IN/OUT parameters

Dave Cramer <pg@fastcrypt.com> writes:

On a related note, drop function needs to support the in/out direction.

It does ... do you see a problem?

regards, tom lane

Hmmmm ....

It seems to me that the bind message needs to support the notion of
direction in order for this to work cleanly.

Alternatively we could punt and use SQL Server's mechanism where they
only support IN, and INOUT, which would require all parameters to be
sent to the procedure.

Does SQL2003 cover IN/OUT parameters? I don't see anything in SQL99.

--
--Josh

Josh Berkus
Aglio Database Solutions
San Francisco

Dave Cramer <pg@fastcrypt.com> writes:

create function foo( out p1 int, in p2 int, out p3 int) ....
then a subsequent

"{call = select foo(?,?,?)}

This would need to be currently transformed into select foo(?), with the
other two being discarded.

It seems to me that the bind message needs to support the notion of
direction in order for this to work cleanly.

How would it help for BIND to incorporate direction? What would it even
*mean* for BIND to incorporate direction --- it's a client-to-server
message, and can hardly be expected to transmit data in the reverse
direction.

regards, tom lane

Ok, then I guess the documentation needs to indicate that. Or the
version I am reading is not up to date.

Dave

Tom Lane wrote:

Dave Cramer <pg@fastcrypt.com> writes:

On a related note, drop function needs to support the in/out direction.

It does ... do you see a problem?

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 2: you can get off all lists at once with the unregister command
(send "unregister YourEmailAddressHere" to majordomo@postgresql.org)

--
Dave Cramer
http://www.postgresintl.com
519 939 0336
ICQ#14675561

At this point I'd just like the backend to ignore the OUT parameter if
it were set that way, but allow
it to be sent.

If you consider that a function could have for arguments sake 10
parameters, and they
can be ordered in any fashion; clients have to essentially parse out the
OUT parameters and not send them. So you have an incoming client call
with 10 parameters but some number less than or equal to get sent to the
backend.

At this point I'm leaning towards just using IN, or INOUT params, as
jdbc has a register outparam, and this could be bound to null if were
not explicitly set.

Dave

Tom Lane wrote:

Dave Cramer <pg@fastcrypt.com> writes:

create function foo( out p1 int, in p2 int, out p3 int) ....
then a subsequent

"{call = select foo(?,?,?)}

This would need to be currently transformed into select foo(?), with the
other two being discarded.

It seems to me that the bind message needs to support the notion of
direction in order for this to work cleanly.

How would it help for BIND to incorporate direction? What would it even
*mean* for BIND to incorporate direction --- it's a client-to-server
message, and can hardly be expected to transmit data in the reverse
direction.

regards, tom lane

--
Dave Cramer
http://www.postgresintl.com
519 939 0336
ICQ#14675561

Tom,

How would it help for BIND to incorporate direction? What would it even
*mean* for BIND to incorporate direction --- it's a client-to-server
message, and can hardly be expected to transmit data in the reverse
direction.

Where directionality comes in is with OUT-only parameters. Which, IMHO,
aren't terribly important unless SQL2003 demands them; MSSQL didn't even
bother to implement them. Anyone know what SQL3 says?

--
--Josh

Josh Berkus
Aglio Database Solutions
San Francisco

Dave Cramer <pg@fastcrypt.com> writes:

At this point I'd just like the backend to ignore the OUT parameter if
it were set that way, but allow it to be sent.

I think you're thinking at the wrong level. AIUI the issue occurs at
the Parse stage, long before any parameter value is sent (or not sent).

What you really want, if I'm understanding correctly, is to be able to
send SQL that looks like this:

SELECT * FROM myfunc($1,$2,$3);

and then have some of the Param symbols be simply ignored while looking
up myfunc().

This is pretty ugly in itself, and I'd not want to add a protocol change
into the mix to make it happen. But possibly we could do it without any
protocol-level change. What would you say to specifying that Params
that are declared as type VOID are ignored in a function lookup? What
this would mean is that you'd need to do the following:

1. The SQL string can look as above.

2. When sending the Parse message, you'd specify the parameter types
as, say, INT4/VOID/VOID (or possibly UNKNOWN/VOID/VOID).

3. When sending Bind, you'd have to specify dummy values (probably
nulls) for the VOID parameter positions.

While I haven't looked at the backend code yet, I think we could drop
VOID-type parameters out of the argument list of a function during
parse analysis without too much trouble. This would affect *all*
function calls, not only those appearing in SELECT * FROM, but
I don't see any legitimate use of VOID-type values that this would
interfere with.

Klugy, isn't it? Anyone have a better idea? Or should we just tell
Dave that the JDBC driver ought to handle it?

regards, tom lane

Dave Cramer <pg@fastcrypt.com> writes:

Ok, then I guess the documentation needs to indicate that. Or the
version I am reading is not up to date.

Oh, you're right, I overlooked adding that to the documentation of
the other statement types that work with function names.

Come to think of it, the regprocedure type may need some work too.
It'd probably be nice if it understood IN/OUT decorations, at
least for input.

regards, tom lane

People:

OK, found it in SQL99:

<SQL parameter declaration> ::=
[ <parameter mode> ] [ <SQL parameter name> ] <parameter type> [ RESULT ]
<parameter mode> ::=
IN
| OUT
| INOUT

... so this is something we need to support, apparently both for Functions and
Procedures (when we get the latter), in the backend, not just JDBC. As you
can imagine, though, SQL03 does nothing to clarify calling rules for IN/OUT
params.

--
--Josh

Josh Berkus
Aglio Database Solutions
San Francisco

I think this is the driver's problem. It appears that we are following
the spec, so lets leave this alone.

Regarding what an OUT parameter might mean. We already have some
facility in the FE/BE to indicate the types of the returning columns.
Directionality is implied by where in the bind message these parameters
are sent. I can see the extra overhead of putting directionality in for
each column.
I'd agree that changing the FE/BE at this point is not a good idea.
However, if we ever do get real stored procs, this may be an issue.

I've looked at the driver some more and it appears we parse the sql more
than once due to the fact that we are supporting v2, and v3 backend
protocols. At this point I've got to consider the TODO on the jdbc list
of putting a proper parser in. Are there any javacc wizards around ?

Dave
Tom Lane wrote:

Dave Cramer <pg@fastcrypt.com> writes:

At this point I'd just like the backend to ignore the OUT parameter if
it were set that way, but allow it to be sent.

I think you're thinking at the wrong level. AIUI the issue occurs at
the Parse stage, long before any parameter value is sent (or not sent).

What you really want, if I'm understanding correctly, is to be able to
send SQL that looks like this:

SELECT * FROM myfunc($1,$2,$3);

and then have some of the Param symbols be simply ignored while looking
up myfunc().

This is pretty ugly in itself, and I'd not want to add a protocol change
into the mix to make it happen. But possibly we could do it without any
protocol-level change. What would you say to specifying that Params
that are declared as type VOID are ignored in a function lookup? What
this would mean is that you'd need to do the following:

1. The SQL string can look as above.

2. When sending the Parse message, you'd specify the parameter types
as, say, INT4/VOID/VOID (or possibly UNKNOWN/VOID/VOID).

3. When sending Bind, you'd have to specify dummy values (probably
nulls) for the VOID parameter positions.

While I haven't looked at the backend code yet, I think we could drop
VOID-type parameters out of the argument list of a function during
parse analysis without too much trouble. This would affect *all*
function calls, not only those appearing in SELECT * FROM, but
I don't see any legitimate use of VOID-type values that this would
interfere with.

Klugy, isn't it? Anyone have a better idea? Or should we just tell
Dave that the JDBC driver ought to handle it?

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 7: don't forget to increase your free space map settings

--
Dave Cramer
http://www.postgresintl.com
519 939 0336
ICQ#14675561