BUG #17428: last_value incorrect for uninitialized sequence
The following bug has been logged on the website:
Bug reference: 17428
Logged by: Glen Edmonds
Email address: glen.edmonds@gmail.com
PostgreSQL version: 14.2
Operating system: MacOS Monterey 12.2.1
Description:
For a freshly defined sequence, the following:
select last_value from mytable_id_seq
should return 0, but returns 1.
Reasoning: In every case except the uninitialised case, last_value is the
same as the current value of the sequence, which is the same as the number
of rows in the table (assuming no deletions). Logically, if there are no
rows in the table, last_value should return 0 to be consistent.
To reproduce:
create table mytable (
id serial,
other int
);
select (select count(*) from mytable), last_value from mytable_id_seq; -- 0,
1
insert into mytable (other) values (0);
select (select count(*) from mytable), last_value from mytable_id_seq; -- 1,
1
insert into mytable (other) values (0);
select (select count(*) from mytable), last_value from mytable_id_seq; -- 2,
2
insert into mytable (other) values (0);
select (select count(*) from mytable), last_value from mytable_id_seq; -- 3,
3
-- etc
As you can see, only the first row returns different results for the same
expression.
This isn't just theoretical. I was writing some DB units tests and this
caused my code to break, but was also a surprise.
IMHO this is a bug.
On Mon, Mar 7, 2022 at 12:15 AM PG Bug reporting form <
noreply@postgresql.org> wrote:
The following bug has been logged on the website:
Bug reference: 17428
Logged by: Glen Edmonds
Email address: glen.edmonds@gmail.com
PostgreSQL version: 14.2
Operating system: MacOS Monterey 12.2.1
Description:For a freshly defined sequence, the following:
select last_value from mytable_id_seq
should return 0, but returns 1.
One seems as good a choice as zero if a non-null value is to be returned.
"Also, last_value will reflect the latest value reserved by any session,
whether or not it has yet been returned by nextval."
I don't understand how last_value can return anything if nextval hasn't
been called at least one. So if there is a bug here it's that the value
one is returned instead of null.
Zero is a valid value for the sequence, just change the start value for a
newly created sequence to zero.
Reasoning: In every case except the uninitialised case, last_value is the
same as the current value of the sequence,
which is the same as the number
of rows in the table (assuming no deletions).
This point doesn't seem relevant. That isn't what the value represents so
basing the premise of a bug report on it is wrong.
David J.
"David G. Johnston" <david.g.johnston@gmail.com> writes:
On Mon, Mar 7, 2022 at 12:15 AM PG Bug reporting form <
noreply@postgresql.org> wrote:For a freshly defined sequence, the following:
select last_value from mytable_id_seq
should return 0, but returns 1.
One seems as good a choice as zero if a non-null value is to be returned.
The larger point here is that any change is much more likely to
break applications expecting the historical behavior than it is
to make anyone's life better. In a green field I'd tend to
agree that returning NULL (and dispensing with is_called) would
be a better design, but that opportunity was missed decades ago.
regards, tom lane
David, Tom,
The problem really is that the two states of not initialized and post first
usage are indistinguishable (both return 1), but the sequence next value is
different. ie it’s a reasonable expectation that the next value is
last_value + 1. IMHO violating this makes it a bug. Further, this situation
means there’s no query that can be run on the sequence that can determine
if the next value is 1 or 2.
As for existing code relying of current behaviour, given that current
behaviour can’t be relied upon to predict the next value, I don’t think
it’s possible for any usage to rely on 1 being returned when uninitialized.
I agree though that returning null is a better choice for uninitialized
sequences than returning 0.
This whole line of enquiry came about because currval('mytable_id_seq')
explodes when called on an uninitialized sequence (another surprise - I was
expecting null); how about fixing that too? If that was fixed is would
provide a work around for the lack of distinction between the two states of
next value being 1 and next value being 2.
Regards,
Glen
On Tue, 8 Mar 2022 at 1:48 am, Tom Lane <tgl@sss.pgh.pa.us> wrote:
Show quoted text
"David G. Johnston" <david.g.johnston@gmail.com> writes:
On Mon, Mar 7, 2022 at 12:15 AM PG Bug reporting form <
noreply@postgresql.org> wrote:For a freshly defined sequence, the following:
select last_value from mytable_id_seq
should return 0, but returns 1.One seems as good a choice as zero if a non-null value is to be returned.
The larger point here is that any change is much more likely to
break applications expecting the historical behavior than it is
to make anyone's life better. In a green field I'd tend to
agree that returning NULL (and dispensing with is_called) would
be a better design, but that opportunity was missed decades ago.regards, tom lane
Glen Edmonds <glen.edmonds@gmail.com> writes:
The problem really is that the two states of not initialized and post first
usage are indistinguishable (both return 1), but the sequence next value is
different.
They're not indistinguishable: if you look at the is_called flag
you'll see that it changes.
ie it’s a reasonable expectation that the next value is
last_value + 1. IMHO violating this makes it a bug.
You can call it that if you like, but it's operating as designed
and documented. I fear it's about twenty years too late to propose
a redesign.
regards, tom lane
On Mon, Mar 7, 2022 at 8:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Glen Edmonds <glen.edmonds@gmail.com> writes:
The problem really is that the two states of not initialized and post
first
usage are indistinguishable (both return 1), but the sequence next value
is
different.
They're not indistinguishable: if you look at the is_called flag
you'll see that it changes.Correct, the boolean is a superior solution to defining null to mean
something. A bit more verbose but equally functional and easier to read.
ie it’s a reasonable expectation that the next value is
last_value + 1. IMHO violating this makes it a bug.
You can call it that if you like, but it's operating as designed
and documented. I fear it's about twenty years too late to propose
a redesign.
And next_value isn't last_value + 1 in any case since we have a block
caching feature. One cannot predict (i.e., without consuming) the
next_value from current state.
David J.