如何在大查询中应用：count(distinct ...) over (partition by ... order by)？

Question

我目前有这个 source table。

我正在尝试从 GCP BigQuery 上 SQL 中的第一个 table 中获取 this second table。

我的查询如下：

        SELECT
            SE.MARKET_ID,
            SE.LOCAL_POS_ID,
            SE.BC_ID,
            LEFT(SE.SALE_CREATION_DATE,6) AS DATE_ID_MONTH,

            COUNT(DISTINCT
                CASE
                    WHEN FLAG
                    THEN SE.CUST_ID
                END)
            OVER (PARTITION BY SE.MARKET_ID, SE.LOCAL_POS_ID, SE.BC_ID, LEFT(SE.SALE_CREATION_DATE,4) ORDER BY LEFT(SE.SALE_CREATION_DATE,6)) AS NB_ACTIVE_CUSTOMERS

        FROM
            SE
        GROUP BY
            SE.MARKET_ID, SE.LOCAL_POS_ID, SE.BC_ID, LEFT(SE.SALE_CREATION_DATE,6)

但是，我得到这个错误，我没有成功绕过：

Window ORDER BY is not allowed if DISTINCT is specified at [12:107]

我无法使用以下请求创建以前的 table :

SELECT DISTINCT
        SE.MARKET_ID,
        SE.LOCAL_POS_ID,
        SE.BC_ID,
        LEFT(SE.SALE_CREATION_DATE,6) AS DATE_ID_MONTH,
        CASE
            WHEN FLAG
            THEN SE.CUST_ID
            ELSE NULL
        END AS VALID_CUST_ID
FROM
        SE

为了在那之后使用 dense_rank() 因为我还有 50 个其他指标（和 500M 行）要添加到这个 table（基于其他标志的指标）并且我可以'显然没有为它们中的每一个创建一个 WITH，我只需要几个 WITH 或 none（就像我当前的查询应该做的那样）。

有人知道我该如何处理吗？

Answer 1

我认为您的某些示例数据不正确，但我确实使用过它并得到了匹配的结果，至少对于 MPE 数据是这样。您可以通过首先在 CUST_ID 上用额外分区标记“不同计数”的行，然后首先在 FLAG DESC 上排序来完成此操作。然后你会以你希望应用 count(distinct <expr>) over ...

的相同方式总结它

WITH SE AS (
    SELECT  1 LINE_ID, 'TW' MARKET_ID, 'X' LOCAL_POS_ID, 'MPE' BC_ID,
            1 CUST_ID, '20200201' SALE_CREATION_DATE, 1 FLAG UNION ALL
    SELECT  2, 'TW', 'X', 'MPE', 2, '20201005', 1 UNION ALL
    SELECT  3, 'TW', 'X', 'MPE', 3, '20200415', 0 UNION ALL
    SELECT  4, 'TW', 'X', 'MPE', 1, '20200223', 1 UNION ALL
    SELECT  5, 'TW', 'X', 'MPE', 6, '20200217', 1 UNION ALL
    SELECT  6, 'TW', 'X', 'MPE', 9, '20200715', 1 UNION ALL
    SELECT  7, 'TW', 'X', 'MPE', 4, '20200223', 1 UNION ALL
    SELECT  8, 'TW', 'X', 'MPE', 1, '20201008', 1 UNION ALL
    SELECT  9, 'TW', 'X', 'MPE', 2, '20201019', 1 UNION ALL
    SELECT 10, 'TW', 'X', 'MPE', 1, '20200516', 1 UNION ALL
    SELECT 11, 'TW', 'X', 'MPE', 1, '20200129', 1 UNION ALL
    SELECT 12, 'TW', 'X', 'MPE', 1, '20201007', 1 UNION ALL
    SELECT 13, 'TW', 'X', 'MPE', 2, '20201005', 1 UNION ALL
    SELECT 14, 'TW', 'X', 'MPE', 3, '20200505', 1 UNION ALL
    SELECT 15, 'TW', 'X', 'MPE', 8, '20201103', 1 UNION ALL
    SELECT 16, 'TW', 'X', 'MPE', 9, '20200820', 1
),
DATA AS (
    SELECT *,
        LEFT(SALE_CREATION_DATE, 6) AS SALE_MONTH,
        LEFT(SALE_CREATION_DATE, 4) AS SALE_YEAR,
        CASE ROW_NUMBER() OVER (
            PARTITION BY MARKET_ID, LOCAL_POS_ID, BC_ID,
                         LEFT(SALE_CREATION_DATE, 4), CUST_ID
            ORDER BY FLAG DESC, LEFT(SALE_CREATION_DATE, 6)
        ) WHEN 1 THEN FLAG END AS COUNTER /* assumes possible to have no flagged row */
    FROM SE
)
SELECT MARKET_ID, LOCAL_POS_ID, BC_ID, SALE_MONTH,
    SUM(SUM(COUNTER)) OVER (
            PARTITION BY MARKET_ID, LOCAL_POS_ID, BC_ID, SALE_YEAR
            ORDER BY SALE_MONTH
    ) AS NB_ACTIVE_CUSTOMERS
FROM DATA
GROUP BY MARKET_ID, LOCAL_POS_ID, BC_ID, SALE_YEAR, SALE_MONTH
ORDER BY MARKET_ID, LOCAL_POS_ID, BC_ID, SALE_YEAR, SALE_MONTH

Answer 2

考虑以下方法

select * except(ids), 
  array_length(array(
    select distinct id
    from unnest(split(ids)) id
  )) as nb_active_customers, 
  format('%t', array(
    select distinct id
    from unnest(split(ids)) id
  )) as distinct_values
from (
  select market_id, local_pos_id, bc_id, date_id_month,
    string_agg('' || ids) over(partition by market_id order by date_id_month) ids
  from (
    select market_id, local_pos_id, bc_id, left(sale_creation_date,6) AS date_id_month,
      string_agg('' || cust_id) ids
    from se
    where flag = 1
    group by market_id, local_pos_id, bc_id, date_id_month
  )
) t

如果应用于您问题中的示例数据 - 输出为

如何在大查询中应用：count(distinct ...) over (partition by ... order by)？

How to apply: count(distinct ...) over (partition by ... order by) in big query?

sql

google-bigquery

partitioning

rolling-computation