order by, for custom types

Started by Grzegorz Jaskiewiczabout 20 years ago16 messages
#1Grzegorz Jaskiewicz
gj@pointblue.com.pl

Hi folks

I have all operators required for b-tree, gist up here, and gist
index defined. But still "order by custom_type" won't work.
I have kind of wild card masks in my type definition, so b-tree won't
work. But still, do I need to define b-tree index as such for "order
by" to work ?
Perhaps gist should be expanded so it would take care of "order by"
and "distinct" ?

Thanks.

--
GJ

"If we knew what we were doing, it wouldn't be called Research, would
it?" - AE

#2Oleg Bartunov
oleg@sai.msu.su
In reply to: Grzegorz Jaskiewicz (#1)
Re: order by, for custom types

On Fri, 18 Nov 2005, Grzegorz Jaskiewicz wrote:

Hi folks

I have all operators required for b-tree, gist up here, and gist index
defined. But still "order by custom_type" won't work.
I have kind of wild card masks in my type definition, so b-tree won't work.
But still, do I need to define b-tree index as such for "order by" to work ?
Perhaps gist should be expanded so it would take care of "order by" and
"distinct" ?

look contrib/ltree for reference (sql/ltree.sql)

Thanks.

Regards,
Oleg
_____________________________________________________________
Oleg Bartunov, sci.researcher, hostmaster of AstroNet,
Sternberg Astronomical Institute, Moscow University (Russia)
Internet: oleg@sai.msu.su, http://www.sai.msu.su/~megera/
phone: +007(095)939-16-83, +007(095)939-23-83

#3Greg Stark
gsstark@mit.edu
In reply to: Grzegorz Jaskiewicz (#1)
Re: order by, for custom types

Grzegorz Jaskiewicz <gj@pointblue.com.pl> writes:

Hi folks

I have all operators required for b-tree, gist up here, and gist index
defined. But still "order by custom_type" won't work.

I think you need to create an "operator class" for ORDER BY to work. Someone
else may answer with more details.

I have kind of wild card masks in my type definition, so b-tree won't work.
But still, do I need to define b-tree index as such for "order by" to work ?
Perhaps gist should be expanded so it would take care of "order by" and
"distinct" ?

This I don't understand. If ORDER BY will work then b-tree indexes will work
too. If your type is such that b-tree indexes don't make sense then neither
does ORDER BY.

--
greg

#4Tom Lane
tgl@sss.pgh.pa.us
In reply to: Grzegorz Jaskiewicz (#1)
Re: order by, for custom types

Grzegorz Jaskiewicz <gj@pointblue.com.pl> writes:

I have all operators required for b-tree, gist up here, and gist
index defined. But still "order by custom_type" won't work.

Define "won't work" ... what happens?

I have kind of wild card masks in my type definition, so b-tree won't
work. But still, do I need to define b-tree index as such for "order
by" to work ?

You don't need an index, but a b-tree operator class is a good idea.
Still, it should be possible to sort with only a "<" operator --- at
the moment anyway. (I've been thinking about some ideas that would
effectively require a b-tree opclass to do sorting, so this might not
still be true in 8.2 ...)

regards, tom lane

#5Grzegorz Jaskiewicz
gj@pointblue.com.pl
In reply to: Tom Lane (#4)
Re: order by, for custom types

On 2005-11-18, at 22:53, Tom Lane wrote:

Grzegorz Jaskiewicz <gj@pointblue.com.pl> writes:

I have all operators required for b-tree, gist up here, and gist
index defined. But still "order by custom_type" won't work.

Define "won't work" ... what happens?

Wildcards cause things not to work as they should

consider everything in [] brackets to be a possible choice and those
three:

a = 1.2.3.4
b = 1.[2,3].3.4
c = 1.3.3.4

a = b, b = c, but a <> c, I was told that because of that btree won't
work on my type. (on irc, that was AndrewSN as I recall).

You don't need an index, but a b-tree operator class is a good idea.
Still, it should be possible to sort with only a "<" operator --- at
the moment anyway. (I've been thinking about some ideas that would
effectively require a b-tree opclass to do sorting, so this might not
still be true in 8.2 ...)

I do have all operators required for btree, no operator class
defined, every single operator. Btree requires some function apart
from operators, this one is not defined, but I do have = operator as
well.

--
GJ

"If we knew what we were doing, it wouldn't be called Research, would
it?" - AE

#6Tom Lane
tgl@sss.pgh.pa.us
In reply to: Grzegorz Jaskiewicz (#5)
Re: order by, for custom types

Grzegorz Jaskiewicz <gj@pointblue.com.pl> writes:

Wildcards cause things not to work as they should

consider everything in [] brackets to be a possible choice and those
three:

a = 1.2.3.4
b = 1.[2,3].3.4
c = 1.3.3.4

a = b, b = c, but a <> c, I was told that because of that btree won't
work on my type. (on irc, that was AndrewSN as I recall).

Well, neither will sorting then. If you can define a consistent sort
order, btree will work; if you can't, then you can't sort either.
AFAICS, with rules like the above you can't define a consistent <
operator.

regards, tom lane

#7Andrew - Supernews
andrew+nonews@supernews.com
In reply to: Grzegorz Jaskiewicz (#1)
Re: order by, for custom types

On 2005-11-19, Grzegorz Jaskiewicz <gj@pointblue.com.pl> wrote:

Wildcards cause things not to work as they should

consider everything in [] brackets to be a possible choice and those
three:

a = 1.2.3.4
b = 1.[2,3].3.4
c = 1.3.3.4

a = b, b = c, but a <> c, I was told that because of that btree won't
work on my type. (on irc, that was AndrewSN as I recall).

Probably. But nothing stops you defining equality and ordering operators
that _do_ work for btree, and hence sorting, it's just that those operators
won't be any use for the matching semantics.

It's clear that for your data type that there is a concept of "equality"
in which all three of your values a,b,c above are unequal. My advice would
be (and I'm sure I suggested this at the time) that you reserve the '='
operator for a true equality operation, and use some other operator such as
~ or @ for the "matches" semantics that you want for your application.
Having an intransitive '=' operator violates the POLA, even if it doesn't
actively break anything otherwise (I have no idea if it does).

I do have all operators required for btree, no operator class
defined, every single operator. Btree requires some function apart
from operators, this one is not defined, but I do have = operator as
well.

You still don't seem to understand that what btree requires is not an
operator _called_ '=', but an operator with the logical semantics of
"equality". That operator can be called anything you please (it doesn't
have to have the name '=').

Sorting doesn't need an equality operator, since it can fabricate one if
given a suitable < operator, i.e. one that constitutes a strict weak
ordering over the elements to be sorted; it can rely on the fact that
NOT(a < b) AND NOT(b < a) implies that a and b are equivalent for sorting
purposes. (The requirement that < constitute a strict weak ordering is
enough to ensure that this is an equivalence relation, and therefore
transitive; if < does not meet this requirement then sorting may give wrong
answers, loop forever, or possibly crash.)

--
Andrew, Supernews
http://www.supernews.com - individual and corporate NNTP services

#8Grzegorz Jaskiewicz
gj@pointblue.com.pl
In reply to: Andrew - Supernews (#7)
Re: order by, for custom types

aye aye Sir.

--
GJ

"If we knew what we were doing, it wouldn't be called Research, would
it?" - AE

#9Kevin McArthur
Kevin@StormTide.ca
In reply to: Grzegorz Jaskiewicz (#1)
Re: order by, for custom types

Take the query.

select a,b from dupa where b::text in (select b::text from dupa group by
b::text having count(b) > 2);

This is acceptable to create a unique constraint, however, we cannot mark
the column unique, without defining btree operators, which clearly are not
possible for sorting. Is there any way to base the operators based on the
text representation of the type for strict equality (not to be confused with
same or equivilent) and thus use that not as an ordering method, but as a
simple equality for uniqueness.

Kevin McArthur

----- Original Message -----
From: "Andrew - Supernews" <andrew+nonews@supernews.com>
To: <pgsql-hackers@postgresql.org>
Sent: Saturday, November 19, 2005 10:54 PM
Subject: Re: [HACKERS] order by, for custom types

Show quoted text

On 2005-11-19, Grzegorz Jaskiewicz <gj@pointblue.com.pl> wrote:

Wildcards cause things not to work as they should

consider everything in [] brackets to be a possible choice and those
three:

a = 1.2.3.4
b = 1.[2,3].3.4
c = 1.3.3.4

a = b, b = c, but a <> c, I was told that because of that btree won't
work on my type. (on irc, that was AndrewSN as I recall).

Probably. But nothing stops you defining equality and ordering operators
that _do_ work for btree, and hence sorting, it's just that those
operators
won't be any use for the matching semantics.

It's clear that for your data type that there is a concept of "equality"
in which all three of your values a,b,c above are unequal. My advice would
be (and I'm sure I suggested this at the time) that you reserve the '='
operator for a true equality operation, and use some other operator such
as
~ or @ for the "matches" semantics that you want for your application.
Having an intransitive '=' operator violates the POLA, even if it doesn't
actively break anything otherwise (I have no idea if it does).

I do have all operators required for btree, no operator class
defined, every single operator. Btree requires some function apart
from operators, this one is not defined, but I do have = operator as
well.

You still don't seem to understand that what btree requires is not an
operator _called_ '=', but an operator with the logical semantics of
"equality". That operator can be called anything you please (it doesn't
have to have the name '=').

Sorting doesn't need an equality operator, since it can fabricate one if
given a suitable < operator, i.e. one that constitutes a strict weak
ordering over the elements to be sorted; it can rely on the fact that
NOT(a < b) AND NOT(b < a) implies that a and b are equivalent for sorting
purposes. (The requirement that < constitute a strict weak ordering is
enough to ensure that this is an equivalence relation, and therefore
transitive; if < does not meet this requirement then sorting may give
wrong
answers, loop forever, or possibly crash.)

--
Andrew, Supernews
http://www.supernews.com - individual and corporate NNTP services

---------------------------(end of broadcast)---------------------------
TIP 2: Don't 'kill -9' the postmaster

#10Tom Lane
tgl@sss.pgh.pa.us
In reply to: Kevin McArthur (#9)
Re: order by, for custom types

"Kevin McArthur" <Kevin@StormTide.ca> writes:

This is acceptable to create a unique constraint, however, we cannot mark
the column unique, without defining btree operators, which clearly are not
possible for sorting. Is there any way to base the operators based on the
text representation of the type for strict equality (not to be confused with
same or equivilent) and thus use that not as an ordering method, but as a
simple equality for uniqueness.

Translation: you do know how to define a sortable order (ie, generate
the text version and compare); you're just too lazy to create the
operators to do it ...

regards, tom lane

#11Grzegorz Jaskiewicz
gj@pointblue.com.pl
In reply to: Tom Lane (#10)
Re: order by, for custom types

On 2005-11-22, at 15:45, Tom Lane wrote:

"Kevin McArthur" <Kevin@StormTide.ca> writes:

This is acceptable to create a unique constraint, however, we
cannot mark
the column unique, without defining btree operators, which clearly
are not
possible for sorting. Is there any way to base the operators based
on the
text representation of the type for strict equality (not to be
confused with
same or equivilent) and thus use that not as an ordering method,
but as a
simple equality for uniqueness.

Translation: you do know how to define a sortable order (ie, generate
the text version and compare); you're just too lazy to create the
operators to do it ...

We do have WORKING < , > , etc operators, and ::text cast already.
Thing is, can I have btree and gist indexes at the same time ?

--
GJ

"If we knew what we were doing, it wouldn't be called Research, would
it?" - AE

#12Martijn van Oosterhout
kleptog@svana.org
In reply to: Grzegorz Jaskiewicz (#11)
Re: order by, for custom types

On Tue, Nov 22, 2005 at 04:24:21PM +0100, Grzegorz Jaskiewicz wrote:

Translation: you do know how to define a sortable order (ie, generate
the text version and compare); you're just too lazy to create the
operators to do it ...

We do have WORKING < , > , etc operators, and ::text cast already.
Thing is, can I have btree and gist indexes at the same time ?

Sure, did you look at the ltree example someone pointed you to?

Have a nice day,
--
Martijn van Oosterhout <kleptog@svana.org> http://svana.org/kleptog/

Show quoted text

Patent. n. Genius is 5% inspiration and 95% perspiration. A patent is a
tool for doing 5% of the work and then sitting around waiting for someone
else to do the other 95% so you can sue them.

#13Oleg Bartunov
oleg@sai.msu.su
In reply to: Grzegorz Jaskiewicz (#11)
Re: order by, for custom types

On Tue, 22 Nov 2005, Grzegorz Jaskiewicz wrote:

Thing is, can I have btree and gist indexes at the same time ?

no, we have contrib/btree_gist for this

Regards,
Oleg
_____________________________________________________________
Oleg Bartunov, sci.researcher, hostmaster of AstroNet,
Sternberg Astronomical Institute, Moscow University (Russia)
Internet: oleg@sai.msu.su, http://www.sai.msu.su/~megera/
phone: +007(095)939-16-83, +007(095)939-23-83

#14Grzegorz Jaskiewicz
gj@pointblue.com.pl
In reply to: Martijn van Oosterhout (#12)
Re: order by, for custom types

On 2005-11-22, at 16:39, Martijn van Oosterhout wrote:

On Tue, Nov 22, 2005 at 04:24:21PM +0100, Grzegorz Jaskiewicz wrote:

Translation: you do know how to define a sortable order (ie,
generate
the text version and compare); you're just too lazy to create the
operators to do it ...

We do have WORKING < , > , etc operators, and ::text cast already.
Thing is, can I have btree and gist indexes at the same time ?

Sure, did you look at the ltree example someone pointed you to?

Have a nice day,

Ok, I hacked btree for my type, and surely I can have both btree and
gist at the same time on the same column.
/me is now going to have a look on btree_gist.

Thanks folks.

--
GJ

"If we knew what we were doing, it wouldn't be called Research, would
it?" - AE

#15Martijn van Oosterhout
kleptog@svana.org
In reply to: Grzegorz Jaskiewicz (#14)
Re: order by, for custom types

On Tue, Nov 22, 2005 at 05:14:35PM +0100, Grzegorz Jaskiewicz wrote:

Ok, I hacked btree for my type, and surely I can have both btree and
gist at the same time on the same column.
/me is now going to have a look on btree_gist.

You don't actually have to have a btree defined on your column for
ORDER BY to work, you just need to define a btree OPERATOR CLASS so
that PostgreSQL knows what you mean by "ORDER BY".

Have a nice day,
--
Martijn van Oosterhout <kleptog@svana.org> http://svana.org/kleptog/

Show quoted text

Patent. n. Genius is 5% inspiration and 95% perspiration. A patent is a
tool for doing 5% of the work and then sitting around waiting for someone
else to do the other 95% so you can sue them.

#16Grzegorz Jaskiewicz
gj@pointblue.com.pl
In reply to: Martijn van Oosterhout (#15)
Re: order by, for custom types

On 2005-11-22, at 17:17, Martijn van Oosterhout wrote:

On Tue, Nov 22, 2005 at 05:14:35PM +0100, Grzegorz Jaskiewicz wrote:

Ok, I hacked btree for my type, and surely I can have both btree and
gist at the same time on the same column.
/me is now going to have a look on btree_gist.

You don't actually have to have a btree defined on your column for
ORDER BY to work, you just need to define a btree OPERATOR CLASS so
that PostgreSQL knows what you mean by "ORDER BY".

yep, I know. Thanks.
Anyhow, I wanted to see if having two indexes will make it faster/
slower. and if it is possible in first place.

--
GJ

"If we knew what we were doing, it wouldn't be called Research, would
it?" - AE