadcast channel in node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
            }
        }
    } else if (node instanceof NAryUnionPlanNode) {
        // only propagate the info down
        try {
            for (Channel channel : node.getInputs()) {
                propagateToChannel(parentSchema, channel, createUtilities);
            }
        } catch (MissingFieldTypeInfoException ex) {
            throw new CompilerPostPassException("Could not set up runtime strategy for the input channel to " + " a union node. Missing type information for field " + ex.getFieldNumber());
        }
    } else // catch the sources of the iterative step functions
    if (node instanceof BulkPartialSolutionPlanNode || node instanceof SolutionSetPlanNode || node instanceof WorksetPlanNode) {
        // get the nodes current schema
        T schema;
        String name;
        if (node instanceof BulkPartialSolutionPlanNode) {
            BulkPartialSolutionPlanNode psn = (BulkPartialSolutionPlanNode) node;
            if (psn.postPassHelper == null) {
                schema = createEmptySchema();
                psn.postPassHelper = schema;
            } else {
                schema = (T) psn.postPassHelper;
            }
            name = "partial solution of bulk iteration '" + psn.getPartialSolutionNode().getIterationNode().getOperator().getName() + "'";
        } else if (node instanceof SolutionSetPlanNode) {
            SolutionSetPlanNode ssn = (SolutionSetPlanNode) node;
            if (ssn.postPassHelper == null) {
                schema = createEmptySchema();
                ssn.postPassHelper = schema;
            } else {
                schema = (T) ssn.postPassHelper;
            }
            name = "solution set of workset iteration '" + ssn.getSolutionSetNode().getIterationNode().getOperator().getName() + "'";
        } else if (node instanceof WorksetPlanNode) {
            WorksetPlanNode wsn = (WorksetPlanNode) node;
            if (wsn.postPassHelper == null) {
                schema = createEmptySchema();
                wsn.postPassHelper = schema;
            } else {
                schema = (T) wsn.postPassHelper;
            }
            name = "workset of workset iteration '" + wsn.getWorksetNode().getIterationNode().getOperator().getName() + "'";
        } else {
            throw new CompilerException();
        }
        schema.increaseNumConnectionsThatContributed();
        // add the parent schema to the schema
        addSchemaToSchema(parentSchema, schema, name);
    } else {
        throw new CompilerPostPassException("Unknown node type encountered: " + node.getClass().getName());
    }
}

Example 12 with NAryUnionPlanNode

use of org.apache.flink.optimizer.plan.NAryUnionPlanNode in project flink by apache.

the class UnionBetweenDynamicAndStaticPathTest method testUnionStaticSecond.

@Test
public void testUnionStaticSecond() {
    try {
        ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
        DataSet<Long> input1 = env.generateSequence(1, 10);
        DataSet<Long> input2 = env.generateSequence(1, 10);
        IterativeDataSet<Long> iteration = input1.iterate(10);
        DataSet<Long> iterResult = iteration.closeWith(iteration.union(iteration).union(input2.union(input2)));
        iterResult.output(new DiscardingOutputFormat<Long>());
        iterResult.output(new DiscardingOutputFormat<Long>());
        Plan p = env.createProgramPlan();
        OptimizedPlan op = compileNoStats(p);
        assertEquals(2, op.getDataSinks().size());
        BulkIterationPlanNode iterPlan = (BulkIterationPlanNode) op.getDataSinks().iterator().next().getInput().getSource();
        SingleInputPlanNode noopNode = (SingleInputPlanNode) iterPlan.getRootOfStepFunction();
        BinaryUnionPlanNode mixedUnion = (BinaryUnionPlanNode) noopNode.getInput().getSource();
        NAryUnionPlanNode staticUnion = (NAryUnionPlanNode) mixedUnion.getInput1().getSource();
        NAryUnionPlanNode dynamicUnion = (NAryUnionPlanNode) mixedUnion.getInput2().getSource();
        assertTrue(mixedUnion.unionsStaticAndDynamicPath());
        assertFalse(mixedUnion.getInput1().isOnDynamicPath());
        assertTrue(mixedUnion.getInput2().isOnDynamicPath());
        assertTrue(mixedUnion.getInput1().getTempMode().isCached());
        assertEquals(0.5, iterPlan.getRelativeMemoryPerSubTask(), 0.0);
        assertEquals(0.5, mixedUnion.getInput1().getRelativeTempMemory(), 0.0);
        assertEquals(0.0, mixedUnion.getInput2().getRelativeTempMemory(), 0.0);
        for (Channel c : staticUnion.getInputs()) {
            assertFalse(c.isOnDynamicPath());
        }
        for (Channel c : dynamicUnion.getInputs()) {
            assertTrue(c.isOnDynamicPath());
        }
        new JobGraphGenerator().compileJobGraph(op);
    } catch (Exception e) {
        e.printStackTrace();
        fail(e.getMessage());
    }
}

Example 13 with NAryUnionPlanNode

use of org.apache.flink.optimizer.plan.NAryUnionPlanNode in project flink by apache.

the class UnionBetweenDynamicAndStaticPathTest method testUnionStaticFirst.

@Test
public void testUnionStaticFirst() {
    try {
        ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
        DataSet<Long> input1 = env.generateSequence(1, 10);
        DataSet<Long> input2 = env.generateSequence(1, 10);
        IterativeDataSet<Long> iteration = input1.iterate(10);
        DataSet<Long> result = iteration.closeWith(input2.union(input2).union(iteration.union(iteration)));
        result.output(new DiscardingOutputFormat<Long>());
        result.output(new DiscardingOutputFormat<Long>());
        Plan p = env.createProgramPlan();
        OptimizedPlan op = compileNoStats(p);
        assertEquals(2, op.getDataSinks().size());
        BulkIterationPlanNode iterPlan = (BulkIterationPlanNode) op.getDataSinks().iterator().next().getInput().getSource();
        SingleInputPlanNode noopNode = (SingleInputPlanNode) iterPlan.getRootOfStepFunction();
        BinaryUnionPlanNode mixedUnion = (BinaryUnionPlanNode) noopNode.getInput().getSource();
        NAryUnionPlanNode staticUnion = (NAryUnionPlanNode) mixedUnion.getInput1().getSource();
        NAryUnionPlanNode dynamicUnion = (NAryUnionPlanNode) mixedUnion.getInput2().getSource();
        assertTrue(mixedUnion.unionsStaticAndDynamicPath());
        assertFalse(mixedUnion.getInput1().isOnDynamicPath());
        assertTrue(mixedUnion.getInput2().isOnDynamicPath());
        assertTrue(mixedUnion.getInput1().getTempMode().isCached());
        for (Channel c : staticUnion.getInputs()) {
            assertFalse(c.isOnDynamicPath());
        }
        for (Channel c : dynamicUnion.getInputs()) {
            assertTrue(c.isOnDynamicPath());
        }
        assertEquals(0.5, iterPlan.getRelativeMemoryPerSubTask(), 0.0);
        assertEquals(0.5, mixedUnion.getInput1().getRelativeTempMemory(), 0.0);
        assertEquals(0.0, mixedUnion.getInput2().getRelativeTempMemory(), 0.0);
        new JobGraphGenerator().compileJobGraph(op);
    } catch (Exception e) {
        e.printStackTrace();
        fail(e.getMessage());
    }
}

Example 14 with NAryUnionPlanNode

use of org.apache.flink.optimizer.plan.NAryUnionPlanNode in project flink by apache.

the class UnionPropertyPropagationTest method testUnion2.

@Test
public void testUnion2() {
    final int NUM_INPUTS = 4;
    // construct the plan it will be multiple flat maps, all unioned
    // and the "unioned" inputDataSet will be grouped
    final ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
    DataSet<String> source = env.readTextFile(IN_FILE);
    DataSet<Tuple2<String, Integer>> lastUnion = source.flatMap(new DummyFlatMap());
    for (int i = 1; i < NUM_INPUTS; i++) {
        lastUnion = lastUnion.union(source.flatMap(new DummyFlatMap()));
    }
    DataSet<Tuple2<String, Integer>> result = lastUnion.groupBy(0).aggregate(Aggregations.SUM, 1);
    result.writeAsText(OUT_FILE);
    // return the plan
    Plan plan = env.createProgramPlan("Test union on new java-api");
    OptimizedPlan oPlan = compileNoStats(plan);
    JobGraphGenerator jobGen = new JobGraphGenerator();
    // Compile plan to verify that no error is thrown
    jobGen.compileJobGraph(oPlan);
    oPlan.accept(new Visitor<PlanNode>() {

        @Override
        public boolean preVisit(PlanNode visitable) {
            /* Test on the union output connections
				 * It must be under the GroupOperator and the strategy should be forward
				 */
            if (visitable instanceof SingleInputPlanNode && visitable.getProgramOperator() instanceof GroupReduceOperatorBase) {
                final Channel inConn = ((SingleInputPlanNode) visitable).getInput();
                Assert.assertTrue("Union should just forward the Partitioning", inConn.getShipStrategy() == ShipStrategyType.FORWARD);
                Assert.assertTrue("Union Node should be under Group operator", inConn.getSource() instanceof NAryUnionPlanNode);
            }
            /* Test on the union input connections
				 * Must be NUM_INPUTS input connections, all FlatMapOperators with a own partitioning strategy(propably hash)
				 */
            if (visitable instanceof NAryUnionPlanNode) {
                int numberInputs = 0;
                for (Iterator<Channel> inputs = visitable.getInputs().iterator(); inputs.hasNext(); numberInputs++) {
                    final Channel inConn = inputs.next();
                    PlanNode inNode = inConn.getSource();
                    Assert.assertTrue("Input of Union should be FlatMapOperators", inNode.getProgramOperator() instanceof FlatMapOperatorBase);
                    Assert.assertTrue("Shipment strategy under union should partition the data", inConn.getShipStrategy() == ShipStrategyType.PARTITION_HASH);
                }
                Assert.assertTrue("NAryUnion should have " + NUM_INPUTS + " inputs", numberInputs == NUM_INPUTS);
                return false;
            }
            return true;
        }

        @Override
        public void postVisit(PlanNode visitable) {
        // DO NOTHING
        }
    });
}

Example 15 with NAryUnionPlanNode

use of org.apache.flink.optimizer.plan.NAryUnionPlanNode in project flink by apache.

the class UnionReplacementTest method testConsecutiveUnionsWithRangePartitioning.

/**
	 *
	 * Checks that a plan with consecutive UNIONs followed by PARTITION_RANGE is correctly translated.
	 *
	 * The program can be illustrated as follows:
	 *
	 * Src1 -\
	 *        >-> Union12--<
	 * Src2 -/              \
	 *                       >-> Union123 -> PartitionByRange -> Output
	 * Src3 ----------------/
	 *
	 * In the resulting plan, the range partitioning must be
	 * pushed to the inputs of the unions (Src1, Src2, Src3).
	 *
	 */
@Test
public void testConsecutiveUnionsWithRangePartitioning() throws Exception {
    // -----------------------------------------------------------------------------------------
    // Build test program
    // -----------------------------------------------------------------------------------------
    ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
    env.setParallelism(DEFAULT_PARALLELISM);
    DataSet<Tuple2<Long, Long>> src1 = env.fromElements(new Tuple2<>(0L, 0L));
    DataSet<Tuple2<Long, Long>> src2 = env.fromElements(new Tuple2<>(0L, 0L));
    DataSet<Tuple2<Long, Long>> src3 = env.fromElements(new Tuple2<>(0L, 0L));
    DataSet<Tuple2<Long, Long>> union12 = src1.union(src2);
    DataSet<Tuple2<Long, Long>> union123 = union12.union(src3);
    union123.partitionByRange(1).output(new DiscardingOutputFormat<Tuple2<Long, Long>>()).name("out");
    // -----------------------------------------------------------------------------------------
    // Verify optimized plan
    // -----------------------------------------------------------------------------------------
    OptimizedPlan optimizedPlan = compileNoStats(env.createProgramPlan());
    OptimizerPlanNodeResolver resolver = getOptimizerPlanNodeResolver(optimizedPlan);
    SingleInputPlanNode sink = resolver.getNode("out");
    // check partitioning is correct
    assertEquals("Sink input should be range partitioned.", PartitioningProperty.RANGE_PARTITIONED, sink.getInput().getGlobalProperties().getPartitioning());
    assertEquals("Sink input should be range partitioned on 1", new Ordering(1, null, Order.ASCENDING), sink.getInput().getGlobalProperties().getPartitioningOrdering());
    SingleInputPlanNode partitioner = (SingleInputPlanNode) sink.getInput().getSource();
    assertTrue(partitioner.getDriverStrategy() == DriverStrategy.UNARY_NO_OP);
    assertEquals("Partitioner input should be range partitioned.", PartitioningProperty.RANGE_PARTITIONED, partitioner.getInput().getGlobalProperties().getPartitioning());
    assertEquals("Partitioner input should be range partitioned on 1", new Ordering(1, null, Order.ASCENDING), partitioner.getInput().getGlobalProperties().getPartitioningOrdering());
    assertEquals("Partitioner input channel should be forwarding", ShipStrategyType.FORWARD, partitioner.getInput().getShipStrategy());
    NAryUnionPlanNode union = (NAryUnionPlanNode) partitioner.getInput().getSource();
    // all union inputs should be range partitioned
    for (Channel c : union.getInputs()) {
        assertEquals("Union input should be range partitioned", PartitioningProperty.RANGE_PARTITIONED, c.getGlobalProperties().getPartitioning());
        assertEquals("Union input channel should be forwarded", ShipStrategyType.FORWARD, c.getShipStrategy());
        // range partitioning is executed as custom partitioning with prior sampling
        SingleInputPlanNode partitionMap = (SingleInputPlanNode) c.getSource();
        assertEquals(DriverStrategy.MAP, partitionMap.getDriverStrategy());
        assertEquals(ShipStrategyType.PARTITION_CUSTOM, partitionMap.getInput().getShipStrategy());
    }
}