opt: memo cache is not invalidated when a user-defined type's schema is renamed

[mgartner]

We don't invalidate a cached memo referencing user-defined types when the schema owning the UDT is renamed.

defaultdb> create type sc1.t as enum ('foo', 'bar');
CREATE TYPE

defaultdb> select 'foo'::sc1.t;
t
-------
foo
(1 row)

defaultdb> alter schema sc1 rename to sc2;
ALTER SCHEMA

defaultdb> select 'foo'::sc1.t;
t
-------
foo
(1 row)

The last query should fail because sc1 no longer exists.

Jira issue: CRDB-24272

[ajwerner]

Won't this fall out of drew's work on checking whether the memo is stale? If it doesn't the reason it doesn't would be that we're not tracking a reference to the schema.

[mgartner]

This is caused by the fact that we stored UDT OID's in the memo and lookup the types by OID to determine if they have changed.

cockroach/pkg/sql/opt/metadata.go

Line 335 in e62ac1f

toCheck, err := catalog.ResolveTypeByOID(ctx, typ.Oid())

We need to either:

• Store the original type reference in the memo and look up by that to determine if we have a stale memo.
• Or, track schemas referenced in a memo and check their version to determine if the memo is stale.

[DrewKimball]

This is a problem for tables referenced by IDs, too... I'm going to try adding schemas to the list of objects tracked during memo validation so we can just track function OIDs. I think it should allow us to make a similar simplification for data sources as well. @chengxiong-ruan added a fix for the switching databases problem in functions here; I think we could do something similar for resolving types, schemas, and data sources.

[chengxiong-ruan]

This is a problem for tables referenced by IDs, too... I'm going to try adding schemas to the list of objects tracked during memo validation so we can just track function OIDs. I think it should allow us to make a similar simplification for data sources as well. @chengxiong-ruan added a fix for the switching databases problem in functions here; I think we could do something similar for resolving types, schemas, and data sources.

Yes, we'd better have defense on those OID references as well. Though it's only fun when users know the numeric representation for some reason and spend effort on finding the IDs heh.