ddl: fix placement rules compatibility with tiflash

[xhebox]

What problem does this PR solve?

Issue: close #22171

Problem Summary: When PD scheduels a partition to tiflash, TiFlash will validate the keyrange if there is a _r suffix. Using GenTableRecordPrefix instead of GenTablePrefix can fix this problem.

But placement rules are not compatible with the old set tiflash replica syntax, so the second commit simply forbids using +engine=tiflash, and will add -engine=tiflash to every rule to avoid scheduels to tiflash instance. Maybe that will be reverted in the future.

Check List

Tests

• Unit test
• Integration test
• Manual test (add detailed scripts or steps below)

set global tidb_enable_alter_placement=on;
drop table if exists test.t1;
create table test.t1 (a int) partition by hash(a) partitions 2;
alter table test.t1 alter partition p1 alter placement policy CONSTRAINTS='["+engine=tiflash"]' ROLE=follower REPLICAS=1;
(8234, 'Invalid placement policy \'alter placement policy constraints=\'["+engine=tiflash"]\' role=follower replicas=1\': unsupported label constraint \'+engine=tiflash\'')

Release note

• Forbid adding tiflash replica by placement rules in SQL

[djshow832]

Add some cases to placement_sql_test.go to test if the placement rules generated by SQL are right in these scenarios:

1. Whether the placement rules all contain "-engine=tiflash".
2. When the user also defines "-engine=tiflash", the placement rules should only contain one "-engine=tiflash" constraint.
3. CONSTRAINTS='{"+zone=bj":1}' REPLICAS=2, the extra count should also contain "-engine=tiflash" constraint. (Maybe?)

[xhebox]

Add some cases to placement_sql_test.go to test if the placement rules generated by SQL are right in these scenarios.

@djshow832 I can not add tests in placement_sql_test.go. Mock test wont update rule cache. Putting it into placement_rule_test.go needs too many modifications(and can not cover the test of 3). I would like to leave it to automated-test.

Or maybe manual test in patition_test.go?

[djshow832]

LGTM

[wjhuang2016]

Please create an issue for this problem.

[wjhuang2016]

The first commit fixed that. It is not convenient for a reviewer to find your first commit.
You can write: "Using GenTableRecordPrefix instead of GenTablePrefix can fix this problem."
Besides, is there any test for this change?

[xhebox]

@wjhuang2016 I will try if TestDDLCallback can test it. The problem is that rule cache will not refresh without a working PD. I can not check the result after executing placement rules SQL DDL jobs.

[wjhuang2016]

@wjhuang2016 I will try if TestDDLCallback can test it. The problem is that rule cache will not refresh without a working PD.

You can do a manual test. Start a cluster to test it.

[xhebox]

You can do a manual test. Start a cluster to test it

A manual test is already done, of course.

[wjhuang2016]

LGTM

[wjhuang2016]

/merge

[ti-srebot]

/run-all-tests