Date: Mon, 23 Oct 2023 06:07:10 -0700
From: Noah Misch <noah@leadboat.com>
To: Donghang Lin <donghanglin@gmail.com>
Cc: Peter Geoghegan <pg@bowt.ie>, pgsql-hackers@postgresql.org
Subject: Re: interval_ops shall stop using btequalimage (deduplication)
Message-ID: <20231023130710.00.nmisch@google.com>
References: <20231011013317.22.nmisch@google.com>
 <CAH2-Wzm_OFH4Aip3uinAJ-AoBgX0wGjB9XBhrhkTnjCoZ_i8NA@mail.gmail.com>
 <20231011032908.35@rfd.leadboat.com>
 <CAH2-WzmDbyMj_7Ev5vPmt7UaKjuRg6m79sRguTjehyZtuBPMgw@mail.gmail.com>
 <CAH2-WzmCppEF8UFUVLbL3SXLD3UwB-5sDV2YC99soN6O0ScWpg@mail.gmail.com>
 <20231011183835.32@rfd.leadboat.com>
 <CAH2-WzmcUFVB91ARCU5Dg8TPEifxYG_0iaOwsiDiu5Y01b8_cQ@mail.gmail.com>
 <20231012231009.bf.nmisch@google.com>
 <CAH2-Wzm2ydcgC0x-=tOpJc5jYbaw63ntG5Z2Ro1fnPEphdzDeg@mail.gmail.com>
 <CAA=D8a0J2N-waEPmbqT7xkySNZUbFnX5O80mnLxTxALBY2VcfQ@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: 
 <CAA=D8a0J2N-waEPmbqT7xkySNZUbFnX5O80mnLxTxALBY2VcfQ@mail.gmail.com>
User-Agent: Mutt/2.2.12 (2023-09-09)
Archived-At: 
 <https://www.postgresql.org/message-id/20231023130710.00.nmisch%40google.com>
Precedence: bulk

On Mon, Oct 16, 2023 at 11:21:20PM -0700, Donghang Lin wrote:
> > I've also caught btree posting lists where one TID refers to a '1d' heap
> > tuple, while another TID refers to a '24h' heap tuple.  amcheck complains.
> Index-only scans can return the '1d' bits where the actual tuple had the
> '24h'
> bits.
> 
> Have a build without the patch. I can't reproduce amcheck complaints in
> release mode
> where all these statements succeed.

The queries I shared don't create the problematic structure, just an assertion
failure.

> >  * Generic "equalimage" support function.
> > *
> > * B-Tree operator classes whose equality function could safely be
> replaced by
> > * datum_image_eq() in all cases can use this as their "equalimage" support
> > * function.
> It seems to me that as long as a data type has a deterministic sort but not
> necessarily be equalimage,
> it should be able to support deduplication.  e.g for interval type, we add
> a byte wise tie breaker
> after '24h' and '1day' are compared equally. In the btree, '24h' and '1day'
> are still adjacent,
> '1day' is always sorted before '24h' in a btree page, can we do dedup for
> each value
> without problem?

Yes.  I'm not aware of correctness obstacles arising if one did that.