You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@hivemall.apache.org by Walaa Eldin Moustafa <wa...@gmail.com> on 2020/01/14 03:17:32 UTC
Hivemall & Transport UDFs
Hi Hivemall users and devs,
I was not sure if this post would be more suitable to the dev or user
list, so I am copying both. Hivemall sounds like a pretty interesting
project with a lot of potential.
At LinkedIn, we have been working on a relevant project for defining
portable SQL user-defined functions: Transport UDFs
https://github.com/linkedin/transport. This framework enable users to
define UDFs once and have them automatically translated to UDFs that
can be executed natively in multiple engines, such as Spark, Hive and
Presto. The framework can also evaluate UDFs on multiple data formats
such as Avro, or just plain Java objects. Do users of Hivemall think
it is beneficial to implement some of the UDFs in the Transport API so
they can run analytics on multiple engines? Looking forward to your
thoughts.
Thanks,
Walaa.
Re: Hivemall & Transport UDFs
Posted by Walaa Eldin Moustafa <wa...@gmail.com>.
Hi Makoto,
Yes, this is a sharp observation! ConstantObjectInspector is not
currently supported, and it is one of the features I was recently
talking to my teammates about. The limitation is mainly driven by the
lack of this concept in Presto UDFs, but I think we have some
workarounds.
For Spark SQL, Transport UDFs operate directly on Spark InternalRows,
and do not go through the ObjectInspector transformation, and thus
have better performance profile compared to Hive UDFs in Spark. We
have ran benchmarks that show these performance improvements.
Thanks,
Walaa.
On Mon, Jan 13, 2020 at 10:48 PM Makoto Yui <yu...@gmail.com> wrote:
>
> Hi Walaa,
>
> I'm concern that const values (ConstantObjectInspector) in UDF would
> not be handled.
>
> Hivemall use const string for UDF options (e.g., hyperparameters).
> https://github.com/apache/incubator-hivemall/blob/master/core/src/main/java/hivemall/utils/hadoop/HiveUtils.java#L585
> http://hivemall.apache.org/userguide/geospatial/latlon.html#map-url-function
> ('-help')
>
> Also, expecting UDF API package is pushed on Maven central.
> It's welcome if the API package's Java dependencies become Java7 or
> later (I know it's too deprecated).
> https://github.com/linkedin/transport/tree/master/transportable-udfs-api
>
> For Spark, SparkSQL supports Hive UDFs/UDAFs/UDTFs execution in my
> understanding.
> So, presto UDF bridge is only required for our project.
>
> Thanks,
> Makoto
>
> 2020年1月14日(火) 14:03 Walaa Eldin Moustafa <wa...@gmail.com>:
> >
> > Also, wanted to mention that meanwhile, we welcome all kinds of
> > contributions to the framework (support for additional engines ,
> > support for new kinds of UDFs such as UDAFs or UDTFs), or
> > contributions to add new UDFs (such as ML UDFs, geospatial UDFs, JSON,
> > etc).
> >
> > Thanks,
> > Walaa.
> >
> > On Mon, Jan 13, 2020 at 8:55 PM Walaa Eldin Moustafa
> > <wa...@gmail.com> wrote:
> > >
> > > Hi Makoto,
> > >
> > > Thanks for your reply!
> > >
> > > It does not support UDAFs or UDTFs at the moment, but this is one of
> > > the areas we currently want to expand into. For integration with
> > > Hivemall, we can start with some scalar UDFs as a proof of concept.
> > >
> > > Thanks,
> > > Walaa.
> > >
> > > On Mon, Jan 13, 2020 at 8:43 PM Makoto Yui <yu...@gmail.com> wrote:
> > > >
> > > > Hi,
> > > >
> > > > It is definitely interesting since my company uses both Hive and Presto.
> > > >
> > > > Then, does your projects supports UDAFs and UDTFs?
> > > > Cannot find examples at a glance.
> > > >
> > > > Thanks,
> > > > Makoto
> > > >
> > > > 2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
> > > > >
> > > > > Hi Hivemall users and devs,
> > > > >
> > > > > I was not sure if this post would be more suitable to the dev or user
> > > > > list, so I am copying both. Hivemall sounds like a pretty interesting
> > > > > project with a lot of potential.
> > > > >
> > > > > At LinkedIn, we have been working on a relevant project for defining
> > > > > portable SQL user-defined functions: Transport UDFs
> > > > > https://github.com/linkedin/transport. This framework enable users to
> > > > > define UDFs once and have them automatically translated to UDFs that
> > > > > can be executed natively in multiple engines, such as Spark, Hive and
> > > > > Presto. The framework can also evaluate UDFs on multiple data formats
> > > > > such as Avro, or just plain Java objects. Do users of Hivemall think
> > > > > it is beneficial to implement some of the UDFs in the Transport API so
> > > > > they can run analytics on multiple engines? Looking forward to your
> > > > > thoughts.
> > > > >
> > > > > Thanks,
> > > > > Walaa.
Re: Hivemall & Transport UDFs
Posted by Walaa Eldin Moustafa <wa...@gmail.com>.
Hi Makoto,
Yes, this is a sharp observation! ConstantObjectInspector is not
currently supported, and it is one of the features I was recently
talking to my teammates about. The limitation is mainly driven by the
lack of this concept in Presto UDFs, but I think we have some
workarounds.
For Spark SQL, Transport UDFs operate directly on Spark InternalRows,
and do not go through the ObjectInspector transformation, and thus
have better performance profile compared to Hive UDFs in Spark. We
have ran benchmarks that show these performance improvements.
Thanks,
Walaa.
On Mon, Jan 13, 2020 at 10:48 PM Makoto Yui <yu...@gmail.com> wrote:
>
> Hi Walaa,
>
> I'm concern that const values (ConstantObjectInspector) in UDF would
> not be handled.
>
> Hivemall use const string for UDF options (e.g., hyperparameters).
> https://github.com/apache/incubator-hivemall/blob/master/core/src/main/java/hivemall/utils/hadoop/HiveUtils.java#L585
> http://hivemall.apache.org/userguide/geospatial/latlon.html#map-url-function
> ('-help')
>
> Also, expecting UDF API package is pushed on Maven central.
> It's welcome if the API package's Java dependencies become Java7 or
> later (I know it's too deprecated).
> https://github.com/linkedin/transport/tree/master/transportable-udfs-api
>
> For Spark, SparkSQL supports Hive UDFs/UDAFs/UDTFs execution in my
> understanding.
> So, presto UDF bridge is only required for our project.
>
> Thanks,
> Makoto
>
> 2020年1月14日(火) 14:03 Walaa Eldin Moustafa <wa...@gmail.com>:
> >
> > Also, wanted to mention that meanwhile, we welcome all kinds of
> > contributions to the framework (support for additional engines ,
> > support for new kinds of UDFs such as UDAFs or UDTFs), or
> > contributions to add new UDFs (such as ML UDFs, geospatial UDFs, JSON,
> > etc).
> >
> > Thanks,
> > Walaa.
> >
> > On Mon, Jan 13, 2020 at 8:55 PM Walaa Eldin Moustafa
> > <wa...@gmail.com> wrote:
> > >
> > > Hi Makoto,
> > >
> > > Thanks for your reply!
> > >
> > > It does not support UDAFs or UDTFs at the moment, but this is one of
> > > the areas we currently want to expand into. For integration with
> > > Hivemall, we can start with some scalar UDFs as a proof of concept.
> > >
> > > Thanks,
> > > Walaa.
> > >
> > > On Mon, Jan 13, 2020 at 8:43 PM Makoto Yui <yu...@gmail.com> wrote:
> > > >
> > > > Hi,
> > > >
> > > > It is definitely interesting since my company uses both Hive and Presto.
> > > >
> > > > Then, does your projects supports UDAFs and UDTFs?
> > > > Cannot find examples at a glance.
> > > >
> > > > Thanks,
> > > > Makoto
> > > >
> > > > 2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
> > > > >
> > > > > Hi Hivemall users and devs,
> > > > >
> > > > > I was not sure if this post would be more suitable to the dev or user
> > > > > list, so I am copying both. Hivemall sounds like a pretty interesting
> > > > > project with a lot of potential.
> > > > >
> > > > > At LinkedIn, we have been working on a relevant project for defining
> > > > > portable SQL user-defined functions: Transport UDFs
> > > > > https://github.com/linkedin/transport. This framework enable users to
> > > > > define UDFs once and have them automatically translated to UDFs that
> > > > > can be executed natively in multiple engines, such as Spark, Hive and
> > > > > Presto. The framework can also evaluate UDFs on multiple data formats
> > > > > such as Avro, or just plain Java objects. Do users of Hivemall think
> > > > > it is beneficial to implement some of the UDFs in the Transport API so
> > > > > they can run analytics on multiple engines? Looking forward to your
> > > > > thoughts.
> > > > >
> > > > > Thanks,
> > > > > Walaa.
Re: Hivemall & Transport UDFs
Posted by Makoto Yui <yu...@gmail.com>.
Hi Walaa,
I'm concern that const values (ConstantObjectInspector) in UDF would
not be handled.
Hivemall use const string for UDF options (e.g., hyperparameters).
https://github.com/apache/incubator-hivemall/blob/master/core/src/main/java/hivemall/utils/hadoop/HiveUtils.java#L585
http://hivemall.apache.org/userguide/geospatial/latlon.html#map-url-function
('-help')
Also, expecting UDF API package is pushed on Maven central.
It's welcome if the API package's Java dependencies become Java7 or
later (I know it's too deprecated).
https://github.com/linkedin/transport/tree/master/transportable-udfs-api
For Spark, SparkSQL supports Hive UDFs/UDAFs/UDTFs execution in my
understanding.
So, presto UDF bridge is only required for our project.
Thanks,
Makoto
2020年1月14日(火) 14:03 Walaa Eldin Moustafa <wa...@gmail.com>:
>
> Also, wanted to mention that meanwhile, we welcome all kinds of
> contributions to the framework (support for additional engines ,
> support for new kinds of UDFs such as UDAFs or UDTFs), or
> contributions to add new UDFs (such as ML UDFs, geospatial UDFs, JSON,
> etc).
>
> Thanks,
> Walaa.
>
> On Mon, Jan 13, 2020 at 8:55 PM Walaa Eldin Moustafa
> <wa...@gmail.com> wrote:
> >
> > Hi Makoto,
> >
> > Thanks for your reply!
> >
> > It does not support UDAFs or UDTFs at the moment, but this is one of
> > the areas we currently want to expand into. For integration with
> > Hivemall, we can start with some scalar UDFs as a proof of concept.
> >
> > Thanks,
> > Walaa.
> >
> > On Mon, Jan 13, 2020 at 8:43 PM Makoto Yui <yu...@gmail.com> wrote:
> > >
> > > Hi,
> > >
> > > It is definitely interesting since my company uses both Hive and Presto.
> > >
> > > Then, does your projects supports UDAFs and UDTFs?
> > > Cannot find examples at a glance.
> > >
> > > Thanks,
> > > Makoto
> > >
> > > 2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
> > > >
> > > > Hi Hivemall users and devs,
> > > >
> > > > I was not sure if this post would be more suitable to the dev or user
> > > > list, so I am copying both. Hivemall sounds like a pretty interesting
> > > > project with a lot of potential.
> > > >
> > > > At LinkedIn, we have been working on a relevant project for defining
> > > > portable SQL user-defined functions: Transport UDFs
> > > > https://github.com/linkedin/transport. This framework enable users to
> > > > define UDFs once and have them automatically translated to UDFs that
> > > > can be executed natively in multiple engines, such as Spark, Hive and
> > > > Presto. The framework can also evaluate UDFs on multiple data formats
> > > > such as Avro, or just plain Java objects. Do users of Hivemall think
> > > > it is beneficial to implement some of the UDFs in the Transport API so
> > > > they can run analytics on multiple engines? Looking forward to your
> > > > thoughts.
> > > >
> > > > Thanks,
> > > > Walaa.
Re: Hivemall & Transport UDFs
Posted by Makoto Yui <yu...@gmail.com>.
Hi Walaa,
I'm concern that const values (ConstantObjectInspector) in UDF would
not be handled.
Hivemall use const string for UDF options (e.g., hyperparameters).
https://github.com/apache/incubator-hivemall/blob/master/core/src/main/java/hivemall/utils/hadoop/HiveUtils.java#L585
http://hivemall.apache.org/userguide/geospatial/latlon.html#map-url-function
('-help')
Also, expecting UDF API package is pushed on Maven central.
It's welcome if the API package's Java dependencies become Java7 or
later (I know it's too deprecated).
https://github.com/linkedin/transport/tree/master/transportable-udfs-api
For Spark, SparkSQL supports Hive UDFs/UDAFs/UDTFs execution in my
understanding.
So, presto UDF bridge is only required for our project.
Thanks,
Makoto
2020年1月14日(火) 14:03 Walaa Eldin Moustafa <wa...@gmail.com>:
>
> Also, wanted to mention that meanwhile, we welcome all kinds of
> contributions to the framework (support for additional engines ,
> support for new kinds of UDFs such as UDAFs or UDTFs), or
> contributions to add new UDFs (such as ML UDFs, geospatial UDFs, JSON,
> etc).
>
> Thanks,
> Walaa.
>
> On Mon, Jan 13, 2020 at 8:55 PM Walaa Eldin Moustafa
> <wa...@gmail.com> wrote:
> >
> > Hi Makoto,
> >
> > Thanks for your reply!
> >
> > It does not support UDAFs or UDTFs at the moment, but this is one of
> > the areas we currently want to expand into. For integration with
> > Hivemall, we can start with some scalar UDFs as a proof of concept.
> >
> > Thanks,
> > Walaa.
> >
> > On Mon, Jan 13, 2020 at 8:43 PM Makoto Yui <yu...@gmail.com> wrote:
> > >
> > > Hi,
> > >
> > > It is definitely interesting since my company uses both Hive and Presto.
> > >
> > > Then, does your projects supports UDAFs and UDTFs?
> > > Cannot find examples at a glance.
> > >
> > > Thanks,
> > > Makoto
> > >
> > > 2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
> > > >
> > > > Hi Hivemall users and devs,
> > > >
> > > > I was not sure if this post would be more suitable to the dev or user
> > > > list, so I am copying both. Hivemall sounds like a pretty interesting
> > > > project with a lot of potential.
> > > >
> > > > At LinkedIn, we have been working on a relevant project for defining
> > > > portable SQL user-defined functions: Transport UDFs
> > > > https://github.com/linkedin/transport. This framework enable users to
> > > > define UDFs once and have them automatically translated to UDFs that
> > > > can be executed natively in multiple engines, such as Spark, Hive and
> > > > Presto. The framework can also evaluate UDFs on multiple data formats
> > > > such as Avro, or just plain Java objects. Do users of Hivemall think
> > > > it is beneficial to implement some of the UDFs in the Transport API so
> > > > they can run analytics on multiple engines? Looking forward to your
> > > > thoughts.
> > > >
> > > > Thanks,
> > > > Walaa.
Re: Hivemall & Transport UDFs
Posted by Walaa Eldin Moustafa <wa...@gmail.com>.
Also, wanted to mention that meanwhile, we welcome all kinds of
contributions to the framework (support for additional engines ,
support for new kinds of UDFs such as UDAFs or UDTFs), or
contributions to add new UDFs (such as ML UDFs, geospatial UDFs, JSON,
etc).
Thanks,
Walaa.
On Mon, Jan 13, 2020 at 8:55 PM Walaa Eldin Moustafa
<wa...@gmail.com> wrote:
>
> Hi Makoto,
>
> Thanks for your reply!
>
> It does not support UDAFs or UDTFs at the moment, but this is one of
> the areas we currently want to expand into. For integration with
> Hivemall, we can start with some scalar UDFs as a proof of concept.
>
> Thanks,
> Walaa.
>
> On Mon, Jan 13, 2020 at 8:43 PM Makoto Yui <yu...@gmail.com> wrote:
> >
> > Hi,
> >
> > It is definitely interesting since my company uses both Hive and Presto.
> >
> > Then, does your projects supports UDAFs and UDTFs?
> > Cannot find examples at a glance.
> >
> > Thanks,
> > Makoto
> >
> > 2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
> > >
> > > Hi Hivemall users and devs,
> > >
> > > I was not sure if this post would be more suitable to the dev or user
> > > list, so I am copying both. Hivemall sounds like a pretty interesting
> > > project with a lot of potential.
> > >
> > > At LinkedIn, we have been working on a relevant project for defining
> > > portable SQL user-defined functions: Transport UDFs
> > > https://github.com/linkedin/transport. This framework enable users to
> > > define UDFs once and have them automatically translated to UDFs that
> > > can be executed natively in multiple engines, such as Spark, Hive and
> > > Presto. The framework can also evaluate UDFs on multiple data formats
> > > such as Avro, or just plain Java objects. Do users of Hivemall think
> > > it is beneficial to implement some of the UDFs in the Transport API so
> > > they can run analytics on multiple engines? Looking forward to your
> > > thoughts.
> > >
> > > Thanks,
> > > Walaa.
Re: Hivemall & Transport UDFs
Posted by Walaa Eldin Moustafa <wa...@gmail.com>.
Also, wanted to mention that meanwhile, we welcome all kinds of
contributions to the framework (support for additional engines ,
support for new kinds of UDFs such as UDAFs or UDTFs), or
contributions to add new UDFs (such as ML UDFs, geospatial UDFs, JSON,
etc).
Thanks,
Walaa.
On Mon, Jan 13, 2020 at 8:55 PM Walaa Eldin Moustafa
<wa...@gmail.com> wrote:
>
> Hi Makoto,
>
> Thanks for your reply!
>
> It does not support UDAFs or UDTFs at the moment, but this is one of
> the areas we currently want to expand into. For integration with
> Hivemall, we can start with some scalar UDFs as a proof of concept.
>
> Thanks,
> Walaa.
>
> On Mon, Jan 13, 2020 at 8:43 PM Makoto Yui <yu...@gmail.com> wrote:
> >
> > Hi,
> >
> > It is definitely interesting since my company uses both Hive and Presto.
> >
> > Then, does your projects supports UDAFs and UDTFs?
> > Cannot find examples at a glance.
> >
> > Thanks,
> > Makoto
> >
> > 2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
> > >
> > > Hi Hivemall users and devs,
> > >
> > > I was not sure if this post would be more suitable to the dev or user
> > > list, so I am copying both. Hivemall sounds like a pretty interesting
> > > project with a lot of potential.
> > >
> > > At LinkedIn, we have been working on a relevant project for defining
> > > portable SQL user-defined functions: Transport UDFs
> > > https://github.com/linkedin/transport. This framework enable users to
> > > define UDFs once and have them automatically translated to UDFs that
> > > can be executed natively in multiple engines, such as Spark, Hive and
> > > Presto. The framework can also evaluate UDFs on multiple data formats
> > > such as Avro, or just plain Java objects. Do users of Hivemall think
> > > it is beneficial to implement some of the UDFs in the Transport API so
> > > they can run analytics on multiple engines? Looking forward to your
> > > thoughts.
> > >
> > > Thanks,
> > > Walaa.
Re: Hivemall & Transport UDFs
Posted by Walaa Eldin Moustafa <wa...@gmail.com>.
Hi Makoto,
Thanks for your reply!
It does not support UDAFs or UDTFs at the moment, but this is one of
the areas we currently want to expand into. For integration with
Hivemall, we can start with some scalar UDFs as a proof of concept.
Thanks,
Walaa.
On Mon, Jan 13, 2020 at 8:43 PM Makoto Yui <yu...@gmail.com> wrote:
>
> Hi,
>
> It is definitely interesting since my company uses both Hive and Presto.
>
> Then, does your projects supports UDAFs and UDTFs?
> Cannot find examples at a glance.
>
> Thanks,
> Makoto
>
> 2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
> >
> > Hi Hivemall users and devs,
> >
> > I was not sure if this post would be more suitable to the dev or user
> > list, so I am copying both. Hivemall sounds like a pretty interesting
> > project with a lot of potential.
> >
> > At LinkedIn, we have been working on a relevant project for defining
> > portable SQL user-defined functions: Transport UDFs
> > https://github.com/linkedin/transport. This framework enable users to
> > define UDFs once and have them automatically translated to UDFs that
> > can be executed natively in multiple engines, such as Spark, Hive and
> > Presto. The framework can also evaluate UDFs on multiple data formats
> > such as Avro, or just plain Java objects. Do users of Hivemall think
> > it is beneficial to implement some of the UDFs in the Transport API so
> > they can run analytics on multiple engines? Looking forward to your
> > thoughts.
> >
> > Thanks,
> > Walaa.
Re: Hivemall & Transport UDFs
Posted by Walaa Eldin Moustafa <wa...@gmail.com>.
Hi Makoto,
Thanks for your reply!
It does not support UDAFs or UDTFs at the moment, but this is one of
the areas we currently want to expand into. For integration with
Hivemall, we can start with some scalar UDFs as a proof of concept.
Thanks,
Walaa.
On Mon, Jan 13, 2020 at 8:43 PM Makoto Yui <yu...@gmail.com> wrote:
>
> Hi,
>
> It is definitely interesting since my company uses both Hive and Presto.
>
> Then, does your projects supports UDAFs and UDTFs?
> Cannot find examples at a glance.
>
> Thanks,
> Makoto
>
> 2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
> >
> > Hi Hivemall users and devs,
> >
> > I was not sure if this post would be more suitable to the dev or user
> > list, so I am copying both. Hivemall sounds like a pretty interesting
> > project with a lot of potential.
> >
> > At LinkedIn, we have been working on a relevant project for defining
> > portable SQL user-defined functions: Transport UDFs
> > https://github.com/linkedin/transport. This framework enable users to
> > define UDFs once and have them automatically translated to UDFs that
> > can be executed natively in multiple engines, such as Spark, Hive and
> > Presto. The framework can also evaluate UDFs on multiple data formats
> > such as Avro, or just plain Java objects. Do users of Hivemall think
> > it is beneficial to implement some of the UDFs in the Transport API so
> > they can run analytics on multiple engines? Looking forward to your
> > thoughts.
> >
> > Thanks,
> > Walaa.
Re: Hivemall & Transport UDFs
Posted by Makoto Yui <yu...@gmail.com>.
Hi,
It is definitely interesting since my company uses both Hive and Presto.
Then, does your projects supports UDAFs and UDTFs?
Cannot find examples at a glance.
Thanks,
Makoto
2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
>
> Hi Hivemall users and devs,
>
> I was not sure if this post would be more suitable to the dev or user
> list, so I am copying both. Hivemall sounds like a pretty interesting
> project with a lot of potential.
>
> At LinkedIn, we have been working on a relevant project for defining
> portable SQL user-defined functions: Transport UDFs
> https://github.com/linkedin/transport. This framework enable users to
> define UDFs once and have them automatically translated to UDFs that
> can be executed natively in multiple engines, such as Spark, Hive and
> Presto. The framework can also evaluate UDFs on multiple data formats
> such as Avro, or just plain Java objects. Do users of Hivemall think
> it is beneficial to implement some of the UDFs in the Transport API so
> they can run analytics on multiple engines? Looking forward to your
> thoughts.
>
> Thanks,
> Walaa.
Re: Hivemall & Transport UDFs
Posted by Makoto Yui <yu...@gmail.com>.
Hi,
It is definitely interesting since my company uses both Hive and Presto.
Then, does your projects supports UDAFs and UDTFs?
Cannot find examples at a glance.
Thanks,
Makoto
2020年1月14日(火) 12:17 Walaa Eldin Moustafa <wa...@gmail.com>:
>
> Hi Hivemall users and devs,
>
> I was not sure if this post would be more suitable to the dev or user
> list, so I am copying both. Hivemall sounds like a pretty interesting
> project with a lot of potential.
>
> At LinkedIn, we have been working on a relevant project for defining
> portable SQL user-defined functions: Transport UDFs
> https://github.com/linkedin/transport. This framework enable users to
> define UDFs once and have them automatically translated to UDFs that
> can be executed natively in multiple engines, such as Spark, Hive and
> Presto. The framework can also evaluate UDFs on multiple data formats
> such as Avro, or just plain Java objects. Do users of Hivemall think
> it is beneficial to implement some of the UDFs in the Transport API so
> they can run analytics on multiple engines? Looking forward to your
> thoughts.
>
> Thanks,
> Walaa.