Re: Use read streams in CREATE DATABASE command when the strategy is wal_log - Mailing list pgsql-hackers

From Noah Misch
Subject Re: Use read streams in CREATE DATABASE command when the strategy is wal_log
Date
Msg-id 20240720181405.5a.nmisch@google.com
Whole thread Raw
In response to Re: Use read streams in CREATE DATABASE command when the strategy is wal_log  (Nazir Bilal Yavuz <byavuz81@gmail.com>)
Responses Re: Use read streams in CREATE DATABASE command when the strategy is wal_log
List pgsql-hackers
On Sat, Jul 20, 2024 at 03:01:31PM +0300, Nazir Bilal Yavuz wrote:
> On Sat, 20 Jul 2024 at 14:27, Noah Misch <noah@leadboat.com> wrote:
> >
> > On Thu, Jul 18, 2024 at 02:11:13PM +0300, Nazir Bilal Yavuz wrote:
> > > v4 is attached.
> >
> > Removal of the PinBufferForBlock() comment about the "persistence =
> > RELPERSISTENCE_PERMANENT" fallback started to feel like a loss.  I looked for
> > a way to re-add a comment about the fallback.
> > https://coverage.postgresql.org/src/backend/storage/buffer/bufmgr.c.gcov.html
> > shows no test coverage of that fallback, and I think the fallback is
> > unreachable.  Hence, I've removed the fallback in a separate commit.  I've
> > pushed that and your three patches.  Thanks.
> 
> Thanks for the separate commit and push!
> 
> With the separate commit (e00c45f685), does it make sense to rename
> the smgr_persistence parameter of the ReadBuffer_common() to
> persistence? Because, ExtendBufferedRelTo() calls ReadBuffer_common()
> with rel's persistence now, not with smgr's persistence.

BMR_REL() doesn't set relpersistence, so bmr.relpersistence is associated with
bmr.smgr and is unset if bmr.rel is set.  That is to say, bmr.relpersistence
is an smgr_persistence.  It could make sense to change ReadBuffer_common() to
take a BufferManagerRelation instead of the three distinct arguments.

On a different naming topic, my review missed that field name
copy_storage_using_buffer_read_stream_private.last_block doesn't fit how the
field is used.  Code uses it like an nblocks.  So let's either rename the
field or change the code to use it as a last_block (e.g. initialize it to
nblocks-1, not nblocks).



pgsql-hackers by date:

Previous
From: vignesh C
Date:
Subject: Re: Slow catchup of 2PC (twophase) transactions on replica in LR
Next
From: Pavel Stehule
Date:
Subject: Re: proposal: schema variables