Обсуждение: glibc 2.35-2.39 upgrade requirements
Hi everyone, we're planning to upgrade the OS running Postgres from ubuntu jammy to ubuntu noble. As part of the OS change, the glibc version will be changed from glibc 2.35 to glibc 2.39..
Can someone confirm if changing the glibc between those versions will require a full reindex of the Postgres cluster?
Thanks in advance.
Can someone confirm if changing the glibc between those versions will require a full reindex of the Postgres cluster?
Thanks in advance.
On Fri, Jan 17, 2025 at 1:12 AM Kamen Kalchev <kalchev035@gmail.com> wrote:
Hi everyone, we're planning to upgrade the OS running Postgres from ubuntu jammy to ubuntu noble. As part of the OS change, the glibc version will be changed from glibc 2.35 to glibc 2.39..
Can someone confirm if changing the glibc between those versions will require a full reindex of the Postgres cluster?
You never have to reindex _everything_. Only (for some definition of "only") indices with text/char/varchar/name columns need to be rebuilt.
This is how I find such indices:
create or replace view dba.all_indices_types as
select tbcl.relnamespace::regnamespace::text||'.'||tbcl.relname as table_name
, ndcl.relname as index_name
, array_agg(ty.typname order by att.attnum) as index_types
from pg_class ndcl
inner join pg_index nd
on (ndcl.oid = nd.indexrelid and ndcl.relkind in ('i', 'I')
inner join pg_class tbcl
on (nd.indrelid = tbcl.oid and tbcl.relkind in ('r', 'R', 'm'))
inner join pg_attribute att
on att.attrelid = nd.indexrelid
inner join pg_type ty
on att.atttypid = ty.oid
where tbcl.relnamespace::regnamespace::text != 'pg_catalog'
group by tbcl.relnamespace::regnamespace::text||'.'||tbcl.relname
, ndcl.relname
order by 1, 2;
select tbcl.relnamespace::regnamespace::text||'.'||tbcl.relname as table_name
, ndcl.relname as index_name
, array_agg(ty.typname order by att.attnum) as index_types
from pg_class ndcl
inner join pg_index nd
on (ndcl.oid = nd.indexrelid and ndcl.relkind in ('i', 'I')
inner join pg_class tbcl
on (nd.indrelid = tbcl.oid and tbcl.relkind in ('r', 'R', 'm'))
inner join pg_attribute att
on att.attrelid = nd.indexrelid
inner join pg_type ty
on att.atttypid = ty.oid
where tbcl.relnamespace::regnamespace::text != 'pg_catalog'
group by tbcl.relnamespace::regnamespace::text||'.'||tbcl.relname
, ndcl.relname
order by 1, 2;
select * from dba.all_indices_types where index_types && '{"text","varchar","char","text"}';
Death to <Redacted>, and butter sauce.
Don't boil me, I'm still alive.
<Redacted> lobster!
You will want to ingest
https://www.joeconway.com/presentations/glibc-PostgresConfSEA-2024.pdf
Karsten
--
GPG 40BE 5B0E C98E 1713 AFA6 5BC0 3BEA AC80 7D4F C89B
Kamen Kalchev <kalchev035@gmail.com> writes:
> Hi everyone, we're planning to upgrade the OS running Postgres from ubuntu
> jammy to ubuntu noble. As part of the OS change, the glibc version will be
> changed from glibc 2.35 to glibc 2.39..
> Can someone confirm if changing the glibc between those versions will
> require a full reindex of the Postgres cluster?
Maybe, maybe not. According to [1], the last glibc collation change
that the PG community really noticed was in glibc 2.28. So maybe
there weren't any significant changes between 2.35 and 2.39. The
conservative path would certainly be to reindex all textual columns
(though you can skip any that have collation "C").
If you feel a need to try to avoid that, you could dump some of your
textual columns into files and sort those using sort(1) on both
old and new systems. (Be sure that the LANG/LC_xxx environment
matches what you use for the database.) If the results are different
then you definitely need to reindex; if they are the same then maybe
you're okay. Pay particular attention to columns containing
punctuation or non-ASCII characters, as those are the areas most
likely to see changes.
regards, tom lane
[1] https://wiki.postgresql.org/wiki/Locale_data_changes
On Fri, 17 Jan 2025 10:27:04 -0500 Tom Lane <tgl@sss.pgh.pa.us> wrote: > Kamen Kalchev <kalchev035@gmail.com> writes: > > Hi everyone, we're planning to upgrade the OS running Postgres from > > ubuntu jammy to ubuntu noble. As part of the OS change, the glibc > > version will be changed from glibc 2.35 to glibc 2.39.. > > Can someone confirm if changing the glibc between those versions > > will require a full reindex of the Postgres cluster? > > Maybe, maybe not. According to [1], the last glibc collation change > that the PG community really noticed was in glibc 2.28. So maybe > there weren't any significant changes between 2.35 and 2.39. The > conservative path would certainly be to reindex all textual columns > (though you can skip any that have collation "C"). I haven't run 2.39 through the scan yet [1]; I should do that because someone was asking the same question on postgres slack. But note that every single ubuntu LTS and every single RHEL major release in the last 10 years has had collation changes, except for ubuntu 14.04 ... so it's worth being cautious. Collations are a bit like time zones - small changes are always happening, but you might not always notice. Jeff Davis and I did a talk at the last pgconf about this, the recording is online [2]. Personally I would recommend using the builtin C collation as database default starting in pg17, and using ICU to do linguistic collation at the table or query level when needed. With ICU there's at least the option to rebuild old versions on new operating system majors, if needed. (Though rebuilding objects - not just indexes, but anything depending on the collation - is the best course.) And be careful about hot standbys, FDWs, and other places where you can get little surprises with different OS majors. The YouTube recording has lots of info. -Jeremy 1: https://github.com/ardentperf/glibc-unicode-sorting 2: https://www.youtube.com/watch?v=KTA6oau7tl8