如何使用avro在镶木地板文件架构中创建REPEATED类型？

Question

We are creating a dataflow pipeline, we will read the data from postgres and write it to a parquet file. ParquetIO.Sink allows you to write a PCollection of GenericRecord into a Parquet file (from here https://beam.apache.org/releases/javadoc/2.5.0/org/apache/beam/sdk/io/parquet/ParquetIO.html ). But the parquet file schema is not like what i expected

here is my schema:

schema = new org.apache.avro.Schema.Parser().parse("{\n" +
         "     \"type\": \"record\",\n" +
         "     \"namespace\": \"com.example\",\n" +
         "     \"name\": \"Patterns\",\n" +
         "     \"fields\": [\n" +
         "       { \"name\": \"id\", \"type\": \"string\" },\n" +
         "       { \"name\": \"name\", \"type\": \"string\" },\n" +
         "       { \"name\": \"createdAt\", \"type\": {\"type\":\"string\",\"logicalType\":\"timestamps-millis\"} },\n" +
         "       { \"name\": \"updatedAt\", \"type\": {\"type\":\"string\",\"logicalType\":\"timestamps-millis\"} },\n" +
         "       { \"name\": \"steps\", \"type\": [\"null\",{\"type\":\"array\",\"items\":{\"type\":\"string\",\"name\":\"json\"}}] },\n" +
         "     ]\n" +
         "}");

this is my code so far:

Pipeline p = Pipeline.create(
        PipelineOptionsFactory.fromArgs(args).withValidation().create());

p.apply(JdbcIO.<GenericRecord> read()
       .withDataSourceConfiguration(JdbcIO.DataSourceConfiguration.create(
             "org.postgresql.Driver", "jdbc:postgresql://localhost:port/database")
             .withUsername("username")
             .withPassword("password"))
       .withQuery("select * from table limit(10)")
       .withCoder(AvroCoder.of(schema))
       .withRowMapper((JdbcIO.RowMapper<GenericRecord>) resultSet -> {
            GenericRecord record = new GenericData.Record(schema);
            ResultSetMetaData metadata = resultSet.getMetaData();
            int columnsNumber = metadata.getColumnCount();
            for(int i=0; i<columnsNumber; i++) {
                Object columnValue = resultSet.getObject(i+1);
                if(columnValue instanceof UUID) columnValue=columnValue.toString();
                if(columnValue instanceof Timestamp) columnValue=columnValue.toString();
                if(columnValue instanceof PgArray) {
                    Object[] array = (Object[]) ((PgArray) columnValue).getArray();
                    List list=new ArrayList();
                    for (Object d : array) {
                        if(d instanceof PGobject) {
                            list.add(((PGobject) d).getValue());
                        }
                    }
                    columnValue = list;
                 }
                 record.put(i, columnValue);
            }
            return record;
        }))
  .apply(FileIO.<GenericRecord>write()
        .via(ParquetIO.sink(schema).withCompressionCodec(CompressionCodecName.SNAPPY))
        .to("something.parquet")
  );

p.run();

this is what i get:

message com.example.table {
  required binary id (UTF8);
  required binary name (UTF8);
  required binary createdAt (UTF8);
  required binary updatedAt (UTF8);
  optional group someArray (LIST) {
    repeated binary array (UTF8);
  }
}

this is what i expected:

message com.example.table {
  required binary id (UTF8);
  required binary name (UTF8);  
  required binary createdAt (UTF8);
  required binary updatedAt (UTF8);
  optional repeated binary someArray(UTF8);
}

please help

Answer 1

Is it a protobuf message you used to describe the expected schema? I think what you got is correctly generated from the specified JSON schema. optional repeated does not make sense in the protobuf language specification: https://developers.google.com/protocol-buffers/docs/reference/proto2-spec

You can remove null and square bracket to generate simply repeated field and it's semantically equivalent to optional repeated (since repeated means zero or more times).

Answer 2

I did not find a way to create a repeated element from Avro that isn't in a GroupType.

The ParquetIO in Beam uses a "standard" avro conversion defined in the parquet-mr project, which is implemented here .

It appears that there are two ways to turn an Avro ARRAY field to a Parquet message -- but neither of them create what you are looking for.

Currently, the avro conversion is the only way to interact with ParquetIO at the moment. I saw this JIRA Use Beam schema in ParquetIO that extend this to Beam Rows, which might permit a different parquet message strategy.

Alternatively, you could create a JIRA feature request for ParquetIO to support thrift structures, which should allow finer control over the parquet structure.