PingCAP-Infra-Meetup-94-sunruoxi-TiFlash

1 . Welcome! 加入 Infra Meetup No.XX 交流群 和大家一起讨论吧～ 我们将分享本期 Meetup 资料 

2 .TiDB with TiFlash Extension A Venture Towards the True HTAP Platform 

3 .About Me ● Sun Ruoxi 孙若曦 ● Senior expert, Analytical Product@PingCAP ● Previously ○ Tech lead, SQL on Hadoop team@Transwarp ○ Tech lead, Arch Infra team@NVIDIA ● Mainly working on Big-data / SQL related things 

4 .Traditional Data Platform Data Warehouse OLTP OLTP Hadoop OLTP Data Lake Big Data ETL Compute Engine NoSQL Analytical Database Traditional data platform relies on complex architecture moving data around via ETL. This introduces maintenance cost and delay of data arrival in data warehouse. 

5 .Why VS 

6 .Fundamental Conflicts ● Large / batch process vs point / short access ○ Row format for OLTP ○ Columnar format for OLAP ● Workload Interference ○ A single large analytical query might cause disaster for your OLTP workload 

7 .A Popular Solution ● Use different types of databases ○ For live and fast data, use an OLTP specialized database or NoSQL ○ For historical data, use Hadoop / analytical database ● Offload data via the ETL process into your Hadoop cluster or analytical database ○ Maybe per day 

8 .Good enough, really? 

9 .Complexity or 

10 .Freshness or 

11 .Consistency 

12 .TiFlash Extension 

13 .What’s TiFlash Extension ● TiFlash is an extended analytical engine for TiDB ● Powered by columnar storage and vectorized compute engine ○ Speed up OLAP workload for both TiSpark and TiDB ○ Partially based on ClickHouse with tons of proprietary modifications ● Data sync via raft learner ○ Strong consistency ○ Trivial overhead ● Clear Isolation of workload not impacting OLTP ● Tightly integrated with TiDB 

14 .Architecture Spark Cluster TiDB TiSpark TiSpark TiDB Worker Worker TiFlash Node 2 TiFlash Node 1 TiKV Node 1 TiKV Node 2 TiKV Node 3 Store 1 Store 2 Store 3 Region 1 Region 4 Region 2 Region 2 Region 3 Region 3 Region 3 Region 2 Region 4 Region 4 Region 1 Region 1 TiFlash Extension Cluster TiKV Cluster 

15 .Columnstore vs Rowstore ● Columnar Storage stores data in columns instead of rows ○ Suitable for analytical workload ■ Possible for column pruning ○ Compression made possible and further IO reduction ■ ⅕ on average storage requirement ○ Bad small random read and write ■ Which is the typical workload for OLTP ● Rowstore is the classic format for databases ○ Researched and optimized for OLTP scenario for decades ○ Cumbersome in analytical use cases 

16 .Columnstore vs Rowstore Rowstore SELECT avg(age) from emp; id name age 0962 Jane 30 7658 John 45 3589 Jim 20 5523 Susan 52 Columnstore id name age 0962 Jane 30 7658 John 45 Usually you don’t read all columns in a table performing analytics. 3589 Jim 20 5523 Susan 52 In columnstore, you avoid unnecessary IO while you have to read them all in rowstore. 

17 .Columnstore vs Rowstore “If your mother and your wife fell into a river at the same time, who would you save?” “Why not both?” 

18 .Raft Learner TiFlash synchronizes data in columnstore via Raft Learner ● Strong consistency on read enabled by the Raft protocol ● Introduce almost zero overhead for the OLTP workload ○ Except the network overhead for sending extra replicas ○ Slight overhead on read (check Raft index for each region in 96 MB by default) ○ Possible for multiple learners to speed up hot data read 

19 .Raft Learner TiKV TiFlash Region A R e g i o n A Instead of connecting as a Raft Follower, TiKV TiKV regions in TiFlash act as Raft Learner. Region A Region A When data is written, Raft leader does not wait for learner to finish writing. Therefore, TiFlash introduces almost no overhead replicating data. 

20 .Raft Learner Raft Leader 4 Raft Learner When being read, Raft Learner sends request to check the Raft log index with 3 Leader to see if its data is up-to-date. 

21 .Raft Learner Raft Leader 4 Raft Learner After data catches up via Raft log, Learner serves the read request 4 then. 

22 .Scalability ● An HTAP database needs to store huge amount of data ● Scalability is very important ● TiDB relies on multi-raft for scalability ○ One command to add / remove node ○ Scaling is fully automatic ○ Smooth and painless data rebalance ● TiFlash adopts the same design 

23 .Isolation ● Perfect Resource Isolation ● Data rebalance based on the “label” mechanism ○ Dedicated nodes for TiFlash / Columnstore ○ Nodes are differentiated by “label” ■ Raft Learner only exists within labeled nodes ■ Rebalance between labeled nodes ● Computation Isolation is possible by nature ○ Use a different set of compute nodes ○ Read only from nodes with AP label 

24 .Isolation Region 1 AP label constrained Peer 1 Label: TP TiDB / Peer 2 Label: TP TiSpark Peer 3 Label: TP Peer 4 Label: AP Label: AP Label: TP TiFlash Node 2 TiFlash Node 1 TiKV Node 1 TiKV Node 2 TiKV Node 3 Store 1 Store 2 Store 3 Region 1 Region 4 Region 2 Region 2 Region 3 Region 3 Region 3 Region 2 Region 4 Region 4 Region 1 Region 1 TiFlash Extension Cluster TiKV Cluster 

25 .Integration ● Tightly Integrated Interaction ○ TiDB / TiSpark might choose to read from either side ■ Based on cost ○ Upon TiFlash replica failure, read TiKV replica transparently ○ Join data from both sides in a single query 

26 . Integration SELECT AVG(s.price) FROM prod p, sales s TiDB WHERE p.pid = s.pid / AND p.batch_id = ‘B1328’; TiSpark Index Scan(batch_id = B1328) TableScan(price,pid) TiFlash Node 2 TiFlash Node 1 TiKV Node 1 TiKV Node 2 TiKV Node 3 Store 1 Store 2 Store 3 Region 1 Region 4 Region 2 Region 2 Region 3 Region 3 Region 3 Region 2 Region 4 Region 4 Region 1 Region 1 TiFlash Extension Cluster TiKV Cluster 

27 .MPP Support ● TiFlash nodes form a MPP cluster by themselves ● Full computation support on MPP layer ○ Speed up TiDB since it is not MPP design ○ Speed up TiSpark by avoiding writing disk during shuffle 

28 .MPP Support TiDB / Coordinator TiSpark TiFlash nodes exchange data and enable complex operators like distributed join. Plan Segment TiFlash Node 1 TiFlash Node 2 TiFlash Node 3 MPP Worker MPP Worker MPP Worker 

29 .Performance ● Comparable performance against vanilla Spark on Hadoop + Parquet ○ Benchmarked with Pre-Alpha version of TiFlash + Spark (without MPP support) ○ TPC-H 100