Spark中集合类型ArrayType/MapType与explode方法的使用

版权声明：本文为博主原创文章，未经博主允许不得转载。

背景：需要对DataFrame中部分字段聚合，再通过udf对聚合的字段列表进行处理，返回列表，再把返回的列表字段列转行，如下：

group_id	feature_1	feature_2	feature_3
1	1.11	1.21	1.31
1	1.12	1.22	1.32
2	2.11	2.21	2.31
2	2.12	2.22	2.32

按 “group_id” 聚合（如：[{“feature_1”: 1.11, “feature_2”: 1.21, “feature_3”: 1.31}, {“feature_1”: 1.12, “feature_2”: 1.22, “feature_3”: 1.32}]），再通过udf对聚合的字段列表进行处理，返回列表[{“feature_1”: 11.1, “feature_2”: 11.2, “feature_3”: 11.3}, {“feature_1”: 12.1, “feature_2”: 12.2, “feature_3”: 12.3}, {“feature_1”: 13.1, “feature_2”: 13.2, “feature_3”: 13.3}]，再把返回的列表字段列转行，最终得到结果如下：

group_id	feature_1	feature_2	feature_3
1	11.1	11.2	11.3
1	12.1	12.2	12.3
1	13.1	13.2	13.3
2	22.11	22.21	22.31

代码如下：

from pyspark.sql import functions as F
from pyspark.sql.functions import explode
from pyspark.sql.types import MapType, StringType, ArrayType, DoubleType

data_before = [
    (1, 1.11, 1.21, 1.31),
    (1, 1.12, 1.22, 1.32),
    (2, 2.11, 2.21, 2.31),
    (2, 2.12, 2.22, 2.32)
]

df = spark.createDataFrame(data_before, schema=['group_id', 'feature_1', 'feature_2', 'feature_3'])


def get_feature_dict(feature_1, feature_2, feature_3):
    # 聚合为dict
    feature_dict = dict()
    feature_dict['feature_1'] = feature_1
    feature_dict['feature_2'] = feature_2
    feature_dict['feature_3'] = feature_3
    return feature_dict

# 转换dict为MapType(StringType(), DoubleType())
udf_feature_dict = F.udf(get_feature_dict, returnType=MapType(StringType(), DoubleType()))


def calc_feature(group_id, feature_list):
    '''
    :return: List
    '''
    if group_id == 1:
        return [{"feature_1": 11.1, "feature_2": 11.2, "feature_3": 11.3},
                {"feature_1": 12.1, "feature_2": 12.2, "feature_3": 12.3},
                {"feature_1": 13.1, "feature_2": 13.2, "feature_3": 13.3}]
    else:
        return [{"feature_1": 22.11, "feature_2": 22.21, "feature_3": 22.31}]


# 转换List 为ArrayType(MapType(StringType(), DoubleType()))
udf_calc = F.udf(calc_feature, returnType=ArrayType(MapType(StringType(), DoubleType())))


df_calc = df.withColumn("feature_dict", udf_feature_dict(F.col("feature_1"), F.col("feature_2"), F.col("feature_3"))) \
    .groupBy('group_id') \
    .agg(F.collect_list("feature_dict").alias("feature_dict_list")) \
    .withColumn("calc_feature_list", udf_calc(F.col("group_id"), F.col("feature_dict_list"))) \
    .withColumn('calc_feature_detail', explode("calc_feature_list")) \
    .select('group_id',
            F.col("calc_feature_detail")["feature_1"].alias("feature_1"),
            F.col("calc_feature_detail")["feature_2"].alias("feature_2"),
            F.col("calc_feature_detail")["feature_3"].alias("feature_3")
            )

---

--------------------------文档信息--------------------------
版权声明：本文为博主原创文章，未经博主允许不得转载
署名(BY) ：dkjkls（dkj卡洛斯）
文章出处：http://blog.csdn.net/dkjkls