你的 AI SQL 工具没有“撤销逻辑”按钮，这才是真正的问题

原创

析言

发布于 2026-08-03 14:41:29

820

文章被收录于专栏：SQLazySQLazy

你有没有过这种经历：对着 AI SQL 工具改了五六版提示词，点了无数次撤销重写，最后跑出来的结果还是不对。

你能撤销输入的文字，能删掉不满意的代码，但你撤销不了 AI 悄悄跑偏的逻辑。这才是 AI 写 SQL 最隐蔽的坑。

就拿用户行为分析里的「间隔超 1 小时重置会话编号」来说，需求一句话就能说清，AI 生成的 SQL 永远语法工整：

WITH lagged AS (
  SELECT *, LAG(dt) OVER (PARTITION BY account_number ORDER BY dt) AS prev_dt
  FROM numEvents
), grouped AS (
  SELECT *,
    SUM(CASE WHEN TIMESTAMPDIFF(SECOND, prev_dt, dt) > 3600 THEN 1 ELSE 0 END)
    OVER (PARTITION BY account_number ORDER BY dt) AS grp
  FROM lagged
)
SELECT *, ROW_NUMBER() OVER (PARTITION BY account_number, grp ORDER BY dt) AS seq
FROM grouped
WITH lagged AS (
  SELECT *, LAG(dt) OVER (PARTITION BY account_number ORDER BY dt) AS prev_dt
  FROM numEvents
), grouped AS (
  SELECT *,
    SUM(CASE WHEN TIMESTAMPDIFF(SECOND, prev_dt, dt) > 3600 THEN 1 ELSE 0 END)
    OVER (PARTITION BY account_number ORDER BY dt) AS grp
  FROM lagged
)
SELECT *, ROW_NUMBER() OVER (PARTITION BY account_number, grp ORDER BY dt) AS seq
FROM grouped

三层嵌套，挑不出语法毛病。可要是结果不对呢？你不知道是 LAG 的边界处理错了，还是累加条件写反了，或是行号的分区不对。想修正？只能重新组织提示词，整段重生成，相当于全盘推倒重来。你的撤销键，只能删掉整段代码，撤不掉逻辑里的偏差。

但在 SQLazy 里，同样的逻辑拆成清晰的 3 步：