数据库｜TiDB资源管控升级版！实现集群资源最大化

前言

在早期听过很多客户一直在讨论TiDB有没有多租户的功能，一直是对标其他分布式数据库的热点话题。TiDB v7.1.0支持基于资源组的资源管控，为同一集群中的不同工作负载分配并隔离资源。该功能显著提升了多应用集群的稳定性，并为多租户奠定了基础。

TiDB 资源管控特性提供了两层资源管理能力，包括在 TiDB 层的流控能力和 TiKV 层的优先级调度的能力。将用户绑定到某个资源组后，TiDB 层会根据用户所绑定资源组设定的配额对用户的读写请求做流控，TiKV 层会根据配额映射的优先级来对请求做调度。通过流控和调度这两层控制，可以实现用户、会话、语句级别的应用资源隔离，满足服务质量 (QoS) 要求。

资源管控流程图

资源管控(Resource Control)测试

一、测试集群环境

[tidb@tidb80 tidb]$ tiup cluster display tidb-test 
tiup is checking updates for component cluster ...
Starting component `cluster`: /home/tidb/.tiup/components/cluster/v1.12.2/tiup-cluster display tidb-test
Cluster type: tidb
Cluster name: tidb-test
Cluster version: v7.1.0 Deploy user: tidb
SSH type: builtin
Dashboard URL: http://192.168.2.81:2379/dashboard Grafana URL: http://192.168.2.80:3000 ID Role Host Ports OS/Arch Status Data Dir Deploy Dir
-- ---- ---- ----- ------- ------ -------- ---------- 192.168.2.80:9093 alertmanager 192.168.2.80 9093/9094 linux/x86_64 Up /data/tidb-data/alertmanager-9093 /data/tidb-deploy/alertmanager-9093 192.168.2.80:3000 grafana 192.168.2.80 3000 linux/x86_64 Up - /data/tidb-deploy/grafana-3000 192.168.2.81:2379 pd 192.168.2.81 2379/2380 linux/x86_64 Up|L|UI /data/tidb-data/pd-2379 /data/tidb-deploy/pd-2379 192.168.2.82:2379 pd 192.168.2.82 2379/2380 linux/x86_64 Up /data/tidb-data/pd-2379 /data/tidb-deploy/pd-2379 192.168.2.83:2379 pd 192.168.2.83 2379/2380 linux/x86_64 Up /data/tidb-data/pd-2379 /data/tidb-deploy/pd-2379 192.168.2.80:9090 prometheus 192.168.2.80 9090/12020 linux/x86_64 Up /data/tidb-data/prometheus-9090 /data/tidb-deploy/prometheus-9090 192.168.2.80:4000 tidb 192.168.2.80 4000/10080 linux/x86_64 Up - /data/tidb-deploy/tidb-4000 192.168.2.81:20160 tikv 192.168.2.81 20160/20180 linux/x86_64 Up /data/tidb-data/tikv-20160 /data/tidb-deploy/tikv-20160 192.168.2.82:20160 tikv 192.168.2.82 20160/20180 linux/x86_64 Up /data/tidb-data/tikv-20160 /data/tidb-deploy/tikv-20160 192.168.2.83:20160 tikv 192.168.2.83 20160/20180 linux/x86_64 Up /data/tidb-data/tikv-20160 /data/tidb-deploy/tikv-20160 Total nodes: 10

二、Request Unit(RU)概念

Request Unit(RU)是TiDB对CPU、IO等系统资源的统一抽象的单位，目前包括CPU、IOPS和IO带宽三个指标。这三个指标的消耗会按照的比例统一到RU单位上。

下表是用户请求对TKV存储层CPU和IO资源的消耗以及对应的RU权重：

基于上表，假设这个资源组消耗的TiKV时间是c毫秒，r1次请求读取了r2KB数据，w1次写请求写入了w2KB数据，集群中非witness

TKV节点数是n,则该资源组消耗的总RU的公式如下：

三、资源管控参数

四、评估实际负载所需容量

注意事项:

评估窗口范围为 10 分钟至 24 小时，TiDB 与 TiKV 的 CPU 利用率不能过低，否则无法进行容量估算。

权限：

TiDB 与 TiKV 的 CPU 利用率不能过低，否则无法进行容量估算

MySQL [(none)]> CALIBRATE RESOURCE START_TIME '2023-06-26 17:25:00' DURATION '10m'; 
+-------+ | QUOTA | +-------+ | 14129 | +-------+ 1 row in set (0.20 sec)

MySQL [(none)]> CALIBRATE RESOURCE START_TIME '2023-06-26 17:25:00' END_TIME '2023-06-26 17:35:00';
+-------+ | QUOTA | +-------+ | 14129 | +-------+ 1 row in set (0.53 sec)

MySQL [(none)]> CALIBRATE RESOURCE;
+-------+ | QUOTA | +-------+ | 9690 | +-------+ 1 row in set (0.21 sec)
MySQL [(none)]> CALIBRATE RESOURCE WORKLOAD OLTP_WRITE_ONLY;
+-------+ | QUOTA | +-------+ | 9148 | +-------+ 1 row in set (0.36 sec)
MySQL [(none)]> CALIBRATE RESOURCE WORKLOAD OLTP_READ_ONLY; 
+-------+ | QUOTA | +-------+ | 1746 | +-------+ 1 row in set (0.23 sec)
MySQL [(none)]> CALIBRATE RESOURCE WORKLOAD OLTP_READ_WRITE;
+-------+ | QUOTA | +-------+ | 3721 | +-------+ 1 row in set (0.01 sec)

五、创建资源组

1.创建 rg1 资源组用于重要应用，限额是每秒 3000 RU，并且允许这个资源组的应用超额占用资源,设置绝对优先级为 HIGH。
MySQL [(none)]> CREATE RESOURCE GROUP IF NOT EXISTS rg1 RU_PER_SEC = 2000 BURSTABLE PRIORITY = HIGH; 2.创建 rg2 资源组，RU 的回填速度是每秒 300 RU。在系统资源充足的时候，不允许这个资源组的应用超额占用资源,设置绝对优先级为 LOW。
MySQL [(none)]> CREATE RESOURCE GROUP IF NOT EXISTS rg2 RU_PER_SEC = 200 PRIORITY = HIGH; 3.创建 rg3 资源组，RU 的回填速度是每秒 300 RU。在系统资源充足的时候，不允许这个资源组的应用超额占用资源,设置绝对优先级为 MEDIUM。
MySQL [(none)]> CREATE RESOURCE GROUP IF NOT EXISTS rg3 RU_PER_SEC = 500 PRIORITY = LOW; 4.创建 rg4 资源组，RU 的回填速度是每秒 300 RU。在系统资源充足的时候，不允许这个资源组的应用超额占用资源,设置绝对优先级为 MEDIUM。
MySQL [(none)]> CREATE RESOURCE GROUP IF NOT EXISTS rg4 RU_PER_SEC = 1000 PRIORITY = MEDIUM; 5.展示所有资源组 (resource group) 的信息
MySQL [(none)]> SELECT * FROM information_schema.resource_groups; 
+---------+------------+----------+-----------+ | NAME | RU_PER_SEC | PRIORITY | BURSTABLE |
+---------+------------+----------+-----------+
| default | UNLIMITED | MEDIUM | YES | | rg1 | 2000 | HIGH | YES |
| rg2 | 200 | HIGH | NO | | rg3 | 500 | LOW | NO |
| rg4 | 1000 | MEDIUM | NO | +---------+------------+----------+-----------+ 5 rows in set (0.01 sec)

六、绑定资源组

MySQL [(none)]> CREATE USER 'test1'@'%' IDENTIFIED BY 'test1' RESOURCE GROUP rg1;
Query OK, 0 rows affected (0.09 sec)
MySQL [(none)]> CREATE USER 'test2'@'%' IDENTIFIED BY 'test2' RESOURCE GROUP rg2;
Query OK, 0 rows affected (0.04 sec)
MySQL [(none)]> CREATE USER 'test3'@'%' IDENTIFIED BY 'test3' RESOURCE GROUP rg3;
Query OK, 0 rows affected (0.06 sec)
MySQL [(none)]> CREATE USER 'test4'@'%' IDENTIFIED BY 'test4' RESOURCE GROUP rg4;
Query OK, 0 rows affected (0.04 sec)
MySQL [(none)]> grant all privileges on *.* to 'test1'@'%';
Query OK, 0 rows affected (0.05 sec)
MySQL [(none)]> grant all privileges on *.* to 'test2'@'%';
Query OK, 0 rows affected (0.04 sec)
MySQL [(none)]> grant all privileges on *.* to 'test3'@'%';
Query OK, 0 rows affected (0.03 sec)
MySQL [(none)]> grant all privileges on *.* to 'test4'@'%';
Query OK, 0 rows affected (0.03 sec)
MySQL [(none)]> select user,host,User_attributes from mysql.user;
+-------+------+---------------------------+
| user | host | User_attributes |
+-------+------+---------------------------+
| root | % | NULL |
| test1 | % | {"resource_group": "rg1"} |
| test2 | % | {"resource_group": "rg2"} |
| test3 | % | {"resource_group": "rg3"} |
| test4 | % | {"resource_group": "rg4"} |
+-------+------+---------------------------+ 5 rows in set (0.02 sec)

[tidb@tidb80 tidb]$ cat config_test1
mysql-host=192.168.2.80 mysql-port=4000 mysql-user=test1
mysql-password=test1
mysql-db=sbtest
time=60 threads=16 report-interval=10 db-driver=mysql
[tidb@tidb80 tidb]$ cat config_test2
mysql-host=192.168.2.80 mysql-port=4000 mysql-user=test2
mysql-password=test2
mysql-db=sbtest
time=60 threads=16 report-interval=10 db-driver=mysql
[tidb@tidb80 tidb]$ cat config_test3
mysql-host=192.168.2.80 mysql-port=4000 mysql-user=test3
mysql-password=test3
mysql-db=sbtest
time=60 threads=16 report-interval=10 db-driver=mysql
[tidb@tidb80 tidb]$ cat config_test4
mysql-host=192.168.2.80 mysql-port=4000 mysql-user=test4
mysql-password=test4
mysql-db=sbtest
time=60 threads=16 report-interval=10 db-driver=mysql
[tidb@tidb80 tidb]$ sysbench --config-file=/tidb/config_test1 /tidb/sysbench/share/sysbench/oltp_write_only.lua --tables=1 --table-size=10000000 --time=600 --mysql-db=rc1 run 
[tidb@tidb80 tidb]$ sysbench --config-file=/tidb/config_test2 /tidb/sysbench/share/sysbench/oltp_write_only.lua --tables=1 --table-size=2000000 --time=600 --mysql-db=rc2 run
[tidb@tidb80 tidb]$ sysbench --config-file=/tidb/config_test3 /tidb/sysbench/share/sysbench/oltp_write_only.lua --tables=1 --table-size=5000000 --time=600 --mysql-db=rc3 run
[tidb@tidb80 tidb]$ sysbench --config-file=/tidb/config_test4 /tidb/sysbench/share/sysbench/oltp_write_only.lua --tables=1 --table-size=10000000 --time=600 --mysql-db=rc4 run

由于rg1设置了应用超额占用资源，rg1消耗RU可达到2K以上，由于rg2，rg3，rg4没有设置超额分配，就算在系统资源充足，也不会让这个资源组的应用超额占用资源。

当系统资源超过限制时，TiDB 会优先满足高优先级 (PRIORITY) 资源组的请求。如果同一优先级的请求无法全部满足，TiDB 会根据用量 (RU_PER_SEC) 的大小按比例分配。

通过把当前会话绑定到资源组，会话对资源的占用会受到指定用量 (RU) 的限制。

分别打开4个会话绑定不同的资源组，由于rg1设置了应用超额占用资源，rg1消耗RU可达到8K以上，由于rg2，rg3，rg4没有设置超额分配，在限制的RU内运行。

当集群遇到突发的 SQL 性能问题，可以结合 SQL Binding 和资源组，临时限制某个 SQL 的资源消耗

将rg2资源组设置ru为1进行语句限制测试，可将某条资源消耗高的语句进行临时限制，从而让集群正常运行

MySQL [(none)]> alter RESOURCE GROUP rg2 RU_PER_SEC = 1; 
MySQL [(none)]> SELECT * FROM information_schema.resource_groups; 
+---------+------------+----------+-----------+
| NAME | RU_PER_SEC | PRIORITY | BURSTABLE |
+---------+------------+----------+-----------+
| default | UNLIMITED | MEDIUM | YES |
| rg1 | 2000 | HIGH | YES |
| rg2 | 1 | MEDIUM | NO |
| rg3 | 500 | LOW | NO |
| rg4 | 1000 | MEDIUM | NO |
+---------+------------+----------+-----------+
MySQL [(none)]> select count(*) from rc1.sbtest1 where pad like '%22%'; +----------+
| count(*) |
+----------+
| 3713861 |
+----------+ create GLOBAL binding for select count(*) from rc1.sbtest1 where pad like '%22%' using select /*+ RESOURCE_GROUP(small_rg1) */ count(*) from rc1.sbtest1 where pad like '%22%';
MySQL [(none)]> show global bindings;
+--------------------------------------------------------------+--------------------------------------------------------------+------------+---------+-------------------------+-------------------------+---------+-----------------+--------+------------------------------------------------------------------+-------------+
| Original_sql | Bind_sql | Default_db | Status | Create_time | Update_time | Charset | Collation | Source | Sql_digest | Plan_digest |
+--------------------------------------------------------------+--------------------------------------------------------------+------------+---------+-------------------------+-------------------------+---------+-----------------+--------+------------------------------------------------------------------+-------------+
| select count ( ? ) from `rc1` . `sbtest1` where `pad` like ? | SELECT count(1) FROM `rc1`.`sbtest1` WHERE `pad` LIKE '%22%' | | enabled | 2023-07-03 11:49:48.436 | 2023-07-03 11:49:48.436 | utf8 | utf8_general_ci | manual | e2231bf96e1696343b39b1897a118cdef02f8d12a651b067b14c887c2611fcfe | |
+--------------------------------------------------------------+--------------------------------------------------------------+------------+---------+-------------------------+-------------------------+---------+-----------------+--------+------------------------------------------------------------------+-------------+
MySQL [(none)]> select count(*) from rc1.sbtest1 where pad like '%22%';
ERROR 8252 (HY000): Exceeded resource group quota limitation

总结