[
https://issues.apache.org/jira/browse/DRILL-3972?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14974941#comment-14974941
]
Julien Le Dem commented on DRILL-3972:
--------------------------------------
Thanks [~andrew.musselman]
I think you should start by looking at [ParquetWriterBatchCreator
here|https://github.com/apache/drill/blob/b4d47c56b6484eedbf3f44516338da4cdbfbb7b4/exec/java-exec/src/main/java/org/apache/drill/exec/store/parquet/ParquetWriterBatchCreator.java#L35]
It needs to return a CloseableRecordBatch which is the class that gets the
record batches in Columns.
Currently it returns a
[WriterRecordBatch|https://github.com/apache/drill/blob/b4d47c56b6484eedbf3f44516338da4cdbfbb7b4/exec/java-exec/src/main/java/org/apache/drill/exec/store/parquet/ParquetFormatPlugin.java#L158]
which writes records one row at a time to a RecordWriter.
{noformat}return new WriterRecordBatch(writer, incoming, context,
getRecordWriter(context, writer));{noformat}
WriterRecordBatch is the common class to write to a row oriented format.
But in this case we'd want to directly convert the columns and not go through a
row representation in between
> Vectorize Parquet Writer
> ------------------------
>
> Key: DRILL-3972
> URL: https://issues.apache.org/jira/browse/DRILL-3972
> Project: Apache Drill
> Issue Type: Bug
> Components: Storage - Parquet
> Reporter: Julien Le Dem
> Labels: pick-me-up
>
> Currently the
> [ParquetRecordWriter|https://github.com/apache/drill/blob/a98da39dd5a8fa368afd8765f4e981826bbfcc0f/exec/java-exec/src/main/java/org/apache/drill/exec/store/parquet/ParquetRecordWriter.java]
> receives one record at a time and then turns that into columns.
> Which means we convert from Drill columns to rows and then to Parquet columns.
> Instead we could directly convert the Drill columns into Parquet columns in a
> vectorized manner.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
