本节介绍了APP在涉及locks时需要注意的地方. 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 ``` **参考资料** [When Postgres blocks: 7 tips for dealing with locks](https://www.citusdata.com/blog/2018/02/22/seven-tips-for-dealing-with-postgres-locks/)