PostgreSQL BRIN索引详解

pg9.5增加的brin索引，存储了被索引字段在块级别的边界值（最大值、最小值）以及其他统计信息。
BRIN索引的扫描原理很简单，扫描BRIN的元数据，根据元数据和用户输入的条件进行比较，过滤不符合条件的HEAP PAGE，只扫描需要扫描的HEAP PAGE。
所以brin索引仅仅适合存储与值线性相关性很好的列。

例子：
我们创建两张表，一张表插入线性相关很好的数据，另一张表插入离散的数据来测试brin索引的性能。

–建表插入数据：

bill=# create table t_brin1(id int);
CREATE TABLE
bill=# create table t_brin2(id int); 
CREATE TABLE
bill=# insert into t_brin1 select random()*1000000 from generate_series(1,10000000);
INSERT 0 10000000
bill=# insert into t_brin2 select generate_series(1,10000000);
INSERT 0 10000000

–在两张表上都创建brin索引

bill=# create index idx_t_brin1 on t_brin1 using brin(id);
CREATE INDEX
bill=# create index idx_t_brin2 on t_brin2 using brin(id);  
CREATE INDEX

–通过pg_stats表查看两张表索引列的线性相关性
可以发现t_brin1表中数据比较离散，而t_brin2中数据线性相关性很好。

bill=# select correlation from pg_stats where tablename='t_brin1';
  correlation  
---------------
 -0.0061385944
(1 row)

bill=# select correlation from pg_stats where tablename='t_brin2';
 correlation 
-------------
           1
(1 row)

–测试性能

t_brin1表：
bill=# explain (analyze ,verbose,timing,costs,buffers) select * from t_brin1 where id between 1 and 100000;
                                                            QUERY PLAN                                                             
-----------------------------------------------------------------------------------------------------------------------------------
 Bitmap Heap Scan on bill.t_brin1  (cost=262.18..194509.84 rows=988836 width=4) (actual time=1.332..1244.478 rows=1000823 loops=1)
   Output: id
   Recheck Cond: ((t_brin1.id >= 1) AND (t_brin1.id <= 100000))
   Rows Removed by Index Recheck: 8999177
   Heap Blocks: lossy=44248
   Buffers: shared hit=16324 read=27932 written=31
   ->  Bitmap Index Scan on idx_t_brin1  (cost=0.00..14.97 rows=9999977 width=0) (actual time=1.299..1.299 rows=442880 loops=1)
         Index Cond: ((t_brin1.id >= 1) AND (t_brin1.id <= 100000))
         Buffers: shared hit=8
 Planning Time: 0.083 ms
 Execution Time: 1302.285 ms
(11 rows)

t_brin2表：

bill=# explain (analyze ,verbose,timing,costs,buffers) select * from t_brin2 where id between 1 and 100000;
                                                         QUERY PLAN                                                          
-----------------------------------------------------------------------------------------------------------------------------
 Bitmap Heap Scan on bill.t_brin2  (cost=26.92..46009.03 rows=91581 width=4) (actual time=0.166..16.945 rows=100000 loops=1)
   Output: id
   Recheck Cond: ((t_brin2.id >= 1) AND (t_brin2.id <= 100000))
   Rows Removed by Index Recheck: 15712
   Heap Blocks: lossy=512
   Buffers: shared hit=1 read=513
   ->  Bitmap Index Scan on idx_t_brin2  (cost=0.00..4.03 rows=115607 width=0) (actual time=0.142..0.142 rows=5120 loops=1)
         Index Cond: ((t_brin2.id >= 1) AND (t_brin2.id <= 100000))
         Buffers: shared hit=1 read=1
 Planning Time: 0.274 ms
 Execution Time: 22.391 ms
(11 rows)

可以发现t_brin2表上时间少了60多倍！其原因也很简单，因为t_brin1表上数据分散比较离散，所以使用brin索引扫描时必然要过滤掉大量的数据(t_brin1过滤了8999177条记录)，自然性能就差很多了。这就好比我们用渔网去捕鱼，t_brin1的情况就类似我们收网上来一看渔网中鱼没几条，而大部分都是不需要的垃圾，而t_brin2的情况恰恰相反，一渔网拉上来全部是鱼。

看到这你可能会感慨brin索引带来的性能提升十分巨大，但是不要急，我们还能继续优化上面的这个查询。

BRIN索引有一个参数pages_per_range可以用来近一步提升brin索引的性能。这个参数用来决定每多少个数据块统计一次边界，例如pages_per_range默认值是128即表示每128个数据块统计一次边界。
1、BRIN索引的精确度。pages_per_range=1，说明边界精确到1个数据块。pages_per_range越小，精度越高，过滤性就越好（注意过滤性越好取决于列的线性相关性很好的情况下，否则就是白瞎）。
2、BRIN索引本身的大小。pages_per_range越小，BRIN索引本身就越大。BRIN越大，单次走BRIN索引扫描BRIN块的成本就越高。

所以我们可以设置合适的pages_per_range来优化上面查询。

bill=# drop index idx_t_brin2;
DROP INDEX
bill=# create index idx_t_brin2 on t_brin2 using brin(id) with (pages_per_range=16);
CREATE INDEX
bill=# explain (analyze ,verbose,timing,costs,buffers) select * from t_brin2 where id between 1 and 100000;
                                                         QUERY PLAN                                                          
-----------------------------------------------------------------------------------------------------------------------------
 Bitmap Heap Scan on bill.t_brin2  (cost=34.50..44354.49 rows=82814 width=4) (actual time=0.662..14.559 rows=100000 loops=1)
   Output: id
   Recheck Cond: ((t_brin2.id >= 1) AND (t_brin2.id <= 100000))
   Rows Removed by Index Recheck: 1248
   Heap Blocks: lossy=448
   Buffers: shared hit=462
   ->  Bitmap Index Scan on idx_t_brin2  (cost=0.00..13.79 rows=83153 width=0) (actual time=0.648..0.648 rows=4480 loops=1)
         Index Cond: ((t_brin2.id >= 1) AND (t_brin2.id <= 100000))
         Buffers: shared hit=14
 Planning Time: 0.151 ms
 Execution Time: 20.102 ms
(11 rows)

可以看到查询时间又近一步缩短了！