Question

我希望找到给定时间范围内存在多少个分组间隙。

starting range: 2012-01-12 00:00:00
  ending range: 2012-01-18 59:59:59

大致翻译为：

type  10 11 12 13 14 15 16 17 18 19 20 
a        |--========]
a                             |==------]
b                 |==============--]
c     |-----===========]
d        |--=====================------]

按类型分组的相同数据：

a        |--========]         |==------]
b                 |==============--]
c     |-----===========]
d        |--=====================------]

导致：

type  gap
---------
a     1  (yes)
b     1  (yes)
c     1  (yes)
d     0  (no)

最终......

SUM(gap) AS gaps
----------------
3

更新澄清：

数据以每种类型的开始和结束时间戳存储。例如：

id  type  start_datetime       end_datetime
--------------------------------------------------
1   a     2012-01-11 00:00:00  2012-01-14 59:59:59
2   a     2012-01-18 00:00:00  2012-01-20 59:59:59
3   b     2012-01-14 00:00:00  2012-01-19 59:59:59
4   c     2012-01-10 00:00:00  2012-01-15 59:59:59
5   d     2012-01-11 00:00:00  2012-01-20 59:59:59

Answer 1

为了避免双重工作，这里是数据（我将包含的上边界湾替换为独占的，这是更常见的，恕我直言）：

-- CREATE SCHEMA tmp;
DROP TABLE tmp.gaps CASCADE;
CREATE TABLE tmp.gaps
        ( id INTEGER NOT NULL PRIMARY KEY       -- surrogate key
        , ztype CHAR(1) NOT NULL
        , start_datetime TIMESTAMP NOT NULL     -- lower boundary := inclusive
        , end_datetime TIMESTAMP NOT NULL       -- upper boundary := exclusive
        );
CREATE UNIQUE INDEX gaps_forward ON tmp.gaps(ztype,start_datetime);
CREATE UNIQUE INDEX gaps_backward ON tmp.gaps(ztype,end_datetime);

INSERT INTO tmp.gaps(id,ztype,start_datetime,end_datetime) VALUES
 (1,'a', '2012-01-11 00:00:00', '2012-01-15 00:00:00' )
,(2,'a', '2012-01-18 00:00:00', '2012-01-21 00:00:00' )
,(3,'b', '2012-01-14 00:00:00', '2012-01-20 00:00:00' )
,(4,'c', '2012-01-10 00:00:00', '2012-01-16 00:00:00' )
,(5,'d', '2012-01-11 00:00:00', '2012-01-21 00:00:00' )
,(6,'e', '2012-01-11 00:00:00', '2012-01-15 00:00:00' ) -- added this
,(7,'e', '2012-01-15 00:00:00', '2012-01-21 00:00:00' ) -- and this
        ;
-- SELECT * FROM tmp.gaps;

更新：这是CTE。在第一个UNION中，我在想要的（1月12日 - 19日 - 1月）间隔的左侧和右侧添加了两个假的间隔。

每个ztype我计算间隔的总数。如果没有孔，这应该是一个，如果有一个孔，则应该是两个，等等。这也将为ztype找到在所需间隔中没有任何记录的空白。

-- EXPLAIN ANALYZE
WITH RECURSIVE meuk(ztype,start_datetime,end_datetime) AS (
        -- For every possible "ztype" add two dummie records
        -- just before and just after our wanted interval.
        WITH plus2 AS (
                SELECT g0.ztype,g0.start_datetime,g0.end_datetime FROM tmp.gaps g0
                WHERE (g0.start_datetime <= '2012-01-12 00:00:00' AND g0.end_datetime >= '2012-01-12 00:00:00')
                   OR (g0.start_datetime >= '2012-01-12 00:00:00' AND g0.end_datetime <= '2012-01-19 00:00:00')
                   OR (g0.start_datetime <= '2012-01-19 00:00:00' AND g0.end_datetime >= '2012-01-19 00:00:00')
                UNION ALL SELECT DISTINCT g1.ztype, '1900-01-01 00:00:00'::timestamp, '2012-01-12 00:00:00'::timestamp FROM tmp.gaps g1
                UNION ALL SELECT DISTINCT g2.ztype, '2012-01-19 00:00:00'::timestamp, '2100-01-01 00:00:00'::timestamp FROM tmp.gaps g2
                )
        SELECT p0.ztype,p0.start_datetime,p0.end_datetime
        FROM plus2 p0
                -- the start of a stretch: there is no older overlapping 
                -- (or touching) interval
        WHERE NOT EXISTS (SELECT *
                FROM plus2 nx
                WHERE nx.ztype = p0.ztype
                AND nx.start_datetime < p0.start_datetime -- older
                AND nx.end_datetime >= p0.start_datetime  -- touching or overlapping
                )
        UNION
        SELECT mk.ztype
                , LEAST(mk.start_datetime,p1.start_datetime)
                , GREATEST(mk.end_datetime,p1.end_datetime)
        FROM plus2 p1
        , meuk mk
        WHERE p1.ztype = mk.ztype
        AND (p1.start_datetime >= mk.start_datetime AND p1.start_datetime <= mk.end_datetime AND p1.end_datetime > mk.end_datetime)
        )
SELECT ztype, COUNT(*)-1 AS ngap
FROM meuk mk
WHERE NOT EXISTS (SELECT *
        FROM meuk  nx
        WHERE nx.ztype = mk.ztype
        AND (nx.start_datetime,nx.end_datetime) OVERLAPS( mk.start_datetime,mk.end_datetime)
        AND (nx.end_datetime - nx.start_datetime) > (mk.end_datetime - mk.start_datetime)
        )
GROUP BY ztype
ORDER BY ztype
        ;

创建最终总和留给读者的练习; - ）

结果：

 ztype | ngap 
-------+------
 a     |    1
 b     |    1
 c     |    1
 d     |    0
 e     |    0
(5 rows)

Answer 2

这是wildplasser使用Windows而不是CTE的答案的变体。基于相同的测试夹具：

select ztype, count(*) as gaps
from (
    select ztype, datetime, sum(n) over(partition by ztype order by datetime asc) as level
    from (
        select id, ztype, start_datetime as datetime, 1 as n from tmp.gaps
        union all
        select id, ztype, end_datetime, -1 from tmp.gaps
        union all
        select 0, ztype, '2012-01-12 00:00:00', 0 from (select distinct ztype from tmp.gaps) z
        union all
        select 0, ztype, '2012-01-19 00:00:00', 0 from (select distinct ztype from tmp.gaps) z
    ) x
) x
where level = 0 and datetime >= '2012-01-12 00:00:00' and datetime < '2012-01-19 00:00:00'
group by ztype
;

这是基于使用sum（）作为窗口聚合，为范围开始加1，为范围结束减1，然后在目标范围内查找运行总和为0的点。我必须做与wildplasser相同的事情，添加一些额外的条目，这些条目在边界的端点没有贡献任何东西，以便找到没有任何东西覆盖边界的组...

这似乎在测试数据上花费较少，但我认为它可能高度依赖于表中没有太多数据要经过。通过一些重新排列（这会使它更难以阅读），它只能对tmp.gaps进行两次完整扫描（其中一次只是获得不同的ztypes）。

计算时间范围内的分组间隙

2 个答案: