问题描述
我正在尝试编写一个查询，该查询将搜索我的数据库并每周查找设备的总唯一序列号。我当前的代码是：
SELECT date_part('week',"timestamp"),count(disTINCT serialno)
FROM eddi_minute em
GROUP BY date_part('week',"timestamp")
不幸的是，我正在搜索的数据集非常大（约 600Gb），因此搜索所需的时间非常长。我希望能够每周搜索一次，每周搜索一次很短的时间，即 1 分钟，也就是 1 分钟
select count(distinct serialno) as Devices
from eddi_minute em where "timestamp" >= '2021-06-23 00:01:00' and "timestamp" < '2021-06-23 00:02:00';
但是对于一整年的每周，这样我就可以按一次 Enter 键，它会为整个数据库执行此操作，并避免不必要的计数。
在理想的世界中，我的想法是创建一个我想要搜索的时间表，然后与它和我的数据库进行左连接以减少我正在搜索的数据，但我只有读取权限到服务器，所以这不是一个选项。有没有一种简单的方法可以做到这一点？如果这里有任何不清楚的地方，请道歉，如果有任何不正确的解释，我会详细说明。
表的索引是
CREATE UNIQUE INDEX "PK_4c94f05e4de575488f4a0c2905d" ON ONLY public.eddi_minute USING btree (serialno,"timestamp")
解释分析结果是：
GroupAggregate (cost=41219561.55..90787854.96 rows=200 width=16) (actual time=7065790.406..8172419.446 rows=53 loops=1)
Group Key: (date_part('week'::text,em."timestamp"))
-> Gather Merge (cost=41219561.55..88747442.16 rows=408082059 width=16) (actual time=7052726.256..7834672.575 rows=408057194 loops=1)
Workers Planned: 2
Workers Launched: 2
-> Sort (cost=41218561.53..41643646.99 rows=170034187 width=16) (actual time=6956066.331..7201252.404 rows=136019065 loops=3)
Sort Key: (date_part('week'::text,em."timestamp"))
Sort Method: external merge disk: 3368720kB
Worker 0: Sort Method: external merge disk: 3640792kB
Worker 1: Sort Method: external merge disk: 3371808kB
-> Parallel Append (cost=0.00..9256242.79 rows=170034187 width=16) (actual time=0.435..2825202.379 rows=136019065 loops=3)
-> Parallel Seq Scan on eddi_minute_p2021_05 em_11 (cost=0.00..1725776.58 rows=34898767 width=16) (actual time=0.011..1722528.987 rows=83740195 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_06 em_12 (cost=0.00..1488905.33 rows=30102507 width=16) (actual time=1.266..1488189.219 rows=72252984 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_04 em_10 (cost=0.00..1428581.36 rows=28905149 width=16) (actual time=149.934..1290294.249 rows=69366177 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_03 em_9 (cost=0.00..1290438.50 rows=26110040 width=16) (actual time=69.475..483281.530 rows=20887814 loops=3)
-> Parallel Seq Scan on eddi_minute_p2021_02 em_8 (cost=0.00..922294.02 rows=18661202 width=16) (actual time=195.734..931653.840 rows=44786882 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_01 em_7 (cost=0.00..823415.96 rows=16660557 width=16) (actual time=102.708..834900.144 rows=39985282 loops=1)
-> Parallel Seq Scan on eddi_minute_p2020_12 em_6 (cost=0.00..293130.95 rows=5931036 width=16) (actual time=182.465..296634.818 rows=14234537 loops=1)
-> Parallel Seq Scan on eddi_minute_p2020_11 em_5 (cost=0.00..111271.35 rows=2251388 width=16) (actual time=195.367..110910.685 rows=5403366 loops=1)
-> Parallel Seq Scan on eddi_minute_p2020_10 em_4 (cost=0.00..105311.10 rows=2130808 width=16) (actual time=146.920..109340.586 rows=5113938 loops=1)
-> Parallel Seq Scan on eddi_minute_p2020_09 em_3 (cost=0.00..93692.39 rows=1895711 width=16) (actual time=87.456..94169.812 rows=4549714 loops=1)
-> Parallel Seq Scan on eddi_minute_p2020_08 em_2 (cost=0.00..86189.97 rows=1743918 width=16) (actual time=0.007..88029.891 rows=4185403 loops=1)
-> Parallel Seq Scan on eddi_minute_p2020_07 em_1 (cost=0.00..33400.45 rows=675796 width=16) (actual time=1.046..14190.279 rows=1621911 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_07 em_13 (cost=0.00..3438.66 rows=88773 width=16) (actual time=0.006..51.229 rows=150887 loops=1)
-> Parallel Seq Scan on eddi_minute_default em_26 (cost=0.00..45.20 rows=1456 width=16) (actual time=0.016..0.639 rows=2477 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_08 em_14 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.000 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_09 em_15 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.515 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_10 em_16 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.000 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_11 em_17 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.000 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2021_12 em_18 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.000 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2022_01 em_19 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.000 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2022_02 em_20 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.000 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2022_03 em_21 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.001 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2022_04 em_22 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.000 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2022_05 em_23 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.000 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2022_06 em_24 (cost=0.00..15.00 rows=400 width=16) (actual time=0.000..0.000 rows=0 loops=1)
-> Parallel Seq Scan on eddi_minute_p2022_07 em_25 (cost=0.00..15.00 rows=400 width=16) (actual time=0.002..0.003 rows=0 loops=1)
Planning Time: 35.809 ms
Execution Time: 8172556.078 ms
解决方法
一些想法：
尽管 "timestamp" 是有效的列名，但为对象使用保留名称被认为是不好的做法。这可能看起来无害，但从长远来看可能会很烦人。
我认为 "timestamp" 列中的索引应该会显着提高第二个查询的性能：
CREATE INDEX idx_timestamp ON eddi_minute ("timestamp");
关于第一个查询：考虑到您有一个 600GB (!) 的表，在 "timestamp" 列中创建一个 partial index 可能会很有趣，这样时间戳就会按照您将要使用的值进行索引在您的查询中使用，例如周：
CREATE INDEX idx_timestamp_week ON eddi_minute (date_part('week',"timestamp"));
注意：虽然索引可以加快查询速度，但会减慢其他操作，例如插入、更新和删除。如果创建新索引，请测试所有相关操作的性能。
演示：db<>fiddle