devbox
爱喝兽奶帝天荒
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

SQL优化选对执行计划，查询速度提升1000倍 | OceanBase 应用实践_oceanbase复制sql查询效率

作者：爱喝兽奶帝天荒 | 2024-07-18 14:11:44

赞

踩

oceanbase复制sql查询效率
作者：爱可生数据库高级工程师任仲禹，擅长故障分析和性能优化。

本文通过一个案例，分享使用OceanBase时，SQL走错执行计划，而导致慢SQL的排查方法论。

案例背景

在使用OceanBase 3.2.3 版本的过程中，项目组反映某个 SELECT 语句在指定时间内的查询响应速度异常缓慢，其耗时远超正常情况的1000倍以上。具体细节如下：

  • 慢 SELECT
SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31' ;
  • 关键表结构、记录数信息如下
  1. -- 脱敏处理
  2. show create table renzy\G                  
  3. *************************** 1. row ***************************
  4.        Table: renzy
  5. Create Table: CREATE TABLE `renzy` (
  6.   `ID` char(18) COLLATE utf8mb4_bin NOT NULL COMMENT ,
  7. ...
  8.   `ACCT_NO` char(40) COLLATE utf8mb4_bin NOT NULL COMMENT ,
  9. ...
  10.   `ACCTG_DT` date DEFAULT NULL COMMENT ,
  11. ...
  12.   PRIMARY KEY (`ID`),
  13. ...
  14.   KEY `renzy_I2` (`ACCT_NO`) BLOCK_SIZE 16384 LOCAL,
  15. ...
  16.   KEY `renzy_I5` (`ACCTG_DT`, `ENQ_INST_CD`, `BLON_INST_CD`, `EMRG_STPY_SRC_CD`) BLOCK_SIZE 16384 LOCAL,
  17. ...
  18. ) DEFAULT CHARSET = utf8mb4;
  19. 1 row in set (0.01 sec)
  20.  
  21. MySQL > SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882');
  22. +---------+
  23. | TOT_CNT |
  24. +---------+
  25. |       1 |
  26. +---------+
  27. 1 row in set (0.02 sec)
  28.  
  29. MySQL > SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31';
  30. +----------+
  31. | TOT_CNT  |
  32. +----------+
  33. | 25432155 |
  34. +----------+
  35. 1 row in set (12.42 sec)
  36.  
  37. MySQL > SELECT COUNT(*) AS TOT_CNT FROM renzy;                                                            
  38. +----------+
  39. | TOT_CNT  |
  40. +----------+
  41. | 25435024 |
  42. +----------+
  43. 1 row in set (10.65 sec)

排查过程

正常执行不慢

  1. MySQL > select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31') as orginal limit 2000; select last_trace_id();
  2. +---------+
  3. | TOT_CNT |
  4. +---------+
  5. |       1 |
  6. +---------+
  7. 1 row in set (0.02 sec)

以下是执行计划，从中可见，索引I2是最高效的选择，它在进行等值匹配时仅需要执行一次回表操作。

  1. *************************** 1. row ***************************
  2. Query Plan: ========================================================
  3. |ID|OPERATOR        |NAME               |EST. ROWS|COST|
  4. --------------------------------------------------------
  5. |0 |LIMIT           |                   |1        |92  |
  6. |1 | SCALAR GROUP BY|                   |1        |92  |
  7. |2 |  TABLE SCAN    |renzy(renzy_I2)|1        |92  |
  8. ========================================================
  9. ...
  10. Outline Data:
  11. -------------------------------------
  12.   /*+
  13.       BEGIN_OUTLINE_DATA
  14.       INDEX(@"SEL$2" "gabsdb.renzy"@"SEL$2" "renzy_I2")
  15.       END_OUTLINE_DATA
  16.   */
  17. ...
  18. renzy:table_rows:25419080, physical_range_rows:1, logical_range_rows:1, index_back_rows:1, output_rows:0, est_method:local_storage, optimization_method=cost_based, avaiable_index_name[renzy_I2,renzy_I5], pruned_index_name[renzy_I1,renzy_I3,renzy_I4,renzy_I6], unstable_index_name[renzy], estimation info[table_id:1105009185965290, (table_type:1, version:0-1699898410195654-1699898410195654, logical_rc:1, physical_rc:1), (table_type:7, version:1699898401860480-1699898401860480-1699898433101378, logical_rc:0, physical_rc:0), (table_type:7, version:1699898433101378-1699904137032515-1699905915658079, logical_rc:0, physical_rc:0), (table_type:5, version:1699898433101378-1699904137032515-1699905915658079, logical_rc:0, physical_rc:0), (table_type:0, version:1699905915658079-1699905915658079-9223372036854775807, logical_rc:0, physical_rc:0)]
  19. ...

通过 OCP 的 SQL 诊断获取慢 SQL 的 plan_id，检查慢 SQL 实际命中的 plan。

  1. MySQL [oceanbase]> select * from gv$plan_cache_plan_stat where plan_id=7288229 \G                                                                         
  2. *************************** 1. row ***************************
  3. ...
  4.               plan_id: 7288229
  5. ...
  6.             statement: select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = ?) AND LAWENF_NTIST_TP_CD NOT LIKE ? AND LAWENF_NTIST_TP_CD NOT LIKE ? AND EMRG_STPY_SRC_CD != ? AND ACCTG_DT >= ? AND ACCTG_DT <= ?) as orginal limit 2000
  7.             query_sql: select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '' AND ACCTG_DT <= '') as orginal limit 2000
  8.        special_params: 2000
  9.           param_infos: {1,0,0,-1,22},{1,0,0,-1,22},{1,0,0,-1,22},{1,0,0,-1,22},{1,0,0,-1,22},{1,0,0,-1,22},{1,0,0,-1,22},{1,0,0,-1,22},{1,0,0,-1,22},{1,0,0,-1,22},{1,0,0,-1,17},{1,0,0,-1,17}
  10.              sys_vars: 45,45,12582912,2,4,1,0,0,32,3,1,0,1,1,0,10485760,1,1,0,1,BINARY,BINARY,AL32UTF8,AL16UTF16,BYTE,FALSE,1,100,64,200,0,13,NULL,1,1,1,1
  11.             plan_hash: 10428103352368081688
  12.       first_load_time: 2023-11-14 10:14:11.578250
  13.        schema_version: 1699927892190832
  14.        merged_version: 287
  15.      last_active_time: 2023-11-14 11:04:58.127020
  16.          avg_exe_usec: 35858760
  17.      slowest_exe_time: 2023-11-14 11:04:58.127020
  18.      slowest_exe_usec: 171575101
  19.            slow_count: 2
  20.             hit_count: 7
  21.             plan_size: 81984
  22.            executions: 8
  23.            disk_reads: 1136285
  24.         direct_writes: 0
  25.           buffer_gets: 18067948
  26. application_wait_time: 0
  27. concurrency_wait_time: 0
  28.     user_io_wait_time: 0
  29.        rows_processed: 8
  30.          elapsed_time: 286870087
  31.              cpu_time: 229807460
  32.          large_querys: 2
  33.  delayed_large_querys: 1
  34.     delayed_px_querys: 0
  35.       outline_version: 0
  36.            outline_id: -1
  37.          outline_data: /*+ BEGIN_OUTLINE_DATA INDEX(@"SEL$2" "gabsdb.renzy"@"SEL$2" "renzy_I5") END_OUTLINE_DATA*/
  38. ....
  39. 1 row in set (0.09 sec)
  40.  
  41. MySQL [oceanbase]> select * from oceanbase.gv$plan_cache_plan_explain where tenant_id=1005 and port=2882 and plan_id=7288229 and ip='12.240.26.70'\G
  42. ....
  43. PLAN_LINE_ID: 2
  44.     OPERATOR:   PHY_TABLE_SCAN
  45.         NAME: renzy(renzy_I5)
  46.         ROWS: 0
  47.         COST: 91
  48.     PROPERTY: table_rows:25419080, physical_range_rows:1, logical_range_rows:1, index_back_rows:0, output_rows:0, est_method:local_storage, avaiable_index_name[renzy_I2,renzy_I5]
  49. ...

上述结果的关键信息如下

1.query_sql ：为该plan第一次执行时的SQL语句。

2.first_load_time ：缓存该plan并hit的时间。

3.slowest_exe_usec ：该计划的最慢耗时。

4.outline_id ： 是否命中了绑定的outline，-1即未命中。

5.statement ：参数化后的SQL语句。

6.name : 该plan走的索引。

分析下第一次的SQL为啥要走 I5 索引

通过下面执行计划和执行耗时可知，第一次执行的语句因为字段 ACCTG_DT 检索不到数据，所以走 I5 效率最高。

  1. MySQL > explain extended select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '' AND ACCTG_DT <= '') as orginal limit 2000\G
  2. *************************** 1. row ***************************
  3. Query Plan: ========================================================
  4. |ID|OPERATOR        |NAME               |EST. ROWS|COST|
  5. --------------------------------------------------------
  6. |0 |LIMIT           |                   |1        |92  |
  7. |1 | SCALAR GROUP BY|                   |1        |92  |
  8. |2 |  TABLE SCAN    |renzy(renzy_I5)|0        |92  |
  9. ========================================================
  10. Outline Data:
  11. -------------------------------------
  12.   /*+
  13.       BEGIN_OUTLINE_DATA
  14.       INDEX(@"SEL$2" "gabsdb.renzy"@"SEL$2" "renzy_I5")
  15.       END_OUTLINE_DATA
  16.   */
  17. renzy:table_rows:25419080, physical_range_rows:1, logical_range_rows:1, index_back_rows:0, output_rows:0, est_method:local_storage, optimization_method=cost_based, avaiable_index_name[renzy_I2,renzy_I5], pruned_index_name[renzy_I1,renzy_I3,renzy_I4,renzy_I6], unstable_index_name[renzy]
  18.  
  19. MySQL > select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '' AND ACCTG_DT <= '') as orginal limit 2000;
  20. +---------+
  21. | TOT_CNT |
  22. +---------+
  23. |       0 |
  24. +---------+
  25. 1 row in set, 2 warnings (0.02 sec)

分析下后续SQL为何不淘汰该plan

我们知道，SQL查询并不需要每次生成查询计划，因为这样涉及到硬解析等耗费性能的操作，所以默认每次会先查询 Plan Cache （硬解析操作包含词法/语法/语义解析，优化器统计信息查询等步骤，参考下图）。

1716357409

本案例中，后续的SQL命中该 Plan 就可以理解，因为要走 I5 索引，range太大基本为全索引扫描，所以耗时太慢。，

ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31'

什么时候淘汰这个计划呢？

  1. // 关键代码段如下（410bp1社区版，这里的逻辑和323bp8企业版类似，企业版代码不便贴出）
  2. if (sample_count < SLOW_QUERY_SAMPLE_SIZE) {
  3.         // do nothing when query execution samples are not enough
  4.       } else {
  5.         if (stat_.cpu_time_ <= SLOW_QUERY_TIME_FOR_PLAN_EXPIRE * stat_.execute_times_) {
  6.         // do nothing for fast query
  7.         } else if (is_plan_unstable(sample_count, sample_exec_row_count, sample_exec_usec)) {
  8.           set_is_expired(true);
  9.         }
  10.         ATOMIC_STORE(&(stat_.sample_times_), 0);
  11.       }
  12.     }
  13.  
  14. bool ObPhysicalPlan::is_plan_unstable(const int64_t sample_count,
  15.                                       const int64_t sample_exec_row_count,
  16.                                       const int64_t sample_exec_usec)
  17. {
  18.   bool bret = false;
  19.   if (sample_exec_usec <= SLOW_QUERY_TIME_FOR_PLAN_EXPIRE * sample_count) {
  20.     // sample query is fast query in the average
  21.   } else if (OB_PHY_PLAN_LOCAL == plan_type_) {
  22.     int64_t first_query_range_rows = ATOMIC_LOAD(&stat_.first_exec_row_count_);
  23.     if (sample_exec_row_count <= SLOW_QUERY_ROW_COUNT_THRESOLD * sample_count) {
  24.       // the sample query does not accesses too many rows in the average
  25.     } else if (sample_exec_row_count / sample_count > first_query_range_rows * 10) {
  26.       // the average sample query range row count increases great
  27.       bret = true;
  28.       LOG_INFO("local query plan is expired due to unstable performance",
  29.                K(bret), K(stat_.execute_times_),
  30.                K(first_query_range_rows), K(sample_exec_row_count), K(sample_count));
  31.     }
  32.   } else if ( OB_PHY_PLAN_DISTRIBUTED == plan_type_) {
  33.     int64_t first_exec_usec = ATOMIC_LOAD(&stat_.first_exec_usec_);
  34.     if (sample_exec_usec / sample_count > first_exec_usec * 2) {
  35.       // the average sample query execute time increases great
  36.       bret = true;
  37.       LOG_INFO("distribute query plan is expired due to unstable performance",
  38.                K(bret), K(stat_.execute_times_), K(first_exec_usec),
  39.                K(sample_exec_usec), K(sample_count));
  40.     }
  41.   } else {
  42.     // do nothing
  43.   }
  44.   return bret;
  45. }

这里淘汰一个 Plan 需要满足的条件有2个：

  • sample_count < SLOW_QUERY_SAMPLE_SIZE)
  • sample_exec_row_count / sample_count > first_query_range_rows * 10

这里的 SLOW_QUERY_SAMPLE_SIZE 是常量，OB410的定义是 20；sample_count（采样次数）实质为Plan的SQL执行次数。

static const int64_t SLOW_QUERY_SAMPLE_SIZE = 20; // smaller than ObPlanStat::MAX_SCAN_STAT_SIZE

结合上下文代码来看，意思是满足如下情况就会淘汰Plan：

  • 命中该Plan的SQL执行大于等于20次。
  • （执行的SQL扫描总行数 / 执行次数） 大于  （第一次SQL执行扫描的行数 * 10）

复现以验证

1.清空 plan cache，执行业务第一次生成 Plan 的 SQL。

  1. MySQL > alter system flush plan cache;
  2. Query OK, 0 rows affected (0.13 sec)
  3.  
  4. MySQL > select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '' AND ACCTG_DT <= '') as orginal limit 2000;
  5. +---------+
  6. | TOT_CNT |
  7. +---------+
  8. |       0 |
  9. +---------+
  10. 1 row in set, 2 warnings (0.02 sec)

2.执行业务 SQL，复现慢的情况。

  1. MySQL > select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31') as orginal limit 2000;
  2. +---------+
  3. | TOT_CNT |
  4. +---------+
  5. |       1 |
  6. +---------+
  7. 1 row in set (2 min 51.61 sec)
  8.  
  9. MySQL > select last_trace_id();
  10. +-----------------------------------+
  11. | last_trace_id()                   |
  12. +-----------------------------------+
  13. | YB420CF01A46-0006009AD91C51ED-0-0 |
  14. +-----------------------------------+
  15. 1 row in set (0.04 sec)
  16.  
  17. MySQL > select * from oceanbase.gv$sql_audit where trace_id='YB420CF01A46-0006009AD91C51ED-0-0'\G                                                                                                         ...
  18.                TRACE_ID: YB420CF01A46-0006009AD91C51ED-0-0
  19. ...
  20.                  SQL_ID: 2B53F4C1C330C2C089C7518CD71D667A
  21.               QUERY_SQL: select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31') as orginal limit 2000
  22. ...
  23.            ELAPSED_TIME: 171575101
  24. ...
  25.            EXECUTE_TIME: 171574843
  26. ...
  27. MEMSTORE_READ_ROW_COUNT: 25416176
  28.  SSSTORE_READ_ROW_COUNT: 50832349
  29. ...

这里通过 sql_audit 可以观测到重要的信息：

  • ELAPSED_TIME ： 执行耗时。
  • MEMSTORE_READ_ROW_COUNT / SSSTORE_READ_ROW_COUNT : 这条SQL扫描的行数。
  1. MySQL [oceanbase]> select * from gv$plan_cache_plan_stat where plan_id=7289113 \G                                                                         
  2. *************************** 1. row ***************************
  3. ...
  4.                sql_id: 2B53F4C1C330C2C089C7518CD71D667A
  5. ...
  6.             statement: select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = ?) AND LAWENF_NTIST_TP_CD NOT LIKE ? AND LAWENF_NTIST_TP_CD NOT LIKE ? AND EMRG_STPY_SRC_CD != ? AND ACCTG_DT >= ? AND ACCTG_DT <= ?) as orginal limit 2000
  7.             query_sql: select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '' AND ACCTG_DT <= '') as orginal limit 2000
  8. ...
  9.            outline_id: -1
  10.          outline_data: /*+ BEGIN_OUTLINE_DATA INDEX(@"SEL$2" "gabsdb.renzy"@"SEL$2" "renzy_I5") END_OUTLINE_DATA*/
  11. ...

通过 plan_cache_plan_stat 可看到这条SQL命中了第一次SQL执行时生成的 Plan（不符合预期）。

3.继续通过脚本执行多次。

  1. #!/bin/bash
  2. for i in `seq 1 30`
  3. do
  4. echo ">>> do  $i"
  5. mysql -h12.240.68.36 -P3306 -uroot@tgabsua2g00#obcdcbsuat01 -pOceanBase_123# -Dgabsdb -A -c -NBe "select now();select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31') as orginal limit 2000; select last_trace_id();select now();"
  6. done
  7.  
  8. # ./1.sh
  9. ...
  10. >>> do  17       # 耗时 2分钟27s 命中
  11. 2023-11-14 16:05:36
  12. 1
  13. YB420CF01A46-0006009B016C66EE-0-0
  14. 2023-11-14 16:08:03
  15. >>> do  18       # 耗时 2min12s  命中
  16. 2023-11-14 16:08:03
  17. 1
  18. YB420CF01A46-0006009AFF8FF46D-0-0
  19. 2023-11-14 16:10:15
  20. >>> do  19       # 耗时 2min36s  命中
  21. 2023-11-14 16:10:15
  22. 1
  23. YB420CF01A46-0006009B012FF1D0-0-0
  24. 2023-11-14 16:12:51
  25. >>> do  20        # 耗时 1s内     未命中，恢复正常
  26. 2023-11-14 16:12:51
  27. 1
  28. YB420CF01A46-0006009AFEBDA7C6-0-0
  29. 2023-11-14 16:12:51
  30. >>> do  21
  31. 2023-11-14 16:12:51
  32. 1
  33. YB420CF01A46-0006009B016F1561-0-0
  34. 2023-11-14 16:12:52
  35. ...

可以观察到，命中该 Plan 的SQL 执行次数大于 20 次（含手工执行）后，该"不符合预期的" Plan 被淘汰。

4. 再次执行的SQL的 sql_audit 和 plan_cache_plan_stat，可看到重新生成了 Plan。

  1. MySQL > select * from oceanbase.gv$sql_audit where trace_id='YB420CF01A46-0006009AFEBDA7C6-0-0'\G                                                                                                         
  2.                  SQL_ID: 2B53F4C1C330C2C089C7518CD71D667A
  3.               QUERY_SQL: select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31') as orginal limit 2000
  4. ...
  5.           ELAPSED_TIME: 207
  6.           PLAN_ID: 7334178
  7. ...
  8. MEMSTORE_READ_ROW_COUNT: 1
  9.  SSSTORE_READ_ROW_COUNT: 2               
  10.  
  11. MySQL [oceanbase]> select * from gv$plan_cache_plan_stat where plan_id=7334178 \G       
  12. *************************** 1. row ***************************
  13. ...
  14.               plan_id: 7334178
  15.                sql_id: 2B53F4C1C330C2C089C7518CD71D667A
  16. ...
  17.             statement: select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = ?) AND LAWENF_NTIST_TP_CD NOT LIKE ? AND LAWENF_NTIST_TP_CD NOT LIKE ? AND EMRG_STPY_SRC_CD != ? AND ACCTG_DT >= ? AND ACCTG_DT <= ?) as orginal limit 2000
  18.             query_sql: select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31') as orginal limit 2000
  19. ...
  20.       first_load_time: 2023-11-14 16:12:51.547434
  21. ...
  22.      slowest_exe_time: 2023-11-14 16:12:51.547618
  23.      slowest_exe_usec: 4139
  24. ...
  25.          elapsed_time: 8279
  26. ...
  27.            outline_id: -1
  28.          outline_data: /*+ BEGIN_OUTLINE_DATA INDEX(@"SEL$2" "gabsdb.renzy"@"SEL$2" "renzy_I2") END_OUTLINE_DATA*/
  29. ...

5.obs日志关键信息

  1. #grep YB420CF01A46-0006009B012FF1D0-0-0 observer.log.20231114161*|less
  2. observer.log.20231114161017:[2023-11-14 16:10:15.813150] INFO  [SQL] ob_sql.cpp:1769 [86881][0][YB420CF01A46-0006009B012FF1D0-0-0] [lt=17] [dc=0] It is a large query, need delay, do not need disconnect(avg_process_time=123860984, exec_cnt=20, large_query_threshold=5000000, plan->get_plan_id()=7328133, ret=-4023)
  3. observer.log.20231114161017:[2023-11-14 16:10:15.813208] TRACE [TRACE]obmp_base.cpp:156 [86881][0][YB420CF01A46-0006009B012FF1D0-0-0] [lt=18] [dc=0] [packet retry query](TRACE=begin_ts=1699949415813080 2023-11-14 08:10:15.813080|[start_sql] u=0 addr:{ip:"12.241.29.28", port:16606}|[process_begin] u=0 addr:{ip:"12.241.29.28", port:16606}, in_queue_time:13, receive_ts:1699949415813066, enqueue_ts:1699949415813067, trace_id:YB420CF01A46-0006009B012FF1D0-0-0|[session] u=3 sid:3221784053, tenant_id:1005|[parse_begin] u=10 stmt:"select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31') as orginal limit 2000", stmt_len:287|[process_end] u=85 run_ts:1699949415813082|total_timeu=98)
  4. observer.log.20231114161302:[2023-11-14 16:12:51.412696] INFO  [SQL.ENG] ob_physical_plan.cpp:736 [86881][0][YB420CF01A46-0006009B012FF1D0-0-0] [lt=15] [dc=0] local query plan is expired due to unstable performance(bret=true, stat_.execute_times_=21, first_query_range_rows=0, sample_exec_row_count=1525906500, sample_count=20)
  5. observer.log.20231114161302:[2023-11-14 16:12:51.412725] WARN  [SHARE.SCHEMA] revert (ob_schema_mgr_cache.cpp:131) [86881][0][YB420CF01A46-0006009B012FF1D0-0-0] [lt=11] [dc=0] long time to hold one guard(schema_mgr=0x7ee87934c610, tenant_id=1, version=1697523399752200, cur_timestamp=1699949571412714, ref_timestamp=1699949415812628, lbt()="0xf51231f 0x6158f04 0x4f5992c 0x50c61cc 0x4ed2f6f 0x4ecf518 0x4ecc8ef 0x4ecaa6e 0xb8c71f1 0x4ec9c90 0xb8c4d31 0x4ec58f6 0xb8c52a7 0xf3f17f3 0xf3f164f 0xf6901df")
  6. observer.log.20231114161302:[2023-11-14 16:12:51.412738] WARN  [SHARE.SCHEMA] revert (ob_schema_mgr_cache.cpp:131) [86881][0][YB420CF01A46-0006009B012FF1D0-0-0] [lt=8] [dc=0] long time to hold one guard(schema_mgr=0x7edddba39170, tenant_id=1005, version=1699931135472584, cur_timestamp=1699949571412734, ref_timestamp=1699949415812628, lbt()="0xf51231f 0x6158f04 0x4f5992c 0x50c61cc 0x4ed2f6f 0x4ecf518 0x4ecc8ef 0x4ecaa6e 0xb8c71f1 0x4ec9c90 0xb8c4d31 0x4ec58f6 0xb8c52a7 0xf3f17f3 0xf3f164f 0xf6901df")
  7. observer.log.20231114161302:[2023-11-14 16:12:51.412798] TRACE [TRACE]obmp_base.cpp:147 [86881][0][YB420CF01A46-0006009B012FF1D0-0-0] [lt=5] [dc=0] [slow query](TRACE=begin_ts=1699949415813229 2023-11-14 08:10:15.813229|[start_sql] u=0 addr:{ip:"12.241.29.28", port:16606}|[process_begin] u=0 addr:{ip:"12.241.29.28", port:16606}, in_queue_time:162, receive_ts:1699949415813066, enqueue_ts:1699949415813225, trace_id:YB420CF01A46-0006009B012FF1D0-0-0|[session] u=2 sid:3221784053, tenant_id:1005|[parse_begin] u=6 stmt:"select * from (SELECT COUNT(*) AS TOT_CNT FROM renzy WHERE (ACCT_NO = '6222620117273900882') AND LAWENF_NTIST_TP_CD NOT LIKE '12%' AND LAWENF_NTIST_TP_CD NOT LIKE '05%' AND EMRG_STPY_SRC_CD != 'JZ05' AND ACCTG_DT >= '1900-01-01' AND ACCTG_DT <= '2025-03-31') as orginal limit 2000", stmt_len:287|[exec_begin] u=29 arg1:false, end_trans_cb:false, plan_id:7328133|[do_open_plan_begin] u=8 |[sql_start_stmt_begin] u=1 |[sql_start_participant_begin] u=5 |[storage_table_scan_begin] u=56 |[storage_table_scan_end] u=116 |[get_row] u=155437570 |[result_set_close] u=161554 |[close_plan_begin] u=0 |[revert_scan_iter] u=96 |[end_participant_begin] u=3 |[start_end_stmt] u=1 |[affected_rows] u=0 affected_rows:-1|[store_found_rows] u=1 found_rows:0, return_rows:1|[auto_end_plan_begin] u=0 |[process_end] u=86 run_ts:1699949415813230|total_timeu=155599534)

第21次执行的SQL的关键日志信息：

[2023-11-14 16:12:51.412696] INFO  [SQL.ENG] ob_physical_plan.cpp:736 [86881][0][YB420CF01A46-0006009B012FF1D0-0-0] [lt=15] [dc=0] local query plan is expired due to unstable performance(bret=true, stat_.execute_times_=21, first_query_range_rows=0, sample_exec_row_count=1525906500, sample_count=20)

由该日志，关键信息如下：

1. sample_exec_row_count=1525906500

2. sample_count=20

3. first_query_range_rows=0

结合代码可知该结果满足 Plan 淘汰条件，从而 plan expire。

  1. sample_exec_row_count / sample_count > first_query_range_rows * 10
  2.  
  3. 1525906500 / 20  > 0 * 10  
  4. // 这里 1525906500 的结果，不难得知，是单次SQL扫描行数 * 20. 
  5. // 即（25416176 + 50832349） * 20 = 1524970500 约等于 1525906500 
  6.  MEMSTORE_READ_ROW_COUNT: 25416176
  7.  SSSTORE_READ_ROW_COUNT: 50832349

结论

1.本例主要是想分享SQL走错 Plan 而SQL慢的排查方法论，问题原因还是比较简单，重点是和大家分享处理OB遇到类似问题的思路等。

2.本例问题在当前OB 323版本中没有好的优化方式，给到的建议是：

  • 如果 I5 索引业务上未使用场景，则删除。
  • 绑定 outline，使该SQL走 I2 索引。

3.分享下OB中 Plan Cache 清理策略：

  • 手工清理
  1. -- 租户内执⾏，清除当前租户中所有 Plan Cache。⽣产慎⽤。
  2. ALTER SYSTEM FLUSH PLAN CACHE;
  3. -- sys租户下执⾏，不同粒度。
  4. ALTER SYSTEM FLUSH PLAN CACHE TENANT = 'T_MySQL';
  5. ALTER SYSTEM FLUSH PLAN CACHE sql_id='B601070DFC14CB85FDA3766A69A9E1B3'
  6. databases='myob1' tenant='tenant1' GLOBAL;
  • 自动清理 ob_plan_cache_percentage 参数控制 Plan Cache占用租户内存的百分比。 本例中提到

1. sample_count < SLOW_QUERY_SAMPLE_SIZE) ：命中该Plan的SQL执行大于等于20次。 

2.sample_exec_row_count / sample_count > first_query_range_rows * 10 ：（执行的SQL扫描总行数 / 执行次数） 大于 （第一次SQL执行扫描的行数 * 10）

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/爱喝兽奶帝天荒/article/detail/846061
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号