Re: [PATCH] pg_upgrade: add --initdb option to create the new cluster automatically

From: Heikki Linnakangas <hlinnaka(at)iki(dot)fi>
To: shihao zhong <zhong950419(at)gmail(dot)com>, Bohyun Lee <bohyun(dot)lee(at)databricks(dot)com>
Cc: harrisontbooth(at)gmail(dot)com, Greg Sabino Mullane <htamfids(at)gmail(dot)com>, Daniel Gustafsson <daniel(at)yesql(dot)se>, pgsql-hackers(at)lists(dot)postgresql(dot)org
Subject: Re: [PATCH] pg_upgrade: add --initdb option to create the new cluster automatically
Date: 2026-11-20 11:13:47
Message-ID: b65af562-b6f8-434a-879c-513b598e956b@iki.fi
Views: Whole Thread | Raw Message | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

On 28/10/2026 02:41, shihao zhong wrote:
> Hi Bohyun,
>
> Thanks for v10. I think this patch is LGTM

I stepped back a little to think harder about the usability aspect of
this. I think for anyone running pg_upgrade directly, this is a clear
improvement. On that front though, I think we should document the new
"pg_upgrade --initdb" option as the easy, default way to run pg_upgrade.
Maybe even make it the default. Or have pg_upgrade check if the new
cluster already exists and run initdb automatically if it doesn't.

But a bigger issue is this thing that Daniel commented earlier:

> I am not convinced that all this complexity and overhead isn't better suited
> for external wrappers to pg_upgrade like Debian's pg_upgradecluster etc.

I think this can be, and really should be, useful for pg_upgradecluster.
(And I mean also other scripts in other distributions, docker images
etc., but I'll use pg_upgradecluster to stand in for all of those). It
currently has logic to read the options of the old cluster, like locale
and whether checksums are enabled, and build the initdb options to use
based on that. It seems silly that the calling script needs to know how
to do all that. pg_ugprade already has the knowledge, because it already
has checks to verify that you used the right options. So this is a case
of "I know exactly what options you need to use, and I will complain if
you get them wrong, but I won't tell you what those options are until
you try."

This new "pg_upgrade --initdb" option should be able to replace that
logic. If it cannot, if pg_upgradecluster still needs to duplicate the
logic to check the old cluster's options, it makes this patch a lot less
interesting. Unfortunately it seems difficult for pg_upgradecluster to
use this in its current form. I think we need to step back and figure
out how to make it usable for pg_upgradecluster.

The reason pg_upgradecluster cannot readily use this is that it doesn't
fit its workflow very well. It does basically this:

1. Read old cluster's options to determine the initdb options needed
2. Call pg_createcluster with with those options. pg_createcluster calls
initdb
3. Copy config files from the old cluster
4. Run some other upgrade scripts from pg_upgradecluster.d. (Not sure
what those might do)
5. Run pg_upgrade

Using "pg_upgrade --initdb" would require big changes to that workflow.
How can we avoid that?

One idea is to have another option, "pg_upgrade --print-initdb-options",
that just prints the initdb options you need to use to initialize the
cluster. That would replace the current pg_upgradecluster's logic to
figure out those option, but keep the workflow otherwise the same.

Another idea is to have an option like "pg_upgrade --initdb-only", which
just runs initdb with the right options, and stops. You can call that
first, to create the new cluster, do all the configuration steps, and
then call pg_upgrade again to perform the upgrade. That requires a
little more changes to the workflow, but I think it wouldn't be too bad.

I think the current "pg_upgrade --initdb" mode where it just creates the
cluster for you is still useful and convenient for anyone running
pg_upgrade directly, and for simpler scripts that don't have extra
configuration steps like pg_upgradecluster does. So I would keep that
too. But one way or another, more complicated scripts need to have the
control back between running initdb and performing the upgrade.

- Heikki

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Tomas Vondra 2026-11-20 11:14:49 Re: A bug in recovery prefetcher, triggering asserts / FATAL: checkpoint request failed
Previous Message Hayato Kuroda (Fujitsu) 2026-11-20 11:08:25 RE: Proposal: Conflict log history table for Logical Replication