对于PostgreSQL的 "create index concurrently". 我个人认为其中存在一个bug。
我的验证过程如下:
我有两个表,tab01和 tab02,这两个表之间没有任何关联。
我认为 对 tab02执行 "create index concurrently" 不会对 访问tab01的事务有任何影响,然而事实并非尽然。
我第一程序的表现: 通过ecpg执行事务,再通过 "create index concurrently" 给tab02建立索引,成功。
我第二程序的表现:通过ecpg执行事务,再通过 "create index concurrently" 给tab02建立索引,被阻塞。
我第三个测试: 通过psql发起事务, 另一个psql客户端执行 "create index concurrently" 成功。
我第四个测试: 通过psql发起事务 另一个psql客户端执行 "create index concurrently",被阻塞。
无论 PostgreSQL9.1.2,还是PostgreSQL9.2.4,结果是一样的。
数据准备:
[postgres@server bin]$ ./psql -U tester -d tester psql (9.1.2) Type "help" for help. tester=> \d tab01; Table "public.tab01" Column | Type | Modifiers --------+----------------------+----------- id | integer | cd | character varying(4) | tester=> \d tab02; Table "public.tab02" Column | Type | Modifiers --------+---------+----------- id | integer | value | integer | tester=> select * from tab01; id | cd ----+---- 1 | 14 2 | 15 3 | 14 (3 rows) tester=> select * from tab02; id | value ----+------- 1 | 100 2 | 200 3 | 300 (3 rows) tester=>
我的测试方法:
对第一个程序和第二个程序:
当我的eccp程序正在睡眠的时候,我另外开一个终端,执行:
"create index concurrently idx_tab02_id_new on tab02(id)"
结果是:
第一个程序执行中,我可成功建立索引。
第二个程序执行中,我无法建立索引,会被阻塞
而我的tab01和tab02之间,没有任何关联。而且我也不认为我的ecpg程序会有潜在的可能去使用tab02的索引。
事实上,如果我去看ecpg预编译后得到的c程序,我可以看到:
{ ECPGdo(__LINE__, 0, 1, "db_conn", 0, ECPGst_normal, "select count ( * ) from tab01 where cd = $1 ", ECPGt_char,(vcd),(long)4 + 1,(long)1,(4 + 1)*sizeof(char), ECPGt_NO_INDICATOR, NULL , 0L, 0L, 0L, ECPGt_EOIT, ECPGt_int,&(vCount),(long)1,(long)1,sizeof(int), ECPGt_NO_INDICATOR, NULL , 0L, 0L, 0L, ECPGt_EORT);}
当我给$1加入引号后,我就可以成功地建立索引了。
{ ECPGdo(__LINE__, 0, 1, "db_conn", 0, ECPGst_normal, "select count ( * ) from tab01 where cd = '$1' ", ECPGt_char,(vcd),(long)4 + 1,(long)1,(4 + 1)*sizeof(char), ECPGt_NO_INDICATOR, NULL , 0L, 0L, 0L, ECPGt_EOIT, ECPGt_int,&(vCount),(long)1,(long)1,sizeof(int), ECPGt_NO_INDICATOR, NULL , 0L, 0L, 0L, ECPGt_EORT);}
下面是我测试程序的源代码:
第一个:
[root@server soft]# cat ./test01/test01.pc int main() { EXEC SQL BEGIN DECLARE SECTION; int vCount; char vcd[4+1]; EXEC SQL END DECLARE SECTION;
EXEC SQL CONNECT TO '[email protected]:5432' AS db_conn USER tester IDENTIFIED BY tester; EXEC SQL AT db_conn SELECT COUNT(*) INTO :vCount FROM tab01; fprintf(stderr,"count is:%d\n",vCount); fprintf(stderr,"Before disconnect,sleep for 500 seconds\n"); sleep(500); EXEC SQL DISCONNECT db_conn; fprintf(stderr,"After disconnect,sleep for 600 seconds\n"); sleep(600); return 0; } [root@server soft]#
第二个:
[root@server soft]# cat ./test02/test02.pc int main() { EXEC SQL BEGIN DECLARE SECTION; int vCount; char vcd[4+1]; EXEC SQL END DECLARE SECTION; EXEC SQL CONNECT TO '[email protected]:5432' AS db_conn USER tester IDENTIFIED BY tester; char *pCd="14"; memset(vcd,'\0',5); strncpy(vcd, pCd,4); EXEC SQL AT db_conn SELECT COUNT(*) INTO :vCount FROM tab01 WHERE cd = :vcd; fprintf(stderr,"count is:%d\n",vCount);
fprintf(stderr,"Before disconnect,sleep for 500 seconds\n"); sleep(500); EXEC SQL DISCONNECT db_conn; fprintf(stderr,"After disconnect,sleep for 600 seconds\n"); sleep(600); return 0; } [root@server soft]#
而且,通过 psql,还可以发现一个与 create index concurrently 相关的现象:
我的第三个测试:
客户端1:
[postgres@server pgsql]$ ./bin/psql -d tester -U tester psql (9.1.2) Type "help" for help. tester=> begin; BEGIN tester=> select * from tab01 where cd = '14'; id | cd ----+---- 1 | 14 3 | 14 (2 rows) tester=>
客户端2:
[postgres@server pgsql]$ ./bin/psql -d tester -U tester psql (9.1.2) Type "help" for help. tester=> create index concurrently idx_tab02_id_new on tab02(id);
可以很快就成功创建索引。
我的第四个测试:
客户端1:
[postgres@server pgsql]$ ./bin/psql -d tester -U tester psql (9.1.2) Type "help" for help. tester=> begin; BEGIN tester=> select * from tab01 where cd = '14'; id | cd ----+---- 1 | 14 3 | 14 (2 rows) tester=> select pg_sleep(500); pg_sleep ---------- (1 row) tester=>
客户端2:
[postgres@server pgsql]$ ./bin/psql -d tester -U tester psql (9.1.2) Type "help" for help. tester=> create index concurrently idx_tab02_id_new on tab02(id);
客户端2的创建索引会被阻塞
根据我对PostgreSQL的源代码的跟踪,可以看到有如下的调用关系:
PortalRunMulti--> PortalRunUtility-->Standard_ProcessUtility-->DefineIndex
而我对DefineIndex作简化后,可以看到:
{
… old_snapshots = GetCurrentVirtualXIDs(snapshot->xmin, true, false, PROC_IS_AUTOVACUUM | PROC_IN_VACUUM, &n_old_snapshots); for (i = 0; i < n_old_snapshots; i++) { … if (VirtualTransactionIdIsValid(old_snapshots[i])) VirtualXactLockTableWait(old_snapshots[i]); } … }
对于我的第一个测试程序,GetCurrentVirtualXIDs 函数执行后,n_old_snapshots 的值为0 ,
for (i = 0; i < n_old_snapshots; i++) 循环不会被执行,索引的生成不会被阻塞。
对我的第二个测试程序,GetCurrentVirtualXIDs 函数执行后,n_old_snapshots 的值为1,
for (i = 0; i < n_old_snapshots; i++) 循环会被执行。
VirtualXactLockTableWait(old_snapshots[i]) 的执行,导致等待一个锁,所以索引生成被阻塞。
再往下分析:
VirtualTransactionId * GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0, bool allDbs, int excludeVacuum, int *nvxids) { VirtualTransactionId *vxids; ProcArrayStruct *arrayP = procArray; int count = 0; int index; /* allocate what's certainly enough result space */ vxids = (VirtualTransactionId *) palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs); LWLockAcquire(ProcArrayLock, LW_SHARED); for (index = 0; index < arrayP->numProcs; index++) { volatile PGPROC *proc = arrayP->procs[index]; if (proc == MyProc) continue; if (excludeVacuum & proc->vacuumFlags) continue; if (allDbs || proc->databaseId == MyDatabaseId) { /* Fetch xmin just once - might change on us */ TransactionId pxmin = proc->xmin;
if (excludeXmin0 && !TransactionIdIsValid(pxmin)) continue; /* * InvalidTransactionId precedes all other XIDs, so a proc that * hasn't set xmin yet will not be rejected by this test. */ if (!TransactionIdIsValid(limitXmin) || TransactionIdPrecedesOrEquals(pxmin, limitXmin)) { VirtualTransactionId vxid; GET_VXID_FROM_PGPROC(vxid, *proc);
if (VirtualTransactionIdIsValid(vxid)) vxids[count++] = vxid; } } } LWLockRelease(ProcArrayLock); *nvxids = count; return vxids; }
对于我的第一个程序,测试结果显示:pxmin 为零,TransactionIdIsValid(pxmin) 为假。所以如下代码导致跳过循环一次。
if (excludeXmin0 && !TransactionIdIsValid(pxmin)) continue;
没有机会执行 vxids[count++]=vxid 这一行。
那么pxmin是如何来的?
看这句: TransactionId pxmin = proc->xmin;
而xmin的含义是:当我们执行程序中对数据进行增删改的时候,会将当前transaction id 赋予给 xmin。
写记录的时候,把这个xmin写入该行记录头。
如此,每个进程看来,它只关心 xmin 小于自己的transaction id的。PostgreSQL用这种方式来保证MVCC。
但此处,proc->xmin为零是很不合理的。
此时,
if (allDbs || proc->databaseId == MyDatabaseId) 里的:
if (excludeXmin0 && !TransactionIdIsValid(pxmin))就会成立,所以会直接continue调回循环开始处,也就没有机会去
vxids[count++] = vxid;
在我的第二个程序里,proc->xmin根本就不为零。故此说,这是一个bug。
另外的佐证:对我的三个测试,运行下列SQL文:
pgsql=# select l.pid, l.mode, sa.procpid, sa.current_query from pg_locks l inner join pg_stat_activity sa on l.pid = sa.procpid where l.mode like '%xclusive%';
一开始在pg_sleep(100)执行期间,可以看到:
pgsql=# select l.pid, l.mode, sa.procpid, sa.current_query from pg_locks l inner join pg_stat_activity sa on l.pid = sa.procpid where l.mode like '%xclusive%'; pid | mode | procpid | current_query ------+---------------+---------+---------------------------------------------------- 5356 | ExclusiveLock | 5356 | select l.pid, l.mode, sa.procpid, sa.current_query+ | | | from pg_locks l + | | | inner join pg_stat_activity sa + | | | on l.pid = sa.procpid + | | | where l.mode like '%xclusive%'; 5517 | ExclusiveLock | 5517 | select pg_sleep(100); (2 rows)
我开另外的终端,执行 "create index concurrently"的时候,再看:
pgsql=# select l.pid, l.mode, sa.procpid, sa.current_query from pg_locks l inner join pg_stat_activity sa on l.pid = sa.procpid where l.mode like '%xclusive%'; pid | mode | procpid | current_query ------+--------------------------+---------+---------------------------------------------------------- 5356 | ExclusiveLock | 5356 | select l.pid, l.mode, sa.procpid, sa.current_query + | | | from pg_locks l + | | | inner join pg_stat_activity sa + | | | on l.pid = sa.procpid + | | | where l.mode like '%xclusive%'; 5517 | ExclusiveLock | 5517 | select pg_sleep(100); 5527 | ExclusiveLock | 5527 | create index concurrently idx_tab02_id_new on tab02(id); 5527 | RowExclusiveLock | 5527 | create index concurrently idx_tab02_id_new on tab02(id); 5527 | ShareUpdateExclusiveLock | 5527 | create index concurrently idx_tab02_id_new on tab02(id); (5 rows)
等到 pg_sleep执行完毕的时候:
pgsql=# select l.pid, l.mode, sa.procpid, sa.current_query from pg_locks l inner join pg_stat_activity sa on l.pid = sa.procpid where l.mode like '%xclusive%'; pid | mode | procpid | current_query ------+--------------------------+---------+---------------------------------------------------------- 5356 | ExclusiveLock | 5356 | select l.pid, l.mode, sa.procpid, sa.current_query + | | | from pg_locks l + | | | inner join pg_stat_activity sa + | | | on l.pid = sa.procpid + | | | where l.mode like '%xclusive%'; 5517 | ExclusiveLock | 5517 | <IDLE> in transaction 5527 | ExclusiveLock | 5527 | create index concurrently idx_tab02_id_new on tab02(id); 5527 | RowExclusiveLock | 5527 | create index concurrently idx_tab02_id_new on tab02(id); 5527 | ShareUpdateExclusiveLock | 5527 | create index concurrently idx_tab02_id_new on tab02(id); (5 rows)