On 7/27/21, 6:05 PM, "Alvaro Herrera" <alvherre@alvh.no-ip.org> wrote:
> On 2021-Feb-19, Bossart, Nathan wrote:
>
>> 0002 adds logic for persisting the last notified segment through
>> crashes. This is needed because a poorly-timed crash could otherwise
>> cause us to skip marking segments as ready-for-archival altogether.
>> This file is only used for primary servers, as there exists a separate
>> code path for marking segments as ready-for-archive for standbys.
>
> I'm not sure I understand what's the reason not to store this value in
> pg_control; I feel like I'm missing something. Can you please explain?
Thanks for taking a look.
The only reason I can think of is that it could make back-patching
difficult. I don't mind working on a version of the patch that uses
pg_control. Back-patching this fix might be a stretch, anyway.
> There were some comments earlier in the thread about the maximum size of
> a record. As I recall, you can have records of arbitrary size if you
> have COMMIT with a large number of relation invalidation messages being
> included in the xlog record, or a large number of XIDs of
> subtransactions in the transaction. Spanning several segments is
> possible, AFAIU.
This is my understanding, too.
Nathan
