TiSpark Benchmark

Benchmark with TPC-H

Environment

Machine * 10
* CPU: 8 Intel Xeon Processor (Icelake)
* Memory: 32G
* Disk: 500G

TiDB 5.4.0: 3 TiDB + 3 TiKV + 1PD (TiDB and PD are in the same machine)

Spark 3.0.3 StandAlone: 1 master + 3 worker

Parallel Number

Parallel number depends on the total number of executor cores = 3*8 = 24

Write Benchmark

Write data from HDFS to TiDB with Data generated by TPC-H (ORDERS table)

TiSpark Write bechmark

Count(*)	Data size	Tasknumber	Time(s)
1,500,000	164M	9	62
15,000,000	1.7G	23	396
150,000,000	17G	226	4722

Spark JDBC Write benchmark

Count(*)	Data size	Tasknumber	Time(s)
1,500,000	164M	24	23
15,000,000	1.7G	24	244
150,000,000	17G	133	2483

Delete Benchmark

Delete data from TiDB with TiSpark (ORDERS table)

Count(*)	Data size	Tasknumber	Time(s)
1,500,000	164M	3	31
15,000,000	1.7G	5	269
150,000,000	17G	33	3225

Select Benchmark

Select with TPCH 22 queries and table scan

• Spark JDBC uses default config without partitionColumn, lowerBound, upperBound to partition the table
• TiSpark will partition the table for us automatically

Query	DataSize	TiSpark(s)	Spark JDBC(s)
TPC-H 22 queries	1G	131	157
TPC-H 22 queries	10G	424	1793 ( q21 OOM )
select * from orders	164M	5	10
select * from orders	1.7G	14	89

---

If you want to do a benchmark for TiSpark，here is a reference (Chinese only for now)

Benchmark with TPC-DS

Environment

Machine * 2
* CPU: 48c
* Memory: 187G

TiDB v6.0.0: 3 TiKV

Spark v3.1.3: Local Mode

the first machine run 2 TiKV, the second machine run 1 TiKV and 1 spark

Data

Load 50G TPC-DS Data to TiDB. See here for the detail of data load

Query

Some queries are not compatible with Spark SQL

• change all the date_add(start_date, interval 30 day) to date_add(start_date, 30)
• change alias from 'name' to `name`

Select BenchMark

Execute 99 TPC-DS query on 50G Data

storage	total time(s)
TiSpark on TiKV	7504
TiSpark on TiFlash	2928 (Q5 Fail)