Index creation in TiDB 5.0 is much slower than 4.0.

[LittleFall]

The schrodinger/sqllogic test is about 4 times slower than 4.0 on 5.0.

https://internal.pingcap.net/idc-jenkins/blue/organizations/jenkins/tiflash_schrodinger_test/activity/

https://pingcap.slack.com/archives/CQ3EL6Q95/p1616368188001100

I tested in idc81, for the same sqllogic binary, the 4.0 cluster is many times faster than the 5.0 cluster, no matter mpp is enabled.

[LittleFall]

I tested in idc81 with the modified test binary:

1. without create index or create unique index statement: commit https://github.com/LittleFall/schrodinger-test/commit/971ef6c94881f847468e6dc995f73df46b642f87

There is little difference between 4.0 and 5.0 qps, both several hundred(with fluctuation over time)

noindex4.log
noindex5.log

2. without tiflash replica, without query, only ddl and insert statement: commit https://github.com/LittleFall/schrodinger-test/commit/c2c423d9a1027641a43a2c22d424da3eef0f499f

4.0 qps 100+
5.0 qps 2

noflash-onlyindex4.log
noflash-onlyindex5.log

[LittleFall]

A minimal and generally reproduce

tiup playground v5.0.0 --tiflash 0

mysql> create table t(a int);
Query OK, 0 rows affected (0.12 sec)

mysql> create index i on t(a);
Query OK, 0 rows affected (2.83 sec)

tiup playground v4.0.12 --tiflash 0

mysql> create table t(a int);
Query OK, 0 rows affected (0.08 sec)

mysql> create index i on t(a);
Query OK, 0 rows affected (0.36 sec)

[AilinKid]

grep sleep before DDL finishes to make async commit and 1PC safe tidb.log

[LittleFall]

sleep added in pingcap/tidb#23364

[zimulala]

I think we can explain roughly in the document, at what-is-the-processing-time-of-a-ddl-operation. Related to pingcap/docs-cn#6020