[PARQUET-577] mandatory status of avro columns ignored

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 1.6.0
• Component/s: parquet-avro
• Assignee: Unassigned
• Reporter: Remek Zajac
• Created: 31/Mar/16 08:56
• Updated: 31/Mar/16 09:04

Description

Avro spec schema [resolution rules ](https://avro.apache.org/docs/1.7.7/spec.html#schema_record) say:
"if the reader's record schema has a field with no default value, and writer's schema does not have a field with the same name, an error is signalled."

I can't find the implementation of this aspect in parquet.avro and indeed observe this rule seemingly ignored. I am using 1.6.0 because that's what we can get off maven.

My writer's schema:

{
  "type" : "record",
  "name" : "SampleSchema_v1",
  "namespace" : "com.xxxx.spark",
  "fields" : [ {
    "name" : "stringField",
    "type" : "string",
    "doc"  : "Sample string field"
  },{
    "name" : "longField",
    "type" : "long",
    "doc"  : "Sample long field"
  } ],
  "doc:" : "A sample/test schema"
}

My reader schema:

{
  "type" : "record",
  "name" : "SampleSchema_newDefaultlessCol",
  "namespace" : "com.xxxx.spark",
  "fields" : [ {
    "name" : "stringField",
    "type" : "string",
    "doc"  : "Sample string field"
  },{
    "name" : "longField",
    "type" : "long",
    "doc"  : "Sample long field"
  },{
    "name" : "mandatoryIntField",
    "type" : "int",
    "doc"  : "Sample mandatory! int field"
  }],
  "doc:" : "v1 + one extra column that has no default"
}

This is my test case:

    "accept new column w/o a default [schema-evolution, undesired]" in new MockAvroParquetGrid {
      //TODO: the behaviour this test case exercises is UNDESIRED, i.e.: a new column with no default value should
      //TODO: Ticket to track this: https://jira.xxxx.io/browse/ADR-610
      //constitute an incompatible schema break, instead, this thing uses 0 for the default
      val inputSampleRecordsV1  = Seq(new SampleSchema_v1(s"string", 1))
      dao.writeParquet[SampleSchema_v1](
        SparkBase.sc.parallelize(inputSampleRecordsV1),
        SampleSchema_v1.SCHEMA$,
        parquetFolder
      )

      dao
        .readParquet[SampleSchema_newDefaultlessCol](parquetFolder, SampleSchema_newDefaultlessCol.SCHEMA$)
        .collect().toSeq.head
        .getMandatoryIntField must equalTo(0) //TODO: zero is an unwelcome guess
    }

This is the implementation of writeParquet and readParquet

  def writeParquet[C](source: RDD[C], schema: org.apache.avro.Schema, dstPath: String)
                     (implicit ctag: ClassTag[C]): Unit = {
    val hadoopJob = Job.getInstance()
    ParquetOutputFormat.setWriteSupportClass(hadoopJob, classOf[AvroWriteSupport])
    ParquetOutputFormat.setCompression(hadoopJob, CompressionCodecName.GZIP)
    AvroWriteSupport.setSchema(hadoopJob.getConfiguration, schema)

    new PairRDDFunctions[Void,C](
      source.map(sourceRecord => (null, sourceRecord))
    ).saveAsNewAPIHadoopFile(
      bucketDAO.uri(dstPath),
      classOf[Void],                            //K
      ctag.runtimeClass.asInstanceOf[Class[C]], //V
      classOf[AvroParquetOutputFormat],
      hadoopJob.getConfiguration
    )
  }

  def readParquet[C](srcPath: String, schema: org.apache.avro.Schema)(implicit ctag: ClassTag[C]): RDD[C] = {
    val hadoopJob = Job.getInstance()
    ParquetInputFormat.setReadSupportClass(hadoopJob, classOf[AvroReadSupport[C]])
    AvroReadSupport.setAvroReadSchema(hadoopJob.getConfiguration, schema)
    sc.newAPIHadoopFile(
      bucketDAO.uri(srcPath),
      classOf[ParquetInputFormat[C]],
      classOf[Void],                            //K
      ctag.runtimeClass.asInstanceOf[Class[C]], //V
      hadoopJob.getConfiguration
    ).map { _._2 }
  }

We use avro-tools to generate java classes from our avro schemas.
java -jar /path/to/avro-tools-1.8.0.jar compile schema <schema file> <destination>

The test case harvests zeroes as values of mandatoryIntField

Naively, I see a problem in the [indexed revord converter](https://git-wip-us.apache.org/repos/asf?p=parquet-mr.git;a=blob;f=parquet-avro/src/main/java/org/apache/parquet/avro/AvroIndexedRecordConverter.java;h=06c66d692571da08298ae1da4f9967446c4864ee;hb=HEAD#l105) in that it cheerfully accepts a condition doomed to fail. The condition being: the reader schema has a column with no default value that is absent in the writer schema.

I am writing predominantly to confirm my diagnosis and to get the intell on why is it implemented the way it is. Is it fixable (or other depend on it as on a feature)? Can people think of a workaround?