subselects

Bruce Momjian wrote:

Vadim, I know you are still thinking about subselects, but I have some
more clarification that may help.

We have to add phantom range table entries to correlated subselects so
they will pass the parser. We might as well add those fields to the
target list of the subquery at the same time:

select *
from taba
where col1 = (select col2
from tabb
where taba.col3 = tabb.col4)

becomes:

select *
from taba
where col1 = (select col2, tabb.col4 <---
from tabb, taba <---
where taba.col3 = tabb.col4)

We add a field to TargetEntry and RangeTblEntry to mark the fact that it
was entered as a correlation entry:

bool isCorrelated;

No, I don't like to add anything in parser. Example:

select *
from tabA
where col1 = (select col2
from tabB
where tabA.col3 = tabB.col4
and exists (select *
from tabC
where tabB.colX = tabC.colX and
tabC.colY = tabA.col2)
)

: a column of tabA is referenced in sub-subselect
(is it allowable by standards ?) - in this case it's better
to don't add tabA to 1st subselect but add tabA to second one
and change tabA.col3 in 1st to reference col3 in 2nd subquery temp table -
this gives us 2-tables join in 1st subquery instead of 3-tables join.
(And I'm still not sure that using temp tables is best of what can be
done in all cases...)

Instead of using isCorrelated in TE & RTE we can add

Index varlevel;

to Var node to reflect (sub)query from where this Var is come
(where is range table to find var's relation using varno). Upmost query
will have varlevel = 0, all its (dirrect) children - varlevel = 1 and so on.
^^^ ^^^^^^^^^^^^
(I don't see problems with distinguishing Vars of different children
on the same level...)

Second, we need to hook the subselect to the main query. I recommend we
add two fields to Query for this:

Query *parentQuery;
List *subqueries;

Agreed. And maybe Index queryLevel.

In the parent query, to parse the WHERE clause, we create a new operator
type, called IN or NOT_IN, or ALL, where the left side is a Var, and the

^^^^^^^^^^^^^^^^^^
No. We have to handle (a,b,c) OP (select x, y, z ...) and
'_a_constant_' OP (select ...) - I don't know is last in standards,
Sybase has this.

Well,

typedef enum OpType
{
OP_EXPR, FUNC_EXPR, OR_EXPR, AND_EXPR, NOT_EXPR

+ OP_EXISTS, OP_ALL, OP_ANY

} OpType;

typedef struct Expr
{
NodeTag type;
Oid typeOid; /* oid of the type of this expr */
OpType opType; /* type of the op */
Node *oper; /* could be Oper or Func */
List *args; /* list of argument nodes */
} Expr;

OP_EXISTS: oper is NULL, lfirst(args) is SubSelect (index in subqueries
List, following your suggestion)

OP_ALL, OP_ANY:

oper is List of Oper nodes. We need in list because of data types of
a, b, c (above) can be different and so Oper nodes will be different too.

lfirst(args) is List of expression nodes (Const, Var, Func ?, a + b ?) -
left side of subquery' operator.
lsecond(args) is SubSelect.

Note, that there are no OP_IN, OP_NOTIN in OpType-s for Expr. We need in
IN, NOTIN in A_Expr (parser node), but both of them have to be transferred
by parser into corresponding ANY and ALL. At the moment we can do:

IN --> = ANY, NOT IN --> <> ALL

but this will be "known bug": this breaks OO-nature of Postgres, because of
operators can be overrided and '=' can mean s o m e t h i n g (not equality).
Example: box data type. For boxes, = means equality of _areas_ and =~
means that boxes are the same ==> =~ ANY should be used for IN.

right side is an index to a slot in the subqueries List.

Vadim

Bruce Momjian wrote:

Vadim, I know you are still thinking about subselects, but I have some
more clarification that may help.

We have to add phantom range table entries to correlated subselects so
they will pass the parser. We might as well add those fields to the
target list of the subquery at the same time:

select *
from taba
where col1 = (select col2
from tabb
where taba.col3 = tabb.col4)

becomes:

select *
from taba
where col1 = (select col2, tabb.col4 <---
from tabb, taba <---
where taba.col3 = tabb.col4)

We add a field to TargetEntry and RangeTblEntry to mark the fact that it
was entered as a correlation entry:

bool isCorrelated;

No, I don't like to add anything in parser. Example:

select *
from tabA
where col1 = (select col2
from tabB
where tabA.col3 = tabB.col4
and exists (select *
from tabC
where tabB.colX = tabC.colX and
tabC.colY = tabA.col2)
)

: a column of tabA is referenced in sub-subselect

This is a strange case that I don't think we need to handle in our first
implementation.

(is it allowable by standards ?) - in this case it's better
to don't add tabA to 1st subselect but add tabA to second one
and change tabA.col3 in 1st to reference col3 in 2nd subquery temp table -
this gives us 2-tables join in 1st subquery instead of 3-tables join.
(And I'm still not sure that using temp tables is best of what can be
done in all cases...)

I don't see any use for temp tables in subselects anymore. After having
implemented UNIONS, I now see how much can be done in the upper
optimizer. I see you just putting the subquery PLAN into the proper
place in the plan tree, with some proper JOIN nodes for IN, NOT IN.

Instead of using isCorrelated in TE & RTE we can add

Index varlevel;

OK. Sounds good.

to Var node to reflect (sub)query from where this Var is come
(where is range table to find var's relation using varno). Upmost query
will have varlevel = 0, all its (dirrect) children - varlevel = 1 and so on.
^^^ ^^^^^^^^^^^^
(I don't see problems with distinguishing Vars of different children
on the same level...)

Second, we need to hook the subselect to the main query. I recommend we
add two fields to Query for this:

Query *parentQuery;
List *subqueries;

Agreed. And maybe Index queryLevel.

Sure. If it helps.

In the parent query, to parse the WHERE clause, we create a new operator
type, called IN or NOT_IN, or ALL, where the left side is a Var, and the

^^^^^^^^^^^^^^^^^^
No. We have to handle (a,b,c) OP (select x, y, z ...) and
'_a_constant_' OP (select ...) - I don't know is last in standards,
Sybase has this.

I have never seen this in my eight years of SQL. Perhaps we can leave
this for later, maybe much later.

Well,

typedef enum OpType
{
OP_EXPR, FUNC_EXPR, OR_EXPR, AND_EXPR, NOT_EXPR

+ OP_EXISTS, OP_ALL, OP_ANY

} OpType;

typedef struct Expr
{
NodeTag type;
Oid typeOid; /* oid of the type of this expr */
OpType opType; /* type of the op */
Node *oper; /* could be Oper or Func */
List *args; /* list of argument nodes */
} Expr;

OP_EXISTS: oper is NULL, lfirst(args) is SubSelect (index in subqueries
List, following your suggestion)

OP_ALL, OP_ANY:

oper is List of Oper nodes. We need in list because of data types of
a, b, c (above) can be different and so Oper nodes will be different too.

lfirst(args) is List of expression nodes (Const, Var, Func ?, a + b ?) -
left side of subquery' operator.
lsecond(args) is SubSelect.

Note, that there are no OP_IN, OP_NOTIN in OpType-s for Expr. We need in
IN, NOTIN in A_Expr (parser node), but both of them have to be transferred
by parser into corresponding ANY and ALL. At the moment we can do:

IN --> = ANY, NOT IN --> <> ALL

but this will be "known bug": this breaks OO-nature of Postgres, because of
operators can be overrided and '=' can mean s o m e t h i n g (not equality).
Example: box data type. For boxes, = means equality of _areas_ and =~
means that boxes are the same ==> =~ ANY should be used for IN.

That is interesting, to use =~ for ANY.

Yes, but how many operators take a SUBQUERY as an operand. This is a
special case to me.

I think I see where you are trying to go. You want subselects to behave
like any other operator, with a subselect type, and you do all the
subselect handling in the optimizer, with special Nodes and actions.

I think this may be just too much of a leap. We have such clean query
logic for single queries, I can't imagine having an operator that has a
Query operand, and trying to get everything to properly handle it.
UNIONS were very easy to implement as a List off of Query, with some
foreach()'s in rewrite and the high optimizer.

Subselects are SQL standard, and are never going to be over-ridden by a
user. Same with UNION. They want UNION, they get UNION. They want
Subselect, we are going to spin through the Query structure and give
them what they want.

The complexities of subselects and correlated queries and range tables
and stuff is so bizarre that trying to get it to work inside the type
system could be a huge project.

right side is an index to a slot in the subqueries List.

I guess the question is what can we have by February 1?

I have been reading some postings, and it seems to me that subselects
are the litmus test for many evaluators when deciding if a database
engine is full-featured.

Sorry to be so straightforward, but I want to keep hashing this around
until we get a conclusion, so coding can start.

My suggestions have been, I believe, trying to get subselects working
with the fullest functionality by adding the least amount of code, and
keeping the logic clean.

Have you checked out the UNION code? It is very small, but it works. I
think it could make a good sample for subselects.

--
Bruce Momjian
maillist@candle.pha.pa.us

Bruce Momjian wrote:

No, I don't like to add anything in parser. Example:

select *
from tabA
where col1 = (select col2
from tabB
where tabA.col3 = tabB.col4
and exists (select *
from tabC
where tabB.colX = tabC.colX and
tabC.colY = tabA.col2)
)

: a column of tabA is referenced in sub-subselect

This is a strange case that I don't think we need to handle in our first
implementation.

I don't know is this strange case or not :)
But I would like to know is this allowed by standards - can someone
comment on this ?
And I don't see problems with handling this...

(is it allowable by standards ?) - in this case it's better
to don't add tabA to 1st subselect but add tabA to second one
and change tabA.col3 in 1st to reference col3 in 2nd subquery temp table -
this gives us 2-tables join in 1st subquery instead of 3-tables join.
(And I'm still not sure that using temp tables is best of what can be
done in all cases...)

I don't see any use for temp tables in subselects anymore. After having
implemented UNIONS, I now see how much can be done in the upper
optimizer. I see you just putting the subquery PLAN into the proper
place in the plan tree, with some proper JOIN nodes for IN, NOT IN.

When saying about temp tables, I meant tables created by node Material
for subquery plan. This is one of two ways - run subquery once for all
possible upper plan tuples and then just join result table with upper
query. Another way is re-run subquery for each upper query tuple,
without temp table but may be with caching results by some ways.
Actually, there is special case - when subquery can be alternatively
formulated as joins, - but this is just special case.

In the parent query, to parse the WHERE clause, we create a new operator
type, called IN or NOT_IN, or ALL, where the left side is a Var, and the

^^^^^^^^^^^^^^^^^^
No. We have to handle (a,b,c) OP (select x, y, z ...) and
'_a_constant_' OP (select ...) - I don't know is last in standards,
Sybase has this.

I have never seen this in my eight years of SQL. Perhaps we can leave
this for later, maybe much later.

Are you saying about (a, b, c) or about 'a_constant' ?
Again, can someone comment on are they in standards or not ?
Tom ?
If yes then please add parser' support for them now...

Note, that there are no OP_IN, OP_NOTIN in OpType-s for Expr. We need in
IN, NOTIN in A_Expr (parser node), but both of them have to be transferred
by parser into corresponding ANY and ALL. At the moment we can do:

IN --> = ANY, NOT IN --> <> ALL

but this will be "known bug": this breaks OO-nature of Postgres, because of
operators can be overrided and '=' can mean s o m e t h i n g (not equality).
Example: box data type. For boxes, = means equality of _areas_ and =~
means that boxes are the same ==> =~ ANY should be used for IN.

That is interesting, to use =~ for ANY.

Yes, but how many operators take a SUBQUERY as an operand. This is a
special case to me.

I think I see where you are trying to go. You want subselects to behave
like any other operator, with a subselect type, and you do all the
subselect handling in the optimizer, with special Nodes and actions.

I think this may be just too much of a leap. We have such clean query
logic for single queries, I can't imagine having an operator that has a
Query operand, and trying to get everything to properly handle it.
UNIONS were very easy to implement as a List off of Query, with some
foreach()'s in rewrite and the high optimizer.

Subselects are SQL standard, and are never going to be over-ridden by a
user. Same with UNION. They want UNION, they get UNION. They want
Subselect, we are going to spin through the Query structure and give
them what they want.

The complexities of subselects and correlated queries and range tables
and stuff is so bizarre that trying to get it to work inside the type
system could be a huge project.

PostgreSQL is a robust, next-generation, Object-Relational DBMS (ORDBMS),
derived from the Berkeley Postgres database management system. While
PostgreSQL retains the powerful object-relational data model, rich data types and
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
easy extensibility of Postgres, it replaces the PostQuel query language with an
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
extended subset of SQL.
^^^^^^^^^^^^^^^^^^^^^^

Should we say users that subselect will work for standard data types only ?
I don't see why subquery can't be used with ~, ~*, @@, ... operators, do you ?
Is there difference between handling = ANY and ~ ANY ? I don't see any.
Currently we can't get IN working properly for boxes (and may be for others too)
and I don't like to try to resolve these problems now, but hope that someday
we'll be able to do this. At the moment - just convert IN into = ANY and
NOT IN into <> ALL in parser.

(BTW, do you know how DISTINCT is implemented ? It doesn't use = but
use type_out funcs and uses strcmp()... DISTINCT is standard SQL thing...)

right side is an index to a slot in the subqueries List.

I guess the question is what can we have by February 1?

I have been reading some postings, and it seems to me that subselects
are the litmus test for many evaluators when deciding if a database
engine is full-featured.

Sorry to be so straightforward, but I want to keep hashing this around
until we get a conclusion, so coding can start.

My suggestions have been, I believe, trying to get subselects working
with the fullest functionality by adding the least amount of code, and
keeping the logic clean.

Have you checked out the UNION code? It is very small, but it works. I
think it could make a good sample for subselects.

There is big difference between subqueries and queries in UNION -
there are not dependences between UNION queries.

Ok, opened issues:

1. Is using upper query' vars in all subquery levels in standard ?
2. Is (a, b, c) OP (subselect) in standard ?
3. What types of expressions (Var, Const, ...) are allowed on the left
side of operator with subquery on the right ?
4. What types of operators should we support (=, >, ..., like, ~, ...) ?
(My vote for all boolean operators).

And - did we get consensus on presentation subqueries stuff in Query,
Expr and Var ?
I would like to have something done in parser near Jan 17 to get
subqueries working by Feb 1. I vote for support of all standard
things (1. - 3.) in parser right now - if there will be no time
to implement something like (a, b, c) then optimizer will call
elog(WARN) (oh, sorry, - elog(ERROR)).

Vadim

Bruce Momjian wrote:

OK, a few questions:

Should we use sortmerge, so we can use our psort as temp tables,
or do we use hashunique?

How do we pass the query to the optimizer? How do we represent
the range table for each, and the links between them in correlated
subqueries?

My suggestion is just use varlevel in Var and don't put upper query'
relations into subquery range table.

Vadim

Vadim B. Mikheev wrote:

Bruce Momjian wrote:

OK, a few questions:

Should we use sortmerge, so we can use our psort as temp tables,
or do we use hashunique?

How do we pass the query to the optimizer? How do we represent
the range table for each, and the links between them in correlated
subqueries?

My suggestion is just use varlevel in Var and don't put upper query'
relations into subquery range table.

Hmm... Sorry, it seems that I did reply to very old message - forget it.

Vadim

Note, that there are no OP_IN, OP_NOTIN in OpType-s for Expr. We need in
IN, NOTIN in A_Expr (parser node), but both of them have to be transferred
by parser into corresponding ANY and ALL. At the moment we can do:

IN --> = ANY, NOT IN --> <> ALL

but this will be "known bug": this breaks OO-nature of Postgres, because of
operators can be overrided and '=' can mean s o m e t h i n g (not equality).
Example: box data type. For boxes, = means equality of _areas_ and =~
means that boxes are the same ==> =~ ANY should be used for IN.

That is interesting, to use =~ for ANY.

If I understand the discussion, I would think is is fine to make an assumption about
which operator is used to implement a subselect expression. If someone remaps an
operator to mean something different, then they will get a different result (or a
nonsensical one) from a subselect.

I'd be happy to remap existing operators to fit into a convention which would work
with subselects (especially if I got to help choose :).

Subselects are SQL standard, and are never going to be over-ridden by a
user. Same with UNION. They want UNION, they get UNION. They want
Subselect, we are going to spin through the Query structure and give
them what they want.

PostgreSQL is a robust, next-generation, Object-Relational DBMS (ORDBMS),
derived from the Berkeley Postgres database management system. While
PostgreSQL retains the powerful object-relational data model, rich data types and
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
easy extensibility of Postgres, it replaces the PostQuel query language with an
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
extended subset of SQL.
^^^^^^^^^^^^^^^^^^^^^^

Should we say users that subselect will work for standard data types only ?
I don't see why subquery can't be used with ~, ~*, @@, ... operators, do you ?
Is there difference between handling = ANY and ~ ANY ? I don't see any.
Currently we can't get IN working properly for boxes (and may be for others too)
and I don't like to try to resolve these problems now, but hope that someday
we'll be able to do this. At the moment - just convert IN into = ANY and
NOT IN into <> ALL in parser.

(BTW, do you know how DISTINCT is implemented ? It doesn't use = but
use type_out funcs and uses strcmp()... DISTINCT is standard SQL thing...)

?? I didn't know that. Wouldn't we want it to eventually use "=" through a sorted
list? That would give more consistant behavior...

I have been reading some postings, and it seems to me that subselects
are the litmus test for many evaluators when deciding if a database
engine is full-featured.

Sorry to be so straightforward, but I want to keep hashing this around
until we get a conclusion, so coding can start.

My suggestions have been, I believe, trying to get subselects working
with the fullest functionality by adding the least amount of code, and
keeping the logic clean.

Have you checked out the UNION code? It is very small, but it works. I
think it could make a good sample for subselects.

There is big difference between subqueries and queries in UNION -
there are not dependences between UNION queries.

Ok, opened issues:

1. Is using upper query' vars in all subquery levels in standard ?

I'm not certain. Let me know if you do not get an answer from someone else and I will
research it.

2. Is (a, b, c) OP (subselect) in standard ?

Yes. In fact, it _is_ the standard, and "a OP (subselect)" is a special case where
the parens are allowed to be omitted from a one element list.

3. What types of expressions (Var, Const, ...) are allowed on the left
side of operator with subquery on the right ?

I think most expressions are allowed. The "constant OP (subselect)" case you were
asking about is just a simplified case since "(a, b, constant) OP (subselect)" where
a and b are column references should be allowed. Of course, our optimizer could
perhaps change this to "(a, b) OP (subselect where x = constant)", or for the first
example "EXISTS (subselect where x = constant)".

4. What types of operators should we support (=, >, ..., like, ~, ...) ?
(My vote for all boolean operators).

Sounds good. But I'll vote with Bruce (and I'll bet you already agree) that it is
important to get an initial implementation for v6.3 which covers a little, some, or
all of the usual SQL subselect constructs. If we have to revisit this for v6.4 then
we will have the benefit of feedback from others in practical applications which
always uncovers new things to consider.

And - did we get consensus on presentation subqueries stuff in Query,
Expr and Var ?
I would like to have something done in parser near Jan 17 to get
subqueries working by Feb 1. I vote for support of all standard
things (1. - 3.) in parser right now - if there will be no time
to implement something like (a, b, c) then optimizer will callelog(WARN) (oh,
sorry, - elog(ERROR)).

Great. I'd like to help with the remaining parser issues; at the moment "row_expr"
does the right thing with expression comparisions but just parses then ignores
subselect expressions. Let me know what structures you want passed back and I'll put
them in, or if you prefer put in the first one and I'll go through and clean up and
add the rest.

- Tom

On Sun, 11 Jan 1998, Vadim B. Mikheev wrote:

No, I don't like to add anything in parser. Example:

select *
from tabA
where col1 = (select col2
from tabB
where tabA.col3 = tabB.col4
and exists (select *
from tabC
where tabB.colX = tabC.colX and
tabC.colY = tabA.col2)
)

: a column of tabA is referenced in sub-subselect

This is a strange case that I don't think we need to handle in our first
implementation.

I don't know is this strange case or not :)
But I would like to know is this allowed by standards - can someone
comment on this ?
And I don't see problems with handling this...

I don't know about "the standards", but in my mind, the above should
work if subselects work...so what if you add a third or fourth level subselect
to the overall query? IMHO, the "outer most" (inner most?) subselect should
be resolved to provide the "EXISTS" list, the the next should be resolved,
etc...

Hell...looking at this, I'd almost think that you could use subselects to
force a pseudo-ordering onto a large complex JOIN (ya ya, really messy though)

Marc G. Fournier
Systems Administrator @ hub.org
primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org

Are you saying about (a, b, c) or about 'a_constant' ?
Again, can someone comment on are they in standards or not ?
Tom ?
If yes then please add parser' support for them now...

As I mentioned a few minutes ago in my last message, I parse the row descriptors and
the subselects but for subselect expressions (e.g. "(a,b) OP (subselect)" I currently
ignore the result. I didn't want to pass things back as lists until something in the
backend was ready to receive them.

If it is OK, I'll go ahead and start passing back a list of expressions when a row
descriptor is present. So, what you will find is lexpr or rexpr in the A_Expr node
being a list rather than an atomic node.

Also, I can start passing back the subselect expression as the rexpr; right now the
parser calls elog() and quits.

btw, to implement "(a,b,c) OP (d,e,f)" I made a new routine in the parser called
makeRowExpr() which breaks this up into a sequence of "and" and/or "or" expressions.
If lists are handled farther back, this routine should move to there also and the
parser will just pass the lists. Note that some assumptions have to be made about the
meaning of "(a,b) OP (c,d)", since usually we only have knowledge of the behavior of
"a OP c". Easy for the standard SQL operators, unknown for others, but maybe it is OK
to disallow those cases or to look for specific appearance of the operator to guess
the behavior (e.g. if the operator has "<" or "=" or ">" then build as "and"s and if
it has "<>" or "!" then build as "or"s.

Let me know what you want...

- Tom

I would like to have something done in parser near Jan 17 to get
subqueries working by Feb 1.

Here are some changes to gram.y and to keywords.c which start to pass through
subselect constructs. I won't commit until/unless you have a chance to look at it and
agree that this is something close to the right direction to head.

- Tom

postgres=> create table x (i int);
CREATE
postgres=> insert into x values (1);
INSERT 18121 1
postgres=> select i from x where i = 1;
i
-
1
(1 row)

postgres=> select i from x where i in (select i from x);
ERROR: transformExpr: does not know how to transform node 604
postgres=> select i from x where (i, 1) in (select i, 1 from x);
ERROR: transformExpr: does not know how to transform node 501
postgres=>

This is a multi-part message in MIME format.
--------------130974A8F3C8025EB9E3F25C
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit

I would like to have something done in parser near Jan 17 to get
subqueries working by Feb 1.

Here are some changes to gram.y and to keywords.c which start to pass through
subselect constructs. I won't commit until/unless you have a chance to look at it and
agree that this is something close to the right direction to head.

Do you realize these are the files, and not context diffs?

--
Bruce Momjian
maillist@candle.pha.pa.us

I would like to have something done in parser near Jan 17 to get
subqueries working by Feb 1.

Here are some changes to gram.y and to keywords.c which start to pass through
subselect constructs. I won't commit until/unless you have a chance to look at it and
agree that this is something close to the right direction to head.

Do you realize these are the files, and not context diffs?

Yup. Thought it would be easier for you, but probably should have sent a diff. Sorry.

- Tom

Bruce Momjian wrote:

OK, a few questions:

Should we use sortmerge, so we can use our psort as temp tables,
or do we use hashunique?

How do we pass the query to the optimizer? How do we represent
the range table for each, and the links between them in correlated
subqueries?

My suggestion is just use varlevel in Var and don't put upper query'
relations into subquery range table.

Vadim

OK.

--
Bruce Momjian
maillist@candle.pha.pa.us

Here are context diffs of gram.y and keywords.c; sorry about sending the full files.
These start sending lists of arguments toward the backend from the parser to
implement row descriptors and subselects.

They should apply OK even over Bruce's recent changes...

- Tom

I would like to have something done in parser near Jan 17 to get
subqueries working by Feb 1. I vote for support of all standard
things (1. - 3.) in parser right now - if there will be no time
to implement something like (a, b, c) then optimizer will call
elog(WARN) (oh, sorry, - elog(ERROR)).

First, let me say I am glad we are still on schedule for Feb 1. I was
panicking because I thought we wouldn't make it in time.

(is it allowable by standards ?) - in this case it's better
to don't add tabA to 1st subselect but add tabA to second one
and change tabA.col3 in 1st to reference col3 in 2nd subquery temp table -
this gives us 2-tables join in 1st subquery instead of 3-tables join.
(And I'm still not sure that using temp tables is best of what can be
done in all cases...)

I don't see any use for temp tables in subselects anymore. After having
implemented UNIONS, I now see how much can be done in the upper
optimizer. I see you just putting the subquery PLAN into the proper
place in the plan tree, with some proper JOIN nodes for IN, NOT IN.

When saying about temp tables, I meant tables created by node Material
for subquery plan. This is one of two ways - run subquery once for all
possible upper plan tuples and then just join result table with upper
query. Another way is re-run subquery for each upper query tuple,
without temp table but may be with caching results by some ways.
Actually, there is special case - when subquery can be alternatively
formulated as joins, - but this is just special case.

This is interesting. It really only applies for correlated subqueries,
and certainly it may help sometimes to just evaluate the subquery for
valid values that are going to come from the upper query than for all
possible values. Perhaps we can use the 'cost' value of each query to
decide how to handle this.

In the parent query, to parse the WHERE clause, we create a new operator
type, called IN or NOT_IN, or ALL, where the left side is a Var, and the

^^^^^^^^^^^^^^^^^^
No. We have to handle (a,b,c) OP (select x, y, z ...) and
'_a_constant_' OP (select ...) - I don't know is last in standards,
Sybase has this.

I have never seen this in my eight years of SQL. Perhaps we can leave
this for later, maybe much later.

Are you saying about (a, b, c) or about 'a_constant' ?
Again, can someone comment on are they in standards or not ?
Tom ?
If yes then please add parser' support for them now...

OK, Thomas says it is, so we will put in as much code as we can to handle
it.

Should we say users that subselect will work for standard data types only ?
I don't see why subquery can't be used with ~, ~*, @@, ... operators, do you ?
Is there difference between handling = ANY and ~ ANY ? I don't see any.
Currently we can't get IN working properly for boxes (and may be for others too)
and I don't like to try to resolve these problems now, but hope that someday
we'll be able to do this. At the moment - just convert IN into = ANY and
NOT IN into <> ALL in parser.

OK.

(BTW, do you know how DISTINCT is implemented ? It doesn't use = but
use type_out funcs and uses strcmp()... DISTINCT is standard SQL thing...)

I did not know that either.

There is big difference between subqueries and queries in UNION -
there are not dependences between UNION queries.

Yes, I know UNIONS are trivial compared to subselects.

Ok, opened issues:

1. Is using upper query' vars in all subquery levels in standard ?
2. Is (a, b, c) OP (subselect) in standard ?
3. What types of expressions (Var, Const, ...) are allowed on the left
side of operator with subquery on the right ?
4. What types of operators should we support (=, >, ..., like, ~, ...) ?
(My vote for all boolean operators).

And - did we get consensus on presentation subqueries stuff in Query,
Expr and Var ?

OK, here are my concrete ideas on changes and structures.

I think we all agreed that Query needs new fields:

Query *parentQuery;
List *subqueries;

Maybe query level too, but I don't think so (see later ideas on Var).

We need a new Node structure, call it Sublink:

int linkType (IN, NOTIN, ANY, EXISTS, OPERATOR...)
Oid operator /* subquery must return single row */
List *lefthand; /* parent stuff */
Node *subquery; /* represents nodes from parser */
Index Subindex; /* filled in to index Query->subqueries */

Of course, the names are just suggestions. Every time we run through
the parsenodes of a query to create a Query* structure, when we do the
WHERE clause, if we come upon one of these Sublink nodes (created in the
parser), we move the supplied Query* in Sublink->subquery to a local
List variable, and we set Subquery->subindex to equal the index of the
new query, i.e. is it the first subquery we found, 1, or the second, 2,
etc.

After we have created the parent Query structure, we run through our
local List variable of subquery parsenodes we created above, and add
Query* entries to Query->subqueries. In each subquery Query*, we set
the parentQuery pointer.

Also, when parsing the subqueries, we need to keep track of correlated
references. I recommend we add a field to the Var structure:

Index sublevel; /* range table reference:
= 0 current level of query
< 0 parent above this many levels

0 index into subquery list

*/

This way, a Var node with sublevel 0 is the current level, and is true
in most cases. This helps us not have to change much code. sublevel =
-1 means it references the range table in the parent query. sublevel =
-2 means the parent's parent. sublevel = 2 means it references the range
table of the second entry in Query->subqueries. Varno and varattno are
still meaningful. Of course, we can't reference variables in the
subqueries from the parent in the parser code, but Vadim may want to.

When doing a Var lookup in the parser, we look in the current level
first, but if not found, if it is a subquery, we can look at the parent
and parent's parent to set the sublevel, varno, and varatno properly.

We create no phantom range table entries in the subquery, and no phantom
target list entries. We can leave that all for the upper optimizer.

--
Bruce Momjian
maillist@candle.pha.pa.us

Bruce Momjian wrote:

We need a new Node structure, call it Sublink:

int linkType (IN, NOTIN, ANY, EXISTS, OPERATOR...)
Oid operator /* subquery must return single row */
List *lefthand; /* parent stuff */
Node *subquery; /* represents nodes from parser */
Index Subindex; /* filled in to index Query->subqueries */

Ok, I agreed that it's better to have new node and don't put subquery stuff
into Expr node.

int linkType
is one of EXISTS, ANY, ALL, EXPR. EXPR is for the case of expression
subqueries (following Sybase naming) which must return single row -
(a, b, c) = (subquery).
Note again, that there are no linkType for IN and NOTIN here.
User' IN and NOT IN must be converted to = ANY and <> ALL by parser.

We need not in Oid operator! In all cases we need in

List *oper
list of Oper nodes for each of a, b, c, ... and operator (=, ...)
corresponding to data type of a, b, c, ...

List *lefthand
is list of Var/Const nodes - representation of (a, b, c, ...)

What is Node *subquery ?
In optimizer we need either in Subindex (to get subquery from Query->subqueries
when beeing in Sublink) or in Node *subquery inside Sublink itself.
BTW, after some thought I don't see how Query->subqueries will be usefull.
So, may be just add bool hassubqueries to Query (and Query *parentQuery)
and use Query *subquery in Sublink, but not subindex ?

Also, when parsing the subqueries, we need to keep track of correlated
references. I recommend we add a field to the Var structure:

Index sublevel; /* range table reference:
= 0 current level of query
< 0 parent above this many levels

0 index into subquery list

*/

This way, a Var node with sublevel 0 is the current level, and is true
in most cases. This helps us not have to change much code. sublevel =
-1 means it references the range table in the parent query. sublevel =
-2 means the parent's parent. sublevel = 2 means it references the range
table of the second entry in Query->subqueries. Varno and varattno are
still meaningful. Of course, we can't reference variables in the
subqueries from the parent in the parser code, but Vadim may want to.

^^^^^^^^^^^^^^^^^
No. So, just use sublevel >= 0: 0 - current level, 1 - one level up, ...
sublevel is for optimizer only - executor will not use it.

When doing a Var lookup in the parser, we look in the current level
first, but if not found, if it is a subquery, we can look at the parent
and parent's parent to set the sublevel, varno, and varatno properly.

We create no phantom range table entries in the subquery, and no phantom
target list entries. We can leave that all for the upper optimizer.

Ok.

Vadim

Thomas G. Lockhart wrote:

btw, to implement "(a,b,c) OP (d,e,f)" I made a new routine in the parser called
makeRowExpr() which breaks this up into a sequence of "and" and/or "or" expressions.
If lists are handled farther back, this routine should move to there also and the
parser will just pass the lists. Note that some assumptions have to be made about the
meaning of "(a,b) OP (c,d)", since usually we only have knowledge of the behavior of
"a OP c". Easy for the standard SQL operators, unknown for others, but maybe it is OK
to disallow those cases or to look for specific appearance of the operator to guess
the behavior (e.g. if the operator has "<" or "=" or ">" then build as "and"s and if
it has "<>" or "!" then build as "or"s.

Oh, god! I never thought about this!
Ok, I have to agree:

1. Only <, <=, =, >, >=, <> is allowed with subselects
2. Use OR's for <>, and so - we need in bool useor in SubLink
for <>, <> ANY and <> ALL:

typedef struct SubLink {
NodeTag type;
int linkType; /* EXISTS, ALL, ANY, EXPR */
bool useor; /* TRUE for <> */
List *lefthand; /* List of Var/Const nodes on the left */
List *oper; /* List of Oper nodes */
Query *subquery; /* */
} SubLink;

Vadim

Thomas G. Lockhart wrote:

btw, to implement "(a,b,c) OP (d,e,f)" I made a new routine in the parser called
makeRowExpr() which breaks this up into a sequence of "and" and/or "or" expressions.
If lists are handled farther back, this routine should move to there also and the
parser will just pass the lists. Note that some assumptions have to be made about the
meaning of "(a,b) OP (c,d)", since usually we only have knowledge of the behavior of
"a OP c". Easy for the standard SQL operators, unknown for others, but maybe it is OK
to disallow those cases or to look for specific appearance of the operator to guess
the behavior (e.g. if the operator has "<" or "=" or ">" then build as "and"s and if
it has "<>" or "!" then build as "or"s.

Sorry, I forgot something: is (a, b) OP (x, y) in standard ?
If not then I suggest to don't implement it at all and allow
(a, b) OP [ANY|ALL] (subselect) only.

Vadim

btw, to implement "(a,b,c) OP (d,e,f)" I made a new routine in the parser called
makeRowExpr() which breaks this up into a sequence of "and" and/or "or" expressions.
If lists are handled farther back, this routine should move to there also and the
parser will just pass the lists. Note that some assumptions have to be made about the
meaning of "(a,b) OP (c,d)", since usually we only have knowledge of the behavior of
"a OP c". Easy for the standard SQL operators, unknown for others, but maybe it is OK
to disallow those cases or to look for specific appearance of the operator to guess
the behavior (e.g. if the operator has "<" or "=" or ">" then build as "and"s and if
it has "<>" or "!" then build as "or"s.

Sorry, I forgot something: is (a, b) OP (x, y) in standard ?

Yes. The problem wouldn't be very interesting otherwise :)

- Tom

Show quoted text

If not then I suggest to don't implement it at all and allow
(a, b) OP [ANY|ALL] (subselect) only.

Thomas G. Lockhart wrote:

btw, to implement "(a,b,c) OP (d,e,f)" I made a new routine in the parser called
makeRowExpr() which breaks this up into a sequence of "and" and/or "or" expressions.
If lists are handled farther back, this routine should move to there also and the
parser will just pass the lists. Note that some assumptions have to be made about the
meaning of "(a,b) OP (c,d)", since usually we only have knowledge of the behavior of
"a OP c". Easy for the standard SQL operators, unknown for others, but maybe it is OK
to disallow those cases or to look for specific appearance of the operator to guess
the behavior (e.g. if the operator has "<" or "=" or ">" then build as "and"s and if
it has "<>" or "!" then build as "or"s.

Oh, god! I never thought about this!
Ok, I have to agree:

1. Only <, <=, =, >, >=, <> is allowed with subselects
2. Use OR's for <>, and so - we need in bool useor in SubLink
for <>, <> ANY and <> ALL:

Ah, but this is just a problem when there are multiple fields on the
left.

typedef struct SubLink {
NodeTag type;
int linkType; /* EXISTS, ALL, ANY, EXPR */
bool useor; /* TRUE for <> */
List *lefthand; /* List of Var/Const nodes on the left */
List *oper; /* List of Oper nodes */
Query *subquery; /* */
} SubLink;

--
Bruce Momjian
maillist@candle.pha.pa.us