MIME-Version: 1.0
References: <20250422181006.dd6f9d1d81299f5b2ad55e1a@sraoss.co.jp>
 <kp43uuykq64fyqxnlihq2rzntarextvzibuzturajfucl6bayo@6fcnflboemtt>
 <20250624170533.3caeec7d6034cd5ddf4f00d2@sraoss.co.jp>
 <20250801002830.143b25971fb9594b89d96aee@sraoss.co.jp>
 <20250808122125.e9eaff938f8c83556f337e50@sraoss.co.jp>
 <20250820141028.dc06e0d1d787ce1151f69ed6@sraoss.co.jp>
 <20250902163341.c174d0e83eedfc54d68b8e9c@sraoss.co.jp>
 <20251231194455.5975fda68ed8e941e4afb805@sraoss.co.jp>
In-Reply-To: <20251231194455.5975fda68ed8e941e4afb805@sraoss.co.jp>
From: Dean Rasheed <dean.a.rasheed@gmail.com>
Date: Tue, 24 Mar 2026 09:18:06 +0000
Message-ID: 
 <CAEZATCV6=q0-5HFBMg4jFS-j1ZCwANhp1sLXBQZYsK2ZLwq0AQ@mail.gmail.com>
Subject: Re: Allow to collect statistics on virtual generated columns
To: Yugo Nagata <nagata@sraoss.co.jp>
Cc: Andres Freund <andres@anarazel.de>, pgsql-hackers@postgresql.org
Content-Type: text/plain; charset="UTF-8"
Archived-At: 
 <https://www.postgresql.org/message-id/CAEZATCV6%3Dq0-5HFBMg4jFS-j1ZCwANhp1sLXBQZYsK2ZLwq0AQ%40mail.gmail.com>
Precedence: bulk

On Wed, 31 Dec 2025 at 10:45, Yugo Nagata <nagata@sraoss.co.jp> wrote:
>
> I've attached an updated patch that fixes the broken test since 10c4fe074a.
>

I took a look at this, and the patch appears to work as intended.
However, I have my doubts as to whether this is the best approach.

Building stats on a virtual generated column is potentially quite
expensive, and not something that everyone will want, so I think this
really should be an optional feature that people can selectively
enable, if they want.

Therefore, I think that the previous approach was probably better. If
I'm understanding it correctly, that allowed

  CREATE STATISTICS stat_name ON virt_col FROM tbl;

as well as allowing statistics to be built on expressions including
virtual generated columns, making it more flexible.

The problem with this previous approach was that it didn't work
correctly if a virtual generated column's expression was changed using
ALTER TABLE ... SET EXPRESSION. I think that could be solved by
expanding generated column expressions at ANALYZE time, rather than at
CREATE STATISTICS time. So then the expression stored in the catalogs
would be one referring to virtual generated columns, not their
expanded forms (compare a CHECK constraint referring to a virtual
generated column).

Regards,
Dean