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

On Fri, Feb 18, 2022 at 4:30 AM Robert Haas <robertmhaas@gmail.com> wrote:
>
>
> > This thread is long. Could you summarize what lead you to consider other
> > approaches (e.g. looking in the filesystem for relfilenodes) as not feasible /
> > too ugly / ...?
>
> I don't think it's infeasible to look at the filesystem for files and
> just copy whatever files we find. It's a plausible alternate design. I
> just don't like it as well. I think that relying on the filesystem
> contents to tell us what's going on is kind of hacky. The only
> technical issue I see there is that the WAL logging might require more
> kludgery, since that mechanism is kind of intertwined with
> shared_buffers. You'd have to get the right block references into the
> WAL record, and you have to make sure that checkpoints don't move the
> redo pointer at an inopportune moment.


Actually based on the previous discussion, I also tried to write the
POC with the file system scanning approach to identify the relation to
be copied seet patch 0007 in this thread [1].  And later we identified
one issue [2], i.e. while scanning directly the disk file we will only
know the relfilenode but we can not identify the relation oid that
means we can not lock the relation.  Now, I am not saying that there
is no way to work around that issue but that was also one of the
reasons for not pursuing that approach.

[1] https://www.postgresql.org/message-id/CAFiTN-v1KYsVAhq_fOWFa27LZiw9uK4n4cz5XmQJxJpsVcfq1w%40mail.gmail.com
[2] https://www.postgresql.org/message-id/CAFiTN-v%3DU58by_BeiZruNhykxk1q9XUxF%2BqLzD2LZAsEn2EBkg%40mail.gmail.com

-- 
Regards,
Dilip Kumar
EnterpriseDB: http://www.enterprisedb.com