雪花：window function '范围'不支持，如何查询？

Question

我有一个包含 txn_date 和 cust_id 的交易表。

对于在 12 月进行交易的每个客户，我想知道该客户在给定交易之前的 90 天内进行了多少次交易。

This seems to be a query that I could run with a window function and a RANGE sliding window, but Snowflake doesn't support the RANGE sliding window frame.

如何在 Snowflake 中运行此查询？

Answer 1

像这样的东西怎么样：

WITH T1 AS (
    SELECT CUSTOMER_ID, TX_DATE
    FROM TRANSACTIONS
    WHERE TX_DATE BETWEEN '2020-12-01' AND '2020-12-31')
SELECT T2.CUSTOMER_ID, T2.TX_DATE
FROM TRANSACTIONS T2
INNER JOIN T1 ON T2.CUSTOMER_ID = T2.CUSTOMER_ID
WHERE T2.TX_DATE BETWEEN (T1.TX_DATE - 90) AND T1.TX_DATE

Answer 2

NickW 一开始的回答也差不多。

WITH data AS (
    SELECT txn_date::timestamp_ntz as txn_date, cust_id, txn_id
    FROM VALUES
        ('2020-12-04',0, 0),
        ('2020-12-03',1, 1),
        ('2020-11-04',1, 2),
        ('2020-10-04',1, 3),
        ('2020-09-04',1, 4), -- just on 90 days
        ('2020-09-02',1, 5), -- too far
        ('2021-01-05',1, 6)  -- in the future
        v(txn_date , cust_id, txn_id)
), dec_txn AS (
    SELECT txn_id,
        cust_id,
        DATEADD('day',-90, txn_date) AS win_start,
        txn_date AS win_end
    FROM data 
    WHERE date_trunc('month', txn_date) = '2020-12-01'
)
SELECT dt.*
    ,t.*
    ,datediff('days', dt.win_end, t.txn_date) as win_time
FROM dec_txn AS dt
LEFT JOIN data AS t 
    ON t.cust_id = dt.cust_id 
    AND t.txn_date between dt.win_start and win_end AND t.txn_id != dt.txn_id
    ;

这使：

TXN_ID   CUST_ID    WIN_START                 WIN_END                   TXN_DATE                  CUST_ID   TXN_ID   WIN_TIME
1        1          2020-09-04 00:00:00.000   2020-12-03 00:00:00.000   2020-11-04 00:00:00.000   1         2        -29
1        1          2020-09-04 00:00:00.000   2020-12-03 00:00:00.000   2020-10-04 00:00:00.000   1         3        -60
1        1          2020-09-04 00:00:00.000   2020-12-03 00:00:00.000   2020-09-04 00:00:00.000   1         4        -90
0        0          2020-09-05 00:00:00.000   2020-12-04 00:00:00.000   NULL                      NULL      NULL     NULL

因此，我们要计算：

WITH data AS (
    SELECT txn_date::timestamp_ntz as txn_date, cust_id, txn_id
    FROM VALUES
        ('2020-12-04',0, 0),   
        ('2020-12-03',1, 1),   
        ('2020-11-04',1, 2),
        ('2020-10-04',1, 3),
        ('2020-09-04',1, 4), -- just on 90 days
        ('2020-09-02',1, 5), -- too far
        ('2021-01-05',1, 6) -- in the future
        v(txn_date , cust_id, txn_id)
), dec_txn AS (
    SELECT txn_id,
        cust_id,
        txn_date,
        DATEADD('day',-90, txn_date) AS win_start,
        txn_date AS win_end
    FROM data 
    WHERE date_trunc('month', txn_date) = '2020-12-01'
)
SELECT dt.cust_id
    ,dt.txn_id
    ,dt.txn_date
    ,count(t.txn_id) as c__prior_90_days_transaction
FROM dec_txn AS dt
LEFT JOIN data AS t 
ON t.cust_id = dt.cust_id 
AND t.txn_date >= dt.win_start and t.txn_date < dt.win_end AND t.txn_id != dt.txn_id
GROUP BY 1,2,3
ORDER BY 1,2
;

给予：

CUST_ID   TXN_ID   TXN_DATE                  C__PRIOR_90_DAYS_TRANSACTION
0         0        2020-12-04 00:00:00.000   0
1         1        2020-12-03 00:00:00.000   3

问题中没有明确定义的是，如果一个客户在 12 月有很多请求该怎么办 如果在同一 12 月的同一天有多个交易怎么办。

上面将为每个客户的每个 12 月交易返回一行，它包括在同一天发生的交易。 但是，如果您的日期/时间戳有时间，那么它只会计算当天早些时候的转换。 但是，如果您想要前几天并且 txn_date 只是一个日期，那么

AND t.txn_date >= dt.win_start and t.txn_date < dt.win_end AND t.txn_id != dt.txn_id

应该使用。

如果 txn_date 是时间戳，则dec_txn应更改为：

dec_txn AS (
    SELECT txn_id,
        cust_id,
        DATEADD('day',-90, txn_date::date) AS win_start,
        txn_date::date AS win_end
    FROM data 
    WHERE date_trunc('month', txn_date) = '2020-12-01'

现在 window 时间戳被截断为天，那么如果您希望午夜交易在当天计算，或者如果您没有午夜时间戳，您将不得不锻炼......