使用 pyspark 将 spark 数据帧转换为嵌套 JSON

Posted 2023-04-15

【中文标题】使用 pyspark 将 spark 数据帧转换为嵌套 JSON

【英文标题】：Convert spark dataframe to nested JSON using pyspark

【发布时间】：2021-12-10 08:59:13

【问题描述】：

我正在尝试将 spark 数据帧转换为 JSON。该数据框中大约有 100 万行，示例代码如下，但性能非常糟糕。所需的输出将是一个 member_id 在 JSON 文件中显示一次，对于一个 member_id 下的 tag_name 相同。请让我知道是否有任何可能的方法可以更快地做到这一点。

示例代码：

iresult = sdf.groupBy('member_id','tag_name').agg(ch.collect_list(ch.struct('detail_name','detail_value')).alias('detail')).\

groupBy('member_id').agg(ch.collect_list(ch.struct('tag_name','detail')).alias('tag'))\

.agg(ch.to_json(ch.collect_list(ch.struct('member_id','tag'))).alias('result'))

result.show()

detail.csv：

member_id, tag_name, detail_name, detail_value
-------------------------------------------------------
abc123, m1, Service_A, 20
abc123, m1, Service_B, 20
abc123, m2, Service_C, 10
xyz456, m3, Service A, 5
xyz456, m3, Service A, 10

所需的输出 JSON：

   "member_id": "abc123",
    "tag"：[ "tag_name": "m1",
            "detail":[ "detail_name": "Service_A",
                        "detail_value": "20",
                       "detail_name": "Service_B",
                        "detail_value": "20"],
            "tag_name": "m2",
            "detail":[ "detail_name": "Service_C",
                        "detail_value": "10"]],
   "member_id": "xyz456",
    "tag"：["tag_name": "m3",
            "detail":[ "detail_name": "Service_A",
                        "detail_value": "5",
                       "detail_name": "Service_A",
                        "detail_value": "10"]]

复制.csv：

member_id, tag_name, detail_name, detail_value
-------------------------------------------------------
abc123, m1, problem_no, 'abc123xyz'
abc123, m1, problem_no, 'abc456zzz'
xyz456, m1, problem_no, 'abc123xyz'
xyz456, m1, problem_no, 'abc456zzz'

重复输出 JSON：

   "member_id": "abc123",
    "tag"：[ "tag_name": "m1",
            "detail":[ "detail_name": "problem_no",
                        "detail_value": "abc123xyz",
                       "detail_name": "problem_no",
                        "detail_value": "abc456zzz",
                       "detail_name": "problem_no",
                        "detail_value": "abc123xyz",
                       "detail_name": "problem_no",
                        "detail_value": "abc456zzz"]],
   "member_id": "xyz456",
    "tag"：[ "tag_name": "m1",
            "detail":[ "detail_name": "problem_no",
                        "detail_value": "abc123xyz",
                       "detail_name": "problem_no",
                        "detail_value": "abc456zzz",
                       "detail_name": "problem_no",
                        "detail_value": "abc123xyz",
                       "detail_name": "problem_no",
                        "detail_value": "abc456zzz"]]

【参考方案1】：

介意通过sql语句实现吗？

逐层构造struct，最后使用to_json函数生成json字符串。

df.createOrReplaceTempView('tmp')
sql = """
    select to_json(collect_list(struct(member_id,tag))) as member
    from
        (select member_id,collect_list(struct(tag_name,detail)) as tag
        from
            (select member_id,tag_name,collect_list(struct(detail_name,detail_value)) as detail
            from tmp
            group by member_id,tag_name)
        group by member_id)
"""
df = spark.sql(sql)
df.show(truncate=False)

【讨论】：

总体上看起来不错，但存在重复问题，我无法找出根本原因。我更新了示例数据框和上面的重复结果。

我的测试结果是正确的，没有重复数据。