Обсуждение: BUG #17271: Updating enum columns type fails when constraints exist

The following bug has been logged on the website:

Bug reference:      17271
Logged by:          Amaury Dumoulin
Email address:      amaury@castordoc.com
PostgreSQL version: 12.8
Operating system:   alpine 3.14 on docker (host macOs 11.6)
Description:

Error message upon failure

The following fails

CREATE TYPE direction_state AS ENUM('UP', 'DOWN');
CREATE TABLE example (id SERIAL PRIMARY KEY, direction direction_state NOT
NULL, below BOOLEAN NOT NULL);
ALTER TABLE example ADD CONSTRAINT ck_example_direction CHECK ((direction =
'DOWN' AND below) OR (NOT below));
ALTER TYPE direction_state RENAME TO direction_state_old;
CREATE TYPE direction_state AS ENUM('UP', 'DOWN', 'UNKNOWN');
ALTER TABLE example ALTER COLUMN direction TYPE direction_state USING
direction::text::direction_state;
DROP TYPE direction_state_old

With a sibyllin error message
ERROR:  operator does not exist: direction_state = direction_state_old
HINT:  No operator matches the given name and argument types. You might need
to add explicit type casts.

If we drop and recreate the constraint it works

CREATE TYPE direction_state AS ENUM('UP', 'DOWN');
CREATE TABLE example (id SERIAL PRIMARY KEY, direction direction_state NOT
NULL, below BOOLEAN NOT NULL);
ALTER TABLE example ADD CONSTRAINT ck_example_direction CHECK ((direction =
'DOWN' AND below) OR (NOT below));
ALTER TYPE direction_state RENAME TO direction_state_old;
CREATE TYPE direction_state AS ENUM('UP', 'DOWN', 'UNKNOWN');
ALTER TABLE example DROP CONSTRAINT ck_example_direction;
ALTER TABLE example ALTER COLUMN direction TYPE direction_state USING
direction::text::direction_state;
DROP TYPE direction_state_old
ALTER TABLE example ADD CONSTRAINT ck_example_direction CHECK ((direction =
'DOWN' AND below) OR (NOT below));

"David G. Johnston" <david.g.johnston@gmail.com> writes:
> On Thu, Nov 4, 2021 at 6:54 AM PG Bug reporting form <noreply@postgresql.org>
> wrote:
>> ERROR:  operator does not exist: direction_state = direction_state_old

> This does not surprise me - why do you believe it is a bug?  What should
> happen?

Yeah.  Concretely, what you have in the constraint (after the RENAME
TYPE) is

# \d example
...
Check constraints:
    "ck_example_direction" CHECK (direction = 'DOWN'::direction_state_old AND below OR NOT below)

Simply updating "direction" to some other type doesn't provide any
guidance about changing the type of the constant, so re-parsing the
constraint fails.  It's hard to see what we could do here that
wouldn't involve a lot of guesswork and chances of getting things
wrong.

I concur with David's advice that ALTER TYPE ADD VALUE would
be a better answer, at least for the specific use-case you're
showing here.

            regards, tom lane

Amaury Dumoulin <amaury@castordoc.com> writes:
> I believe it should fail.
> However the error message should state that the constraint is blocking the
> operation.
> The current error message makes it hard to make a connection to the culprit.

Hmm, that's a reasonable concern, though you didn't make it clear that
that was what you were on about.  We could possibly provide an addendum
like "CONTEXT: while processing constraint foo" in this situation.

            regards, tom lane