Message-ID: <b697b506-b3dd-c538-dd67-bd6a735af2f5@illuminatedcomputing.com>
Date: Thu, 14 Sep 2023 09:09:19 -0700
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.13.0
Subject: Re: SQL:2011 application time
Content-Language: en-US
To: jian he <jian.universality@gmail.com>
Cc: Vik Fearing <vik@postgresfriends.org>,
 Peter Eisentraut <peter@eisentraut.org>,
 Corey Huinker <corey.huinker@gmail.com>,
 PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
References: <CA+renyUApHgSZF9-nd-a0+OPGharLQLO=mDHcY4_qQ0+noCUVg@mail.gmail.com>
 <CA+renyWqpNdsi9OaPmXheeQL72adi_iz36VgVsqyyGpQ20FDAg@mail.gmail.com>
 <21964319-46e5-3e47-217b-6ac4169bdf13@enterprisedb.com>
 <CA+renyX0ipvY6A_jUOHeB1q9mL4bEYfAZ5FBB7G7jUo5bykjrA@mail.gmail.com>
 <f42e5be5-cd3e-bc4f-83e4-04a1e0270b17@enterprisedb.com>
 <27361388-f5ab-ea36-ea35-41d68a90e60d@illuminatedcomputing.com>
 <152e4c2c-36b4-bdc0-1b62-e9f9e8b68adc@enterprisedb.com>
 <251B44D1-D3E0-47E4-AB9D-4A848EDB495E@yesql.se>
 <1c674adf-4114-e8d8-cec7-ac10e2f424c7@enterprisedb.com>
 <CA+renyVyMOb+t6QNoq=PBK1CY_oa9wdJrObjj6S1-RK9CxYwKQ@mail.gmail.com>
 <831f17ba-ff36-1380-a475-a7cd2c65a89c@enterprisedb.com>
 <ab4cf49c-9839-3426-5114-f596fa9ee76b@illuminatedcomputing.com>
 <6f010a6e-8e20-658b-dc05-dc9033a694da@eisentraut.org>
 <7d44dade-ff31-5eef-80ac-78ed838066f6@postgresfriends.org>
 <88dadcc1-6652-ff7d-264d-73906a53dee9@illuminatedcomputing.com>
 <CACJufxHD2kv6hCy8yjVk486+d5rcuJS2XxkNH80g9Ui6Rz81Cg@mail.gmail.com>
From: Paul Jungwirth <pj@illuminatedcomputing.com>
In-Reply-To: <CACJufxHD2kv6hCy8yjVk486+d5rcuJS2XxkNH80g9Ui6Rz81Cg@mail.gmail.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Archived-At: <https://www.postgresql.org/message-id/b697b506-b3dd-c538-dd67-bd6a735af2f5%40illuminatedcomputing.com>
Precedence: bulk

On 9/7/23 18:24, jian he wrote:
> for a range primary key, is it fine to expect it to be unique, not
> null and also not overlap? (i am not sure how hard to implement it).
> 
> -----------------------------------------------------------------
> quote from 7IWD2-02-Foundation-2011-12.pdf. 4.18.3.2 Unique
> constraints, page 97 of 1483.
> 
> ...
> -----------------------------------------------------------------
> based on the above, the unique constraint does not specify that the
> column list must be range type. UNIQUE (a, c WITHOUT OVERLAPS).
> Here column "a" can be a range type (that have overlap property) and
> can be not.
> In fact, many of your primary key, foreign key regess test using
> something like '[11,11]' (which make it more easy to understand),
> which in logic is a non-range usage.
> So UNIQUE (a, c WITHOUT OVERLAPS), column "a" be a non-range data type
> does make sense?

I'm not sure I understand this question, but here are a few things that 
might help clarify things:

In SQL:2011, a temporal primary key, unique constraint, or foreign key 
may have one or more "scalar" parts (just like a regular key) followed 
by one "PERIOD" part, which is denoted with "WITHOUT OVERLAPS" (in 
PKs/UNIQUEs) or "PERIOD" (in FKs). Except for this last key part, 
everything is still compared for equality, just as in a traditional key. 
But this last part is compared for overlaps. It's exactly the same as 
`EXCLUDE (id WITH =, valid_at WITH &&)`. The overlap part must come last 
and you can have only one (but you may have more than one scalar part if 
you like).

In the patch, I have followed that pattern, except I also allow a 
regular range column anywhere I allow a PERIOD. In fact PERIODs are 
mostly implemented on top of range types. (Until recently PERIOD support 
was in the first patch, not the last, and there was code all throughout 
for handling both, e.g. within indexes, etc. But at pgcon Peter 
suggested building everything on just range columns, and then having 
PERIODs create an "internal" GENERATED column, and that cleaned up the 
code considerably.)

One possible source of confusion is that in the tests I'm using range 
columns *also* for the scalar key part. So valid_at is a tsrange, and 
int is an int4range. This is not normally how you'd use the feature, but 
you need the btree_gist extension to mix int & tsrange (e.g.), and 
that's not available in the regress tests. We are still comparing the 
int4range for regular equality and the tsrange for overlaps. If you 
search this thread there was some discussion about bringing btree_gist 
into core, but it sounds like it doesn't need to happen. (It might be 
still desirable independently. EXCLUDE constraints are also not really 
something you can use practically without it, and their tests use the 
same trick of comparing ranges for plain equality.)

The piece of discussion you're replying to is about allowing *multiple* 
WITHOUT OVERLAPS modifiers on a PK/UNIQUE constraint, and in any 
position. I think that's a good idea, so I've started adapting the code 
to support it. (In fact there is a lot of code that assumes the overlaps 
key part will be in the last position, and I've never really been happy 
with that, so it's an excuse to make that more robust.) Here I'm saying 
(1) you will still need at least one scalar key part, (2) if there are 
no WITHOUT OVERLAPS parts then you just have a regular key, not a 
temporal one, (3) changing this obliges us to do the same for foreign 
keys and FOR PORTION OF.

I hope that helps! I apologize if I've completely missed the point. If 
so please try again. :-)

Yours,

-- 
Paul              ~{:-)
pj@illuminatedcomputing.com