PostgreSQL中APP在涉及locks时需要注意的地方有哪些-创新互联

这篇文章主要讲解了“PostgreSQL中APP在涉及locks时需要注意的地方有哪些”，文中的讲解内容简单清晰，易于学习与理解，下面请大家跟着小编的思路慢慢深入，一起来研究和学习“PostgreSQL中APP在涉及locks时需要注意的地方有哪些”吧！

测试数据:

[local]:5432 pg12@testdb=# drop table if exists tbl;
DROP TABLE
Time: 36.136 ms
[local]:5432 pg12@testdb=# create table tbl(id int,c1 varchar(20),c2 varchar(20));
CREATE TABLE
Time: 4.903 ms
[local]:5432 pg12@testdb=# 
[local]:5432 pg12@testdb=# insert into tbl select x,'c1'||x,'c2'||x from generate_series(1,1000000) as x;
INSERT 0 1000000
Time: 3677.812 ms (00:03.678)
[local]:5432 pg12@testdb=#

— session 1

[local]:5432 pg12@testdb=# select pg_backend_pid();
 pg_backend_pid 
----------------
           1541
(1 row)

— session 2

[local]:5432 pg12@testdb=# select pg_backend_pid();
 pg_backend_pid 
----------------
           1628
(1 row)
Time: 4.446 ms

1: Never add a column with a default value
表上新增列时获取的锁是AccessExclusiveLock,会阻塞RW(包括SELECT),为了尽快完成列的添加,新增有默认值的列,可拆分为新增列,然后执行UPDATE语句以免出现R阻塞.

-- session 1
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 0.929 ms
[local]:5432 pg12@testdb=#* alter table tbl add column c3 varchar(20) default 'c3';
ALTER TABLE
Time: 32.881 ms
[local]:5432 pg12@testdb=# 
-- session 2
[local]:5432 pg12@testdb=# select * from tbl;
-- 阻塞
-- session 3
[local]:5432 pg12@testdb=# select pid,locktype,relation::regclass,mode,page,tuple,virtualxid,transactionid,virtualtransaction,granted,fastpath from pg_locks where relation='tbl'::regclass;
-[ RECORD 1 ]------+--------------------
pid                | 1541
locktype           | relation
relation           | tbl
mode               | AccessExclusiveLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 3/8
granted            | t
fastpath           | f
Time: 29.088 ms

使用先添加列,后更新默认值的方法

------ session 1
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 0.330 ms
[local]:5432 pg12@testdb=#* alter table tbl add column c4 varchar(20);
ALTER TABLE
Time: 0.460 ms
[local]:5432 pg12@testdb=#* end;
COMMIT
Time: 0.530 ms
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 0.199 ms
[local]:5432 pg12@testdb=#* update tbl set c4 = 'c4';
UPDATE 1000000
Time: 5286.769 ms (00:05.287)
[local]:5432 pg12@testdb=#* 
------ session 2
[local]:5432 pg12@testdb=# select * from tbl limit 1; 
 id | c1  | c2  | c3 | c4 
----+-----+-----+----+----
  1 | c11 | c21 | c3 | 
(1 row)
Time: 2.793 ms
------ session 3
[local]:5432 pg12@testdb=# select pid,locktype,relation::regclass,mode,page,tuple,virtualxid,transactionid,virtualtransaction,granted,fastpath from pg_locks where relation='tbl'::regclass;
-[ RECORD 1 ]------+-----------------
pid                | 1541
locktype           | relation
relation           | tbl
mode               | RowExclusiveLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 3/10
granted            | t
fastpath           | t
Time: 1.062 ms

虽然更新耗费的时间远比直接add column设置默认值要大,但锁等级是RowExclusiveLock,并不会阻塞读

2: Beware of lock queues, use lock timeouts
PG中每一个锁都有一个队列,在获取锁时如需等待存在冲突的其他锁,则会阻塞.可通过设置超时时间避免长时间的等待.这样虽然会失败,但可通过后台查询等方法获取数据库活动,保持数据库可控.

------ session 1
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 1.148 ms
[local]:5432 pg12@testdb=#* alter table tbl add column c5 varchar(20) default 'c3';
ALTER TABLE
Time: 2.726 ms
[local]:5432 pg12@testdb=#* 
------ session 3
[local]:5432 pg12@testdb=# select pid,locktype,relation::regclass,mode,page,tuple,virtualxid,transactionid,virtualtransaction,granted,fastpath from pg_locks where relation='tbl'::regclass;
-[ RECORD 1 ]------+--------------------
pid                | 1541
locktype           | relation
relation           | tbl
mode               | AccessExclusiveLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 3/11
granted            | t
fastpath           | f
Time: 2.751 ms
------ session 2
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 0.861 ms
[local]:5432 pg12@testdb=#* SET lock_timeout TO '1s';
SET
Time: 0.689 ms
[local]:5432 pg12@testdb=#* select * from tbl limit 1;
ERROR:  canceling statement due to lock timeout
LINE 1: select * from tbl limit 1;
                      ^
Time: 1001.031 ms (00:01.001)
[local]:5432 pg12@testdb=#! end;
ROLLBACK
Time: 0.984 ms
[local]:5432 pg12@testdb=#!

3: Create indexes CONCURRENTLY
使用CONCURRENTLY模式创建Index.
新插入1000w数据

[local]:5432 pg12@testdb=# insert into tbl select x,'c1'||x,'c2'||x,'c3'||x from generate_series(1,10000000) as x;
INSERT 0 10000000
Time: 32784.183 ms (00:32.784)

普通模式创建索引

------ session 1
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 29.276 ms
[local]:5432 pg12@testdb=#* create index idx_tbl_id on tbl(id);
CREATE INDEX
Time: 7261.828 ms (00:07.262)
[local]:5432 pg12@testdb=#* 
------ session 2
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 0.358 ms
[local]:5432 pg12@testdb=#* insert into tbl(id) values(0);
-- 阻塞
------ session 3
[local]:5432 pg12@testdb=# select pid,locktype,relation::regclass,mode,page,tuple,virtualxid,transactionid,virtualtransaction,granted,fastpath from pg_locks where relation='tbl'::regclass;
-[ RECORD 1 ]------+-----------------
pid                | 1628
locktype           | relation
relation           | tbl
mode               | RowExclusiveLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 5/13
granted            | f
fastpath           | f
-[ RECORD 2 ]------+-----------------
pid                | 1541
locktype           | relation
relation           | tbl
mode               | ShareLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 3/13
granted            | t
fastpath           | f
Time: 0.795 ms

回滚事务后,使用CONCURRENTLY模式创建索引,要注意的是CONCURRENTLY模式不能用在事务中

[local]:5432 pg12@testdb=#* -- only blocks other DDL
[local]:5432 pg12@testdb=#* create index CONCURRENTLY idx_tbl_id on tbl(id);
ERROR:  CREATE INDEX CONCURRENTLY cannot run inside a transaction block
Time: 0.491 ms
[local]:5432 pg12@testdb=#!

不启动事务,直接执行

------ session 1
[local]:5432 pg12@testdb=# -- only blocks other DDL
create index CONCURRENTLY idx_tbl_id on tbl(id);
CREATE INDEX
Time: 9718.400 ms (00:09.718)
------ session 2
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 0.373 ms
[local]:5432 pg12@testdb=#* insert into tbl(id) values(0);
INSERT 0 1
Time: 0.686 ms
[local]:5432 pg12@testdb=#* 
------ session 3
[local]:5432 pg12@testdb=# select pid,locktype,relation::regclass,mode,page,tuple,virtualxid,transactionid,virtualtransaction,granted,fastpath from pg_locks where relation='tbl'::regclass;
-[ RECORD 1 ]------+-------------------------
pid                | 1541
locktype           | relation
relation           | tbl
mode               | ShareUpdateExclusiveLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 3/21
granted            | t
fastpath           | f
-[ RECORD 2 ]------+-------------------------
pid                | 1701
locktype           | relation
relation           | tbl
mode               | ShareUpdateExclusiveLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 6/71
granted            | t
fastpath           | f
Time: 0.754 ms

使用CONCURRENTLY模式创建索引,获取的lock是ShareUpdateExclusiveLock,不会阻塞INSERT/UPDATE/DELETE操作(请求的锁是RowExclusiveLock)

4: Take aggressive locks as late as possible
这个跟编程中定义变量类似 : 离需要用到的地方越近的地方才定义.文中的例子见仁见智,选择使用.

5: Adding a primary key with minimal locking
重新构建测试数据

[local]:5432 pg12@testdb=# truncate table tbl;
TRUNCATE TABLE
Time: 91.815 ms
[local]:5432 pg12@testdb=# insert into tbl select x,'c1'||x,'c2'||x,'c3'||x from generate_series(1,12000000) as x;
INSERT 0 12000000
Time: 59285.694 ms (00:59.286)

把add primary key这一个动作拆解为先添加唯一索引,再添加primary key constraint这两个动作.

------ session 1
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 1.155 ms
[local]:5432 pg12@testdb=#* alter table tbl add primary key(id);
ALTER TABLE
Time: 10572.201 ms (00:10.572)
[local]:5432 pg12@testdb=#* 
------ session 2
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 0.703 ms
[local]:5432 pg12@testdb=#* insert into tbl(id) values(0);
-- 阻塞
------ session 3
[local]:5432 pg12@testdb=# select pid,locktype,relation::regclass,mode,page,tuple,virtualxid,transactionid,virtualtransaction,granted,fastpath from pg_locks where relation='tbl'::regclass;
-[ RECORD 1 ]------+--------------------
pid                | 1628
locktype           | relation
relation           | tbl
mode               | RowExclusiveLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 5/18
granted            | f
fastpath           | f
-[ RECORD 2 ]------+--------------------
pid                | 1541
locktype           | relation
relation           | tbl
mode               | ShareLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 3/28
granted            | t
fastpath           | f
-[ RECORD 3 ]------+--------------------
pid                | 1541
locktype           | relation
relation           | tbl
mode               | AccessExclusiveLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 3/28
granted            | t
fastpath           | f
-[ RECORD 4 ]------+--------------------
pid                | 1907
locktype           | relation
relation           | tbl
mode               | ShareLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 6/127
granted            | t
fastpath           | f
Time: 1.397 ms

拆解后,使用CONCURRENTLY模式创建索引,与第3点类似

------ session 1
[local]:5432 pg12@testdb=# -- takes a long time, but doesn’t block queries
[local]:5432 pg12@testdb=# CREATE UNIQUE INDEX CONCURRENTLY idx_tbl_id ON tbl (id); 
CREATE INDEX
Time: 9908.405 ms (00:09.908)
[local]:5432 pg12@testdb=# -- blocks queries, but only very briefly
[local]:5432 pg12@testdb=# ALTER TABLE tbl ADD CONSTRAINT pk_tbl PRIMARY KEY USING INDEX idx_tbl_id;  
NOTICE:  ALTER TABLE / ADD CONSTRAINT USING INDEX will rename index "idx_tbl_id" to "pk_tbl"
ALTER TABLE
Time: 4582.013 ms (00:04.582)

6: Never VACUUM FULL

------ session 1
[local]:5432 pg12@testdb=# vacuum full;
------ session 2
------ session 3
[local]:5432 pg12@testdb=# select pid,locktype,relation::regclass,mode,page,tuple,virtualxid,transactionid,virtualtransaction,granted,fastpath from pg_locks where relation='tbl'::regclass;
-[ RECORD 1 ]------+--------------------
pid                | 1541
locktype           | relation
relation           | tbl
mode               | AccessExclusiveLock
page               | 
tuple              | 
virtualxid         | 
transactionid      | 
virtualtransaction | 3/49
granted            | t
fastpath           | f
Time: 0.803 ms

vacuum full请求的锁是AccessExclusiveLock,会阻塞读写,在目前vacuum full并不智能的情况下,手工发起对单个表的vacuum full会保险许多.

7: Avoid deadlocks by ordering commands
注意命令的顺序,避免死锁

------ session 1
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 0.440 ms
[local]:5432 pg12@testdb=#* delete from tbl where id = 1;
DELETE 1
Time: 0.567 ms
[local]:5432 pg12@testdb=#* 
------ session 2
[local]:5432 pg12@testdb=# begin;
BEGIN
Time: 0.960 ms
[local]:5432 pg12@testdb=#* delete from tbl where id = 2;
DELETE 1
Time: 1.783 ms
[local]:5432 pg12@testdb=#* 
------ session 3

产生死锁

------ session 1
[local]:5432 pg12@testdb=#* delete from tbl where id = 2;
------ session 2
[local]:5432 pg12@testdb=#* delete from tbl where id = 1;
ERROR:  deadlock detected
DETAIL:  Process 1628 waits for ShareLock on transaction 623; blocked by process 1541.
Process 1541 waits for ShareLock on transaction 624; blocked by process 1628.
HINT:  See server log for query details.
CONTEXT:  while deleting tuple (0,1) in relation "tbl"
Time: 1004.485 ms (00:01.004)
[local]:5432 pg12@testdb=#! 
------ session 3

感谢各位的阅读，以上就是“PostgreSQL中APP在涉及locks时需要注意的地方有哪些”的内容了，经过本文的学习后，相信大家对PostgreSQL中APP在涉及locks时需要注意的地方有哪些这一问题有了更深刻的体会，具体使用情况还需要大家实践验证。这里是创新互联，小编将为大家推送更多相关知识点的文章，欢迎关注！