On 11/23/21, 1:41 PM, "Tom Lane" <tgl@sss.pgh.pa.us> wrote:
> I wrote:
>> I wonder though if we shouldn't try to improve the existing text.
>> The phrasing "never rolled back" seems like it's too easily
>> misinterpreted. Maybe rewrite the <caution> block like
>> ...
>
> A bit of polishing later, maybe like the attached.
The doc updates look good to me. Yesterday I suggested possibly
adding a way to ensure that nextval() called in an uncommitted
transaction was persistent, but I think we'd have to also ensure that
synchronous replication waits for those records, too. Anyway, I don't
think it is unreasonable to require the transaction to be committed to
avoid duplicates from nextval().
Nathan