PostgreSQL 范围过滤 + 其他字段排序OFFSET LIMIT(多字段区间过滤)的优化与加速
背景
在索引扫描中，如果两个字段扫描都是区间扫描，那么只能用到某个字段的过滤条件，另一个字段需要全扫描。
例如
create table t(id int, c1 int, c2 int);
insert into t select generate_series(1,6000000), random()*10000, random()*10000;
create index idx_t_1 on t(c1, c2);
explain (analyze,verbose,timing,costs,buffers) select * from t where c1 between 1 and 10000 order by c2 limit 1;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=25496.76..25496.76 rows=1 width=12) (actual time=5330.577..5330.578 rows=1 loops=1)
Output: id, c1, c2
Buffers: shared hit=6021360
-> Sort (cost=25496.76..25571.75 rows=29997 width=12) (actual time=5330.576..5330.576 rows=1 loops=1)
Output: id, c1, c2
Sort Key: t.c2
Sort Method: top-N heapsort Memory: 25kB
Buffers: shared hit=6021360
-> Index Scan using idx_t_1 on public.t (cost=0.43..25346.77 rows=29997 width=12) (actual time=0.032..4526.864 rows=5999724 loops=1)
Output: id, c1, c2
Index Cond: ((t.c1 >= 1) AND (t.c1 <= 10000))
Buffers: shared hit=6021360
Planning time: 0.100 ms
Execution time: 5330.734 ms
(14 rows)
explain (analyze,verbose,timing,costs,buffers) select * from t where c1 between 1 and 10000 and c2 between 100000 and 1111110 limit 1;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.43..5.52 rows=1 width=12) (actual time=248.439..248.439 rows=0 loops=1)
Output: id, c1, c2
Buffers: shared hit=21828
-> Index Scan using idx_t_1 on public.t (cost=0.43..763.89 rows=150 width=12) (actual time=248.437..248.437 rows=0 loops=1)
Output: id, c1, c2
Index Cond: ((t.c1 >= 1) AND (t.c1 <= 10000) AND (t.c2 >= 100000) AND (t.c2 <= 1111110))
Buffers: shared hit=21828
Planning time: 0.120 ms
Execution time: 248.578 ms
(9 rows)
原因是在复合索引中，每一个复合KEY是有序的，但是多级下来，每一层并不是有序的。例如:
那么前面的例子SQL2 实际上是层1用到了索引过滤，而层二则需要过滤。
例子SQL2 实际上是层1用到了索引过滤，并取出所有满足条件的行（因为用到了索引外的字段，所以不能用INDEX ONLY SCAN），再按层2排序。
索引结构可以参考：
那么如何优化这类SQL呢？
优化与场景
某个业务，数据包含了：
1、订单支付时间
2、订单号
3、唯一键 = 订单支付时间+订单号 （因为一个时间点，可能有多个订单）
业务需要取出大于某个支付时间，并按唯一键排序，偏移N条后，取一条。
DEMO表如下：
create table test(id int, c1 text, c2 timestamp);
写入6000万订单
insert into test select id, c1::text||id, c1 from (select generate_series(1,10000000) id, clock_timestamp()::timestamp(1) c1) t;
创建后期优化需要用到的索引
create index idx1 on test (c1);
create index idx2 on test (c2);
create index idx3 on test (c3);
查询语句如下:
select * from test where c2 >= '2018-01-22 16:13:25.4'::timestamp order by c1 limit 1 offset 50000;
LOG: duration: 943.851 ms plan:
Query Text: select * from test where c2 >= '2018-01-22 16:13:25.4'::timestamp order by c1 limit 1 offset 50000;
Limit (cost=4714.82..4714.92 rows=1 width=40) (actual time=943.822..943.823 rows=1 loops=1)
Output: id, c1, c2
Buffers: shared hit=141577
-> Index Scan using idx1 on public.test (cost=0.56..339427.36 rows=3600000 width=40) (actual time=928.122..939.856 rows=50001 loops=1)
Output: id, c1, c2
Filter: (test.c2 >= '2018-01-22 16:13:25.4'::timestamp without time zone)
Rows Removed by Filter: 6281766
Buffers: shared hit=141577
id | c1 | c2
---------+------------------------------+-----------------------
6414521 | 2018-01-22 16:13:25.46414521 | 2018-01-22 16:13:25.4
(1 row)
Time: 944.485 ms
性能并不理想。耗费了1秒左右的时间。原因是满足条件1的数据量太大，需要完全取出后排序再OFFSET。
优化方法
由于C1由C2+订单号组成，所以C1的最小值一定出现在C2中的最小值的区间中。那么可以这么优化
create or replace function get_test1(timestamp, int) returns test as $$
declare
v1 timestamp;
v2 text;
res text;
begin
set enable_seqscan=off;
-- 从输入的C2的条件，得到最小的C2
-- 索引精确定位
select c2 into v1 from test where c2 >= $1 order by c2 limit 1;
-- 在最小的C2中，求最小的C1
-- 索引精确定位
select min(c1) into v2 from test where c2 = v1;
-- 大于最小的C1，排序返回
-- 索引offset定位
select t into res from test t where c1 >= v2 order by c1 limit 1 offset $2 ;
return res::test;
end;
$$ language plpgsql strict;
使用以上优化，每一步的开销都是最小的。效果如下：
postgres=# \timing
postgres=# load 'auto_explain';
LOAD
postgres=# set auto_explain.log_nested_statements =on;
SET
Time: 0.165 ms
postgres=# set auto_explain.log_analyze =on;
SET
Time: 0.168 ms
postgres=# set auto_explain.log_buffers =on;
SET
Time: 0.144 ms
postgres=# set auto_explain.log_min_duration =0;
SET
Time: 0.161 ms
postgres=# set auto_explain.log_timing =on;
SET
Time: 0.147 ms
postgres=# set auto_explain.log_verbose =on;
SET
Time: 0.173 ms
postgres=# select * from get_test1('2018-01-22 16:13:25.4'::timestamp, 50000);
LOG: duration: 0.045 ms plan:
Query Text: select c2 from test where c2 >= $1 order by c2 limit 1
Limit (cost=0.43..0.46 rows=1 width=8) (actual time=0.043..0.043 rows=1 loops=1)
Output: c2
Buffers: shared hit=4
-> Index Only Scan using idx2 on public.test (cost=0.43..99814.16 rows=3333333 width=8) (actual time=0.041..0.041 rows=1 loops=1)
Output: c2
Index Cond: (test.c2 >= $1)
Heap Fetches: 1
Buffers: shared hit=4
LOG: duration: 0.024 ms plan:
Query Text: select min(c1) from test where c2 = v1
Result (cost=0.96..0.97 rows=1 width=32) (actual time=0.023..0.023 rows=1 loops=1)
Output: $0
Buffers: shared hit=5
InitPlan 1 (returns $0)
-> Limit (cost=0.56..0.96 rows=1 width=28) (actual time=0.020..0.020 rows=1 loops=1)
Output: test.c1
Buffers: shared hit=5
-> Index Only Scan using idx3 on public.test (cost=0.56..31750.61 rows=79365 width=28) (actual time=0.020..0.020 rows=1 loops=1)
Output: test.c1
Index Cond: ((test.c2 = $4) AND (test.c1 IS NOT NULL))
Heap Fetches: 1
Buffers: shared hit=5
LOG: duration: 57.454 ms plan:
Query Text: select t from test t where c1 >= v2 order by c1 limit 1 offset $2
Limit (cost=1699.64..1699.67 rows=1 width=92) (actual time=57.451..57.451 rows=1 loops=1)
Output: t.*, c1
Buffers: shared hit=774
-> Index Scan using idx1 on public.test t (cost=0.56..125732.85 rows=3700012 width=92) (actual time=0.048..53.365 rows=50001 loops=1)
Output: t.*, c1
Index Cond: (t.c1 >= '2018-01-22 16:13:25.46364521'::text)
Buffers: shared hit=774
LOG: duration: 58.163 ms plan:
Query Text: select * from get_test1('2018-01-22 16:13:25.4'::timestamp, 50000);
Function Scan on public.get_test1 (cost=0.25..0.26 rows=1 width=44) (actual time=58.155..58.156 rows=1 loops=1)
Output: id, c1, c2
Function Call: get_test1('2018-01-22 16:13:25.4'::timestamp without time zone, 50000)
Buffers: shared hit=783
id | c1 | c2
---------+------------------------------+-----------------------
6414521 | 2018-01-22 16:13:25.46414521 | 2018-01-22 16:13:25.4
(1 row)
Time: 58.503 ms
性能提升非常明显。
业务逻辑优化
实际上业务方要的是c2的值（支付时间），因此完全没必要按UK（c2+订单号）来排序，也就是说，SQL可以改写成这样。
postgres=# select * from test where c2 >= '2018-01-22 16:13:25.4'::timestamp order by c2 limit 1 offset 50000;
LOG: duration: 14.945 ms plan:
Query Text: select * from test where c2 >= '2018-01-22 16:13:25.4'::timestamp order by c2 limit 1 offset 50000;
Limit (cost=1497.64..1497.67 rows=1 width=40) (actual time=14.932..14.933 rows=1 loops=1)
Output: id, c1, c2
Buffers: shared hit=608
-> Index Scan using idx2 on public.test (cost=0.43..107799.04 rows=3600000 width=40) (actual time=0.030..10.541 rows=50001 loops=1)
Output: id, c1, c2
Index Cond: (test.c2 >= '2018-01-22 16:13:25.4'::timestamp without time zone)
Buffers: shared hit=608
id | c1 | c2
---------+------------------------------+-----------------------
6414521 | 2018-01-22 16:13:25.46414521 | 2018-01-22 16:13:25.4
(1 row)
Time: 15.442 ms
从业务层出发，改一条SQL，达到了最佳的效果。(而且最后只需要一个索引即可。)
从另一个层面来看这个优化，多个字段都是范围的查询，复合索引并不是最好的选择，更好的选择可能是：分区表+单索引，或者分区索引。需要内核的功能加强。