递归 CTE Table - 计算 EWMA (EMA) - 如何 optimize/refactor 这段代码不会每次都被 TDWM 杀死？

Question

我创建了一个正在制作 CTE 表的查询，其中两个是非递归的，一个是递归的，以便计算指数加权移动平均值 (EMA)。当我运行我在 Teradata 中的代码在一段时间后被 TDWM 杀死。

有什么想法可以 improve/work 解决这个问题吗？

WITH 

smooth AS (
    SELECT CAST(0.741870935 AS NUMERIC (20,5)) AS alpha
),

numbered AS (
    SELECT 
        ROW_NUMBER() OVER (ORDER BY customer_name, closed_date) as rn,  
        customer, 
        closed_date, 
        metric

    FROM my_table
),

recursive EWMA AS (
    SELECT rn, customer, closed_date, metric, CAST(metric AS NUMERIC(20,5)) as EWMA
    FROM numbered
    WHERE rn = 1

    UNION ALL

    SELECT numbered.rn, numbered.customer, numbered.closed_date, numbered.metric,
    smooth.alpha * numbered.metric + (1-smooth.alpha) * EWMA.EWMA
    FROM EWMA
    JOIN numbered
    ON EWMA.rn + 1 = numbered.rn
    CROSS JOIN smooth   
)

SELECT * FROM EWMA
ORDER BY closed_date;

Answer 1

您是否尝试过设置一个 depth 字段来限制递归？类似于：

WITH smooth AS (...),
numbered AS (...),
recursive EWMA AS (
  SELECT 
    rn, customer, closed_date, metric, CAST(metric AS NUMERIC(20,5)) as EWMA, 
    1 AS depth
  FROM numbered
  WHERE rn = 1

  UNION ALL

  SELECT 
    numbered.rn, numbered.customer, numbered.closed_date, numbered.metric,
    smooth.alpha * numbered.metric + (1-smooth.alpha) * EWMA.EWMA, 
    EWMA.Depth + 1 AS Depth
  FROM EWMA
  INNER JOIN numbered ON EWMA.rn + 1 = numbered.rn
  CROSS JOIN smooth   
  WHERE depth <= 10 -- Restrict recursion
)
SELECT * 
FROM EWMA
ORDER BY closed_date;

假定 my_table table 非常大，与 numbered 的递归连接可能会导致此问题。理想情况下，您希望对 PI 列进行直接相等连接——即 table1.pi_col1 = table2.pi_col2。不确定使用 +1 表达式将如何影响连接。

从高层次上看您的查询，您似乎只想在当前行的计算中使用上一行的值。如果是这样，那么您可以完全取消递归 CTE，只需使用 LAG() window 函数：

WITH smooth AS (
  SELECT CAST(0.741870935 AS NUMERIC (20,5)) AS alpha
)
SELECT 
  ROW_NUMBER() OVER (ORDER BY customer_name, closed_date) as rn, -- row number
  customer, 
  closed_date, 
  metric,
  CAST(
    COALESCE(
      (smooth.alpha * metric + (1-smooth.alpha)) * -- current row's value
      LAG((smooth.alpha * metric + (1-smooth.alpha))) OVER(
        ORDER BY customer_name, closed_date) -- previous row's value
      , metric -- handle first row (no previous "EWMA" value)
    )
  AS NUMERIC(20,5)) AS EWMA
FROM my_table;

递归 CTE Table - 计算 EWMA (EMA) - 如何 optimize/refactor 这段代码不会每次都被 TDWM 杀死？

Recursive CTE Table - Calculating EWMA (EMA) - How to optimize/refactor this code so it doesn't get killed by TDWM every time?

sql

statistics

recursion

teradata

common-table-expression