0基础学习PyFlink——用户自定义函数之UDAF

在前面几篇文章中，我们学习了非聚合类的用户自定义函数。这节我们将介绍最简单的聚合函数UDAF。

UDAF

我们对比下UDAF和UDF的定义

def udaf(f: Union[Callable, AggregateFunction, Type] = None,input_types: Union[List[DataType], DataType, str, List[str]] = None,result_type: Union[DataType, str] = None, accumulator_type: Union[DataType, str] = None,deterministic: bool = None, name: str = None,func_type: str = "general") -> Union[UserDefinedAggregateFunctionWrapper, Callable]:

def udf(f: Union[Callable, ScalarFunction, Type] = None,input_types: Union[List[DataType], DataType, str, List[str]] = None,result_type: Union[DataType, str] = None,deterministic: bool = None, name: str = None, func_type: str = "general",udf_type: str = None) -> Union[UserDefinedScalarFunctionWrapper, Callable]:

可以发现：

• udaf比udf多了一个参数accumulator_type
• udaf比udf少了一个参数udf_type

accumulator中文是“累加器”。我们可以将其看成聚合过后（比如GroupBy）的成批数据，每批都要走一次函数。
举一个例子：我们对图中左侧的成绩单，使用人名（name）进行聚类，然后计算出最高分数。即算出每个人考出的最高分数是多少。

如图所示，聚合后的数据每个都会经过accumulator计算。计算出来的值的类型就是accumulator_type。这个类型的数据是中间态，它并不是最终UDAF返回的数据类型——result_type。具体这块的知识我们会在后面讲解。
为了方便讲解，我们就以上面例子来讲解其使用。先贴出准备的代码：

from pyflink.common import Configuration
from pyflink.table import (EnvironmentSettings, TableEnvironment, Schema)
from pyflink.table.types import DataTypes
from pyflink.table.table_descriptor import TableDescriptor
from pyflink.table.expressions import lit, col
from pyflink.common import Row
from pyflink.table.udf import udf,udtf,udaf,udtaf
import pandas as pd
from pyflink.table.udf import UserDefinedFunctiondef word_count():config = Configuration()# write all the data to one fileconfig.set_string('parallelism.default', '1')env_settings = EnvironmentSettings \.new_instance() \.in_batch_mode() \.with_configuration(config) \.build()t_env = TableEnvironment.create(env_settings)row_type_tab_source = DataTypes.ROW([DataTypes.FIELD('name', DataTypes.STRING()), DataTypes.FIELD('score', DataTypes.FLOAT()), DataTypes.FIELD('class', DataTypes.STRING())])students_score = [("张三", 80.0, "English"),("李四", 75.0, "English"),("王五", 90.0, "English"),("赵六", 85.0, "English"),("张三", 60.0, "Math"),("李四", 95.0, "Math"),("王五", 90.0, "Math"),("赵六", 70.0, "Math"),("孙七", 60.0, "Math"),]tab_source = t_env.from_elements(students_score, row_type_tab_source )

我们在tab_source表中录入了学生的成绩信息，其中包括姓名（name）、成绩（score）和科目（class）。

入参并非表中一行（Row）的集合

计算每个人考了几门课

1. 按姓名（name）聚类
2. UDTF统计聚类后集合的个数并返回
3. 别名UDTF返回的列名
4. select出数据

@udaf(result_type=DataTypes.ROW([DataTypes.FIELD("count", DataTypes.BIGINT())]), func_type="pandas")def exam_count(pandas_df: pd.DataFrame):return Row(pandas_df.count())tab_student_exam_count = tab_source.group_by(col('name')) \.aggregate(exam_count(col('name')).alias("count")) \.select(col('name'), col('count')) tab_student_exam_count.execute().print()

+--------------------------------+----------------------+
|                           name |                count |
+--------------------------------+----------------------+
|                           孙七 |                    1 |
|                           张三 |                    2 |
|                           李四 |                    2 |
|                           王五 |                    2 |
|                           赵六 |                    2 |
+--------------------------------+----------------------+
5 rows in set

计算每门课有几个人考试

1. 按姓名（class）聚类
2. UDTF统计聚类后集合的个数并返回
3. 别名UDTF返回的列名
4. select出数据

    @udaf(result_type=DataTypes.ROW([DataTypes.FIELD("count", DataTypes.BIGINT())]), func_type="pandas")def exam_count(pandas_df: pd.DataFrame):return Row(pandas_df.count())tab_class_exam_count = tab_source.group_by(col('class')) \.aggregate(exam_count(col('class')).alias("count")) \.select(col('class'), col('count')) tab_class_exam_count.execute().print()

+--------------------------------+----------------------+
|                          class |                count |
+--------------------------------+----------------------+
|                        English |                    4 |
|                           Math |                    5 |
+--------------------------------+----------------------+
2 rows in set

计算每个人的平均分

1. 按姓名（name）聚类
2. UDTF统计聚类后集合的均值并返回
3. 别名UDTF返回的列名
4. select出数据

    @udaf(result_type=DataTypes.ROW([DataTypes.FIELD("avg", DataTypes.FLOAT())]), func_type="pandas")def avg_score(pandas_df: pd.DataFrame):return Row(pandas_df.mean())tab_student_avg_score = tab_source.group_by(col('name')) \.aggregate(avg_score(col('score')).alias("avg")) \.select(col('name'), col('avg')) tab_student_avg_score.execute().print()

+--------------------------------+--------------------------------+
|                           name |                            avg |
+--------------------------------+--------------------------------+
|                           孙七 |                           60.0 |
|                           张三 |                           70.0 |
|                           李四 |                           85.0 |
|                           王五 |                           90.0 |
|                           赵六 |                           77.5 |
+--------------------------------+--------------------------------+
5 rows in set

计算每课的平均分

1. 按姓名（class）聚类
2. UDTF统计聚类后集合的均值并返回
3. 别名UDTF返回的列名
4. select出数据

    @udaf(result_type=DataTypes.ROW([DataTypes.FIELD("avg", DataTypes.FLOAT())]), func_type="pandas")def avg_score(pandas_df: pd.DataFrame):return Row(pandas_df.mean())tab_class_avg_score = tab_source.group_by(col('class')) \.aggregate(avg_score(col('score')).alias("avg")) \.select(col('class'), col('avg')) tab_class_avg_score.execute().print()

+--------------------------------+--------------------------------+
|                          class |                            avg |
+--------------------------------+--------------------------------+
|                        English |                           82.5 |
|                           Math |                           75.0 |
+--------------------------------+--------------------------------+
2 rows in set

计算每个人的最高分和最低分

1. 按姓名（name）聚类
2. UDTF统计聚类后集合的最大值和最小值，并返回
3. 别名UDTF返回的列名
4. select出数据

    @udaf(result_type=DataTypes.ROW([DataTypes.FIELD("max", DataTypes.FLOAT()), DataTypes.FIELD("min", DataTypes.FLOAT())]), func_type="pandas")def max_min_score(pandas_df: pd.DataFrame):return Row(pandas_df.max(), pandas_df.min())tab_student_max_min_score = tab_source.group_by(col('name')) \.aggregate(max_min_score(col('score')).alias("max", "min")) \.select(col('name'), col('max'), col('min')) tab_student_max_min_score.execute().print()

+--------------------------------+--------------------------------+--------------------------------+
|                           name |                            max |                            min |
+--------------------------------+--------------------------------+--------------------------------+
|                           孙七 |                           60.0 |                           60.0 |
|                           张三 |                           80.0 |                           60.0 |
|                           李四 |                           95.0 |                           75.0 |
|                           王五 |                           90.0 |                           90.0 |
|                           赵六 |                           85.0 |                           70.0 |
+--------------------------------+--------------------------------+--------------------------------+
5 rows in set

入参是表中一行（Row）的集合

计算每个人的最高分、最低分以及所属的课程

1. 按姓名（name）聚类
2. UDTF统计聚类后集合中分数最大值、最小值；分数最大值所在行的课程名，和分数最小值所在行的课程名，并返回
3. 别名UDTF返回的列名
4. select出数据

    @udaf(result_type=DataTypes.ROW([DataTypes.FIELD("max", DataTypes.FLOAT()), DataTypes.FIELD("max tag", DataTypes.STRING()), DataTypes.FIELD("min", DataTypes.FLOAT()), DataTypes.FIELD("min tag", DataTypes.STRING())]), func_type="pandas")def max_min_score_with_class(pandas_df: pd.DataFrame):return Row(pandas_df["score"].max(), pandas_df.loc[pandas_df["score"].idxmax(), "class"], pandas_df["score"].min(), pandas_df.loc[pandas_df["score"].idxmin(), "class"])tab_student_max_min_score = tab_source.group_by(col('name')) \.aggregate(max_min_score_with_class.alias("max", "class(max)", "min", "class(min)")) \.select(col('name'), col('max'), col('class(max)'), col('min'), col('class(min)')) tab_student_max_min_score.execute().print()

+--------------------------------+--------------------------------+--------------------------------+--------------------------------+--------------------------------+
|                           name |                            max |                     class(max) |                            min |                     class(min) |
+--------------------------------+--------------------------------+--------------------------------+--------------------------------+--------------------------------+
|                           孙七 |                           60.0 |                           Math |                           60.0 |                           Math |
|                           张三 |                           80.0 |                        English |                           60.0 |                           Math |
|                           李四 |                           95.0 |                           Math |                           75.0 |                        English |
|                           王五 |                           90.0 |                        English |                           90.0 |                        English |
|                           赵六 |                           85.0 |                        English |                           70.0 |                           Math |
+--------------------------------+--------------------------------+--------------------------------+--------------------------------+--------------------------------+
5 rows in set

计算每课的最高分数、最低分数以及所属人

1. 按姓名（class）聚类
2. UDTF统计聚类后集合中分数最大值、最小值；分数最大值所在行的人名，和分数最小值所在行的人名，并返回
3. 别名UDTF返回的列名
4. select出数据

    @udaf(result_type=DataTypes.ROW([DataTypes.FIELD("max", DataTypes.FLOAT()), DataTypes.FIELD("max tag", DataTypes.STRING()), DataTypes.FIELD("min", DataTypes.FLOAT()), DataTypes.FIELD("min tag", DataTypes.STRING())]), func_type="pandas")def max_min_score_with_name(pandas_df: pd.DataFrame):return Row(pandas_df["score"].max(), pandas_df.loc[pandas_df["score"].idxmax(), "name"], pandas_df["score"].min(), pandas_df.loc[pandas_df["score"].idxmin(), "name"])tab_class_max_min_score = tab_source.group_by(col('class')) \.aggregate(max_min_score_with_name.alias("max", "name(max)", "min", "name(min)")) \.select(col('class'), col('max'), col('name(max)'), col('min'), col('name(min)')) tab_class_max_min_score.execute().print()

+--------------------------------+--------------------------------+--------------------------------+--------------------------------+--------------------------------+
|                          class |                            max |                      name(max) |                            min |                      name(min) |
+--------------------------------+--------------------------------+--------------------------------+--------------------------------+--------------------------------+
|                        English |                           90.0 |                           王五 |                           75.0 |                           李四 |
|                           Math |                           95.0 |                           李四 |                           60.0 |                           张三 |
+--------------------------------+--------------------------------+--------------------------------+--------------------------------+--------------------------------+
2 rows in set

完整代码

入参并非表中一行（Row）的集合

from pyflink.common import Configuration
from pyflink.table import (EnvironmentSettings, TableEnvironment, Schema)
from pyflink.table.types import DataTypes
from pyflink.table.table_descriptor import TableDescriptor
from pyflink.table.expressions import lit, col
from pyflink.common import Row
from pyflink.table.udf import udf,udtf,udaf,udtaf
import pandas as pd
from pyflink.table.udf import UserDefinedFunctiondef word_count():config = Configuration()# write all the data to one fileconfig.set_string('parallelism.default', '1')env_settings = EnvironmentSettings \.new_instance() \.in_batch_mode() \.with_configuration(config) \.build()t_env = TableEnvironment.create(env_settings)row_type_tab_source = DataTypes.ROW([DataTypes.FIELD('name', DataTypes.STRING()), DataTypes.FIELD('score', DataTypes.FLOAT()), DataTypes.FIELD('class', DataTypes.STRING())])students_score = [("张三", 80.0, "English"),("李四", 75.0, "English"),("王五", 90.0, "English"),("赵六", 85.0, "English"),("张三", 60.0, "Math"),("李四", 95.0, "Math"),("王五", 90.0, "Math"),("赵六", 70.0, "Math"),("孙七", 60.0, "Math"),]tab_source = t_env.from_elements(students_score, row_type_tab_source )@udaf(result_type=DataTypes.ROW([DataTypes.FIELD("count", DataTypes.BIGINT())]), func_type="pandas")def exam_count(pandas_df: pd.DataFrame):return Row(pandas_df.count())tab_student_exam_count = tab_source.group_by(col('name')) \.aggregate(exam_count(col('name')).alias("count")) \.select(col('name'), col('count')) tab_student_exam_count.execute().print()tab_class_exam_count = tab_source.group_by(col('class')) \.aggregate(exam_count(col('class')).alias("count")) \.select(col('class'), col('count')) tab_class_exam_count.execute().print()@udaf(result_type=DataTypes.ROW([DataTypes.FIELD("avg", DataTypes.FLOAT())]), func_type="pandas")def avg_score(pandas_df: pd.DataFrame):return Row(pandas_df.mean())tab_student_avg_score = tab_source.group_by(col('name')) \.aggregate(avg_score(col('score')).alias("avg")) \.select(col('name'), col('avg')) tab_student_avg_score.execute().print()tab_class_avg_score = tab_source.group_by(col('class')) \.aggregate(avg_score(col('score')).alias("avg")) \.select(col('class'), col('avg')) tab_class_avg_score.execute().print()@udaf(result_type=DataTypes.ROW([DataTypes.FIELD("max", DataTypes.FLOAT()), DataTypes.FIELD("min", DataTypes.FLOAT())]), func_type="pandas")def max_min_score(pandas_df: pd.DataFrame):return Row(pandas_df.max(), pandas_df.min())tab_student_max_min_score = tab_source.group_by(col('name')) \.aggregate(max_min_score(col('score')).alias("max", "min")) \.select(col('name'), col('max'), col('min')) tab_student_max_min_score.execute().print()if __name__ == '__main__':word_count()

入参是表中一行（Row）的集合

from pyflink.common import Configuration
from pyflink.table import (EnvironmentSettings, TableEnvironment, Schema)
from pyflink.table.types import DataTypes
from pyflink.table.table_descriptor import TableDescriptor
from pyflink.table.expressions import lit, col
from pyflink.common import Row
from pyflink.table.udf import udf,udtf,udaf,udtaf
import pandas as pd
from pyflink.table.udf import UserDefinedFunctiondef word_count():config = Configuration()# write all the data to one fileconfig.set_string('parallelism.default', '1')env_settings = EnvironmentSettings \.new_instance() \.in_batch_mode() \.with_configuration(config) \.build()t_env = TableEnvironment.create(env_settings)row_type_tab_source = DataTypes.ROW([DataTypes.FIELD('name', DataTypes.STRING()), DataTypes.FIELD('score', DataTypes.FLOAT()), DataTypes.FIELD('class', DataTypes.STRING())])students_score = [("张三", 80.0, "English"),("李四", 75.0, "English"),("王五", 90.0, "English"),("赵六", 85.0, "English"),("张三", 60.0, "Math"),("李四", 95.0, "Math"),("王五", 90.0, "Math"),("赵六", 70.0, "Math"),("孙七", 60.0, "Math"),]tab_source = t_env.from_elements(students_score, row_type_tab_source )@udaf(result_type=DataTypes.ROW([DataTypes.FIELD("max", DataTypes.FLOAT()), DataTypes.FIELD("max tag", DataTypes.STRING()), DataTypes.FIELD("min", DataTypes.FLOAT()), DataTypes.FIELD("min tag", DataTypes.STRING())]), func_type="pandas")def max_min_score_with_class(pandas_df: pd.DataFrame):return Row(pandas_df["score"].max(), pandas_df.loc[pandas_df["score"].idxmax(), "class"], pandas_df["score"].min(), pandas_df.loc[pandas_df["score"].idxmin(), "class"])tab_student_max_min_score = tab_source.group_by(col('name')) \.aggregate(max_min_score_with_class.alias("max", "class(max)", "min", "class(min)")) \.select(col('name'), col('max'), col('class(max)'), col('min'), col('class(min)')) tab_student_max_min_score.execute().print()@udaf(result_type=DataTypes.ROW([DataTypes.FIELD("max", DataTypes.FLOAT()), DataTypes.FIELD("max tag", DataTypes.STRING()), DataTypes.FIELD("min", DataTypes.FLOAT()), DataTypes.FIELD("min tag", DataTypes.STRING())]), func_type="pandas")def max_min_score_with_name(pandas_df: pd.DataFrame):return Row(pandas_df["score"].max(), pandas_df.loc[pandas_df["score"].idxmax(), "name"], pandas_df["score"].min(), pandas_df.loc[pandas_df["score"].idxmin(), "name"])tab_class_max_min_score = tab_source.group_by(col('class')) \.aggregate(max_min_score_with_name.alias("max", "name(max)", "min", "name(min)")) \.select(col('class'), col('max'), col('name(max)'), col('min'), col('name(min)')) tab_class_max_min_score.execute().print()if __name__ == '__main__':word_count()