Re: unlogged sequences

On Thu, Jun 20, 2019 at 09:30:34AM +0200, Peter Eisentraut wrote:
> The discussion in bug #15631 revealed that serial/identity sequences of
> temporary tables should really also be temporary (easy), and that
> serial/identity sequences of unlogged tables should also be unlogged.
> But there is no support for unlogged sequences, so I looked into that.

Thanks for doing so.

> If you copy the initial sequence relation file to the init fork, then
> this all seems to work out just fine. Attached is a patch. The
> low-level copying seems to be handled quite inconsistently across the
> code, so I'm not sure what the most appropriate way to do this would be.
> I'm looking for feedback from those who have worked on tableam and
> storage manager to see what the right interfaces are or whether some new
> interfaces might perhaps be appropriate.

But the actual deal is that the sequence meta-data is now in
pg_sequences and not the init forks, no? I have just done a small
test:
1) Some SQL queries:
create unlogged sequence popo;
alter sequence popo increment 2;
select nextval('popo');
select nextval('popo');
2) Then a hard crash:
pg_ctl stop -m immediate
pg_ctl start
3) Again, with a crash:
select nextval('popo');
#2 0x000055ce60f3208d in ExceptionalCondition
(conditionName=0x55ce610f0570 "!(((PageHeader) (page))->pd_special >=
(__builtin_offsetof (PageHeaderData, pd_linp)))",
errorType=0x55ce610f0507 "FailedAssertion",
fileName=0x55ce610f04e0 "../../../src/include/storage/bufpage.h",
lineNumber=317) at assert.c:54
#3 0x000055ce60b43200 in PageValidateSpecialPointer
(page=0x7ff7692b3d80 "") at
../../../src/include/storage/bufpage.h:317
#4 0x000055ce60b459d4 in read_seq_tuple (rel=0x7ff768ad27e0,
buf=0x7ffc5707f0bc, seqdatatuple=0x7ffc5707f0a0) at
sequence.c:1213
#5 0x000055ce60b447ee in nextval_internal (relid=16385,
check_permissions=true) at sequence.c:678
#6 0x000055ce60b44533 in nextval_oid (fcinfo=0x55ce62537570) at sequence.c:607
--
Michael