How to store Protobuf messages in a consolidated Parquet file using Spark

user_5630335 0
Viewed 902

I have a Kafka topic where Protobuf messages arive. I want to store them in blob storage and then be able to query them. Due to the volume of data I want to do this with Spark in order to have a scalable solution. The problem is that Spark does not have native support for Protobuf. I can think of two possible solutions:

  1. Store the raw data as a binary column in some format such as Parquet. When querying, use a UDF to parse the raw data.
  2. Convert the Protobuf to Parquet (this should map 1:1) on write. Then reading/querying with Spark becomes trivial.

I have managed to implement (2) using plain Java using the org.apache.parquet.proto.ProtoParquetWriter, i.e.

        ProtoParquetWriter w = new ProtoParquetWriter(file, MyProtobufClass.class);

        for (Object record : messages) {
            w.write(record);
        }

        w.close();

The problem is how to implement this using Spark?

There is this sparksql-protobuf project but it hasn't been updated in years.

I have found this related question but it is 3,5 years old and has no answer.

apache-spark protocol-buffers
1 Answers

AFAIK there is no easy way to do it using dataset apis. We(in our team/company) use spark hadoop mapreduce apis which uses protobuf-parquet and use spark RDDs to encode/decore protobuf in parquet format. I can show some examples if you have trouble following parquet-protobuf docs.

Working sample

import org.apache.parquet.hadoop.ParquetInputFormat;
import org.apache.parquet.hadoop.ParquetOutputFormat;
import org.apache.parquet.proto.ProtoParquetOutputFormat;
import org.apache.parquet.proto.ProtoReadSupport;
import org.apache.parquet.proto.ProtoWriteSupport;

... 

this.job = Job.getInstance();
ParquetInputFormat.setReadSupportClass(job, ProtoReadSupport.class);
ParquetOutputFormat.setWriteSupportClass(job, ProtoWriteSupport.class);
ProtoParquetOutputFormat.setProtobufClass(job, YourProtoMessage.class); // your proto class

// read
context = new JavaSparkContext(sparkContext)
JavaPairRDD<Void, YourProtoMessage.Builder> input =
                context.newAPIHadoopFile(
                        inputDir,
                        ParquetInputFormat.class,
                        Void.class,
                        YourProtoMessage.Builder.class,
                        job.getConfiguration());


// write

rdd.saveAsNewAPIHadoopFile(OutputDir,
                        Void.class,
                        YourProtoMessage.class,
                        ParquetOutputFormat.class,
                        job.getConfiguration());
user_62796920
Related