Re: Preallocation changes in Postgresql 16

[Riku Iki <riku.iki.x@xxxxxxxxx>]

I did the testing and confirmed that this was the issue.

I run following query:

 create table t as select '1234567890' from generate_series(1, 1000000000);

I commented if (numblocks > 8) codeblock, and see the following results from "compsize /dbdir/" command.


Before my changes:

Processed 1381 files, 90007 regular extents (90010 refs), 15 inline.
Type       Perc     Disk Usage   Uncompressed Referenced  
TOTAL       97%       41G          42G          42G      
none       100%       41G          41G          41G      
zstd        14%      157M         1.0G         1.0G      
prealloc   100%       16M          16M          16M



After the changes:

Processed 1381 files, 347328 regular extents (347331 refs), 15 inline.
Type       Perc     Disk Usage   Uncompressed Referenced  
TOTAL        3%      1.4G          42G          42G      
none       100%       80K          80K          80K      
zstd         3%      1.4G          42G          42G

It is clearly visible that files created with fallocate are not compressed, and disk usage is much larger.

I am wondering if there is a way to have some feature request to have this parameter user configurable..  

On Fri, Apr 26, 2024 at 4:15 PM Riku Iki <riku.iki.x@xxxxxxxxx> wrote:

Thank you, I have such a system. I think my task would be to compile PG from sources(need to learn this), and see how it works with and without that code block.

On Thu, Apr 25, 2024 at 2:25 PM Thomas Munro <thomas.munro@xxxxxxxxx> wrote:

On Fri, Apr 26, 2024 at 4:37 AM Riku Iki <riku.iki.x@xxxxxxxxx> wrote:
> I am wondering if there were preallocation related changes in PG16, and if it is possible to disable preallocation in PostgreSQL 16?

I have no opinion on the btrfs details, but I was wondering if someone
might show up with a system that doesn't like that change.  Here is a
magic 8, tuned on "some filesystems":

        /*
         * If available and useful, use posix_fallocate() (via
         * FileFallocate()) to extend the relation. That's often more
         * efficient than using write(), as it commonly won't cause the kernel
         * to allocate page cache space for the extended pages.
         *
         * However, we don't use FileFallocate() for small extensions, as it
         * defeats delayed allocation on some filesystems. Not clear where
         * that decision should be made though? For now just use a cutoff of
         * 8, anything between 4 and 8 worked OK in some local testing.
         */
        if (numblocks > 8)

I wonder if it wants to be a GUC.