On Tue, Sep 26, 2017 at 4:57 PM, Tom Lane <
tgl@sss.pgh.pa.us> wrote:
>
> Fabrízio de Royes Mello <
fabriziomello@gmail.com> writes:
> > I didn't came with better solution, but for now what I did is inside
> > transformaAlterTableStmt when calling transformColumnDefinition now we pass
> > down "AlterTableStmt->missing_ok" to check and skip CREATE SEQUENCE
> > statements when use SERIAL pseudo-types.
>
> > It's not an elegant solution because during ATExecAddColumn we check it
> > again by calling check_for_column_name_collision... Ideas are very welcome?
>
> I do not think this is a solution at all. It doesn't address the
> fundamental problem that we decide whether to make a serial sequence
> before determining whether we're going to make a column default
> depending on it.
I tried to address only the ALTER TABLE ... ADD COLUMN IF NOT EXISTS statement, and do not touch CREATE TABLE statements...
For example when we add a new SERIAL column to a relation:
ALTER TABLE foo ADD COLUMN bar SERIAL;
What I understood is actually PostgreSQL will convert it to:
1. CREATE SEQUENCE foo_bar_seq;
2. ALTER TABLE foo ADD COLUMN bar INTEGER DEFAULT nextval('foo_bar_seq');
3. ALTER SEQUENCE foo_bar_seq OWNER BY foo.bar;
And what I tried to implement is skip step 1 and 3... and fo step 2 skip the DEFAULT constraint
> What it does do is introduce a different set of failure
> conditions, basically race conditions around the discrepancy between
> parse-time and execution-time state.
>
I didn't understand what you mean here...
> I don't feel like this is exactly a "must fix" problem, and it certainly
> isn't one that we should fix by introducing different oddities of
> behavior.
>
When I see the code I felt the same... :-(
> We could maybe fix things by arranging to create the sequence only after
> we've made the column successfully, but that would be messy.
If we do that we should add more steps to execution queue...