云贝教育 |【技术文章】pg缓存插件介绍

一、pg_buffercache

主要作用是查看pg的共享池中缓存的对象信息

1.1 创建扩展

postgres=# create extension pg_buffercache;
CREATE EXTENSION

1.2 查看视图pg_buffercache

postgres=# \d pg_buffercacheView "public.pg_buffercache"Column      |   Type   | Collation | Nullable | Default 
------------------+----------+-----------+----------+---------bufferid         | integer  |           |          | relfilenode      | oid      |           |          | reltablespace    | oid      |           |          | reldatabase      | oid      |           |          | relforknumber    | smallint |           |          | relblocknumber   | bigint   |           |          | isdirty          | boolean  |           |          | usagecount       | smallint |           |          | pinning_backends | integer  |           |          |

1.3 要看当前模式下的表​​​​​​​

postgres=# \dList of relationsSchema |      Name      | Type  |  Owner   
--------+----------------+-------+----------public | pg_buffercache | view  | postgrespublic | t1             | table | postgres
(2 rows)

1.4 查看t1表在当前pg缓存中的信息

postgres=# select count(1) from pg_buffercache where relfilenode='t1'::regclass;count 
-------0
(1 row)

​​​​​​​0表示没有缓存

1.5 查询一次t1表​​​​​​​

postgres=# select count(1) from t1;count 
-------1
(1 row)postgres=# select count(1) from pg_buffercache where relfilenode='t1'::regclass;count 
-------1
(1 row)

1.6 查看T1表缓存情况​​​​​​​

postgres=# select * from pg_buffercache where relfilenode='t1'::regclass;bufferid | relfilenode | reltablespace | reldatabase | relforknumber | relblocknumber | isdirty | usagecount | pinning_backends 
----------+-------------+---------------+-------------+---------------+----------------+---------+------------+------------------1006 |       16388 |          1663 |           5 |             0 |              0 | f       |          1 |                0
(1 row)

有记录表示被缓存

isdirty ：f表示不是脏块

1.7 修改表t1数据​​​​​​​

postgres=# update t1 set id=22 where id=1;
UPDATE 1

1.8 对比缓存的块是否变脏​​​​​​​

postgres=# select * from pg_buffercache where relfilenode='t1'::regclass;bufferid | relfilenode | reltablespace | reldatabase | relforknumber | relblocknumber | isdirty | usagecount | pinning_backends 
----------+-------------+---------------+-------------+---------------+----------------+---------+------------+------------------1006 |       16388 |          1663 |           5 |             0 |              0 | t       |          2 |                0
(1 row)

isdirty :t表示脏块

二、pg_prevarm

预热功能使用pg_prevarm函数，方便将数据缓存到OS缓存中或PG缓存中

比如生产系统中，数据库重启了，此时发起的业务SQL，就会发生物理读

语法​​​​​​​

pg_prewarm(regclass,   --预热的relationmode text default 'buffer', --使用预热的方法fork text derfault 'main', --relation fork被预热first_block int8 default null, --预热的第一块号last_block int8 default null  --预热的最后一个块号
) return int8;prefetch/read:缓存到os cache
buffer:缓存到os cache和pg shared buffers

参数说明：

regclass ：数据库对像,通常情况为表名

mode ：加载模式

• prefetch：异步地将数据预加载到操作系统缓存

• read：最终结果和 prefetch 一样,但它是同步方式,支持所有平台

• buffer：将数据预加载到数据库缓存

fork

main ：主表

fsm：空间空间地图

vm：可见性地图

first_block ：开始prewarm的数据块

last_block ：最后 prewarm 的数据块

2.1 创建prewarm插件

create EXTENSION pg_prewarm

2.2 在默认shared_buffer参数，创建一张大表

testdb=# show shared_buffers;shared_buffers 
----------------128MB
(1 row)testdb=# create table t1 ( id int,name varchar(100),c1 varchar(200),c2 varchar(200));
CREATE TABLE
testdb=# insert into t1 select id,md5(id::varchar),md5(md5(id::varchar)),md5(md5(md5(id::varchar))) from generate_series(1,10000000) as id;
INSERT 0 10000000

2.3 在没有OS和PG缓存的情况下

（一）不进行prewarm

1、查看表体积

testdb=# \d t1Table "public.t1"Column |          Type          | Collation | Nullable | Default 
--------+------------------------+-----------+----------+---------id     | integer                |           |          | name   | character varying(100) |           |          | c1     | character varying(200) |           |          | c2     | character varying(200) |           |          | testdb=# \dt+ t1List of relationsSchema | Name | Type  |  Owner   | Persistence | Access method |  Size   | Descr
iption 
--------+------+-------+----------+-------------+---------------+---------+-------------public | t1   | table | postgres | permanent   | heap          | 1281 MB | 
(1 row)testdb=# SELECT pg_size_pretty(pg_total_relation_size('t1'));pg_size_pretty 
----------------1281 MB
(1 row)

2、重启数据库并消除OS缓存​​​​​​​

pg_ctl restartecho 3 > /proc/sys/vm/drop_caches

3、查看执行计划​​​​​​​

testdb=# explain analyze select count(*) from t1;QUERY PLAN                                                                 
-------------------------------------------------------------------------------------------------------------------------------------------Finalize Aggregate  (cost=217018.73..217018.74 rows=1 width=8) (actual time=3584.950..3585.012 rows=1 loops=1)->  Gather  (cost=217018.52..217018.73 rows=2 width=8) (actual time=3584.897..3584.981 rows=3 loops=1)Workers Planned: 2Workers Launched: 2->  Partial Aggregate  (cost=216018.52..216018.53 rows=1 width=8) (actual time=3559.160..3559.160 rows=1 loops=3)->  Parallel Seq Scan on t1  (cost=0.00..205601.81 rows=4166681 width=0) (actual time=0.193..3373.351 rows=3333333 loops=3)Planning Time: 4.743 msExecution Time: 3586.400 ms
(8 rows)

缓存1G的数据，耗时3586.400 ms

（二）进行prewarm

分别测试read和buffer模式下的效果

1、重启数据库并消除OS缓存​​​​​​​

pg_ctl restartecho 3 > /proc/sys/vm/drop_caches

2、预热数据到OS缓存

testdb=# select pg_prewarm('t1', 'read', 'main');pg_prewarm 
------------163935
(1 row)

3、查看执行计划

testdb=# explain analyze select count(*) from t1;QUERY PLAN                                                                
------------------------------------------------------------------------------------------------------------------------------------------Finalize Aggregate  (cost=217018.73..217018.74 rows=1 width=8) (actual time=657.884..658.970 rows=1 loops=1)->  Gather  (cost=217018.52..217018.73 rows=2 width=8) (actual time=657.516..658.959 rows=3 loops=1)Workers Planned: 2Workers Launched: 2->  Partial Aggregate  (cost=216018.52..216018.53 rows=1 width=8) (actual time=652.264..652.265 rows=1 loops=3)->  Parallel Seq Scan on t1  (cost=0.00..205601.81 rows=4166681 width=0) (actual time=0.092..405.615 rows=3333333 loops=3)Planning Time: 0.126 msExecution Time: 658.997 ms
(8 rows)

​​​​​​​4、预热到数据库缓存中​​​​​​​

testdb=# select pg_prewarm('t1', 'buffer', 'main');pg_prewarm 
------------163935
(1 row)

5、查看执行计划

testdb=# explain analyze select count(*) from t1;QUERY PLAN                                                                
------------------------------------------------------------------------------------------------------------------------------------------Finalize Aggregate  (cost=217018.73..217018.74 rows=1 width=8) (actual time=681.629..683.325 rows=1 loops=1)->  Gather  (cost=217018.52..217018.73 rows=2 width=8) (actual time=681.485..683.319 rows=3 loops=1)Workers Planned: 2Workers Launched: 2->  Partial Aggregate  (cost=216018.52..216018.53 rows=1 width=8) (actual time=674.079..674.080 rows=1 loops=3)->  Parallel Seq Scan on t1  (cost=0.00..205601.81 rows=4166681 width=0) (actual time=0.025..445.632 rows=3333333 loops=3)Planning Time: 0.039 msExecution Time: 683.353 ms
(8 rows)

缓存1G的数据，耗时683.353 ms

总结：缓存到OS缓存中和PG缓存中，两者性能差异不大。但比不预热的情况下效果提升明显。

预热	参数	耗时
否	无	3586.400 ms
是	read	658.997 ms
是	buffer	683.353 ms

三、pgfincore( )

将数据库对象CACHE到OS层面的缓存

3.1 安装插件​​​​​​​

testdb=# CREATE EXTENSION pgfincore;
CREATE EXTENSION

3.2 查看对象缓存信息​​​​​​

testdb=# select * from pgfincore ('t1'); relpath       | segment | os_page_size | rel_os_pages | pages_mem | group_mem | os_pages_free | databit | pages_dirty | group_dirty 
--------------------+---------+--------------+--------------+-----------+-----------+---------------+---------+-------------+-------------base/16391/33102   |       0 |         4096 |       262144 |    262144 |         1 |       1191325 |         |           0 |           0base/16391/33102.1 |       1 |         4096 |        65726 |     65726 |         1 |       1191325 |         |           0 |           0
(2 rows)

参数说明

• relpath：文件位置及名称

• segment：文件段编号

• os_page_size：OS page或block大小

• rel_os_pages：对象占用系统缓存需要的页面个数

• pages_mem：对象已经占用缓存页面个数

• group_mem：在缓存中连续的页面组的个数

• os_pages_free：OS剩余的page数

• databit：加载信息的位图

pgsysconf与pgsysconf_pretty

查看当前OS块大小及使用情况

testdb=# select * from pgsysconf(); os_page_size | os_pages_free | os_total_pages 
--------------+---------------+----------------4096 |       1190139 |        1997572
(1 row)testdb=# select * from pgsysconf_pretty();os_page_size | os_pages_free | os_total_pages 
--------------+---------------+----------------4096 bytes   | 4649 MB       | 7803 MB
(1 row)

pgfadvise_willneed

将数据库对象缓存到OS CACHE

testdb=# select * from pgfadvise_willneed('t1'); relpath       | os_page_size | rel_os_pages | os_pages_free 
--------------------+--------------+--------------+---------------base/16391/33102   |         4096 |       262144 |       1190033base/16391/33102.1 |         4096 |        65726 |       1190033
(2 rows)

pgfadvise_dontneed

将数据库对象刷出OS CACHE

对当前对象设置dontneed标记。dontneed标记的意思就是当操作系统需要释放内存时优先释放标记为dontneed的pages。

testdb=#  select * from pgfadvise_dontneed('t1');relpath       | os_page_size | rel_os_pages | os_pages_free 
--------------------+--------------+--------------+---------------base/16391/33102   |         4096 |       262144 |       1452085base/16391/33102.1 |         4096 |        65726 |       1517801
(2 rows)testdb=# select * from pgfincore ('t1'); relpath       | segment | os_page_size | rel_os_pages | pages_mem | group_mem | os_pages_free | databit | pages_dirty | group_dirty 
--------------------+---------+--------------+--------------+-----------+-----------+---------------+---------+-------------+-------------base/16391/33102   |       0 |         4096 |       262144 |         0 |         0 |       1517805 |         |           0 |           0base/16391/33102.1 |       1 |         4096 |        65726 |         0 |         0 |       1517805 |         |           0 |           0
(2 rows)