Re: Alter column to type serial

Hi all,

Would it be possible (or reasonable) to add support for changing the type of
a column to serial or bigserial (yes, yes, I know they're not actual
types)? In effect this would mean that users who forgot to set up a
sequence could change it's type so that a new implicit sequence will be
created, set with its current value set to the highest value of whatever
column it was bound to. This thought was triggered by a user on IRC wishing
to migrate from MySQL, but had tables with some sort of ID column without
any associated sequence.

So if you had:

CREATE TABLE stuff (id int, content text);

INSERT INTO stuff (id, content) values (1,'alpha'),(2,'beta'),(5,'gamma');

You could just issue:

ALTER TABLE stuff ALTER COLUMN id TYPE serial;

And continue as so:

INSERT INTO stuff (content) values ('delta');

SELECT id from stuff;

id
----
1
2
5
6
(4 rows)

This would be instead of having to do:

CREATE SEQUENCE id_stuff_seq;

SELECT setval('id_stuff_seq', (SELECT max(id) FROM stuff))

ALTER TABLE stuff ALTER COLUMN id SET DEFAULT
nextval('id_stuff_seq'::regclass);

Which would also mean the sequence would not get dropped with the table.

Abhorrent idea, or acceptable?

--
Thom Brown
Twitter: @darkixion
IRC (freenode): dark_ixion
Registered Linux user: #516935

Thom Brown <thom(at)linux(dot)com> writes:
> Would it be possible (or reasonable) to add support for changing the type of
> a column to serial or bigserial (yes, yes, I know they're not actual
> types)?

We've looked at that in the past and decided there were enough corner
cases that it wasn't clearly a good idea. In particular, what do you do
with the existing data in the column? What do you do if there's already
a DEFAULT expression for the column, throw it away? In particular, what
of the special case that the column is in fact already a serial, so the
default is pointing at an existing sequence?

It is possible to accomplish everything that such a command would do
manually, so the argument for having it boils down to wanting it to
be a bit easier. But unless the command can always do the right thing
automatically, I'm not sure "easy" is a good argument.

There's also the objection that such an operation would actually have
very little to do with ALTER COLUMN TYPE --- most of the things it would
do are not that. The fact that serial was bolted on as a fake type is a
wart that maybe we shouldn't extend in this particular fashion.

regards, tom lane

On 4 November 2010 14:04, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Thom Brown <thom(at)linux(dot)com> writes:
> > Would it be possible (or reasonable) to add support for changing the type
> of
> > a column to serial or bigserial (yes, yes, I know they're not actual
> > types)?
>
> We've looked at that in the past and decided there were enough corner
> cases that it wasn't clearly a good idea. In particular, what do you do
> with the existing data in the column? What do you do if there's already
> a DEFAULT expression for the column, throw it away? In particular, what
> of the special case that the column is in fact already a serial, so the
> default is pointing at an existing sequence?
>
> It is possible to accomplish everything that such a command would do
> manually, so the argument for having it boils down to wanting it to
> be a bit easier. But unless the command can always do the right thing
> automatically, I'm not sure "easy" is a good argument.
>
> There's also the objection that such an operation would actually have
> very little to do with ALTER COLUMN TYPE --- most of the things it would
> do are not that. The fact that serial was bolted on as a fake type is a
> wart that maybe we shouldn't extend in this particular fashion.
>
>
I suspected this may have been discussed previously, I just failed to find
it. And yes, it's purely for simplification, and to auto-clean sequences
when tables are dropped. I didn't think it would be straightforward, but
clearly there are show-stoppers abound.

Thanks for the reply though. :)

--
Thom Brown
Twitter: @darkixion
IRC (freenode): dark_ixion
Registered Linux user: #516935

Thom Brown <thom(at)linux(dot)com> writes:
> I suspected this may have been discussed previously, I just failed to find
> it. And yes, it's purely for simplification, and to auto-clean sequences
> when tables are dropped. I didn't think it would be straightforward, but
> clearly there are show-stoppers abound.

The latest thread I can find on the matter is
http://archives.postgresql.org/pgsql-general/2008-11/msg00825.php

although I seem to remember others.

regards, tom lane

Excerpts from Thom Brown's message of jue nov 04 09:05:01 -0300 2010:

> This would be instead of having to do:
>
> CREATE SEQUENCE id_stuff_seq;
>
> SELECT setval('id_stuff_seq', (SELECT max(id) FROM stuff))
>
> ALTER TABLE stuff ALTER COLUMN id SET DEFAULT
> nextval('id_stuff_seq'::regclass);
>
> Which would also mean the sequence would not get dropped with the table.

You can fix that with an ALTER SEQUENCE OWNED BY.

> Abhorrent idea, or acceptable?

I think the problem is in locking the table against futher insertions
while you do the setval.

--
Álvaro Herrera <alvherre(at)commandprompt(dot)com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support