1.How to build a Self-Driving database An overview Dongxu 

2.Who am I ! Dongxu Huang ! CTO & Cofounder, PingCAP ! Distributed system engineer / Open source advocator ! h@pingcap.com ! TiDB / TiKV / TiSpark 

3.Self-Driving? 

4.

5.

6.Why now? ! Better hardware ! Better tools ! Workload is different 

7.SQL Tuning 

8.

9.

10.Imagine we got a logical plan: Its physical plan could be: or: 

11.Cost estimation Network cost Memory cost CPU cost In TiDB, the default memory factor is 5 and CPU factor is 0.8. For example: Operator Sort(r), its cost would be: 

12.Data placement 

13.Problem: Uneven Data Distribution ! Dealing with hotspot ○ Choose a wrong sharding key ! Inefficient usage ○ Some are busy ○ Some are idle ! Caused by the nature of RDBMS 

14.A little bit about how data is organized in TiDB PD PD TSO/Data location Data location PD PD Cluster Metadata Spark TiDB Driver TiKV TiKV DistSQL API DistSQL API Job TiDB Application TiDB Worker SparkSQ KV API TiKV TiKV L Syncer Worker TiDB TiKV TiKV Worker TiDB Spark Cluster TiDB TiKV Cluster Cluster (Storage) TiSpark 

15. Any ORM which ODBC/JDBC MySQL Client supports MySQL MySQL Network Protocol SQL Parser TiDB Cost-based Optimizer Distributed Executor (Coprocessor) TiKV Node1 Node2 Node3 Node4 

16.Why Raft? ! Saft split/merge ! Self-healing Start Follower ! Easy to implement Election Timeout, Start new election. Receive higher Find leader or term msg receive higher term msg Election, re- Candidat campaign Leader e Receive majority vote 

17. Table mapping: Rows => Key-Value pairs (-∞, a) Dataset [a, b) (b, +∞) 

18.Data movement Step 1: 

19.Data movement Step 2: 

20.Placement Driver: The scheduler 

21.

22.Only by Region count? Size also matters. 

23.Hot Balance 

24.Scheduler - More ! More… ○ Weight Balance - High-weight TiKV will save more data ○ Evict Leader Balance - Some TiKV node can’t have any Raft leader ! OpInfluence - Avoid over frequent balancing 

25.Autonomy 

26.

27.Our work: Oasis 

28.Isolation Forest Start API ! Anomaly detection New Job ! Easy to implement ! Paper IForest Get Metrics From Prometheus Model Features Train Error Predict Alert Report Finish 

29.Related work - OtterTune ! Database Tuning-as-a-Service ○ Automatically generates DBMS knob configurations ○ Reuse data from previous tuning sessions ! Supported systems ○ PostgreSQL ○ MySQL ○ Greenplum ○ Vectorwise