CODEWITHSUNDEEP
javahadoopmapreducecascading
Viswa
Viswa

Reputation: 1377

Build custom join logic in Cascading ensuring MAP_SIDE only

I have 3 cascading pipes (one to join against other two) described as follows,

enter image description here

enter image description here

Psuedocode as follows, This example involves two joins

IF F1DecidingFactor = YES then Join LHSPipe with RHS Lookup#1 BY (LHSPipe.F1Input = RHS Lookup#1.Join#F1) and set the lookup result (SET LHSPipe.F1Output = Result#F1) Otherwise SET LHSPipe.F1Output = N/A

The same logic applies for F2 computation.

The expected output,

enter image description here

This scenario forced me to go with Custom Join operation as IF-ELSE decides whether to Join or not.

Considering the above scenario, I would like to go for MAP-SIDE join (keeping RHSPipe in memory of MAP task node), I was thinking of the below possible solutions, each has its pros and cons. Need your suggestions on these.

Option#1:

CoGroup - We can build custom join logic using CoGroup with BufferJoiner followed by custom join (operation), but that wouldnt ensure MAP-SIDE join.

Option#2:

HashJoin - It ensures MAP-SIDE join, but as far as I see custom join cannot be built using this.

Please correct my understanding and suggest your opinions to work on this requirement.

Thanks in advance.

Upvotes: 10

Views: 252

Answers (1)

Ambrish
Ambrish

Reputation: 3677

The best way to solve this problem (which I can think off) is to modify your smaller dataset. You can add a new field (F1DecidingFactor) to the smaller dataset. The value of F1Result can should like:

Sudo code

if F1DecidingFactor == "Yes" then
    F1Result = ACTUAL_VALUE
else
    F1Result = "N/A"

Result Table

|F1#Join|F1#Result|F1#DecidingFactor|
|    Yes|        0|             True|
|    Yes|        1|            False|
|     No|        0|              N/A|
|     No|        1|              N/A|

You can do above via cascading as well.

After this, you can do your map side join.

If modifying smaller dataset is not possible, then I have 2 options to get the problem solved.

Option 1

Add new fields to your small pipes which is equivalent to you deciding factor (i.e. F1DecidingFactor_RHS = Yes). Then include it to your join criteria. Once your join is done, You will have values to only those rows where this condition is matching. Otherwise it will be null/blank. Sample code:

Main Class

import cascading.operation.Insert;
import cascading.pipe.Each;
import cascading.pipe.HashJoin;
import cascading.pipe.Pipe;
import cascading.pipe.assembly.Discard;
import cascading.pipe.joiner.LeftJoin;
import cascading.tuple.Fields;

public class StackHashJoinTestOption2 {
    public StackHashJoinTestOption2() {
        Fields f1Input = new Fields("F1Input");
        Fields f2Input = new Fields("F2Input");
        Fields f1Join = new Fields("F1Join");
        Fields f2Join = new Fields("F2Join");

        Fields f1DecidingFactor = new Fields("F1DecidingFactor");
        Fields f2DecidingFactor = new Fields("F2DecidingFactor");
        Fields f1DecidingFactorRhs = new Fields("F1DecidingFactor_RHS");
        Fields f2DecidingFactorRhs = new Fields("F2DecidingFactor_RHS");

        Fields lhsJoinerOne = f1DecidingFactor.append(f1Input);
        Fields lhsJoinerTwo = f2DecidingFactor.append(f2Input);

        Fields rhsJoinerOne = f1DecidingFactorRhs.append(f1Join);
        Fields rhsJoinerTwo = f2DecidingFactorRhs.append(f2Join);

        Fields functionFields = new Fields("F1DecidingFactor", "F1Output", "F2DecidingFactor", "F2Output");

        // Large Pipe fields : 
        // F1DecidingFactor F1Input F2DecidingFactor F2Input
        Pipe largePipe = new Pipe("large-pipe");

        // Small Pipe 1 Fields : 
        // F1Join F1Result
        Pipe rhsOne = new Pipe("small-pipe-1");

        // New field to small pipe. Expected Fields:
        // F1Join F1Result F1DecidingFactor_RHS
        rhsOne = new Each(rhsOne, new Insert(f1DecidingFactorRhs, "Yes"), Fields.ALL);

        // Small Pipe 2 Fields : 
        // F2Join F2Result
        Pipe rhsTwo = new Pipe("small-pipe-2");

        // New field to small pipe. Expected Fields:
        // F2Join F2Result F2DecidingFactor_RHS
        rhsTwo = new Each(rhsTwo, new Insert(f1DecidingFactorRhs, "Yes"), Fields.ALL);

        // Joining first small pipe. Expected fields after join: 
        // F1DecidingFactor F1Input F2DecidingFactor F2Input F1Join F1Result F1DecidingFactor_RHS
        Pipe resultsOne = new HashJoin(largePipe, lhsJoinerOne, rhsOne, rhsJoinerOne, new LeftJoin());

        // Joining second small pipe. Expected fields after join: 
        // F1DecidingFactor F1Input F2DecidingFactor F2Input F1Join F1Result F1DecidingFactor_RHS F2Join F2Result F2DecidingFactor_RHS
        Pipe resultsTwo = new HashJoin(resultsOne, lhsJoinerTwo, rhsTwo, rhsJoinerTwo, new LeftJoin());

        Pipe result = new Each(resultsTwo, functionFields, new TestFunction(), Fields.REPLACE);

        result = new Discard(result, f1DecidingFactorRhs);
        result = new Discard(result, f2DecidingFactorRhs);

        // result Pipe should have expected result
    }
}

Option 2

If you want to have default value instead of null/blank, then I would suggest you do the HashJoin first with default Joiners followed by a function to update tuples with appropriate values. Something like:

Main Class

import cascading.pipe.Each;
import cascading.pipe.HashJoin;
import cascading.pipe.Pipe;
import cascading.pipe.joiner.LeftJoin;
import cascading.tuple.Fields;

public class StackHashJoinTest {
    public StackHashJoinTest() {
        Fields f1Input = new Fields("F1Input");
        Fields f2Input = new Fields("F2Input");
        Fields f1Join = new Fields("F1Join");
        Fields f2Join = new Fields("F2Join");

        Fields functionFields = new Fields("F1DecidingFactor", "F1Output", "F2DecidingFactor", "F2Output");

        // Large Pipe fields : 
        // F1DecidingFactor F1Input F2DecidingFactor F2Input
        Pipe largePipe = new Pipe("large-pipe");

        // Small Pipe 1 Fields : 
        // F1Join F1Result
        Pipe rhsOne = new Pipe("small-pipe-1");

        // Small Pipe 2 Fields : 
        // F2Join F2Result
        Pipe rhsTwo = new Pipe("small-pipe-2");

        // Joining first small pipe. 
        // Expected fields after join: 
        // F1DecidingFactor F1Input F2DecidingFactor F2Input F1Join F1Result
        Pipe resultsOne = new HashJoin(largePipe, f1Input, rhsOne, f1Join, new LeftJoin());

        // Joining second small pipe. 
        // Expected fields after join: 
        // F1DecidingFactor F1Input F2DecidingFactor F2Input F1Join F1Result F2Join F2Result
        Pipe resultsTwo = new HashJoin(resultsOne, f2Input, rhsTwo, f2Join, new LeftJoin());

        Pipe result = new Each(resultsTwo, functionFields, new TestFunction(), Fields.REPLACE);

        // result Pipe should have expected result
    }
}

Update Function

import cascading.flow.FlowProcess;
import cascading.operation.BaseOperation;
import cascading.operation.Function;
import cascading.operation.FunctionCall;
import cascading.tuple.Fields;
import cascading.tuple.TupleEntry;

public class TestFunction extends BaseOperation<Void> implements Function<Void> {

    private static final long serialVersionUID = 1L;

    private static final String DECIDING_FACTOR = "No";
    private static final String DEFAULT_VALUE = "N/A";

    // Expected Fields: "F1DecidingFactor", "F1Output", "F2DecidingFactor", "F2Output"
    public TestFunction() {
        super(Fields.ARGS);
    }

    @Override
    public void operate(@SuppressWarnings("rawtypes") FlowProcess process, FunctionCall<Void> call) {
        TupleEntry arguments = call.getArguments();

        TupleEntry result = new TupleEntry(arguments);

        if (result.getString("F1DecidingFactor").equalsIgnoreCase(DECIDING_FACTOR)) {
            result.setString("F1Output", DEFAULT_VALUE);
        }

        if (result.getString("F2DecidingFactor").equalsIgnoreCase(DECIDING_FACTOR)) {
            result.setString("F2Output", DEFAULT_VALUE);
        }

        call.getOutputCollector().add(result);
    }

}

References

This should solve your problem. Let me know if this helps.

Upvotes: 1

Related Questions