Re: [Proposal] Fully WAL logged CREATE DATABASE - No Checkpoints

Andres Freund <andres@anarazel.de> writes:
> On August 4, 2022 4:11:13 PM PDT, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> [pile^2]  Also, what is the rationale for locking the target buffer
>> but not the source buffer?  That seems pretty hard to justify from
>> here, even granting the assumption that we don't expect any other
>> processes to be interested in these buffers (which I don't grant,
>> because checkpointer).

> I'm not arguing it's good or should stay that way, but it's probably okayish that checkpointer / bgwriter have
access,given that they will never modify buffers. They just take a lock to prevent concurrent modifications, which
RelationCopyStorageUsingBufferhopefully doesn't do.  

I'm not arguing that it's actively broken today --- but AFAIR,
every other access to a shared buffer takes a buffer lock.
It does not seem to me to be very future-proof for this code to
decide it's exempt from that rule, without so much as a comment
justifying it.  Furthermore, what's the gain?  We aren't expecting
contention here, I think.  If we were, then it probably *would* be
actively broken.

            regards, tom lane