[
https://issues.apache.org/jira/browse/BEAM-3157?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ismaël Mejía updated BEAM-3157:
-------------------------------
Description:
Currently the Beam SQL transform only supports input and output data
represented as a BeamRecord. This seems to me like an usability limitation
(even if we can do a ParDo to prepare objects before and after the transform).
I suppose this constraint comes from the fact that we need to map
name/type/value from an object field into Calcite so it is convenient to have a
specific data type (BeamRecord) for this. However we can accomplish the same by
using a PCollection of JavaBean (where we know the same information via the
field names/types/values) or by using Avro records where we also have the
Schema information. For the output PCollection we can map the object via a
Reference (e.g. a JavaBean to be filled with the names of an Avro object).
Note: I am assuming for the moment simple mappings since the SQL does not
support composite types for the moment.
A simple API idea would be something like this:
A simple filter:
PCollection<MyPojo> col = BeamSql.query("SELECT * FROM .... WHERE
...").from(MyPojo.class);
A projection:
PCollection<MyNewPojo> newCol = BeamSql.query("SELECT id,
name").from(MyPojo.class).as(MyNewPojo.class);
A first approach could be to just add the extra ParDos + transform DoFns
however I suppose that for memory use reasons maybe mapping directly into
Calcite would make sense.
was:
Currently the Beam SQL transform only supports input and output data
represented as a BeamRecord. This seems to me like an usability limitation
(even if we can do a ParDo to prepare objects before and after the transform).
I suppose this constraint comes from the fact that we need to map
name/type/value from an object field into Calcite so it is convenient to have a
specific data type (BeamRecord) for this. However we can accomplish the same by
using a PCollection of JavaBean (where we know the same information via the
field names/types/values) or by using Avro records where we also have the
Schema information. For the output PCollection we can map the object via a
Reference (e.g. a JavaBean to be filled with the names of an Avro object).
Note: I am assuming for the moment simple mappings since the SQL does not
support composite types for the moment.
A simple API idea would be something like this:
PCollection<MyPojo> col = ...
PCollection<MyNewPojo> newCol = BeamSql.query("SELECT ...", MyNewPojo.class);
A first approach could be to just add the extra ParDos + transform DoFns
however I suppose that for memory use reasons maybe mapping directly into
Calcite would make sense.
> BeamSql transform should support other PCollection types
> --------------------------------------------------------
>
> Key: BEAM-3157
> URL: https://issues.apache.org/jira/browse/BEAM-3157
> Project: Beam
> Issue Type: Improvement
> Components: dsl-sql
> Reporter: Ismaël Mejía
>
> Currently the Beam SQL transform only supports input and output data
> represented as a BeamRecord. This seems to me like an usability limitation
> (even if we can do a ParDo to prepare objects before and after the transform).
> I suppose this constraint comes from the fact that we need to map
> name/type/value from an object field into Calcite so it is convenient to have
> a specific data type (BeamRecord) for this. However we can accomplish the
> same by using a PCollection of JavaBean (where we know the same information
> via the field names/types/values) or by using Avro records where we also have
> the Schema information. For the output PCollection we can map the object via
> a Reference (e.g. a JavaBean to be filled with the names of an Avro object).
> Note: I am assuming for the moment simple mappings since the SQL does not
> support composite types for the moment.
> A simple API idea would be something like this:
> A simple filter:
> PCollection<MyPojo> col = BeamSql.query("SELECT * FROM .... WHERE
> ...").from(MyPojo.class);
> A projection:
> PCollection<MyNewPojo> newCol = BeamSql.query("SELECT id,
> name").from(MyPojo.class).as(MyNewPojo.class);
> A first approach could be to just add the extra ParDos + transform DoFns
> however I suppose that for memory use reasons maybe mapping directly into
> Calcite would make sense.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
