join of array

Pavel Stehule <stehule@kix.fsv.cvut.cz> writes:

Is possible merge two arrays like
array[1,2,3] + array[4,5,6] => array[1,2,3,4,5,6]

I was about to say that || would do it, but I see that's not quite
right:

regression=# SELECT ARRAY[1,2,3] || ARRAY[4,5,6];
?column?
-------------------
{{1,2,3},{4,5,6}}
(1 row)

Offhand, I would think that '{1,2,3,4,5,6}' would be what I'd
intuitively expect to get from "concatenating" these arrays.
Joe, do we really have this implemented per spec?

regards, tom lane

Tom Lane wrote:

Pavel Stehule <stehule@kix.fsv.cvut.cz> writes:

Is possible merge two arrays like
array[1,2,3] + array[4,5,6] => array[1,2,3,4,5,6]

I was about to say that || would do it, but I see that's not quite
right:

regression=# SELECT ARRAY[1,2,3] || ARRAY[4,5,6];
?column?
-------------------
{{1,2,3},{4,5,6}}
(1 row)

Offhand, I would think that '{1,2,3,4,5,6}' would be what I'd
intuitively expect to get from "concatenating" these arrays.
Joe, do we really have this implemented per spec?

Hmmm, it made sense to me, at at least at some point ;-). Here's the
SQL99 guidance (SQL200X doesn't give any more detailed guidance):

4.11.3.2 Operators that operate on array values and return array values
<array concatenation> is an operation that returns the array value made
by joining its array value operands in the order given.

So I guess it ought to be changed.

We also have
ARRAY[1,2] || 3 == '{1,2,3}'
and
ARRAY[[1,2],[3,4]] || ARRAY[5,6] == '{{1,2},{3,4},{5,6}}'
and
ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{{1,2},{3,4}},{{1,2},{3,4}}}'

I think the first two still make sense. I guess the third case ought to be:
ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{1,2},{3,4},{1,2},{3,4}}'
?

If this sounds good, I'll work on a patch for the behavior as well as
the docs.

Joe

Joe Conway <mail@joeconway.com> writes:

Hmmm, it made sense to me, at at least at some point ;-). Here's the
SQL99 guidance (SQL200X doesn't give any more detailed guidance):

4.11.3.2 Operators that operate on array values and return array values
<array concatenation> is an operation that returns the array value made
by joining its array value operands in the order given.

That's about as clear as mud :-( ... but I found a clearer statement in
SQL99 6.31:

2) If <array concatenation> is specified, then:

a) Let AV1 be the value of <array value expression 1> and let
AV2 be the value of <array value expression 2>.

b) If either AV1 or AV2 is the null value, then the result of
the <array concatenate function> is the null value.

c) Otherwise, the result is the array comprising every element
of AV1 followed by every element of AV2.

(c) seems to be pretty clearly what Pavel wants for the 1-D case, but
it's not immediately clear how to apply it to multidimensional arrays.

We also have
ARRAY[1,2] || 3 == '{1,2,3}'
and
ARRAY[[1,2],[3,4]] || ARRAY[5,6] == '{{1,2},{3,4},{5,6}}'
and
ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{{1,2},{3,4}},{{1,2},{3,4}}}'

I think the first two still make sense. I guess the third case ought to be:
ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{1,2},{3,4},{1,2},{3,4}}'
?

Probably. AFAICS this doesn't affect the data copying at all, only the
way in which the result's dimension values are computed, right?

Also, we might want to take another look at the rules for selecting the
lower-bounds of the result array. In the cases where we're joining
N+1-D to N-D (including 1-D to scalar) it still seems to make sense to
preserve the subscripts of the higher-dimensional object, so the lower-
dimensional one is "pushed" onto one end or the other. In the N-D to
N-D case I can't see any really principled way to do it; for lack of
a better idea, I suggest preserving the subscripts of the lefthand
input (ie, using its lower-bound).

regards, tom lane

ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{{1,2},{3,4}},{{1,2},{3,4}}}'

I think the first two still make sense. I guess the third case ought to be:
ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{1,2},{3,4},{1,2},{3,4}}'
?

I do not think this is right. I think the current behaviour
is right. You are effectively dereferencing or flattening
the second array which changes the definition of the second
object.

The ability to do the dereference/flattening is useful,
but it is not the || operator. How about |* which would
flatten 1 level? Of course, that begs the question
of what about n levels and I'm not sure about that.

--elein

elein wrote:

ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{{1,2},{3,4}},{{1,2},{3,4}}}'

I think the first two still make sense. I guess the third case ought to be:
ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{1,2},{3,4},{1,2},{3,4}}'
?

I do not think this is right. I think the current behaviour
is right. You are effectively dereferencing or flattening
the second array which changes the definition of the second
object.

It makes sense in analogy to
ARRAY[1,2] || ARRAY[3,4] == '{1,2,3,4}'

In the case of, e.g. ARRAY[[1,2],[3,4]] || ARRAY[[5,6],[7,8]],
'{1,2}', '{3,4}', '{5,6}', and '{7,8}' are "elements" of the higher
level array, just like 1, 2, 3, & 4 are elements of '{1,2,3,4}'

Joe

I'd like to summarize what I know (or don't know) since this topic has been
hit around a little and I'm new to this. I'm hoping it will clear things up,
at least for me. You are all the experts, I want to make sure I am singing
from the same page.

data sample:
id | fm | ls | addr | city | st | z |c| start|end
----------------------------------------------------------------------------------

191922C,Bob Cobb,D'Obbalina Sr.,312 Elm Street,Yountville,CA,94599,5,062001,082009
339111C,Elma Thelma,Velma,98 Oak Lane,St. Louis,MO,63119-2065,,,

What I wanted to do was to import lots of these from a text file. In the case
where there is an empty string (i.e. no value after a comma) I wanted to
define the column in the table in a way that would accept the empty string but
replace it with the default value for that column. I didn't know that the
copy command is just some C code that stuffs the data into the db ala
fois grois.

What I would really benefit from (and I hope some other new soul would too)
is if someone would outline exactly how they would approach this problem.

Maybe provide the correct table definition and the copy command. Or if that
just won't work an alternate approach. I realize that some of you have
done this partially but there have been too many replies to get into a
single cohesive instruction.

Anyway I suppose my initial frustration in trying to do this may have blinded
me from reason.

create table contact (
id character(7) NOT NULL,
fm character(30) DEFAULT 'xzxzxzxz',
ls character(30) DEFAULT 'xzxzxzxz',
addr character(30) DEFAULT '123 xzxzxzxz',
city character(25) DEFAULT 'xzxzxzxz',
st character(2) DEFAULT 'xz',
c character(1) DEFAULT 'x',
start decimal(6) DEFAULT 122038,
end decimal(6) DEFAULT 122038,
CONSTRAINT handle PRIMARY KEY (id)
) WITHOUT OIDS;

you said we had:

We also have
ARRAY[1,2] || 3 == '{1,2,3}'
and
ARRAY[[1,2],[3,4]] || ARRAY[5,6] == '{{1,2},{3,4},{5,6}}'
and
ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{{1,2},{3,4}},{{1,2},{3,4}}}'

I think the first two still make sense. I guess the third case ought to be:
ARRAY[[1,2],[3,4]] || ARRAY[[1,2],[3,4]] ==
'{{1,2},{3,4},{1,2},{3,4}}'

The second case above makes this case wrong. Or vice versa.
It is dereferencing the right operand in one case and
not in the other. How exactly do you tell the difference?
The only difference I can tell is the depth of the
left hand array. This is confusing and unnecessarily complex.

It makes sense in analogy to
ARRAY[1,2] || ARRAY[3,4] == '{1,2,3,4}'

This case should be '{1,2,{3,4}}'.

Unless we have a flattening operator or dereferencing function.
Or a policy of always flattening the right operand one
level (confusing!).

Consistent treatment of the operands is important. Treating
object (arrays) as they were defined makes the nesting
and usage easier. Not to mention the implementation.

(Good problem :-)

Elein

elein wrote:

you said we had:

We also have

^^^^

There are two variants each of two cases. The first case is what started
this discussion. The newest reading of the SQL99 spec says that we
*must* do this:
1a) ARRAY[1,2] || ARRAY[3,4] == '{1,2,3,4}'

Quoting the paragraph provided by Tom:
"c) Otherwise, the result is the array comprising every element
of AV1 followed by every element of AV2."

The variant is that when you have an "array of arrays", i.e. a
multidimensional array (which Peter E pointed out earlier is part of
SQL99 too), the spec wording implies that we also *must* do this:
1b) ARRAY[[1],[2]] || ARRAY[[3],[4]] == '{{1},{2},{3},{4}'

The second case is not directly addressed by the spec as far as I can
see, i.e. it is a Postgres extension. That is:
2a) ARRAY[1,2] || 3 == '{1,2,3}'

So by analogy the multidimensional variant is:
2b) ARRAY[[1],[2]] || ARRAY[3] == '{{1},{2},{3}}'

Cases 1a and 1b are currently wrong according to the spec., and that's
the change we've been discussing. Cases 2a and 2b currently work as
shown and are correct IMHO (although Tom pointed out a lower bound index
issue that I'll address in my response to him).

Does this help?

Joe

Tom Lane wrote:

That's about as clear as mud :-( ... but I found a clearer statement
in SQL99 6.31:

2) If <array concatenation> is specified, then:

a) Let AV1 be the value of <array value expression 1> and let AV2 be
the value of <array value expression 2>.

b) If either AV1 or AV2 is the null value, then the result of the
<array concatenate function> is the null value.

c) Otherwise, the result is the array comprising every element of AV1
followed by every element of AV2.

(c) seems to be pretty clearly what Pavel wants for the 1-D case, but
it's not immediately clear how to apply it to multidimensional
arrays.

Thanks -- I found the corresponding paragraph in SQL200x (6.35) and it
pretty much reads the same.

Probably. AFAICS this doesn't affect the data copying at all, only
the way in which the result's dimension values are computed, right?

Looks that way to me.

Also, we might want to take another look at the rules for selecting
the lower-bounds of the result array. In the cases where we're
joining N+1-D to N-D (including 1-D to scalar) it still seems to make
sense to preserve the subscripts of the higher-dimensional object, so
the lower- dimensional one is "pushed" onto one end or the other.

This is mostly the way it currently works:

regression=# create table arr(f1 int[]);
CREATE TABLE
regression=# insert into arr values ('{}');
INSERT 2498103 1
regression=# update arr set f1[-2] = 1;
UPDATE 1
regression=# select array_lower(f1,1) from arr;
array_lower
-------------
-2
(1 row)

regression=# select array_lower(f1 || 2, 1) from arr;
array_lower
-------------
-2
(1 row)

regression=# select array_lower(0 || f1, 1) from arr;
array_lower
-------------
-3
(1 row)
regression=# update arr set f1 = ARRAY[[1,2],[3,4]];
UPDATE 1
regression=# select array_lower(f1,1) from arr;
array_lower
-------------
1
(1 row)

regression=# select array_lower(f1 || ARRAY[5,6], 1) from arr;
array_lower
-------------
1
(1 row)

regression=# select array_lower(ARRAY[-1,0] || f1, 1) from arr;
array_lower
-------------
1
(1 row)

It looks like the only "wrong" case is the last one. Will fix.

In the N-D to N-D case I can't see any really principled way to do
it; for lack of a better idea, I suggest preserving the subscripts of
the lefthand input (ie, using its lower-bound).

OK, will do.

Thanks,

Joe

Joe Conway <mail@joeconway.com> writes:

elein wrote:

I do not think this is right. I think the current behaviour
is right. You are effectively dereferencing or flattening
the second array which changes the definition of the second
object.

It makes sense in analogy to
ARRAY[1,2] || ARRAY[3,4] == '{1,2,3,4}'

I agree with Joe. The spec is quite clear about what to do in the
one-dimensional case: the original arrays lose their separate identity.
In the multi-dimensional case, they should lose their separate
identities in the outermost dimension.

I believe the behavior Elein wants can be had by writing
ARRAY[ n_d_array , n_d_array ]
(Joe, would you confirm that's true, and document it? I don't think
either section 8.10 or section 4.2.8 makes clear that you can build
arrays from smaller array values rather than just scalars.) As long as
we have that alternative, it's not necessary that concatenation do the
same thing.

Another argument for doing it this way is that it makes array
concatenation associative, which is a nice bit of symmetry. Consider

ARRAY[[1,1],[2,2]] || ARRAY[[3,3],[4,4]] ||
ARRAY[[5,5],[6,6]] || ARRAY[[7,7],[8,8]]

Right now, with the default left-to-right association you get

{{{1,1},{2,2}},{{3,3},{4,4}},{{5,5},{6,6}},{{7,7},{8,8}}}

but if you parenthesize it differently you can get a different answer:

regression=# select (ARRAY[[1,1],[2,2]] || ARRAY[[3,3],[4,4]]) || (ARRAY[[5,5],[6,6]] || ARRAY[[7,7],[8,8]]);
?column?
---------------------------------------------------------------
{{{{1,1},{2,2}},{{3,3},{4,4}}},{{{5,5},{6,6}},{{7,7},{8,8}}}}
(1 row)

With the flattening approach all the intermediate results will remain
2-D arrays and so you get the same answer for all parenthesizations,
namely {{1,1},{2,2},{3,3},{4,4},{5,5},{6,6},{7,7},{8,8}}. That strikes
me as more nearly the intuitive meaning of "concatenation" than what
we've got now.

(Cases involving some N-D and some N+1-D inputs remain non-associative,
though, which is a tad annoying. Maybe it's okay seeing that the inputs
are of different kinds.)

regards, tom lane

Tom Lane wrote:

I believe the behavior Elein wants can be had by writing
ARRAY[ n_d_array , n_d_array ]
(Joe, would you confirm that's true, and document it? I don't think
either section 8.10 or section 4.2.8 makes clear that you can build
arrays from smaller array values rather than just scalars.) As long as
we have that alternative, it's not necessary that concatenation do the
same thing.

Well this works:
regression=# select ARRAY[ARRAY[[1,2],[3,4]],ARRAY[[5,6],[7,8]]];
array
-------------------------------
{{{1,2},{3,4}},{{5,6},{7,8}}}
(1 row)

But I was disappointed that this doesn't:

regression=# select ARRAY['{{1,2},{3,4}}'::int[],'{{5,6},{7,8}}'::int[]];
ERROR: multidimensional ARRAY[] must be built from nested array expressions

Nor does this:

create table arr(f1 int[], f2 int[]);
insert into arr values (ARRAY[[1,2],[3,4]],ARRAY[[5,6],[7,8]]);
regression=# select ARRAY[f1,f2] from arr;
ERROR: multidimensional ARRAY[] must be built from nested array expressions

It does work for the element to array case:

create table els(f1 int, f2 int);
insert into els values (1,2);
regression=# select ARRAY[f1,f2] from els;
array
-------
{1,2}
(1 row)

Should I try to make the second and third cases work?

Joe

On Fri, 2003-08-15 at 12:41, Joe Conway wrote:

Tom Lane wrote:

That's about as clear as mud :-( ... but I found a clearer statement
in SQL99 6.31:

2) If <array concatenation> is specified, then:

a) Let AV1 be the value of <array value expression 1> and let AV2 be
the value of <array value expression 2>.

b) If either AV1 or AV2 is the null value, then the result of the
<array concatenate function> is the null value.

c) Otherwise, the result is the array comprising every element of AV1
followed by every element of AV2.

(c) seems to be pretty clearly what Pavel wants for the 1-D case, but
it's not immediately clear how to apply it to multidimensional
arrays.

Thanks -- I found the corresponding paragraph in SQL200x (6.35) and it
pretty much reads the same.

Why are arrays even mentioned in the the same breath wrt relations
DBMSs? Aren't they an anathema to all we know and love?

-- 
+---------------------------------------------------------------+
| Ron Johnson, Jr.        Home: ron.l.johnson@cox.net           |
| Jefferson, LA  USA                                            |
|                                                               |
| "Man, I'm pretty.  Hoo Hah!"                                  |
|    Johnny Bravo                                               |
+---------------------------------------------------------------+

I guess I am arguing against the spec. :-)
But given the spec...
The spec is consistent in that it seems to
dereference the right operand one level.

However, that would still make 2b inconsistent
in the evaluation of the right operand.

1a) ARRAY[1,2] || ARRAY[3,4] == '{1,2,3,4}'
1b) ARRAY[[1],[2]] || ARRAY[[3],[4]] == '{{1},{2},{3},{4}}'

and ARRAY[1,2] || ARRAY[[3],[4]] == '{1,2,{3},{4}}'

So by analogy the multidimensional variant is:
2b) ARRAY[[1],[2]] || ARRAY[3] == '{{1},{2},{3}}'

I would think this would be '{{1},{2}, 3}}'
and ARRAY[1,2] || ARRAY[[3],[4]] == '{1,2,{3},{4}}'

I do see the analogy you are making. But I
respectfully disagree (with the spec ?) that
the type/structure of the left operand should be
taken into account when evaluating the right operand.

elein

Joe Conway <mail@joeconway.com> writes:

But I was disappointed that this doesn't:

regression=# select ARRAY['{{1,2},{3,4}}'::int[],'{{5,6},{7,8}}'::int[]];
ERROR: multidimensional ARRAY[] must be built from nested array expressions

Drat, I was assuming that that *would* work.

Should I try to make the second and third cases work?

Could you look at how big a change it'd be, anyway? Offhand I think it
may just mean that the subscript-checking done in parse_expr.c needs to
be done at runtime instead. Remember parse_expr should only be
concerned about determining datatype, and for its purposes all arrays of
a given element type are the same --- subscript checking should happen
at runtime. (It seems likely that having an ndims field in ArrayExpr
is inappropriate.)

regards, tom lane

PostgreSQL is an ORDBMS, not just an RDBMS.

A column holds a type of value. Any kind. The
structure and operands define the type. The data
defines the value. This holds true for simple types
like an integer or complex types like an array.

The database data is relatively "type blind" in an
ORDBMS. It uses the standard overloaded operands
to determine the type of function to perform for
all of the usual RDBMS utilities.
Constraints, triggers, sorting, etc. all apply.

That's what the ORDBMS stuff can give you.
Arrays are a natural extension.

Arrays don't necessarily imply denormalization.
It depends on how you use them. The same rule
applies for integers.

elein

On Friday 15 August 2003 01:13 pm, Ron Johnson wrote:

On Fri, 2003-08-15 at 12:41, Joe Conway wrote:

Tom Lane wrote:

That's about as clear as mud :-( ... but I found a clearer statement
in SQL99 6.31:

2) If <array concatenation> is specified, then:

a) Let AV1 be the value of <array value expression 1> and let AV2 be
the value of <array value expression 2>.

b) If either AV1 or AV2 is the null value, then the result of the
<array concatenate function> is the null value.

c) Otherwise, the result is the array comprising every element of AV1
followed by every element of AV2.

(c) seems to be pretty clearly what Pavel wants for the 1-D case, but
it's not immediately clear how to apply it to multidimensional
arrays.

Thanks -- I found the corresponding paragraph in SQL200x (6.35) and it
pretty much reads the same.

Why are arrays even mentioned in the the same breath wrt relations
DBMSs? Aren't they an anathema to all we know and love?

This gives rise to a couple of good questions:

When and why would you want to use arrays instead of a relational model?

Thanks,

Andrew Gould

Hi,

Tom Lane wrote:

Joe Conway <mail@joeconway.com> writes:

But I was disappointed that this doesn't:

regression=# select ARRAY['{{1,2},{3,4}}'::int[],'{{5,6},{7,8}}'::int[]];
ERROR: multidimensional ARRAY[] must be built from nested array expressions

Drat, I was assuming that that *would* work.

Should I try to make the second and third cases work?

Could you look at how big a change it'd be, anyway? Offhand I think it
may just mean that the subscript-checking done in parse_expr.c needs to
be done at runtime instead. Remember parse_expr should only be
concerned about determining datatype, and for its purposes all arrays of
a given element type are the same --- subscript checking should happen
at runtime. (It seems likely that having an ndims field in ArrayExpr
is inappropriate.)

Wouldn't it be a good idea to just extend the partner arrays? Say
if we concenate array A(Na,..,Xa) || B(Nb,...,Xb)
The resulting array C would be of dimension
C(Na+Nb,max(Oa,Ob),max(Pa,Pb), ... max(Xa,Xb))
So concenation would be an extending and right hand appending (at first
level)

Regards
Tino Wildenhain

On Fri, 2003-08-15 at 13:32, elein wrote:

PostgreSQL is an ORDBMS, not just an RDBMS.

But y'all are talking about the SQL standard here.

A column holds a type of value. Any kind. The
structure and operands define the type. The data
defines the value. This holds true for simple types
like an integer or complex types like an array.

The database data is relatively "type blind" in an
ORDBMS. It uses the standard overloaded operands
to determine the type of function to perform for
all of the usual RDBMS utilities.
Constraints, triggers, sorting, etc. all apply.

That's what the ORDBMS stuff can give you.
Arrays are a natural extension.

Arrays don't necessarily imply denormalization.
It depends on how you use them. The same rule
applies for integers.

I dunno 'bout that...

elein

On Fri, Aug 15, 2003 at 01:13:52PM -0500, Ron Johnson wrote:

Why are arrays even mentioned in the the same breath wrt relations
DBMSs? Aren't they an anathema to all we know and love?

-- 
+---------------------------------------------------------------+
| Ron Johnson, Jr.        Home: ron.l.johnson@cox.net           |
| Jefferson, LA  USA                                            |
|                                                               |
| "Man, I'm pretty.  Hoo Hah!"                                  |
|    Johnny Bravo                                               |
+---------------------------------------------------------------+

In response to both Andrew Gould and Ron Johnson...

If arrays are not natural in the organization of
your data, don't use them. That is the guideline.

If the array defines something specific they are
very natural. The confusion could be that arrays
are abstract types.

Specific implementations which use arrays might
be clearer. For example, a definition of a polygon
is an array of Points. Points, themselves are an
array.

(The actual postgreSQL implementation of polygons and points
doesn't use the newer cleaner array abstraction, I think.
But if I were reimplementing them, I would build on
top of the new array capabilities. The point is to show
an array structured object which makes sense in context.)

Of course you can denomalize via arrays, but it tends
to make things harder for you. And I believe the
same thing is true for denormalized integer columns.

elein
=============================================================
elein@varlena.com www.varlena.com
PostgreSQL Consulting & Support
PostgreSQL General Bits http://www.varlena.com/GeneralBits/
=============================================================
"Free your mind the rest will follow"
-- En Vogue