加窗平均，考虑差距

Question

我需要计算前 4 周的平均值...

SELECT
  *,
  AVG(val) OVER (PARTITION BY some_identifier, day_of_week_column
                     ORDER BY date_column
                  ROW BETWEEN 4 PRECEDING AND 1 PRECEDING
                )
                   AS preceding_4_week_average
FROM
  myTable

然而，数据是"sparse"

在这种情况下，我的 window 函数应该回顾“4 周”而不是“4 行”。
- 丢失的日期不是 0，而是隐含的 NULL

 thing | date       | dow | val | avg
   1   | 2018-01-01 |  1  |  1  | NULL  <= AVG({})
   1   | 2018-01-08 |  1  |  2  | 1     <= AVG({1})
   1   | 2018-01-15 |  1  |  3  | 1.5   <= AVG({1,2})
   1   | 2018-01-22 |  1  |  4  | 2     <= AVG({1,2,3})
   1   | 2018-01-29 |  1  |  5  | 2.5   <= AVG({1,2,3,4})

   1   | 2018-02-12 |  1  |  7  | 4     <= AVG({3,4,5})
   1   | 2018-02-19 |  1  |  8  | 5.33  <= AVG({4,5,7})
   1   | 2018-02-26 |  1  |  9  | 6.66  <= AVG({5,7,8})
   1   | 2018-03-05 |  1  |  10 | 8     <= AVG({7,8,9})
   1   | 2018-03-12 |  1  |  11 | 11.25 <= AVG({7,8,9,10})
   1   | 2018-03-19 |  1  |  12 | 9.5   <= AVG({8,9,10,11})

注意：2018-02-05 没有值

我通常会通过以下两种方式之一处理它...

LEFT JOIN 到模板 "force" 所有日期都存在，并有效地依赖 AVG() "ignoring" NULL。

这不太理想，因为 "things" 的数量很大，而且构建此模板的成本很高。

SELECT
  *,
  AVG(mytable.val) OVER (PARTITION BY things.id, dates.dow
                             ORDER BY dates.date
                          ROW BETWEEN 4 PRECEDING AND 1 PRECEDING
                        )
                           AS preceding_4_week_average
FROM
  things
CROSS JOIN
  dates
LEFT JOIN
  myTable
    ON  myTable.date = dates.date
    AND myTable.id   = things.id

不要使用 window 函数，而是使用自连接

这不太理想，因为 myTable 中有数百列，而 BigQuery 在这方面表现不佳。

SELECT
  myTable.*,
  AVG(hist.val)   AS preceding_4_week_average
FROM
  myTable
LEFT JOIN
  myTable   AS hist
    ON  hist.id    = myTable.id
    AND hist.date >= myTable.date - INTERVAL 28 DAYS
    AND hist.date <  myTable.date
GROUP BY
  myTable.column1,
  myTable.column2,
  etc, etc

真题

有没有其他人有替代方案，最好使用 windowed/analytic 函数来 "look back 4 weeks" 而不是 "look back 4 rows"？

Answer 1

这是蛮力，但应该更快：

select t.*,
       ((case when date_1 >= date_add(date, interval -4 week)
              then val_1 else 0
         end) +
        (case when date_2 >= date_add(date, interval -4 week)
              then val_2 else 0
         end) +
        (case when date_3 >= date_add(date, interval -4 week)
              then val_3 else 0
         end) +
        (case when date_4 >= date_add(date, interval -4 week)
              then val_4 else 0
         end)
        ) /
       ((case when date_1 >= date_add(date, interval -4 week)
              then 1 else 0
         end) +
        (case when date_2 >= date_add(date, interval -4 week)
              then 1 else 0
         end) +
        (case when date_3 >= date_add(date, interval -4 week)
              then 1 else 0
         end) +
        (case when date_4 >= date_add(date, interval -4 week)
              then 1 else 0
         end)
        )        
from (select t.*,
             lag(val, 1) over (partition by id, dow order by date) as val_1,
             lag(val, 2) over (partition by id, dow order by date) as val_2,
             lag(val, 3) over (partition by id, dow order by date) as val_3,
             lag(val, 4) over (partition by id, dow order by date) as val_4,
             lag(date, 1) over (partition by id, dow order by date) as date_1,
             lag(date, 2) over (partition by id, dow order by date) as date_2,
             lag(date, 3) over (partition by id, dow order by date) as date_3,
             lag(date, 4) over (partition by id, dow order by date) as date_4
   from mytable t
     ) t;

可能有一种巧妙的方法可以使用数组来表达这一点，但我在这里有点早。

Answer 2

以下适用于 BigQuery 标准 SQL

如您所见 - 诀窍在于使用 RANGE 而不是 ROW

#standardSQL
SELECT *,
  AVG(val) OVER(
    PARTITION BY id, dow 
    ORDER BY DATE_DIFF(DATE_TRUNC(date, WEEK), DATE_TRUNC(CURRENT_DATE(), WEEK), WEEK) 
    RANGE BETWEEN 4 PRECEDING AND 1 PRECEDING
  ) AVG
FROM `project.dataset.table`

您可以使用您问题中的虚拟数据来测试和玩上面的内容，如下所示

#standardSQL
WITH `project.dataset.table` AS (
  SELECT 1 id, DATE '2018-01-01' date, 1 dow, 1 val UNION ALL
  SELECT 1, '2018-01-08', 1, 2  UNION ALL
  SELECT 1, '2018-01-15', 1, 3  UNION ALL
  SELECT 1, '2018-01-22', 1, 4  UNION ALL
  SELECT 1, '2018-01-29', 1, 5  UNION ALL
  SELECT 1, '2018-02-12', 1, 7  UNION ALL
  SELECT 1, '2018-02-19', 1, 8  UNION ALL
  SELECT 1, '2018-02-26', 1, 9  UNION ALL
  SELECT 1, '2018-03-05', 1, 10 UNION ALL
  SELECT 1, '2018-03-12', 1, 11 UNION ALL
  SELECT 1, '2018-03-19', 1, 12  
)
SELECT *,
  AVG(val) OVER(
    PARTITION BY id, dow 
    ORDER BY DATE_DIFF(DATE_TRUNC(date, WEEK), DATE_TRUNC(CURRENT_DATE(), WEEK), WEEK) 
    RANGE BETWEEN 4 PRECEDING AND 1 PRECEDING
  ) avg
FROM `project.dataset.table`
-- ORDER BY date

结果为

Row id      date    dow val avg 
1   1   2018-01-01  1   1   null     
2   1   2018-01-08  1   2   1.0  
3   1   2018-01-15  1   3   1.5  
4   1   2018-01-22  1   4   2.0  
5   1   2018-01-29  1   5   2.5  
6   1   2018-02-12  1   7   4.0  
7   1   2018-02-19  1   8   5.333333333333333    
8   1   2018-02-26  1   9   6.666666666666667    
9   1   2018-03-05  1   10  8.0  
10  1   2018-03-12  1   11  8.5  
11  1   2018-03-19  1   12  9.5

加窗平均，考虑差距

Windowed Average, accounting for gaps

sql

analytic-functions

google-bigquery

rolling-average