This change is 

@ajwerner this needs more testing before I'm ready to submit it, but can you take a look at let me know what you think of the overall approach?

Changing the crdb_region column to an enum type is OK, but the behavior doesn't seem right. Once the type has been changed, we couldn't compare with a string anymore:

root@localhost:26257/defaultdb> select * from system.sql_instances where crdb_region = 'gcp-us-central1';
ERROR: unsupported comparison: bytes to crdb_internal_region
SQLSTATE: 42804

On the other hand, for a regular regional by row table, I could still filter by crdb_region, and the DB knows what the possible values are:

jay@proxy-serverless-mr-jay-1-3hh.gcp-us-central1.crdb.io:26257/test> select * from users where crdb_region = 'us-central1';
ERROR: invalid input value for enum crdb_internal_region: "us-central1"
SQLSTATE: 22P02
jay@proxy-serverless-mr-jay-1-3hh.gcp-us-central1.crdb.io:26257/test> select * from users where crdb_region = 'gcp-us-central1';
                   id                  |  city   | first_name | last_name | address
---------------------------------------+---------+------------+-----------+----------
  6f0be85d-4d46-48a5-af3d-b3925b7aaa54 | chicago | foo        | bar       | baz
(1 row)

Yes, this approach is sound

Changing the crdb_region column to an enum type is OK, but the behavior doesn't seem right. Once the type has been changed, we couldn't compare with a string anymore:

root@localhost:26257/defaultdb> select * from system.sql_instances where crdb_region = 'gcp-us-central1';
ERROR: unsupported comparison: bytes to crdb_internal_region
SQLSTATE: 42804

On the other hand, for a regular regional by row table, I could still filter by crdb_region, and the DB knows what the possible values are:

jay@proxy-serverless-mr-jay-1-3hh.gcp-us-central1.crdb.io:26257/test> select * from users where crdb_region = 'us-central1';
ERROR: invalid input value for enum crdb_internal_region: "us-central1"
SQLSTATE: 22P02
jay@proxy-serverless-mr-jay-1-3hh.gcp-us-central1.crdb.io:26257/test> select * from users where crdb_region = 'gcp-us-central1';
                   id                  |  city   | first_name | last_name | address
---------------------------------------+---------+------------+-----------+----------
  6f0be85d-4d46-48a5-af3d-b3925b7aaa54 | chicago | foo        | bar       | baz
(1 row)

Andrew and I debugged this. The error is coming from the optimizer. The statistics for the table contain the type of the column and that type is used when planning the query. Regenerating statistics for the table causes the error to go away. Arguably we are breaking an invariant in the optimizer: the type of a column is changing, but the id of the column is not changing. For now we will work around this by clearing stats in the unsafe_optimize_system_database built in. During tenant creation, erasing stats will be a no-op, since stats will not have been generated by the time the tenant is created.

I removed the draft label. This change is ready for review.

bors r+

Build succeeded:

• Bazel Essential CI (Cockroach)