Question

我有一个如下所示的对象状态日志

timestamp, object_id, state, level
2018-01-01 123        f      100 
2018-01-02 123        t      100    
2018-01-02 123        f      100
2018-01-03 123        f      100
2018-01-03 123        f      100
2018-01-06 123        t      90
2018-01-07 123        t      90
2018-01-08 123        f      90

时间戳实际上是一个完整的日期/时间，为简洁起见，我没有包括时间部分。

我要获取的是基于唯一状态和级别的状态转换列表，看起来像这样

start      end        object_id, state, level
2018-01-01 2018-01-02 123        f      100 
2018-01-02 2018-01-02 123        t      100
2018-01-02 2018-01-06 123        f      100
2018-01-06 2018-01-08 123        t      90
2018-01-08 NOW()      123        f      90

我试图想出一种使用窗口函数来完成此操作的方法，例如

SELECT
    timestamp,
    object_id,
    timestamp as start,
    lead(timestamp) OVER (ORDER BY timestamp) as end,
FROM (
    SELECT
        timestamp,
        object_id,
        state,
        evel,
        rank() OVER (PARTITION BY (state, level) ORDER BY timestamp) as rank
    FROM state_log AS l
    WHERE object_id=123 AND timestamp >= DATE '2018-01-01'
    ORDER BY timestamp
) AS states
WHERE rank=1

但是我想我不明白rank（）是如何工作的，它不能满足我的需要。出于某种原因，我认为rank（）会在分区每次更改时重置行计数，但事实并非如此。我该怎么做？

Answer 1

这是一个孤岛问题。一个不错的解决方案使用row_number()：

select object_id, level, state, min(timestamp), max(timestamp)
from (select t.*,
             row_number() over (partition by object_id, level order by timestamp) as seqnum,
             row_number() over (partition by object_id, level, state order by timestamp) as seqnum_2
      from t
     ) t
group by (seqnum - seqnum_2), object_id, level, state;

很难解释为什么这样做。但是，如果查看子查询的结果，则将发现当状态为常数时，两个seqnum之间的差异是恒定的。这就定义了您想要的分组-以及其他列-其余的只是聚合。

这是rextester，显示它正常工作。

Answer 2

这与“差距和孤岛”无关。该技术通过具有某个字段的相应常量值的组进行操作，但是您需要使用此类组的边界进行操作。所以：

create table state_log(timestamp timestamp, object_id int, state boolean, level int);
insert into state_log values
  ('2018-01-01 00:00:01', 123, 'f', 100),
  ('2018-01-02 00:00:02', 123, 't', 100),
  ('2018-01-02 00:00:03', 123, 'f', 100),
  ('2018-01-03 00:00:04', 123, 'f', 100),
  ('2018-01-03 00:00:05', 123, 'f', 100),
  ('2018-01-06 00:00:06', 123, 't', 90),
  ('2018-01-07 00:00:07', 123, 't', 90),
  ('2018-01-08 00:00:08', 123, 'f', 90);

select
  timestamp::date as start,
  coalesce(lead(timestamp) over (order by timestamp), now()::timestamp)::date as end,
  object_id, state, level
from (
  select 
    *,
    coalesce(lag(state) over (order by timestamp) <> state, true) as is_new_group
  from state_log) as t
where
  object_id = 123 and timestamp >= date '2018-01-01' and
  is_new_group
order by timestamp;

结果（我删除了时间部分，使其更像问题中指定的结果）：

┌────────────┬────────────┬───────────┬───────┬───────┐
│   start    │    end     │ object_id │ state │ level │
├────────────┼────────────┼───────────┼───────┼───────┤
│ 2018-01-01 │ 2018-01-02 │       123 │ f     │   100 │
│ 2018-01-02 │ 2018-01-02 │       123 │ t     │   100 │
│ 2018-01-02 │ 2018-01-06 │       123 │ f     │   100 │
│ 2018-01-06 │ 2018-01-08 │       123 │ t     │    90 │
│ 2018-01-08 │ 2018-08-30 │       123 │ f     │    90 │
└────────────┴────────────┴───────────┴───────┴───────┘

如何在PostgreSQL中将状态日志数据聚合到具有相同状态的时间间隔中？

2 个答案: