MySQL [db_stat]> explain delete from t_stat where imp_date<='20200202';
+---------------------+--------------+------+------------------------------------------------------+
| id | count | task | operator info |
+---------------------+--------------+------+------------------------------------------------------+
| TableReader_6 | 220895815.00 | root | data:Selection_5 |
| └─Selection_5 | 220895815.00 | cop | le(db_stat.t_stat.imp_date, "20200202") |
| └─TableScan_4 | 220895815.00 | cop | table:t_stat, range:[-inf,+inf], keep order:false |
+---------------------+--------------+------+------------------------------------------------------+
3 rows in set (0.00 sec)
MySQL [db_stat]> select count(*) from t_stat where imp_date<='20200202';
+-----------+
| count(*) |
+-----------+
| 184340473 |
+-----------+
1 row in set (17.88 sec)
- 大批量清理数据时系统资源消耗高,可能导致 TiDB OOM
- imp_date 字段虽然有索引,但是范围扫描的数据量过大,无论是走 IndexScan 还是 Table Scan,Coprocessor 都要处理大量数据
- TiKV 节点 Coprocessor CPU 飙高
- 执行 Delete 操作的 TiDB 节点内存飙高,因为要将1.8亿条数据加载 tidb server 内存
- 删除数据时,缩小数据筛选范围,或者加上 limit N 每次删除一批数据
- 建议使用 3.0 版本的 Range 分区,按照分区快速删除
MySQL [db_stat]> explain SELECT * FROM `tbl_article_check_result` `t` WHERE (articleid = '20190925A0PYT800') ORDER BY checkTime desc LIMIT 100 ;
+--------------------------+----------+------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| id | count | task | operator info |
+--------------------------+----------+------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| Projection_7 | 100.00 | root | db_stat.t.type, db_stat.t.articleid, db_stat.t.docid, db_stat.t.version, db_stat.t.checkid, db_stat.t.checkstatus, db_stat.t.seclevel, db_stat.t.t1checkstatus, db_stat.t.t2checkstatus, db_stat.t.mdaichannel, db_stat.t.mdaisubchannel, db_stat.t.checkuser, db_stat.t.checktime, db_stat.t.addtime, db_stat.t.havegot, db_stat.t.checkcode |
| └─Limit_12 | 100.00 | root | offset:0, count:100 |
| └─IndexLookUp_34 | 100.00 | root | |
| ├─IndexScan_31 | 30755.49 | cop | table:t, index:checkTime, range:[NULL,+inf], keep order:true, desc |
| └─Selection_33 | 100.00 | cop | eq(db_dayu_1.t.articleid, "20190925A0PYT800") |
| └─TableScan_32 | 30755.49 | cop | table:tbl_article_check_result, keep order:false |
+--------------------------+----------+------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
6 rows in set (0.00 sec)
- articleid 和 checkTime 字段分别建有单列索引,正常情况下走 articleid 索引比较快,有时执行计划不稳定走 checkTime 索引,查询延迟达到分钟级别
- LIMIT 100 限定了获取 100 条记录,checkTime 和 articleid 列之间存在相关性,如果列之间的相关度不高,在独立性假设失效时,优化器估算走 checkTime 索引并满足 articleid 条件时扫描的行数,可能比走 articleid 索引扫描的行数更少
- 业务响应延迟不稳定,偶尔有抖动
- 手动 analyze table,配合 crontab 定期 analyze,维持统计信息准确度
- 自动 auto analyze,调低 analyze ratio 阈值,提高收集频次,并设置运行时间窗口
- set global tidb_auto_analyze_ratio=0.2;
- set global tidb_auto_analyze_start_time='00:00 +0800';
- set global tidb_auto_analyze_end_time='06:00 +0800';
- 业务修改 SQL ,使用 force index 固定使用 articleid 列上的索引
- 3.0 版本下,业务不修改 SQL,使用 create binding 创建 force index 的绑定 SQL
MySQL [db_stat]> explain select * from t_like_list where person_id=1535538061143263;
+---------------------+------------+------+-----------------------------------------------------------------------------------+
| id | count | task | operator info |
+---------------------+------------+------+-----------------------------------------------------------------------------------+
| Selection_5 | 1430690.40 | root | eq(cast(db_stat.t_like_list.person_id), 1.535538061143263e+15) |
| └─TableReader_7 | 1788363.00 | root | data:TableScan_6 |
| └─TableScan_6 | 1788363.00 | cop | table:t_like_list, range:[-inf,+inf], keep order:false |
+---------------------+------------+------+-----------------------------------------------------------------------------------+
3 rows in set (0.00 sec)
MySQL [db_stat]>
- person_id 列上建有索引且选择性较好,但执行计划走了 TableScan
- person_id 是字符串类型,但是存储的值都是数字,业务认为可以直接赋值;而优化器需要在字段上做 cast 类型转换,导致无法使用索引
- where 条件加上引号
MySQL [db_stat]> explain select * from table:t_like_list where person_id='1535538061143263';
+-------------------+-------+------+----------------------------------------------------------------------------------------------------------+
| id | count | task | operator info |
+-------------------+-------+------+----------------------------------------------------------------------------------------------------------+
| IndexLookUp_10 | 0.00 | root | |
| ├─IndexScan_8 | 0.00 | cop | table:t_like_list, index:person_id, range:["1535538061143263","1535538061143263"], keep order:false |
| └─TableScan_9 | 0.00 | cop | table:t_like_list, keep order:false |
+-------------------+-------+------+----------------------------------------------------------------------------------------------------------+
3 rows in set (0.00 sec)
一个数据量不大(600G左右)读多写少的集群,某段时间发现 query summary 监控中的 duration 显著增加,p99 如下图:
p999 如下图:
通过查看监控发现 TiDB 的 KV Duration 变高,KV Errors 仍然很低。其中 KV Request Duration 999 by store 监控是各 TiKV 轮流上涨,如下图:
继续查看 TiKV 监控,Coprocessor Overview 如下:
Coprocessor CPU 如下:
Coprocessor 监控如下:
Coprocessor CPU 几乎打满了 CPU。下面开始分析日志,调查 Duration 和 Coprocessor CPU 高的原因。
通过 pt-query-digest 工具分析 TiDB 慢查询日志:
./pt-query-digest tidb_slow_query.log > result
分析 result 发现 Process keys 多的 SQL 并不一定 Process time 也长,二者并不正相关。比如如下 SQL 的 Process keys 为 22.09M,Process time 为 51s:
但是如下 SQL 的 Process keys 为 12.68M,而 Process time 为 142353s:
过滤 Process time 较多的 SQL,发现了3个典型的 SQL,分析执行计划是否正常:
SQL1: select a.a_id, a.b_id,uqm.p_id from a join hsq on a.b_id=hsq.id join uqm on a.a_id=uqm.id;
SQL2: select distinct g.abc, g.def, g.ghi, h.abcd, hi.jq from ggg g left join ggg_host gh on g.id = gh.ggg_id left join host h on gh.a_id = h.id left join a_jq hi on h.id = hi.hid where h.abcd is not null and h.abcd <> '' and hi.jq is not null and hi.jq <> '';
SQL3: select tb1.mt, tb2.name from tb2 left join tb1 on tb2.mtId=tb1.id where tb2.type=0 and (tb1.mt is not null and tb1.mt != '') and (tb2.name is not null or tb2.name != '');
从执行计划看没有问题,查看表的统计信息也正常,继续分析 TiDB、TiKV日志。
查看日志中标记为 [slow-query] 的日志行中的 region 分布情况:
more tikv.log.2019-10-16-06\:28\:13 |grep slow-query |awk -F ']' '{print $1}' | awk '{print $6}' | sort | uniq -c | sort –n
访问频率最大的3个 region 为:
73 29452
140 33324
757 66625
这些 region 的访问次数远大于其它 region。之后定位这些 region 所属的表名。 首先查看 [slow-query] 行里的 table_id 记录值和 start_ts 记录值,然后查询 TiDB 日志以获取表名。比如 table_id 为1318,start_ts 为411837294180565013,则:
more tidb-2019-10-14T16-40-51.728.log |grep '"/[1318/]"' |grep 411837294180565013
过滤后发现是上述慢查询 SQL 涉及到的表。
尝试对这些 region 做 split 操作,以 66625 为例,需要将 x.x.x.x 替换为实际的 pd 地址:
pd-ctl –u http://x.x.x.x:2379 operator add split-region 66625
查看 PD 日志如下:
[2019/10/16 18:22:56.223 +08:00] [INFO] [operator_controller.go:99] ["operator finish"] [region-id=30796] [operator="\"admin-split-region (kind:admin, region:66625(1668,3), createAt:2019-10-16 18:22:55.888064898 +0800 CST m=+110918.823762963, startAt:2019-10-16 18:22:55.888223469 +0800 CST m=+110918.823921524, currentStep:1, steps:[split region with policy SCAN]) finished\""]
说明 region 分裂完成,之后查看是否 region 仍然是热点:
more tikv.log.2019-10-16-06\:28\:13 |grep slow-query | grep 66625 | more
观察一段时间后确认 66625 已不是热点 region,之后继续处理其它的热点 region。 所有热点 region 处理完成后 query summary 监控中的 duration 显著降低:
不过只保持了一段时间(19:35 后仍有较高的 duration 出现,下图未列出):
观察压力较重的 tikv,移走热点 region 的 leader:
pd-ctl –u http://x.x.x.x:2379 operator add transfer-leader 1 2 //把Region1 的 leader 调度到 Store2
leader 迁走之后 ,原 TiKV 的 duration 立刻下降了,但是 leader 所在的 TiKV duration 随之上升(图中展示了一个 TiKV 的变化过程):
分析后反复分裂热点 region,之后效果如下:
对于分布式数据库读热点的情况,难以通过优化 SQL 本身来解决,需要通过系统分析整个数据库的状态来定位原因,采用分裂 region 的方式来缓解和消除读热点对 SQL 查询效率的影响。读热点实际上限制了分布式数据库的并发读取优势,分裂 region 的方式本质上是恢复分布式数据库的并发读取优势。
- SQL 执行时间突然变长
- SQL 语句:
select count(*)
from tods.bus_jijin_trade_record a, tods.bus_jijin_info b
where a.fund_code=b.fund_code and a.type in ('PURCHASE','APPLY')
and a.status='CANCEL_SUCCESS' and a.pay_confirm_status = 1
and a.cancel_app_no is not null and a.id >= 177045000
and a.updated_at > date_sub(now(), interval 48 hour) ;
执行结果,需要 1 分 3.7s:
mysql> select count(*)
-> from tods.bus_jijin_trade_record a, tods.bus_jijin_info b
-> where a.fund_code=b.fund_code and a.type in ('PURCHASE','APPLY')
-> and a.status='CANCEL_SUCCESS' and a.pay_confirm_status = 1
-> and a.cancel_app_no is not null and a.id >= 177045000
-> and a.updated_at > date_sub(now(), interval 48 hour) ;
+----------+
| count(*) |
+----------+
| 708 |
+----------+
1 row in set (1 min 3.77 sec)
- 索引信息
表名 | 数据行 | 索引名 |
---|---|---|
bus_jijin_trade_record | 176384036 | PRIMARY KEY (ID ) KEY idx_bus_jijin_trade_record_upt (UPDATED_AT ) |
bus_jijin_info | 6442 | PRIMARY KEY (ID ) |
- 查看执行计划
mysql> explain
-> select count(*)
-> from tods.bus_jijin_trade_record a, tods.bus_jijin_info b
-> where a.fund_code=b.fund_code and a.type in ('PURCHASE','APPLY')
-> and a.status='CANCEL_SUCCESS' and a.pay_confirm_status = 1
-> and a.cancel_app_no is not null and a.id >= 177045000
-> and a.updated_at > date_sub(now(), interval 48 hour) ;
+----------------------------+--------------+------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| id | count | task | operator info |
+----------------------------+--------------+------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| StreamAgg_13 | 1.00 | root | funcs:count(1) |
| └─HashRightJoin_27 | 421.12 | root | inner join, inner:TableReader_18, equal:[eq(a.fund_code, b.fund_code)] |
| ├─TableReader_18 | 421.12 | root | data:Selection_17 |
| │ └─Selection_17 | 421.12 | cop | eq(a.pay_confirm_status, 1), eq(a.status, "CANCEL_SUCCESS"), gt(a.updated_at, 2020-03-03 22:31:08), in(a.type, "PURCHASE", "APPLY"), not(isnull(a.cancel_app_no)) |
| │ └─TableScan_16 | 145920790.55 | cop | table:a, range:[177045000,+inf], keep order:false |
| └─TableReader_37 | 6442.00 | root | data:TableScan_36 |
| └─TableScan_36 | 6442.00 | cop | table:b, range:[-inf,+inf], keep order:false |
+----------------------------+--------------+------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------+
TableScan_16,TableScan_36:表示在 KV 端分别对表 a 和 b 的数据进行扫描,其中 TableScan_16 扫描了 1.46 亿的行数;
Selection_17:表示满足表 a 后面 where 条件的数据
TableReader_37: 由于表 b 没有独立的附加条件,所以直接将这部分数据返回给 tidb;
TableReader_18:将各个 coprocessor 满足 a 表条件的结果返回给 tidb;
HashRightJoin_27:将 TableReader_37 和 TableReader_18 上的结果进行 hash join;
StreamAgg_13:进一步统计所有行数,将数据返回给客户端;
可以看到语句中 a 表 ( bus_jijin_trade_record ) 的条件 id >= 177045000, 和 updated_at > date_sub(now(), interval 48 hour ) 上,这两个列分别都有索引,但是 tidb 还是选择了全表扫描;
按照上面两个条件分别查询数据分区情况
mysql> SELECT COUNT(*) FROM tods.bus_jijin_trade_record WHERE id >= 177045000 ;
+-----------+
| COUNT(*) |
+-----------+
| 145917327 |
+-----------+
1 row in set (16.86 sec)
mysql> SELECT COUNT(*) FROM tods.bus_jijin_trade_record WHERE updated_at > date_sub(now(), interval 48 hour) ;
+-----------+
| COUNT(*) |
+-----------+
| 713682 |
+-----------+
可以看到,表 bus_jijin_trade_record 有 1.7 亿的数据量, 应该走 updated_at 字段上的索引;
使用强制 hint 进行执行,6.27 秒就执行完成了,速度从之前 63s 到现在的 6.3s,提升了 10 倍
mysql> select count(*)
-> from tods.bus_jijin_trade_record a use index(idx_bus_jijin_trade_record_upt), tods.bus_jijin_info b
-> where a.fund_code=b.fund_code and a.type in ('PURCHASE','APPLY')
-> and a.status='CANCEL_SUCCESS' and a.pay_confirm_status = 1
-> and a.cancel_app_no is not null and a.id >= 177045000
-> and a.updated_at > date_sub(now(), interval 48 hour) ;
+----------+
| count(*) |
+----------+
| 709 |
+----------+
1 row in set (6.27 sec)
强制 hint 后的执行计划
mysql> explain
-> select count(*)
-> from tods.bus_jijin_trade_record a use index(idx_bus_jijin_trade_record_upt), tods.bus_jijin_info b
-> where a.fund_code=b.fund_code and a.type in ('PURCHASE','APPLY')
-> and a.status='CANCEL_SUCCESS' and a.pay_confirm_status = 1
-> and a.cancel_app_no is not null and a.id >= 177045000
-> and a.updated_at > date_sub(now(), interval 48 hour) ;
+------------------------------+--------------+------+----------------------------------------------------------------------------------------------------------------------------+
| id | count | task | operator info |
+------------------------------+--------------+------+----------------------------------------------------------------------------------------------------------------------------+
| StreamAgg_13 | 1.00 | root | funcs:count(1) |
| └─HashRightJoin_24 | 421.12 | root | inner join, inner:IndexLookUp_20, equal:[eq(a.fund_code, b.fund_code)] |
| ├─IndexLookUp_20 | 421.12 | root | |
| │ ├─Selection_18 | 146027634.83 | cop | ge(a.id, 177045000) |
| │ │ └─IndexScan_16 | 176388219.00 | cop | table:a, index:UPDATED_AT, range:(2020-03-03 23:05:30,+inf], keep order:false |
| │ └─Selection_19 | 421.12 | cop | eq(a.pay_confirm_status, 1), eq(a.status, "CANCEL_SUCCESS"), in(a.type, "PURCHASE", "APPLY"), not(isnull(a.cancel_app_no)) |
| │ └─TableScan_17 | 146027634.83 | cop | table:bus_jijin_trade_record, keep order:false |
| └─TableReader_31 | 6442.00 | root | data:TableScan_30 |
| └─TableScan_30 | 6442.00 | cop | table:b, range:[-inf,+inf], keep order:false |
+------------------------------+--------------+------+----------------------------------------------------------------------------------------------------------------------------+
使用 hint 后的执行计划,预估 updated_at 上的索引会扫描 176388219,索引选择了全表扫描,可以判定是由于错误的统计信息导致执行计划有问题
查看表 bus_jijin_trade_record 上的统计信息
mysql> show stats_meta where table_name like 'bus_jijin_trade_record' and db_name like 'tods';
+---------+------------------------+---------------------+--------------+-----------+
| Db_name | Table_name | Update_time | Modify_count | Row_count |
+---------+------------------------+---------------------+--------------+-----------+
| tods | bus_jijin_trade_record | 2020-03-05 22:04:21 | 10652939 | 176381997 |
+---------+------------------------+---------------------+--------------+-----------+
mysql> show stats_healthy where table_name like 'bus_jijin_trade_record' and db_name like 'tods';
+---------+------------------------+---------+
| Db_name | Table_name | Healthy |
+---------+------------------------+---------+
| tods | bus_jijin_trade_record | 93 |
+---------+------------------------+---------+
根据统计信息,表 bus_jijin_trade_record 有 176381997,修改的行数有 10652939,
该表的健康度为:(176381997-10652939)/176381997 *100=93
- 重新收集统计信息
mysql> set tidb_build_stats_concurrency=10;
Query OK, 0 rows affected (0.00 sec)
#调整收集统计信息的并发度,以便快速对统计信息进行收集
mysql> analyze table tods.bus_jijin_trade_record;
Query OK, 0 rows affected (3 min 48.74 sec)
- 查看没有使用 hint 语句的执行计划
mysql> explain select count(*)
-> from tods.bus_jijin_trade_record a, tods.bus_jijin_info b
-> where a.fund_code=b.fund_code and a.type in ('PURCHASE','APPLY')
-> and a.status='CANCEL_SUCCESS' and a.pay_confirm_status = 1
-> and a.cancel_app_no is not null and a.id >= 177045000
-> and a.updated_at > date_sub(now(), interval 48 hour) ;;
+------------------------------+-----------+------+----------------------------------------------------------------------------------------------------------------------------+
| id | count | task | operator info |
+------------------------------+-----------+------+----------------------------------------------------------------------------------------------------------------------------+
| StreamAgg_13 | 1.00 | root | funcs:count(1) |
| └─HashRightJoin_27 | 1.99 | root | inner join, inner:IndexLookUp_23, equal:[eq(a.fund_code, b.fund_code)] |
| ├─IndexLookUp_23 | 1.99 | root | |
| │ ├─Selection_21 | 626859.65 | cop | ge(a.id, 177045000) |
| │ │ └─IndexScan_19 | 757743.08 | cop | table:a, index:UPDATED_AT, range:(2020-03-03 23:28:14,+inf], keep order:false |
| │ └─Selection_22 | 1.99 | cop | eq(a.pay_confirm_status, 1), eq(a.status, "CANCEL_SUCCESS"), in(a.type, "PURCHASE", "APPLY"), not(isnull(a.cancel_app_no)) |
| │ └─TableScan_20 | 626859.65 | cop | table:bus_jijin_trade_record, keep order:false |
| └─TableReader_37 | 6442.00 | root | data:TableScan_36 |
| └─TableScan_36 | 6442.00 | cop | table:b, range:[-inf,+inf], keep order:false |
+------------------------------+-----------+------+----------------------------------------------------------------------------------------------------------------------------+
9 rows in set (0.00 sec)
可以看到,收集完统计信息后,现在的执行计划走了索引扫描,与手动添加 hint 的行为一致,且扫描的行数 757743 符合预期;
此时执行时间变为 1.69s ,在执行计划没变的情况下,应该是由于缓存命中率上升带来的提升。
mysql> select count(*)
-> from tods.bus_jijin_trade_record a, tods.bus_jijin_info b
-> where a.fund_code=b.fund_code and a.type in ('PURCHASE','APPLY')
-> and a.status='CANCEL_SUCCESS' and a.pay_confirm_status = 1
-> and a.cancel_app_no is not null and a.id >= 177045000
-> and a.updated_at > date_sub(now(), interval 48 hour) ;
+----------+
| count(*) |
+----------+
| 712 |
+----------+
1 row in set (1.69 sec)
可以看出该 SQL 执行效率变差是由于统计信息不准确造成的,在通过收集统计信息之后得到了正确的执行计划。
从最终结果 712 行记录来看,创建联合索引可以更大的降低扫描数据的量,更进一步提升性能。在性能已经满足业务要求情况下,联合索引会有额外的成本,留待以后尝试。