From: Tom Lane <tgl@sss.pgh.pa.us>
To: bokanist@gmail.com
cc: pgsql-hackers@postgresql.org
Subject: Re: [Feature Request] INSERT FROZEN to Optimize Large Cold Data Imports and Migrations
In-reply-to: <CANtq+vS0KH9GPSjkQ65VmTcaWwPr3D+C2HNK1dN=C06xdvsL=g@mail.gmail.com>
References: <CANtq+vTuAsz9kX83rcO0-xJRN-rX2APW8kFmKO0p-JfQSsFMKA@mail.gmail.com> <CANtq+vS0KH9GPSjkQ65VmTcaWwPr3D+C2HNK1dN=C06xdvsL=g@mail.gmail.com>
Comments: In-reply-to =?UTF-8?Q?S=C3=A9bastien?= <bokanist@gmail.com>
	message dated "Thu, 13 Feb 2025 10:52:31 +0100"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <3588398.1739459324.1@sss.pgh.pa.us>
Content-Transfer-Encoding: quoted-printable
Date: Thu, 13 Feb 2025 10:08:44 -0500
Message-ID: <3588399.1739459324@sss.pgh.pa.us>
Archived-At: <https://www.postgresql.org/message-id/3588399.1739459324%40sss.pgh.pa.us>
Precedence: bulk

=3D?UTF-8?Q?S=3DC3=3DA9bastien?=3D <bokanist@gmail.com> writes:
> Implementation details:

>    - A new INSERT FROZEN option could be introduced, similar to COPY FRE=
EZE,
>    allowing direct insertion of tuples in a frozen state.
>    - This would likely require changes in heap storage logic to ensure
>    tuples are written with a frozen XID at insert time.
>    - Consideration should be given to transaction semantics and WAL logg=
ing
>    to ensure consistency and crash recovery integrity.

That last is exactly why this won't happen.  A frozen tuple would be
considered committed and visible the instant it appears in the table,
thus completely breaking both atomicity and integrity of the
transaction.

There has been work going on recently to reduce the impact of freezing
massive amounts of data by spreading the work more effectively [1].
I don't say that that particular commit has completely solved the
problem, but I think that continued effort in that direction is more
likely to yield usable results than what you're suggesting.

BTW, this might or might not be usable in your particular workflow,
but: there have long been some optimizations for data load into a
table created in the same transaction.  The idea there is that if the
transaction rolls back, the table will never have been visible to any
other transaction at all, so that maintaining atomicity/integrity of
its contents is moot.

			regards, tom lane

[1] https://git.postgresql.org/gitweb/?p=3Dpostgresql.git&a=3Dcommitdiff&h=
=3D052026c9b