如果去掉partition by，你可以用你的初始方法通过算术来实现这一点：demo

SELECT at_timestamp, 
       sum(users_count*(2*event_type-1)) OVER (ORDER BY at_timestamp)
FROM users_events
ORDER BY at_timestamp;

那是因为

• event_type似乎是integer
• 2 * 0 - 1是-1，并且
• 2 * 1 - 1是1

将其用作乘数，0的event_type导致从步进和中减去相应的users_count，1的event_type导致users_count的相加

简明扼要：CASE WHEN乘数，从@shawnt00开始，在运行求和表达式中使用窗口子句。我使用一个 * 命名窗口 *，因为我发现它更可读，PostgreSQL允许它。
最后，通过运行和降序并应用LIMIT 1进行排序：

marco ~/1/Vertica/supp $ psql -af rsum.sql
WITH
-- Your in data ...
indata(at_timestamp,users_count,event_type) AS (
          SELECT 100000,2,1
UNION ALL SELECT 100001,4,1
UNION ALL SELECT 100003,5,0
UNION ALL SELECT 100005,1,1
UNION ALL SELECT 100006,3,1
UNION ALL SELECT 100008,2,0
UNION ALL SELECT 100008,1,1
)
-- real query starts here ...
SELECT 
  *
, SUM(users_count * CASE event_type WHEN 1 THEN 1 ELSE -1 END) OVER w AS runsum
FROM indata
WINDOW w AS (ORDER BY at_timestamp)
ORDER BY 4 DESC LIMIT 1;
 at_timestamp | users_count | event_type | runsum 
--------------+-------------+------------+--------
       100001 |           4 |          1 |      6

with tallied as (
    select *,
        sum(user_count
              * case when event_type = 1
                     then 1 else -1 end) over (
            order by at_timestamp) as tally
    from T
), ranked as (
    select *,
        dense_rank() over (
            order by tally desc) as dr
    from tallied
) 
select * from ranked where dr = 1;

由于排序中可能存在并列关系，因此使用dense_rank()过滤最大输出行。