Snowflake cloud data platform 嵌套窗口函数在雪花中不起作用

我正在将spark sql迁移到snowsql。 曾经有一个场景，我在spark sql中使用了嵌套窗口函数。我想将sql查询迁移到snowflake中。但snowflake不支持嵌套窗口函数

Spark sql查询-

SELECT 
    *,
    (case when (
        (
            lead(timestamp -lag(timestamp)
                over (partition by session_id order by timestamp))
                over (partition by session_id order by timestamp)
        ) is not null)
     then  
    (
        lead(timestamp -lag(timestamp)
            over (partition by session_id order by timestamp))
            over (partition by session_id order by timestamp)
    ) 
    else 0 end)/1000 as pg_to_pg
FROM dwell_time_step2

with lagsession as (
SELECT 
    a.*,
    lag(timestamp) over (partition BY session_id   order by timestamp asc) lagsession 
FROM mktg_web_wi.dwell_time_step2 a
)

select 
    a.,
    nvl(lead(a.timestamp - b.lagsession) over (partition BY a.session_id order by   a.timestamp),0)/1000 pg_to_pg
FROM mktg_web_wi.dwell_time_step2 a,
    lagsession b
WHERE a.key=b.key
order by timestamp;

输出- 我已尝试将上述查询转换为雪花，如下所示

转换后的Snowsql-

SELECT 
    *,
    (case when (
        (
            lead(timestamp -lag(timestamp)
                over (partition by session_id order by timestamp))
                over (partition by session_id order by timestamp)
        ) is not null)
     then  
    (
        lead(timestamp -lag(timestamp)
            over (partition by session_id order by timestamp))
            over (partition by session_id order by timestamp)
    ) 
    else 0 end)/1000 as pg_to_pg
FROM dwell_time_step2

with lagsession as (
SELECT 
    a.*,
    lag(timestamp) over (partition BY session_id   order by timestamp asc) lagsession 
FROM mktg_web_wi.dwell_time_step2 a
)

select 
    a.,
    nvl(lead(a.timestamp - b.lagsession) over (partition BY a.session_id order by   a.timestamp),0)/1000 pg_to_pg
FROM mktg_web_wi.dwell_time_step2 a,
    lagsession b
WHERE a.key=b.key
order by timestamp;

输出-

SELECT 
    *,
    (case when (
        (
            lead(timestamp -lag(timestamp)
                over (partition by session_id order by timestamp))
                over (partition by session_id order by timestamp)
        ) is not null)
     then  
    (
        lead(timestamp -lag(timestamp)
            over (partition by session_id order by timestamp))
            over (partition by session_id order by timestamp)
    ) 
    else 0 end)/1000 as pg_to_pg
FROM dwell_time_step2

with lagsession as (
SELECT 
    a.*,
    lag(timestamp) over (partition BY session_id   order by timestamp asc) lagsession 
FROM mktg_web_wi.dwell_time_step2 a
)

select 
    a.,
    nvl(lead(a.timestamp - b.lagsession) over (partition BY a.session_id order by   a.timestamp),0)/1000 pg_to_pg
FROM mktg_web_wi.dwell_time_step2 a,
    lagsession b
WHERE a.key=b.key
order by timestamp;

这里，问题在于Snow sql输出。将时间值分配给不同的URL

期望spark sql查询在snowsql上工作，并且在这两种情况下输出应该相同。

如果有人知道如何解决这个问题，请告诉我

---

谢谢

我认为将其从嵌套窗口函数更改为cte已经改变了滞后和超前记录所指的内容，但这很难让我了解

无论如何，如果我理解这里的代码，我认为有一种更简单的方法，只有一个windows函数

select 
    a.*,
    (nvl(lead(a.timestamp) over (partition BY a.session_id order by a.timestamp) - a.timestamp)/1000,0) pg_to_pg
FROM mktg_web_wi.dwell_time_step2 a
order by timestamp;

---

我认为，将其从嵌套窗口函数更改为cte已经改变了滞后和超前记录所指的内容，但这很难让我理解

无论如何，如果我理解这里的代码，我认为有一种更简单的方法，只有一个windows函数

select 
    a.*,
    (nvl(lead(a.timestamp) over (partition BY a.session_id order by a.timestamp) - a.timestamp)/1000,0) pg_to_pg
FROM mktg_web_wi.dwell_time_step2 a
order by timestamp;