用以前的非空值替换空值

Replace nulls with the previous non-null value

我正在使用基于 Presto 0.172 的 Amazon Athena 引擎版本 1。

考虑示例数据集:

id date_column col1
1 01/03/2021 NULL
1 02/03/2021 1
1 15/03/2021 2
1 16/03/2021 NULL
1 17/03/2021 NULL
1 30/03/2021 NULL
1 30/03/2021 1
1 31/03/2021 NULL

我想用最后一个非 NULL 值替换 table 中的所有 NULL,即我想得到:

id date_column col1
1 01/03/2021 NULL
1 02/03/2021 1
1 15/03/2021 2
1 16/03/2021 2
1 17/03/2021 2
1 30/03/2021 1
1 30/03/2021 1
1 31/03/2021 1

我正在考虑使用带有 IGNORE NULLS 选项的 lag 函数,但不幸的是,Athena 引擎版本 1 不支持 IGNORE NULLS(它也不被 Athena 引擎版本支持2, 基于 Presto 0.217).

如何在不使用 IGNORE NULLS 选项的情况下实现所需的格式?

这里是生成示例的一些模板table:

WITH source1 AS (
SELECT
 *
FROM (
  VALUES
  (1, date('2021-03-01'), NULL),
  (1, date('2021-03-02'), 1),
  (1, date('2021-03-15'), 2),
  (1, date('2021-03-16'), NULL),
  (1, date('2021-03-17'), NULL),
  (1, date('2021-03-30'), NULL), 
  (1, date('2021-03-30'), 1), 
  (1, date('2021-03-31'), NULL)
) AS t (id, date_col, col1)
)

SELECT
  id
, date_col
, col1
-- This doesn't work as IGNORE NULLS is not supported.
-- CASE 
--   WHEN col1 IS NOT NULL THEN col1
--   ELSE lag(col1) OVER IGNORE NULLS (PARTITION BY id ORDER BY date_col) 
-- END AS col1_lag_nulls_ignored
FROM
  source1
ORDER BY
  date_co

在查看关于 SO (here and here) 的类似问题后,以下解决方案适用于所有列类型(包括字符串和日期):

WITH source1 AS (
SELECT
 *
FROM (
  VALUES
  (1, date('2021-03-01'), NULL),
  (1, date('2021-03-02'), 1),
  (1, date('2021-03-15'), 2),
  (1, date('2021-03-16'), NULL),
  (1, date('2021-03-17'), NULL),
  (1, date('2021-03-30'), 1), 
  (1, date('2021-03-31'), NULL)
) AS t (id, date_col, col1)
)

, grouped AS (
SELECT
  id
, date_col
, col1
-- If the row has a value in a column, then this row and all subsequent rows 
-- with a NULL (before the next non-NULL value) will be in the same group.
, sum(CASE WHEN col1 IS NULL THEN 0 ELSE 1 END) OVER (
    PARTITION BY id ORDER BY date_col) AS grp
FROM 
 source1
)

SELECT
  id
, date_col
, col1
-- max is used instead of first_value, since in cases where there will
-- be multiple records with NULL on the same date, the first_value may 
-- still return a NULL.
, max(col1) OVER (PARTITION BY id, grp ORDER BY date_col) AS col1_filled
, grp
FROM
  grouped 
ORDER BY
  date_col