PyFlink - Scala UDF - How to convert Scala Map in Table API?

[Xingbo Huang]

Hi Pierre,

The serialization/deserialization of sparse Row in flink is specially optimized. The principle is that each Row will have a leading mask when serializing to identify whether the field at the specified position is NULL, and one field corresponds to one bit. For example, if you have 10k fields, then there will be a mask of 10k bit / 8 = 1250 byte. In this way, the serialization/deserialization overhead can be omitted for those field values that are NULL.

For specific code optimization logic, you can refer to java logic[1], or python logic[2] and cython logic[3].

[1] https://github.com/apache/flink/blob/master/flink-core/src/main/java/org/apache/flink/api/java/typeutils/runtime/RowSerializer.java#L185
[2] https://github.com/apache/flink/blob/master/flink-python/pyflink/fn_execution/beam/beam_coder_impl_slow.py#L100
[3] https://github.com/apache/flink/blob/master/flink-python/pyflink/fn_execution/coder_impl_fast.pyx#L697

Best,
Xingbo

Pierre Oberholzer <[hidden email]> 于2020年12月3日周四 下午3:08写道：

Hi Xingbo, Community,

Thanks a lot for your support.

May I finally ask to conclude this thread, including wider audience:

- Are serious performance issues to be expected with 100k fields per ROW (i.e. due solely to metadata overhead and independently of queries logic) ?

- In sparse population (say 99% sparsity) already optimized in the ROW object or are sparse types on your roadmap ?

Any experience with sparse Table from other users (including benchmarks vs. other frameworks) are also highly welcome.

Thanks !

Best

Le jeu. 3 déc. 2020 à 02:53, Xingbo Huang <[hidden email]> a écrit :

Hi Pierre,

This example is written based on the syntax of release-1.12 that is about to be released, and the test passed. In release-1.12, input_type can be omitted and expression can be used directly. If you are using release-1.11, you only need to modify the grammar of udf used slightly according to the udf documentation[1].

The flink table connector supports avro format, please refer to the document[2].

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/python/table-api-users-guide/udfs/python_udfs.html#scalar-functions
[2] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/connectors/formats/avro.html#avro-format

Best,
Xingbo

Pierre Oberholzer <[hidden email]> 于2020年12月3日周四 上午2:57写道：

Hi Xingbo,

Nice ! This looks a bit hacky, but shows that it can be done ;)

I just got an exception preventing me running your code, apparently from udf.py:

TypeError: Invalid input_type: input_type should be DataType but contains None

Can you pls check again ?

If the schema is defined is a .avsc file, do we have to parse it and rebuild those syntax (ddl and udf) and or is there an existing component that could be used ?

Thanks a lot !

Best,

Le mer. 2 déc. 2020 à 04:50, Xingbo Huang <[hidden email]> a écrit :

Hi Pierre,

I wrote a PyFlink implementation, you can see if it meets your needs:


from pyflink.datastream import StreamExecutionEnvironment
from pyflink.table import StreamTableEnvironment, EnvironmentSettings, DataTypes
from pyflink.table.udf import udf


def test():
    env = StreamExecutionEnvironment.get_execution_environment()
    env.set_parallelism(1)
    t_env = StreamTableEnvironment.create(env,
                                          environment_settings=EnvironmentSettings.new_instance()
                                          .in_streaming_mode().use_blink_planner().build())
    t_env.get_config().get_configuration().set_string("taskmanager.memory.task.off-heap.size",
                                                          '80m')

    # 10k nested columns
    num_field = 10_000
    fields = ['f%s INT' % i for i in range(num_field)]
    field_str = ','.join(fields)
    t_env.execute_sql(f"""
            CREATE TABLE source_table (
            f0 BIGINT,
            f1 DECIMAL(32,2),
            f2 ROW<${field_str}>,
            f3 TIMESTAMP(3)
        ) WITH (
          'connector' = 'datagen',
          'number-of-rows' = '2'
        )
    """)

    t_env.execute_sql(f"""
        CREATE TABLE print_table (
         f0 BIGINT,
         f1 DECIMAL(32,2),
         f2 ROW<${field_str}>,
         f3 TIMESTAMP(3)
        ) WITH (
         'connector' = 'print'
        )
    """)
    result_type = DataTypes.ROW(
        [DataTypes.FIELD("f%s" % i, DataTypes.INT()) for i in range(num_field)])

    func = udf(lambda x: x, result_type=result_type)

    source = t_env.from_path("source_table")
    result = source.select(source.f0, source.f1, func(source.f2), source.f3)
    result.execute_insert("print_table")


if __name__ == '__main__':
    test()


 Best,
 Xingbo

Pierre Oberholzer <[hidden email]> 于2020年12月1日周二 下午6:10写道：

Hi Xingbo,

That would mean giving up on using Flink (table) features on the content of the parsed JSON objects, so definitely a big loss. Let me know if I missed something.

Thanks !

Le mar. 1 déc. 2020 à 07:26, Xingbo Huang <[hidden email]> a écrit :

Hi Pierre,

Have you ever thought of declaring your entire json as a string field in `Table` and putting the parsing work in UDF?

Best,
Xingbo

Pierre Oberholzer <[hidden email]> 于2020年12月1日周二 上午4:13写道：

Hi Xingbo,

Many thanks for your follow up. Yes you got it right.

So using Table API and a ROW object for the nested output of my UDF, and since types are mandatory, I guess this boils down to:

- How to nicely specify the types for the 100k fields : shall I use TypeInformation [1] or better retrieve it from Schema Registry [2] ?

- Do I have to put NULL values for all the fields that don't have a value in my JSON ?

- Will the resulting Table be "sparse" and suffer performance limitations ?

Let me know if Table API and ROW are the right candidates here, or if other better alternatives exist.

As said I'd be glad to apply some downstream transformations using key,value access (and possibly some Table <-> Pandas operations). Hope that doesn't make it a too long wish list ;)

Thanks a lot !

Best regards,

[1] https://stackoverflow.com/questions/48696875/how-to-attach-schema-to-a-flink-datastream-on-the-fly

[2] https://docs.cloudera.com/csa/1.2.0/datastream-connectors/topics/csa-schema-registry.html

Le sam. 28 nov. 2020 à 04:04, Xingbo Huang <[hidden email]> a écrit :

Hi Pierre,

Sorry for the late reply.

Your requirement is that your `Table` has a `field` in `Json` format and its key has reached 100k, and then you want to use such a `field` as the input/output of `udf`, right? As to whether there is a limit on the number of nested key, I am not quite clear. Other contributors with experience in this area may have answers. On the part of `Python UDF`, if the type of key or value of your `Map` is `Any`, we do not support it now. You need to specify a specific type. For more information, please refer to the related document[1].

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/python/table-api-users-guide/python_types.html

Best,

Xingbo

2020年11月28日 上午12:49，Pierre Oberholzer <[hidden email]> 写道：

Hello Wei, Dian, Xingbo,

Not really sure when it is appropriate to knock on the door of the community ;)

I just wanted to mention that your feedback on the above topic will be highly appreciated as it will condition the choice of framework on our side for the months to come, and potentially help the community to cover sparse data with Flink.

Thanks a lot !

Have a great week-end

Best,

Le ven. 20 nov. 2020 à 10:11, Pierre Oberholzer <[hidden email]> a écrit :

Hi Wei,

Thanks for the hint. May I please follow up by adding more context and ask for your guidance.

In case the bespoken Map[String,Any] object returned by Scala:

- Has a defined schema (incl. nested) with up to 100k (!) different possible keys
- Has only some portion of the keys populated for each record
- Is convertible to JSON
- Has to undergo downstream processing in Flink and/or Python UDF with key value access
- Has to be ultimately stored in a Kafka/AVRO sink

How would you declare the types explicitly in such a case ?

Thanks for your support !

Pierre

Le jeu. 19 nov. 2020 à 03:54, Wei Zhong <[hidden email]> a écrit :

Hi Pierre,

Currently there is no type hint like ‘Map[String, Any]’. The recommended way is declaring your type more explicitly.

If you insist on doing this, you can try to declaring a RAW data type for java.util.HashMap [1], but you may encounter some troubles [2] related to the kryo serializers.

Best,

Wei

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html#raw

[2] https://stackoverflow.com/questions/28157236/kryo-serialization-with-nested-hashmap-with-custom-class

在 2020年11月19日，04:31，Pierre Oberholzer <[hidden email]> 写道：

Hi Wei,

It works ! Thanks a lot for your support.

I hadn't tried this last combination for option 1, and I had wrong syntax for option 2.

So to summarize..

Methods working:

- Current: DataTypeHint in UDF definition + SQL for UDF registering

- Outdated: override getResultType in UDF definition + t_env.register_java_function for UDF registering

Type conversions working:

- scala.collection.immutable.Map[String,String] => org.apache.flink.types.Row => ROW<STRING,STRING>

- scala.collection.immutable.Map[String,String] => java.util.Map[String,String] => MAP<STRING,STRING>

Any hint for Map[String,Any] ?

Best regards,

Le mer. 18 nov. 2020 à 03:26, Wei Zhong <[hidden email]> a écrit :

Hi Pierre,

Those 2 approaches all work in my local machine, this is my code:

Scala UDF:

package com.dummy

import org.apache.flink.api.common.typeinfo.TypeInformation
import org.apache.flink.table.annotation.DataTypeHint
import org.apache.flink.table.api.Types
import org.apache.flink.table.functions.ScalarFunction
import org.apache.flink.types.Row

/**
  * The scala UDF.
  */
class dummyMap extends ScalarFunction {

  // If the udf would be registered by the SQL statement, you need add this typehint
  @DataTypeHint("ROW<s STRING,t STRING>")
  def eval(): Row = {

    Row.of(java.lang.String.valueOf("foo"), java.lang.String.valueOf("bar"))

  }

  // If the udf would be registered by the method 'register_java_function', you need override this
  // method.
  override def getResultType(signature: Array[Class[_]]): TypeInformation[_] = {
    // The type of the return values should be TypeInformation
    Types.ROW(Array("s", "t"), Array[TypeInformation[_]](Types.STRING(), Types.STRING()))
  }
}

Python code:

from pyflink.datastream import StreamExecutionEnvironment

from pyflink.table import StreamTableEnvironment

s_env = StreamExecutionEnvironment.get_execution_environment()

st_env = StreamTableEnvironment.create(s_env)

# load the scala udf jar file, the path should be modified to yours

# or your can also load the jar file via other approaches

st_env.get_config().get_configuration().set_string("pipeline.jars", "file:///Users/zhongwei/the-dummy-udf.jar")

# register the udf via 

st_env.execute_sql("CREATE FUNCTION dummyMap AS 'com.dummy.dummyMap' LANGUAGE SCALA")

# or register via the method

# st_env.register_java_function("dummyMap", "com.dummy.dummyMap")

# prepare source and sink

t = st_env.from_elements([(1, 'hi', 'hello'), (2, 'hi', 'hello')], ['a', 'b', 'c'])

st_env.execute_sql("""create table mySink (

        output_of_my_scala_udf ROW<s STRING,t STRING>

    ) with (

        'connector' = 'print'

    )""")

# execute query

t.select("dummyMap()").execute_insert("mySink").get_job_client().get_job_execution_result().result()

Best,

Wei

在 2020年11月18日，03:28，Pierre Oberholzer <[hidden email]> 写道：

Hi Wei,

True, I'm using the method you mention, but glad to change. 

I tried your suggestion instead, but got a similar error.

Thanks for your support. That is much more tedious than I thought.

Option 1 - SQL UDF

SQL UDF

create_func_ddl = """
CREATE FUNCTION dummyMap
  AS 'com.dummy.dummyMap' LANGUAGE SCALA

"""
 
t_env.execute_sql(create_func_ddl)

Error

Py4JJavaError: An error occurred while calling o672.execute.
: org.apache.flink.table.api.TableException: Result field does not match requested type. Requested: Row(s: String, t: String); Actual: GenericType<org.apache.flink.types.Row>

Option 2 - Overriding getResultType

Back to the old registering method, but overriding getResultType:

t_env.register_java_function("dummyMap","com.dummy.dummyMap")

Scala UDF

class dummyMap() extends ScalarFunction {

  def eval(): Row = {

      Row.of(java.lang.String.valueOf("foo"), java.lang.String.valueOf("bar"))

  }

  override def getResultType(signature: Array[Class[_]]): TypeInformation[_] = DataTypes.ROW(DataTypes.STRING,DataTypes.STRING)
}

Error (on compilation)

[error] dummyMap.scala:66:90: overloaded method value ROW with alternatives:

[error]   (x$1: org.apache.flink.table.api.DataTypes.AbstractField*)org.apache.flink.table.types.UnresolvedDataType <and>

[error]   ()org.apache.flink.table.types.DataType <and>

[error]   (x$1: org.apache.flink.table.api.DataTypes.Field*)org.apache.flink.table.types.DataType

[error]  cannot be applied to (org.apache.flink.table.types.DataType, org.apache.flink.table.types.DataType)

[error]   override def getResultType(signature: Array[Class[_]]): TypeInformation[_] = DataTypes.ROW(DataTypes.STRING,DataTypes.STRING)

[error]                                                                                          ^

[error] one error found

[error] (Compile / compileIncremental) Compilation failed

[error] Total time: 3 s, completed 17 nov. 2020 à 20:00:01

Le mar. 17 nov. 2020 à 14:01, Wei Zhong <[hidden email]> a écrit :

Hi Pierre,

I guess your UDF is registered by the method 'register_java_function' which uses the old type system. In this situation you need to override the 'getResultType' method instead of adding type hint. 

You can also try to register your UDF via the "CREATE FUNCTION" sql statement, which accepts the type hint.

Best,

Wei

在 2020年11月17日，19:29，Pierre Oberholzer <[hidden email]> 写道：

Hi Wei,

Thanks for your suggestion. Same error.

Scala UDF

@FunctionHint(output = new DataTypeHint("ROW<s STRING,t STRING>"))

class dummyMap() extends ScalarFunction {
  def eval(): Row = {
    Row.of(java.lang.String.valueOf("foo"), java.lang.String.valueOf("bar"))
  }
}

Best regards,

Le mar. 17 nov. 2020 à 10:04, Wei Zhong <[hidden email]> a écrit :

Hi Pierre,

You can try to replace the '@DataTypeHint("ROW<s STRING,t STRING>")' with '@FunctionHint(output = new DataTypeHint("ROW<s STRING,t STRING>”))'

Best,

Wei

在 2020年11月17日，15:45，Pierre Oberholzer <[hidden email]> 写道：

Hi Dian, Community,

(bringing the thread back to wider audience)

As you suggested, I've tried to use DataTypeHint with Row instead of Map but also this simple case leads to a type mismatch between UDF and Table API.

I've also tried other Map objects from Flink (table.data.MapData, flink.types.MapValue, flink.table.api.DataTypes.MAP) in addition to Java (java.util.Map) in combination with DataTypeHint, without success.

N.B. I'm using version 1.11.

Am I doing something wrong or am I facing limitations in the toolkit ?

Thanks in advance for your support !

Best regards,

Scala UDF

class dummyMap() extends ScalarFunction {

 @DataTypeHint("ROW<s STRING,t STRING>")

 def eval(): Row = {

    Row.of(java.lang.String.valueOf("foo"), java.lang.String.valueOf("bar"))

  }

}

Table DDL

my_sink_ddl = f"""
    create table mySink (
        output_of_my_scala_udf ROW<s STRING,t STRING>
    ) with (
        ...
    )
"""

Error

Py4JJavaError: An error occurred while calling o2.execute.
: org.apache.flink.table.api.ValidationException: Field types of query result and registered TableSink `default_catalog`.`default_database`.`mySink` do not match.
Query result schema: [output_of_my_scala_udf: GenericType<org.apache.flink.types.Row>]
TableSink schema:    [output_of_my_scala_udf: Row(s: String, t: String)]

Le ven. 13 nov. 2020 à 11:59, Pierre Oberholzer <[hidden email]> a écrit :

Thanks Dian, but same error when using explicit returned type:

class dummyMap() extends ScalarFunction {

  def eval() : util.Map[java.lang.String,java.lang.String] = {
    
    val states = Map("key1" -> "val1", "key2" -> "val2")
    states.asInstanceOf[util.Map[java.lang.String,java.lang.String]]

  }
}

Le ven. 13 nov. 2020 à 10:34, Dian Fu <[hidden email]> a écrit :

You need to explicitly defined the result type the UDF. You could refer to [1] for more details if you are using Flink 1.11. If you are using other versions of Flink, you need to refer to the corresponding documentation.

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/functions/udfs.html#implementation-guide

在 2020年11月13日，下午4:56，Pierre Oberholzer <[hidden email]> 写道：

ScalarFunction

-- 

Pierre

-- 

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

[Pierre Oberholzer]

Hi Xingbo, Wei, Dian,

Many thanks for this plus for the high quality and prompt support overall. Let’s close this thread here. Looking forward trying your approach. Community, feel free to reach out with additional remarks and experiences on structured streaming on complex/sparse objects.

Best regards,

Le jeu. 3 déc. 2020 à 08:47, Xingbo Huang <[hidden email]> a écrit :

Hi Pierre,

The serialization/deserialization of sparse Row in flink is specially optimized. The principle is that each Row will have a leading mask when serializing to identify whether the field at the specified position is NULL, and one field corresponds to one bit. For example, if you have 10k fields, then there will be a mask of 10k bit / 8 = 1250 byte. In this way, the serialization/deserialization overhead can be omitted for those field values that are NULL.

For specific code optimization logic, you can refer to java logic[1], or python logic[2] and cython logic[3].

[1] https://github.com/apache/flink/blob/master/flink-core/src/main/java/org/apache/flink/api/java/typeutils/runtime/RowSerializer.java#L185
[2] https://github.com/apache/flink/blob/master/flink-python/pyflink/fn_execution/beam/beam_coder_impl_slow.py#L100
[3] https://github.com/apache/flink/blob/master/flink-python/pyflink/fn_execution/coder_impl_fast.pyx#L697

Best,
Xingbo

Pierre Oberholzer <[hidden email]> 于2020年12月3日周四 下午3:08写道：

Hi Xingbo, Community,

Thanks a lot for your support.

May I finally ask to conclude this thread, including wider audience:

- Are serious performance issues to be expected with 100k fields per ROW (i.e. due solely to metadata overhead and independently of queries logic) ?

- In sparse population (say 99% sparsity) already optimized in the ROW object or are sparse types on your roadmap ?

Any experience with sparse Table from other users (including benchmarks vs. other frameworks) are also highly welcome.

Thanks !

Best

Le jeu. 3 déc. 2020 à 02:53, Xingbo Huang <[hidden email]> a écrit :

Hi Pierre,

This example is written based on the syntax of release-1.12 that is about to be released, and the test passed. In release-1.12, input_type can be omitted and expression can be used directly. If you are using release-1.11, you only need to modify the grammar of udf used slightly according to the udf documentation[1].

The flink table connector supports avro format, please refer to the document[2].

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/python/table-api-users-guide/udfs/python_udfs.html#scalar-functions
[2] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/connectors/formats/avro.html#avro-format

Best,
Xingbo

Pierre Oberholzer <[hidden email]> 于2020年12月3日周四 上午2:57写道：

Hi Xingbo,

Nice ! This looks a bit hacky, but shows that it can be done ;)

I just got an exception preventing me running your code, apparently from udf.py:

TypeError: Invalid input_type: input_type should be DataType but contains None

Can you pls check again ?

If the schema is defined is a .avsc file, do we have to parse it and rebuild those syntax (ddl and udf) and or is there an existing component that could be used ?

Thanks a lot !

Best,

Le mer. 2 déc. 2020 à 04:50, Xingbo Huang <[hidden email]> a écrit :

Hi Pierre,

I wrote a PyFlink implementation, you can see if it meets your needs:


from pyflink.datastream import StreamExecutionEnvironment
from pyflink.table import StreamTableEnvironment, EnvironmentSettings, DataTypes
from pyflink.table.udf import udf


def test():
    env = StreamExecutionEnvironment.get_execution_environment()
    env.set_parallelism(1)
    t_env = StreamTableEnvironment.create(env,
                                          environment_settings=EnvironmentSettings.new_instance()
                                          .in_streaming_mode().use_blink_planner().build())
    t_env.get_config().get_configuration().set_string("taskmanager.memory.task.off-heap.size",
                                                          '80m')

    # 10k nested columns
    num_field = 10_000
    fields = ['f%s INT' % i for i in range(num_field)]
    field_str = ','.join(fields)
    t_env.execute_sql(f"""
            CREATE TABLE source_table (
            f0 BIGINT,
            f1 DECIMAL(32,2),
            f2 ROW<${field_str}>,
            f3 TIMESTAMP(3)
        ) WITH (
          'connector' = 'datagen',
          'number-of-rows' = '2'
        )
    """)

    t_env.execute_sql(f"""
        CREATE TABLE print_table (
         f0 BIGINT,
         f1 DECIMAL(32,2),
         f2 ROW<${field_str}>,
         f3 TIMESTAMP(3)
        ) WITH (
         'connector' = 'print'
        )
    """)
    result_type = DataTypes.ROW(
        [DataTypes.FIELD("f%s" % i, DataTypes.INT()) for i in range(num_field)])

    func = udf(lambda x: x, result_type=result_type)

    source = t_env.from_path("source_table")
    result = source.select(source.f0, source.f1, func(source.f2), source.f3)
    result.execute_insert("print_table")


if __name__ == '__main__':
    test()


 Best,
 Xingbo

Pierre Oberholzer <[hidden email]> 于2020年12月1日周二 下午6:10写道：

Hi Xingbo,

That would mean giving up on using Flink (table) features on the content of the parsed JSON objects, so definitely a big loss. Let me know if I missed something.

Thanks !

Le mar. 1 déc. 2020 à 07:26, Xingbo Huang <[hidden email]> a écrit :

Hi Pierre,

Have you ever thought of declaring your entire json as a string field in `Table` and putting the parsing work in UDF?

Best,
Xingbo

Pierre Oberholzer <[hidden email]> 于2020年12月1日周二 上午4:13写道：

Hi Xingbo,

Many thanks for your follow up. Yes you got it right.

So using Table API and a ROW object for the nested output of my UDF, and since types are mandatory, I guess this boils down to:

- How to nicely specify the types for the 100k fields : shall I use TypeInformation [1] or better retrieve it from Schema Registry [2] ?

- Do I have to put NULL values for all the fields that don't have a value in my JSON ?

- Will the resulting Table be "sparse" and suffer performance limitations ?

Let me know if Table API and ROW are the right candidates here, or if other better alternatives exist.

As said I'd be glad to apply some downstream transformations using key,value access (and possibly some Table <-> Pandas operations). Hope that doesn't make it a too long wish list ;)

Thanks a lot !

Best regards,

[1] https://stackoverflow.com/questions/48696875/how-to-attach-schema-to-a-flink-datastream-on-the-fly

[2] https://docs.cloudera.com/csa/1.2.0/datastream-connectors/topics/csa-schema-registry.html

Le sam. 28 nov. 2020 à 04:04, Xingbo Huang <[hidden email]> a écrit :

Hi Pierre,

Sorry for the late reply.

Your requirement is that your `Table` has a `field` in `Json` format and its key has reached 100k, and then you want to use such a `field` as the input/output of `udf`, right? As to whether there is a limit on the number of nested key, I am not quite clear. Other contributors with experience in this area may have answers. On the part of `Python UDF`, if the type of key or value of your `Map` is `Any`, we do not support it now. You need to specify a specific type. For more information, please refer to the related document[1].

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/python/table-api-users-guide/python_types.html

Best,

Xingbo

2020年11月28日 上午12:49，Pierre Oberholzer <[hidden email]> 写道：

Hello Wei, Dian, Xingbo,

Not really sure when it is appropriate to knock on the door of the community ;)

I just wanted to mention that your feedback on the above topic will be highly appreciated as it will condition the choice of framework on our side for the months to come, and potentially help the community to cover sparse data with Flink.

Thanks a lot !

Have a great week-end

Best,

Le ven. 20 nov. 2020 à 10:11, Pierre Oberholzer <[hidden email]> a écrit :

Hi Wei,

Thanks for the hint. May I please follow up by adding more context and ask for your guidance.

In case the bespoken Map[String,Any] object returned by Scala:

- Has a defined schema (incl. nested) with up to 100k (!) different possible keys
- Has only some portion of the keys populated for each record
- Is convertible to JSON
- Has to undergo downstream processing in Flink and/or Python UDF with key value access
- Has to be ultimately stored in a Kafka/AVRO sink

How would you declare the types explicitly in such a case ?

Thanks for your support !

Pierre

Le jeu. 19 nov. 2020 à 03:54, Wei Zhong <[hidden email]> a écrit :

Hi Pierre,

Currently there is no type hint like ‘Map[String, Any]’. The recommended way is declaring your type more explicitly.

If you insist on doing this, you can try to declaring a RAW data type for java.util.HashMap [1], but you may encounter some troubles [2] related to the kryo serializers.

Best,

Wei

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html#raw

[2] https://stackoverflow.com/questions/28157236/kryo-serialization-with-nested-hashmap-with-custom-class

在 2020年11月19日，04:31，Pierre Oberholzer <[hidden email]> 写道：

Hi Wei,

It works ! Thanks a lot for your support.

I hadn't tried this last combination for option 1, and I had wrong syntax for option 2.

So to summarize..

Methods working:

- Current: DataTypeHint in UDF definition + SQL for UDF registering

- Outdated: override getResultType in UDF definition + t_env.register_java_function for UDF registering

Type conversions working:

- scala.collection.immutable.Map[String,String] => org.apache.flink.types.Row => ROW<STRING,STRING>

- scala.collection.immutable.Map[String,String] => java.util.Map[String,String] => MAP<STRING,STRING>

Any hint for Map[String,Any] ?

Best regards,

Le mer. 18 nov. 2020 à 03:26, Wei Zhong <[hidden email]> a écrit :

Hi Pierre,

Those 2 approaches all work in my local machine, this is my code:

Scala UDF:

package com.dummy

import org.apache.flink.api.common.typeinfo.TypeInformation
import org.apache.flink.table.annotation.DataTypeHint
import org.apache.flink.table.api.Types
import org.apache.flink.table.functions.ScalarFunction
import org.apache.flink.types.Row

/**
  * The scala UDF.
  */
class dummyMap extends ScalarFunction {

  // If the udf would be registered by the SQL statement, you need add this typehint
  @DataTypeHint("ROW<s STRING,t STRING>")
  def eval(): Row = {

    Row.of(java.lang.String.valueOf("foo"), java.lang.String.valueOf("bar"))

  }

  // If the udf would be registered by the method 'register_java_function', you need override this
  // method.
  override def getResultType(signature: Array[Class[_]]): TypeInformation[_] = {
    // The type of the return values should be TypeInformation
    Types.ROW(Array("s", "t"), Array[TypeInformation[_]](Types.STRING(), Types.STRING()))
  }
}

Python code:

from pyflink.datastream import StreamExecutionEnvironment

from pyflink.table import StreamTableEnvironment

s_env = StreamExecutionEnvironment.get_execution_environment()

st_env = StreamTableEnvironment.create(s_env)

# load the scala udf jar file, the path should be modified to yours

# or your can also load the jar file via other approaches

st_env.get_config().get_configuration().set_string("pipeline.jars", "file:///Users/zhongwei/the-dummy-udf.jar")

# register the udf via 

st_env.execute_sql("CREATE FUNCTION dummyMap AS 'com.dummy.dummyMap' LANGUAGE SCALA")

# or register via the method

# st_env.register_java_function("dummyMap", "com.dummy.dummyMap")

# prepare source and sink

t = st_env.from_elements([(1, 'hi', 'hello'), (2, 'hi', 'hello')], ['a', 'b', 'c'])

st_env.execute_sql("""create table mySink (

        output_of_my_scala_udf ROW<s STRING,t STRING>

    ) with (

        'connector' = 'print'

    )""")

# execute query

t.select("dummyMap()").execute_insert("mySink").get_job_client().get_job_execution_result().result()

Best,

Wei

在 2020年11月18日，03:28，Pierre Oberholzer <[hidden email]> 写道：

Hi Wei,

True, I'm using the method you mention, but glad to change. 

I tried your suggestion instead, but got a similar error.

Thanks for your support. That is much more tedious than I thought.

Option 1 - SQL UDF

SQL UDF

create_func_ddl = """
CREATE FUNCTION dummyMap
  AS 'com.dummy.dummyMap' LANGUAGE SCALA

"""
 
t_env.execute_sql(create_func_ddl)

Error

Py4JJavaError: An error occurred while calling o672.execute.
: org.apache.flink.table.api.TableException: Result field does not match requested type. Requested: Row(s: String, t: String); Actual: GenericType<org.apache.flink.types.Row>

Option 2 - Overriding getResultType

Back to the old registering method, but overriding getResultType:

t_env.register_java_function("dummyMap","com.dummy.dummyMap")

Scala UDF

class dummyMap() extends ScalarFunction {

  def eval(): Row = {

      Row.of(java.lang.String.valueOf("foo"), java.lang.String.valueOf("bar"))

  }

  override def getResultType(signature: Array[Class[_]]): TypeInformation[_] = DataTypes.ROW(DataTypes.STRING,DataTypes.STRING)
}

Error (on compilation)

[error] dummyMap.scala:66:90: overloaded method value ROW with alternatives:

[error]   (x$1: org.apache.flink.table.api.DataTypes.AbstractField*)org.apache.flink.table.types.UnresolvedDataType <and>

[error]   ()org.apache.flink.table.types.DataType <and>

[error]   (x$1: org.apache.flink.table.api.DataTypes.Field*)org.apache.flink.table.types.DataType

[error]  cannot be applied to (org.apache.flink.table.types.DataType, org.apache.flink.table.types.DataType)

[error]   override def getResultType(signature: Array[Class[_]]): TypeInformation[_] = DataTypes.ROW(DataTypes.STRING,DataTypes.STRING)

[error]                                                                                          ^

[error] one error found

[error] (Compile / compileIncremental) Compilation failed

[error] Total time: 3 s, completed 17 nov. 2020 à 20:00:01

Le mar. 17 nov. 2020 à 14:01, Wei Zhong <[hidden email]> a écrit :

Hi Pierre,

I guess your UDF is registered by the method 'register_java_function' which uses the old type system. In this situation you need to override the 'getResultType' method instead of adding type hint. 

You can also try to register your UDF via the "CREATE FUNCTION" sql statement, which accepts the type hint.

Best,

Wei

在 2020年11月17日，19:29，Pierre Oberholzer <[hidden email]> 写道：

Hi Wei,

Thanks for your suggestion. Same error.

Scala UDF

@FunctionHint(output = new DataTypeHint("ROW<s STRING,t STRING>"))

class dummyMap() extends ScalarFunction {
  def eval(): Row = {
    Row.of(java.lang.String.valueOf("foo"), java.lang.String.valueOf("bar"))
  }
}

Best regards,

Le mar. 17 nov. 2020 à 10:04, Wei Zhong <[hidden email]> a écrit :

Hi Pierre,

You can try to replace the '@DataTypeHint("ROW<s STRING,t STRING>")' with '@FunctionHint(output = new DataTypeHint("ROW<s STRING,t STRING>”))'

Best,

Wei

在 2020年11月17日，15:45，Pierre Oberholzer <[hidden email]> 写道：

Hi Dian, Community,

(bringing the thread back to wider audience)

As you suggested, I've tried to use DataTypeHint with Row instead of Map but also this simple case leads to a type mismatch between UDF and Table API.

I've also tried other Map objects from Flink (table.data.MapData, flink.types.MapValue, flink.table.api.DataTypes.MAP) in addition to Java (java.util.Map) in combination with DataTypeHint, without success.

N.B. I'm using version 1.11.

Am I doing something wrong or am I facing limitations in the toolkit ?

Thanks in advance for your support !

Best regards,

Scala UDF

class dummyMap() extends ScalarFunction {

 @DataTypeHint("ROW<s STRING,t STRING>")

 def eval(): Row = {

    Row.of(java.lang.String.valueOf("foo"), java.lang.String.valueOf("bar"))

  }

}

Table DDL

my_sink_ddl = f"""
    create table mySink (
        output_of_my_scala_udf ROW<s STRING,t STRING>
    ) with (
        ...
    )
"""

Error

Py4JJavaError: An error occurred while calling o2.execute.
: org.apache.flink.table.api.ValidationException: Field types of query result and registered TableSink `default_catalog`.`default_database`.`mySink` do not match.
Query result schema: [output_of_my_scala_udf: GenericType<org.apache.flink.types.Row>]
TableSink schema:    [output_of_my_scala_udf: Row(s: String, t: String)]

Le ven. 13 nov. 2020 à 11:59, Pierre Oberholzer <[hidden email]> a écrit :

Thanks Dian, but same error when using explicit returned type:

class dummyMap() extends ScalarFunction {

  def eval() : util.Map[java.lang.String,java.lang.String] = {
    
    val states = Map("key1" -> "val1", "key2" -> "val2")
    states.asInstanceOf[util.Map[java.lang.String,java.lang.String]]

  }
}

Le ven. 13 nov. 2020 à 10:34, Dian Fu <[hidden email]> a écrit :

You need to explicitly defined the result type the UDF. You could refer to [1] for more details if you are using Flink 1.11. If you are using other versions of Flink, you need to refer to the corresponding documentation.

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/functions/udfs.html#implementation-guide

在 2020年11月13日，下午4:56，Pierre Oberholzer <[hidden email]> 写道：

ScalarFunction

-- 

Pierre

-- 

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--

Pierre

--