Apache Flink union operator giving wrong response

Question

I am applying union operator on two DataStreams of Generic record type.

package com.gslab.com.dataSets;
import java.io.File;
import java.util.ArrayList;
import java.util.List;
import org.apache.avro.Schema;
import org.apache.avro.generic.GenericData;
import org.apache.avro.generic.GenericData.Record;
import org.apache.avro.generic.GenericRecord;
import org.apache.flink.api.common.functions.MapFunction;
import org.apache.flink.streaming.api.datastream.DataStream;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;

public class FlinkBroadcast {
    public static void main(String[] args) throws Exception {

        final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
        env.setParallelism(2);

        List<String> controlMessageList = new ArrayList<String>();
        controlMessageList.add("controlMessage1");
        controlMessageList.add("controlMessage2");

        List<String> dataMessageList = new ArrayList<String>();
        dataMessageList.add("Person1");
        dataMessageList.add("Person2");
        dataMessageList.add("Person3");
        dataMessageList.add("Person4");

        DataStream<String> controlMessageStream  = env.fromCollection(controlMessageList);
        DataStream<String> dataMessageStream  = env.fromCollection(dataMessageList);

        DataStream<GenericRecord> controlMessageGenericRecordStream = controlMessageStream.map(new MapFunction<String, GenericRecord>() {
            @Override
            public GenericRecord map(String value) throws Exception {
                 Record gr = new GenericData.Record(new Schema.Parser().parse(new File("src/main/resources/controlMessageSchema.avsc")));
                 gr.put("TYPE", value);
                 return gr;
            }
        });

        DataStream<GenericRecord> dataMessageGenericRecordStream = dataMessageStream.map(new MapFunction<String, GenericRecord>() {
            @Override
            public GenericRecord map(String value) throws Exception {
                 Record gr = new GenericData.Record(new Schema.Parser().parse(new File("src/main/resources/dataMessageSchema.avsc")));
                 gr.put("FIRSTNAME", value);
                 gr.put("LASTNAME", value+": lastname");
                 return gr;
            }
        });

        //Displaying Generic records
        dataMessageGenericRecordStream.map(new MapFunction<GenericRecord, GenericRecord>() {
            @Override
            public GenericRecord map(GenericRecord value) throws Exception {
                System.out.println("data before union: "+ value);
                return value;
            }
        });

        controlMessageGenericRecordStream.broadcast().union(dataMessageGenericRecordStream).map(new MapFunction<GenericRecord, GenericRecord>() {
            @Override
            public GenericRecord map(GenericRecord value) throws Exception {
                System.out.println("data after union: " + value);
                return value;
            }
        });
        env.execute("stream");
    }
}

Output:

05/09/2016 13:02:13 Map(2/2) switched to FINISHED 
data after union: {"TYPE": "controlMessage1"}
data before union: {"FIRSTNAME": "Person2", "LASTNAME": "Person2: lastname"}
data after union: {"TYPE": "controlMessage1"}
data before union: {"FIRSTNAME": "Person1", "LASTNAME": "Person1: lastname"}
data after union: {"TYPE": "controlMessage2"}
data after union: {"TYPE": "controlMessage2"}
data after union: {"FIRSTNAME": "Person1", "LASTNAME": "Person1"}
data before union: {"FIRSTNAME": "Person4", "LASTNAME": "Person4: lastname"}
data before union: {"FIRSTNAME": "Person3", "LASTNAME": "Person3: lastname"}
data after union: {"FIRSTNAME": "Person2", "LASTNAME": "Person2"}
data after union: {"FIRSTNAME": "Person3", "LASTNAME": "Person3"}
05/09/2016 13:02:13 Map -> Map(2/2) switched to FINISHED 
data after union: {"FIRSTNAME": "Person4", "LASTNAME": "Person4"}
05/09/2016 13:02:13 Map -> Map(1/2) switched to FINISHED 
05/09/2016 13:02:13 Map(1/2) switched to FINISHED 
05/09/2016 13:02:13 Map(2/2) switched to FINISHED 
05/09/2016 13:02:13 Job execution switched to status FINISHED.

As you can see records in dataMessageGenericRecordStream are not correct after union. All field values are getting replaced by first fields value.

Answer 1

I have spent a couple days investigating this for a different issue (but still involving GenericRecord) and have found the root cause and the solution.

Root Cause: Within the Apache Avro "Schema.class" the "field" position is a TRANSIENT and does not get serialized by Kryo, and therefore gets initialized as position "0" when deserialized within the Flink pipeline.

See The JIRA AVRO-1476 which describes this and specifically mentions kyro serialization.

This was fixed in Avro 1.7.7

Solution: Flink must use Avro 1.7.7 (or later). I have verified the fix in my local machine by replacing the Avro classes within the flink-dist_2.11-1.1.3.jar and it corrected my issue.

I updated the JIRA issue for this: https://issues.apache.org/jira/browse/FLINK-5039

There is a PR for this now: https://github.com/apache/flink/pull/2953

And I expect it will be included in the Flink 1.1.4 and 1.2.0 builds.

Answer 2

I was facing similar issue in DataSet API. I was reading some Avro files as GenericRecords and saw this weird behavior. I used this workaround- instead of reading them as GenericRecords, I read them as specific records (e.g. MyAvroObject) and then use a map to convert/typecast them as GenericRecords.

I wrote some code to test your use case using DataSet API and it works with the above workaround-

public static void maintest(String[] args) throws Exception {
    ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
    env.setParallelism(2);

    List<String> queryList1 = new ArrayList<String>();
    queryList1.add("query1");
    queryList1.add("query2");

    List<String> queryList2 = new ArrayList<String>();
    queryList2.add("QUERY1");
    queryList2.add("QUERY2");
    queryList2.add("QUERY3");
    queryList2.add("QUERY4");

    DataSet<String> dataset1  = env.fromCollection(queryList1);
    DataSet<String> dataset2  = env.fromCollection(queryList2);

    DataSet<GenericRecord> genericDS1 = dataset1.map(new MapFunction<String, GenericRecord>() {
        @Override
        public GenericRecord map(String value) throws Exception {
            Query query = Query.newBuilder().setQuery(value).build();
            return (GenericRecord) query;
        }
    });

    DataSet<GenericRecord> genericDS2 = dataset2.map(new MapFunction<String, GenericRecord>() {
        @Override
        public GenericRecord map(String value) throws Exception {
            SearchEngineQuery searchEngineQuery = SearchEngineQuery.newBuilder().setSeQuery(value).build();
            return (GenericRecord) searchEngineQuery;
        }
    });

    genericDS2.map(new MapFunction<GenericRecord, GenericRecord>() {
        @Override
        public GenericRecord map(GenericRecord value) throws Exception {
            System.out.println("DEBUG: data before union: " + value);
            return value;
        }
    });

    genericDS1.union(genericDS2).map(new MapFunction<GenericRecord, GenericRecord>() {
        @Override
        public GenericRecord map(GenericRecord value) throws Exception {
            System.out.println("DEBUG: data after union: " + value);
            return value;
        }
    }).print();
}

Where Query and SearchEngine query are my Avro objects (similar to your control message list and data message list).

Output:

{"query": "query1"}
{"se_query": "QUERY1"}
{"se_query": "QUERY3"}
{"query": "query2"}
{"se_query": "QUERY2"}
{"se_query": "QUERY4"}