You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@parquet.apache.org by "Scott Carey (JIRA)" <ji...@apache.org> on 2018/08/30 05:42:00 UTC
[jira] [Comment Edited] (PARQUET-1407) Data loss on duplicate
values with AvroParquetWriter/Reader
[ https://issues.apache.org/jira/browse/PARQUET-1407?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16596941#comment-16596941 ]
Scott Carey edited comment on PARQUET-1407 at 8/30/18 5:41 AM:
---------------------------------------------------------------
adding a call to duplicate() fixes it.
Apparently Spark isn't calling that, if you use its parquet reader to create a DataFrame the same thing happens, values are skipped. Its a Parquet bug, since GenericRecord.get()'s contract does not indicate that two different instances can return the same ByteBuffer instance. Parquet should call return a duplicate() if it is re-using underlying data between records.
was (Author: scottcarey):
adding a call to duplicate() fixes it.
Apparently Spark isn't calling that, if you use its parquet reader to create a DataFrame the same thing happens, values are skipped. So its either a bug in Parquet or one in Spark.
> Data loss on duplicate values with AvroParquetWriter/Reader
> -----------------------------------------------------------
>
> Key: PARQUET-1407
> URL: https://issues.apache.org/jira/browse/PARQUET-1407
> Project: Parquet
> Issue Type: Bug
> Components: parquet-avro
> Affects Versions: 1.8.3
> Reporter: Scott Carey
> Priority: Critical
>
> {code:java}
> public class Blah {
> private static Path parquetFile = new Path("oops");
> private static Schema schema = SchemaBuilder.record("spark_schema")
> .fields().optionalBytes("value").endRecord();
> private static GenericData.Record recordFor(String value) {
> return new GenericRecordBuilder(schema)
> .set("value", value.getBytes()).build();
> }
> public static void main(String ... args) throws IOException {
> try (ParquetWriter<GenericData.Record> writer = AvroParquetWriter
> .<GenericData.Record>builder(parquetFile)
> .withSchema(schema)
> .build()) {
> writer.write(recordFor("one"));
> writer.write(recordFor("two"));
> writer.write(recordFor("three"));
> writer.write(recordFor("three"));
> writer.write(recordFor("two"));
> writer.write(recordFor("one"));
> writer.write(recordFor("zero"));
> }
> try (ParquetReader<GenericRecord> reader = AvroParquetReader
> .<GenericRecord>builder(parquetFile)
> .withConf(new Configuration()).build()) {
> GenericRecord rec;
> int i = 0;
> while ((rec = reader.read()) != null) {
> ByteBuffer buf = (ByteBuffer) rec.get("value");
> byte[] bytes = new byte[buf.remaining()];
> buf.get(bytes);
> System.out.println("rec " + i++ + ": " + new String(bytes));
> }
> }
> }
> }
> {code}
> Expected output:
> {noformat}
> rec 0: one
> rec 1: two
> rec 2: three
> rec 3: three
> rec 4: two
> rec 5: one
> rec 6: zero{noformat}
> Actual:
> {noformat}
> rec 0: one
> rec 1: two
> rec 2: three
> rec 3:
> rec 4:
> rec 5:
> rec 6: zero{noformat}
>
> This was found when we started getting empty byte[] values back in spark unexpectedly. (Spark 2.3.1 and Parquet 1.8.3). I have not tried to reproduce with parquet 1.9.0, but its a bad enough bug that I would like a 1.8.4 release that I can drop-in replace 1.8.3 without any binary compatibility issues.
> Duplicate byte[] values are lost.
>
> A few clues:
> If I do not call ByteBuffer.get, the size of ByteBuffer.remaining does not go to zero. I suspect a ByteBuffer is being recycled, but the call to ByteBuffer.get mutates it. I wonder if an appropriately placed ByteBuffer.duplicate() would fix it.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)