为什么PG会提示增加max_locks_per_transaction的值

这篇文章主要介绍“为什么PG会提示增加max_locks_per_transaction的值”，在日常操作中，相信很多人在为什么PG会提示增加max_locks_per_transaction的值问题上存在疑惑，小编查阅了各式资料，整理出简单好用的操作方法，希望对大家解答”为什么PG会提示增加max_locks_per_transaction的值”的疑惑有所帮助！接下来，请跟着小编一起来学习吧！

有时候我们可能会在PG的日志发现如下信息：

2020-01-09 16:29:19.062 CST,"pg12","testdb",6193,"[local]",5e16dccd.1831,1,"CREATE TABLE",2020-01-09 15:57:01 CST,2/34,1512004206,ERROR,53200,"out of shared memory",,"You might need to increase max_locks_per_transaction.",,,,"CREATE TABLE a13030 (id int);",,,"psql"
2020-01-09 16:29:19.379 CST,"pg12","testdb",6193,"[local]",5e16dccd.1831,2,"CREATE TABLE",2020-01-09 15:57:01 CST,2/0,1512004206,ERROR,25P02,"current transaction is aborted, commands ignored until end of transaction block",,,,,,"CREATE TABLE a13031 (id int);",,,"psql"

直观上来看，OOM似乎与max_locks_per_transaction扯不上什么关系，为什么PG会提示增加max_locks_per_transaction的值呢？在一个事务中，shared lock table最大可以跟踪max_locks_per_transaction * (max_connections + max_prepared_transactions) 个对象（如数据表），超过的会报OOM错误。注意：锁粒度是object(如relation等)，跟行数无关。

OOM场景模拟
下面是一个模拟场景，在同一个事务中创建1w张表：

\pset footer off
\o /tmp/drop.sql
SELECT 'drop table if exists tbl' || id || ' ;' as "--"
       FROM generate_series(1, 20000) AS id;
\i /tmp/drop.sql
\pset footer off
\pset tuples_only
\o /tmp/create.sql
SELECT 'CREATE TABLE tbl' || id || ' (id int);' as "--"
       FROM generate_series(1, 20000) AS id;
\o /tmp/ret.txt
begin;
\i /tmp/create.sql

使用watch监控输出

watch -n1 "psql -c \"select locktype,mode,count(*) from pg_locks group by locktype,mode;\""
Every 1.0s: psql -c "select locktype,mode,count(*) from pg_locks group by locktype,mode;"  Fri Jan 10 14:41:26 2020
Expanded display is used automatically.
   locktype    |        mode         | count
---------------+---------------------+-------
 object        | AccessShareLock     |     1
 relation      | AccessShareLock     |     1
 virtualxid    | ExclusiveLock       |     2
 relation      | AccessExclusiveLock |  3776
 transactionid | ExclusiveLock       |     1
(5 rows)
...
Every 1.0s: psql -c "select locktype,mode,count(*) from pg_locks group by locktype,mode;"  Fri Jan 10 14:41:50 2020
Expanded display is used automatically.
   locktype    |        mode         | count
---------------+---------------------+-------
 object        | AccessShareLock     |     1
 relation      | AccessShareLock     |     1
 virtualxid    | ExclusiveLock       |     2
 relation      | AccessExclusiveLock | 10000
 transactionid | ExclusiveLock       |     1
(5 rows)
...

在执行到tbl13034时报错

2020-01-10 14:44:18.855 CST,"pg12","testdb",32120,"[local]",5e181bea.7d78,3,"CREATE TABLE",2020-01-10 14:38:34 CST,2/106085,1512036258,ERROR,53200,"out of shared memory",,"You might need to increase max_locks_per_transaction.",,,,"CREATE TABLE tbl13034 (id int);",,,"psql"
2020-01-10 14:44:19.202 CST,"pg12","testdb",32120,"[local]",5e181bea.7d78,4,"CREATE TABLE",2020-01-10 14:38:34 CST,2/0,1512036258,ERROR,25P02,"current transaction is aborted, commands ignored until end of transaction block",,,,,,"CREATE TABLE tbl13035 (id int);",,,"psql"

相关源码
搜索You might need to increase max_locks_per_transaction.该错误信息出现在lock.c中

 /*
  * LockAcquireExtended - allows us to specify additional options
  *
  * reportMemoryError specifies whether a lock request that fills the lock
  * table should generate an ERROR or not.  Passing "false" allows the caller
  * to attempt to recover from lock-table-full situations, perhaps by forcibly
  * cancelling other lock holders and then retrying.  Note, however, that the
  * return code for that is LOCKACQUIRE_NOT_AVAIL, so that it's unsafe to use
  * in combination with dontWait = true, as the cause of failure couldn't be
  * distinguished.
  *
  * If locallockp isn't NULL, *locallockp receives a pointer to the LOCALLOCK
  * table entry if a lock is successfully acquired, or NULL if not.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
                     LOCKMODE lockmode,
                     bool sessionLock,
                     bool dontWait,
                     bool reportMemoryError,
                     LOCALLOCK **locallockp)
 {
     ...
     /*
      * If this lock could potentially have been taken via the fast-path by
      * some other backend, we must (temporarily) disable further use of the
      * fast-path for this lock tag, and migrate any locks already taken via
      * this method to the main lock table.
      */
     if (ConflictsWithRelationFastPath(locktag, lockmode))
     {
         uint32      fasthashcode = FastPathStrongLockHashPartition(hashcode);
         BeginStrongLockAcquire(locallock, fasthashcode);
         if (!FastPathTransferRelationLocks(lockMethodTable, locktag,
                                            hashcode))
         {
             AbortStrongLockAcquire();
             if (locallock->nLocks == 0)
                 RemoveLocalLock(locallock);
             if (locallockp)
                 *locallockp = NULL;
             if (reportMemoryError)
                 ereport(ERROR,
                         (errcode(ERRCODE_OUT_OF_MEMORY),
                          errmsg("out of shared memory"),
                          errhint("You might need to increase max_locks_per_transaction.")));
             else
                 return LOCKACQUIRE_NOT_AVAIL;
         }
     }
     ...

到此，关于“为什么PG会提示增加max_locks_per_transaction的值”的学习就结束了，希望能够解决大家的疑惑。理论与实践的搭配能更好的帮助大家学习，快去试试吧！若想继续学习更多相关知识，请继续关注亿速云网站，小编会继续努力为大家带来更多实用的文章！