[
https://issues.apache.org/jira/browse/BEAM-7925?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17161255#comment-17161255
]
Neville Li commented on BEAM-7925:
----------------------------------
No that would only work if all un-projected fields are nullable i.e. with
{{"type": ["null", "SomeType"], "default": null}} in the schema.
Otherwise the result Avro records will not survive serialization. That's why
our Parquet Avro IO requires an extra {{projectionFn}} that maps incomplete
Avro into a user type {{T}}, most commonly a case class.
Also Avro is just one of the many supported in-memory representation for
Parquet. Ideally a generic and actually useful Parquet IO should support
different {{Parquet{Read,Write}Support<T>}} classes and composable. But yeah
each format (thrift, Protobuf, etc.) might impose some challengages.
> ParquetIO supports neither column projection nor filter predicate
> -----------------------------------------------------------------
>
> Key: BEAM-7925
> URL: https://issues.apache.org/jira/browse/BEAM-7925
> Project: Beam
> Issue Type: Improvement
> Components: io-java-parquet
> Affects Versions: 2.14.0
> Reporter: Neville Li
> Priority: P2
>
> Current {{ParquetIO}} supports neither column projection nor filter predicate
> which defeats the performance motivation of using Parquet in the first place.
> That's why we have our own implementation of
> [ParquetIO|https://github.com/spotify/scio/tree/master/scio-parquet/src] in
> Scio.
> Reading Parquet as Avro with column projection has some complications,
> namely, the resulting Avro records may be incomplete and will not survive
> ser/de. A workaround maybe provide a {{TypedRead}} interface that takes a
> {{Function<A, B>}} that maps invalid Avro {{A}} into user defined type {{B}}.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
