Re: head fails to build on SLES 12 (wal_compression=zstd) - Mailing list pgsql-hackers

From Justin Pryzby
Subject Re: head fails to build on SLES 12 (wal_compression=zstd)
Date
Msg-id 20220331224230.GR28503@telsasoft.com
Whole thread Raw
In response to Re: head fails to build on SLES 12 (wal_compression=zstd)  (Tom Lane <tgl@sss.pgh.pa.us>)
List pgsql-hackers
On Thu, Mar 31, 2022 at 11:44:40AM -0400, Tom Lane wrote:
> Justin Pryzby <pryzby@telsasoft.com> writes:
> > Possible responses look like:
> >  - Use 0 which also means "default" (need to verify that works across versions);
> >  - Or #ifndef ZSTD_CLEVEL_DEFAULT #define ZSTD_CLEVEL_DEFAULT 3;
> >  - Add a test for a minimum zstd version v1.3.7.  This may be a good idea for
> >    v15 in any case, since we're using a few different APIs (at least
> >    ZSTD_compress and ZSTD_compressStream2 and execve(zstd)).
> 
> In view of 51c0d186d ("Allow parallel zstd compression"), I agree
> that some clarity about the minimum supported version of zstd
> seems essential.  I don't want to be dealing with threading bugs
> in ancient zstd versions.  However, why do you suggest 1.3.7 in
> particular?

That's where I found that ZSTD_CLEVEL_DEFAULT was added, in their git.

I've just installed a .deb for 1.3.8, and discovered that the APIs used by
basebackup were considered experimental/nonpublic/static-lib-only until 1.4.0
https://github.com/facebook/zstd/releases/tag/v1.4.0
ZSTD_CCtx_setParameter ZSTD_c_compressionLevel ZSTD_c_nbWorkers ZSTD_CCtx_reset ZSTD_reset_session_only
ZSTD_compressStream2ZSTD_e_continue ZSTD_e_end
 

FYI: it looks like parallel, thread workers were also a nonpublic,
"experimental" API until v1.3.7.  Actually, ZSTD_p_nbThreads was renamed to
ZSTD_p_nbWorkers and then (in 1.3.8) renamed to ZSTD_c_nbWorkers.

Versions less than 1.3.8 would have required compile-time conditionals for both
ZSTD_CLEVEL_DEFAULT and ZSTD_p_nbThreads/ZSTD_p_nbWorkers/ZSTD_c_nbWorkers (but
that is moot).

Negative compression levels were added in 1.3.4 (but I think the range of their
levels was originally -1..-7 and now expanded).  And long-distance matching
added in 1.3.2.

cirrus has installed:
linux (debian bullseye) 1.4.8
macos has 1.5.0
freebsd has 1.5.0

debian buster (oldstable) has v1.3.8, which is too old.
ubuntu focal LTS has 1.4.4 (bionic LTS has 1.3.3 which seems too old)
rhel7 has 1.5.2 in EPEL;

SLES 12.5 has zstd 1.3.3, so it won't be supported.  But postgres should fail
gracefully during ./configure rather than during build.

Note that check-world fails if wal_compression is enabled due to two
outstanding issues, so it's not currently possible to set that in CI or
buildfarm...
https://www.postgresql.org/message-id/c86ce84f-dd38-9951-102f-13a931210f52%40dunslane.net

Attachment

pgsql-hackers by date:

Previous
From: Cary Huang
Date:
Subject: Re: pg_receivewal fail to streams when the partial file to write is not fully initialized present in the wal receiver directory
Next
From: Peter Geoghegan
Date:
Subject: Re: should vacuum's first heap pass be read-only?