MySQL派生表查询大数据量无结果问题分析与解决
一、问题发现
在客户现场的一次问题报告中发现某个带有派生表进行查询的时候,数据量少的时候有结果,但是数据量大的时候返回无记录。看下面例子:
1、准备表
CREATE TABLE `cmdb_item` (
`cm_item_id` varchar(350) NOT NULL,
`cm_model_id` varchar(350) NOT NULL,
PRIMARY KEY (`cm_item_id`,`cm_model_id`)
);
CREATE TABLE `sys_auth_role_cmdb` (
`id` varchar(60) NOT NULL,
`auth_type_id` varchar(60) DEFAULT NULL,
`cmdb_item_id` varchar(60) DEFAULT NULL,
`cmdb_model_id` varchar(60) DEFAULT NULL,
`role_id` varchar(60) DEFAULT NULL,
PRIMARY KEY (`id`)
);
2、一条数据场合
如下所示插入一条数据并且进行派生表查询,可以发现结果有一条数据,符合预期。
greatsql> SELECT b.* FROM( SELECT cmdb_item_id FROM sys_auth_role_cmdb where1=0 union SELECT cm_item_id FROM cmdb_item) a LEFT JOIN cmdb_item b ON a.cmdb_item_id = b.cm_item_id WHERE b.cm_model_id='PmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmq' and
cm_item_id='8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J';
+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| cm_item_id | cm_model_id |
+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| 8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J | PmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmq |
+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1 row in set (3.50 sec)
3、插入大量数据的场合
-- 首先设置tmp_table_size参数为合适的值
SET LOCAL tmp_table_size=2000000;
set sql_mode="oracle";
delimiter $$
CREATE or replace PROCEDURE p1() as
BEGIN
for i in 1 .. 5000 loop
insert into cmdb_item values('8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J'|| i,'PmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmq' || i);
end loop;
END$$
delimiter ;
--插入5000条数据
call p1();
--执行跟上面一样的查询,可以发现结果为空,此处为问题
greatsql> SELECT b.* FROM( SELECT cmdb_item_id FROM sys_auth_role_cmdb where1=0 union SELECT cm_item_id FROM cmdb_item) a LEFT JOIN cmdb_item b ON a.cmdb_item_id = b.cm_item_id WHERE b.cm_model_id='PmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmq' and cm_item_id='8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J';
Empty set (0.00 sec)
4、tmp_table_size值最小的场合
-- 设置tmp_table_size参数为最小值
SET LOCAL tmp_table_size=1024;
-- 执行跟上面一样的查询,可以发现结果跟只有一条数据时候一样,出现一条数据
+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| cm_item_id | cm_model_id |
+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| 8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J | PmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmq |
+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1 row in set (3.50 sec)
5、对比执行计划
查看上面几个场合的执行计划,结果如下:
数据量为1的场合执行计划:
greatsql> explain analyze SELECT b.* FROM( SELECT cmdb_item_id FROM sys_auth_role_cmdb where1=0 union SELECT cm_item_id FROM cmdb_item) a LEFT JOIN cmdb_item b ON a.cmdb_item_id = b.cm_item_id WHERE b.cm_model_id='PmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmq' and cm_item_id='8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J';
+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| EXPLAIN |
+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| -> Constant row from a (cost=1207.97..1207.97 rows=1) (actual time=1107.026..1107.027 rows=1 loops=1)
-> Union materialize with deduplication (cost=1207.62..1207.62 rows=4553) (actual time=1106.831..1106.831 rows=5001 loops=1)
-> Zero rows (Impossible WHERE) (cost=0.00..0.00 rows=0) (actual time=0.000..0.000 rows=0 loops=1)
-> Covering index scan on cmdb_item using PRIMARY (cost=752.32 rows=4553) (actual time=0.723..69.643 rows=5001 loops=1)
|
+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1 row in set (1.11 sec)
数据量为5001的场合执行计划,可以看到下面的执行计划跟上面是一样的,不同点在于最后一步Constant row from a
这段的rows=0,说明执行计划没问题,但是执行阶段有问题。
greatdb> explain analyze SELECT b.* FROM ( SELECT cmdb_item_id FROM sys_auth_role_cmdb WHERE 1=0 UNION SELECT cm_item_id FROM cmdb_item) a LEFT JOIN cmdb_item b ON a.cmdb_item_id = b.cm_item_id WHERE b.cm_model_id='PmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmqPmqECfANKxmbpgt3XBWaxzF6dx32dLmq' AND cm_item_id='8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J8b0zIlHsiarv8Ls6NejWji6Cu4h3iN2J';
+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| EXPLAIN |
+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| -> Constant row from a (cost=1053.17..1053.17 rows=1) (actual time=1118.948..1118.948 rows=0 loops=1)这个地方rows=0说明问题在这里
-> Union materialize with deduplication (cost=1052.82..1052.82 rows=4553) (actual time=1118.946..1118.946 rows=5001 loops=1)
-> Zero rows (Impossible WHERE) (cost=0.00..0.00 rows=0) (actual time=0.001..0.001 rows=0 loops=1)
-> Covering index scan on cmdb_item using PRIMARY (cost=597.52 rows=4553) (actual time=2.315..70.077 rows=5001 loops=1)
|
+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1 row in set (1.13 sec)
二、问题调查过程
查询带有派生表的时候内部会创建临时表用于保存临时数据,因此先看一下上面几个场合的临时表情况:
tmp_table_size值 | 数据量 | 查询开始时候临时表情况 | 查询过程中临时表情况 | 说明 |
---|---|---|---|---|
2000000 | 1条数据 | 内存表 | 内存表 | 没问题 |
2000000 | 5001条数据 | 内存表 | 转为落盘表 | 问题 |
1024 | 5001条数据 | 落盘表 | 落盘表 | 没问题 |
先看一下读取数据的代码:
int ConstIterator::Read() {
if (!m_first_record_since_init) {
return -1;
}
m_first_record_since_init = false;
// 这里打印err结果为-1,跳进read_const看一下为何没有数据返回
int err = read_const(table(), m_ref);
if (err == 0 && m_examined_rows != nullptr) {
++*m_examined_rows;
}
table()->const_table = true;
return err;
}
int read_const(TABLE *table, Index_lookup *ref) {
int error;
DBUG_TRACE;
// 数据为一条的时候会进入这个if线正常读数据,但是数据为5001条的时候却没进入这个if线,可见问题在table的这个状态。
// 这里要求table->m_status = STATUS_NOT_STARTED才会进入
if (!table->is_started()) // If first read
这里面读取表数据
} else if (table->has_row() && table->is_nullable()) {}
return table->has_row() ? 0 : -1;
}
考虑上面的现象,数据量少和数据量多的场合结果不一样,这里面的区别在于union在查询阶段会创建临时物化表用于储存union结果,数据量少的时候在内存就可以创建物化表出结果了,但是如果数据量大的话,因为内存不够放,因此需要重新创建落盘表用于储存查询结果。
又考虑到上面<<tmp_table_size值最小的场合>>这节,把内存表值改小结果又正常,因此可以推测是内存临时表改为落盘表后table的is_started()
状态值改变了导致结果为空。这里可以从创建落盘表入手。找到创建落盘表的函数create_ondisk_from_heap()
并且打个断点,看看堆栈:
#0 create_ondisk_from_heap (
thd=0x55555accb1a0 <native_rw_unlock(native_rw_lock_t*)+24>, wtable=0x7fffc83eddc0,
error=0, insert_last_record=false, ignore_last_dup=false, is_duplicate=0x0)
at sql/sql_tmp_table.cc:2777
#1 0x0000555559b98fb2 in MaterializeIterator<DummyIteratorProfiler>::MaterializeQueryBlock (this=0x7fff2023f390, query_block=..., stored_rows=0x7fffc83eef10)
at sql/iterators/composite_iterators.cc:1433
#2 0x0000555559b9649e in MaterializeIterator<DummyIteratorProfiler>::Init (
this=0x7fff2023f390)
at sql/iterators/composite_iterators.cc:975
#3 0x00005555597dae49 in Query_expression::ExecuteIteratorQuery (this=0x7fff2026d310,
thd=0x7fff20001050)
at sql/sql_union.cc:1814
#4 0x00005555597db287 in Query_expression::execute (this=0x7fff2026d310,
thd=0x7fff20001050)
at sql/sql_union.cc:1877
#5 0x0000555559700334 in Sql_cmd_dml::execute_inner (this=0x7fff20224010,
thd=0x7fff20001050)
at sql/sql_select.cc:872
#6 0x00005555596ff502 in Sql_cmd_dml::execute (this=0x7fff20224010, thd=0x7fff20001050)
at sql/sql_select.cc:612
#7 0x00005555596621f0 in mysql_execute_command (thd=0x7fff20001050, first_level=true)
at sql/sql_parse.cc:5227
#8 0x0000555559664838 in dispatch_sql_command (thd=0x7fff20001050,
parser_state=0x7fffc83f0760, update_userstat=false)
at sql/sql_parse.cc:5905
#9 0x0000555559658b03 in dispatch_command (thd=0x7fff20001050, com_data=0x7fffc83f1210,
command=COM_QUERY)
查看MaterializeIterator::MaterializeQueryBlock
函数代码,追踪一下table->is_started()
状态值:
bool MaterializeIterator<Profiler>::MaterializeQueryBlock(
error = t->file->ha_write_row(t->record[0]);
if (error == 0) {
++*stored_rows;
continue;
}
// create_ondisk_from_heap will generate error if needed.
if (!t->file->is_ignorable_error(error)) {
bool is_duplicate;
// 这里table->m_status = STATUS_NOT_STARTED | STATUS_NOT_FOUND,正常初始值,符合预期
if (create_ondisk_from_heap(thd(), t, error,
/*insert_last_record=*/true,
/*ignore_last_dup=*/true, &is_duplicate))
return true; /* purecov: inspected */
// 这里执行完create_ondisk_from_heap以后table->m_status = STATUS_NOT_FOUND,可见在create_ondisk_from_heap函数内部改变了这个m_status
}
查看create_ondisk_from_heap代码,看一下哪个地方改变了table->m_status值。
bool create_ondisk_from_heap(THD *thd, TABLE *wtable, int error,
bool insert_last_record, bool ignore_last_dup,
bool *is_duplicate) {
// 执行下面这句的时候因为调用了table->set_row_status_from_handler(result)改变了m_status值,先从3变为0最后变为2
// 状态迁移情况:STATUS_NOT_STARTED | STATUS_NOT_FOUND ==> 0 ==> STATUS_NOT_FOUND
while (!table->file->ha_rnd_next(new_table.record[1])) {
write_err = new_table.file->ha_write_row(new_table.record[1]);
DBUG_EXECUTE_IF("raise_error", write_err = HA_ERR_FOUND_DUPP_KEY;);
if (write_err) goto err_after_open;
}
}
发现果然是create_ondisk_from_heap
函数改变了table->is_started()
状态值,所以导致了结果为空。
三、问题解决
结合上面分析,我们可以把代码作如下修复,就可以解决这个问题了。
bool MaterializeIterator<Profiler>::MaterializeQueryBlock(
error = t->file->ha_write_row(t->record[0]);
if (error == 0) {
++*stored_rows;
continue;
}
// create_ondisk_from_heap will generate error if needed.
if (!t->file->is_ignorable_error(error)) {
bool is_duplicate;
if (create_ondisk_from_heap(thd(), t, error,
/*insert_last_record=*/true,
/*ignore_last_dup=*/true, &is_duplicate))
return true; /* purecov: inspected */
// 这里把table->is_started()状态复原即可,问题解决。
t->set_not_started();
}
接着执行上面的查询,发现可以查出结果了。
四、问题总结
通过以上分析我们可以发现,带有派生表的查询会在内部创建临时表用于储存中间数据,根据tmp_table_size
值设置会影响临时表存放地方,如果tmp_table_size
设置小了那么一开始就会创建落盘表,如果设置大了那么创建的就是内存表,随着数据的增加后面可能会转为落盘表,因此才有了本文章提及的bug。分析类似问题的时候,需要结合查询计划的执行和临时表创建情况分析,才能更好解决问题。