Reynold Xin commented on SPARK-23325:

Yes perhaps we should do that. It is a lot more work than what you guys think 
though, because as Wenchen said we need to properly define the semantics of all 
the data, similar to all of Hadoop IO (Text, etc) but more, because we have 
more data types.

I'd probably prefer us defining the columnar format first, since if one is 
going after high performance, one'd probably prefer using that one...


> DataSourceV2 readers should always produce InternalRow.
> -------------------------------------------------------
>                 Key: SPARK-23325
>                 URL: https://issues.apache.org/jira/browse/SPARK-23325
>             Project: Spark
>          Issue Type: Sub-task
>          Components: SQL
>    Affects Versions: 2.3.0
>            Reporter: Ryan Blue
>            Priority: Major
> DataSourceV2 row-oriented implementations are limited to producing either 
> {{Row}} instances or {{UnsafeRow}} instances by implementing 
> {{SupportsScanUnsafeRow}}. Instead, I think that implementations should 
> always produce {{InternalRow}}.
> The problem with the choice between {{Row}} and {{UnsafeRow}} is that neither 
> one is appropriate for implementers.
> File formats don't produce {{Row}} instances or the data values used by 
> {{Row}}, like {{java.sql.Timestamp}} and {{java.sql.Date}}. An implementation 
> that uses {{Row}} instances must produce data that is immediately translated 
> from the representation that was just produced by Spark. In my experience, it 
> made little sense to translate a timestamp in microseconds to a 
> (milliseconds, nanoseconds) pair, create a {{Timestamp}} instance, and pass 
> that instance to Spark for immediate translation back.
> On the other hand, {{UnsafeRow}} is very difficult to produce unless data is 
> already held in memory. Even the Parquet support built into Spark 
> deserializes to {{InternalRow}} and then uses {{UnsafeProjection}} to produce 
> unsafe rows. When I went to build an implementation that deserializes Parquet 
> or Avro directly to {{UnsafeRow}} (I tried both), I found that it couldn't be 
> done without first deserializing into memory because the size of an array 
> must be known before any values are written.
> I ended up deciding to deserialize to {{InternalRow}} and use 
> {{UnsafeProjection}} to convert to unsafe. There are two problems with this: 
> first, this is Scala and was difficult to call from Java (it required 
> reflection), and second, this causes double projection in the physical plan 
> (a copy for unsafe to unsafe) if there is a projection that wasn't fully 
> pushed to the data source.
> I think the solution is to have a single interface for readers that expects 
> {{InternalRow}}. Then, a projection should be added in the Spark plan to 
> convert to unsafe and avoid projection in the plan and in the data source. If 
> the data source already produces unsafe rows by deserializing directly, this 
> still minimizes the number of copies because the unsafe projection will check 
> whether the incoming data is already {{UnsafeRow}}.
> Using {{InternalRow}} would also match the interface on the write side.

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org

Reply via email to