Re: Multiple "selects" returned from a single stored procedure - Mailing list pgsql-hackers

Joe Conway wrote:

> I think you're misinterpreting the docs. The CALL keyword has not been 
> implemented in Postgres AFAIK. Where precisely did you see that?

Can't find it any more. I'm pretty sure I read it somewhere. Doesn't matter.

>> I don't see that as a problem.
>
>
> It is THE problem as far as implementing what you're looking for. An 
> SRF will never be able to deal with multiple results sets because the 
> number and types of the returned attributes must be known/resolved 
> when the query is parsed. On the other hand, a stored procedure would 
> not have this restriction if it, *by design could not* participate in 
> a normal SELECT.

What I meant is that I don't see not being able to use stored procedures 
as a FROM source as a problem.

> Sure. I should think you'd allow a normal SELECT statement in your 
> stored procedure, exactly as you would do in MSSQL. The result tuples 
> could be formed and projected in similar fashion to EXPLAIN or SHOW 
> ALL. See, for example, ShowAllGUCConfig() in guc.c.

I'm more into client-side hacking than server-side hacking. I'm 
currently using libpq for client-server communications, so bumping the 
protocol version a notch higher won't affect me much, but it's a point 
to consider.

>
>> I have several ways I can take this solution, and I would like your 
>> opinion:
>> 1. Simply assume that if a command returned a single column of 
>> refcursors, that it meant to return several rowsets, and implement 
>> the corresponding OLE DB interface.
>
>
> This seems cleanest, except you might want a configurable option to 
> turn it off, in case someone really wants the refcursor results.

As soon as I figure out how to pass custom-attributes to OLE DB. Docs 
seem hazy on this. Marek, who originally wrote OLE DB, did manage to 
figure it out. Unfotunetly, I had to dump his code entirely for 
unrelated reasons. I'll be glad if he reintroduced it into this 
iteration (hint hint).

In any case, as OLE DB uses a binary interface to pass data around, I 
need to know of all data types you would want to participate in any 
activity. This means that refcursors will not be supported on their own 
unless I figure out how to return them standalone, which I don't see 
happening.

>
>> 2. Require that the cursors be named a certain way, according to 
>> their intended usage. If the command returned cursors named "<unnamed 
>> portal 9>", to just treat it as is, while if it returned a cursor 
>> named "MultiResult1", treat it as above?
>>
>> The advantage of 2 is that it allows me to simulate out variables. If 
>> the cursor is called "outputvars", I direct it to the output 
>> variables interface.
>> The disadvantage is that I'm not sure what to do if only some of the 
>> rows in the result are named MultiResult.
>
>
> This one seems a bit grotty to me. But the only other kludge I can 
> think of to similate output variables would require some backend 
> hacking, or at least a user defined C function.

I can also assume all refcursors are rowsets EXCEPT if one is named 
"outputvars". Then again, I'm doing this work in the context of the 
needs of a particular client, and he already said it was ok with him to 
pull the output vars from one of the rowsets.

>
> In case you're interested, here is the idea. Implement functions 
> necessary to create, change, and remove session local variables. If 
> the function uses named parameters (recently implemented for functions) 

There's one of the rubs. I'm basing my development on 7.4 backend. This 
has to go into production ASAP, and I can't rely on 7.5 being out by the 
time I'm done.

Also, I can't know (from the client) what the input vars are called 
unless I perform another query. The interface that does this extra query 
is highly recommended against by the OLE DB docs, for that reason.

Interestingly enough, my specific client does this extra query. However, 
I don't believe in writing general code based on a single entity's 
needs. The design must be one capable of expanding, even if the 
implementation is sparse (which it is, at the moment).

> *and* one or more session local variable of the same names are found, 
> assume the value of the variables are your outputvar results.

My backend knowledge is not good enough to understand this in depth, but 
I'll keep this message in case the need arises for later re-reading.

>> Whatever method I'll use, I may have to start a transaction for the 
>> purpose of the command, if we were not already in one. Otherwise, the 
>> ref-cursors are just useless strings.
>
>
> Cursors can now outlive transactions (DECLARE ... WITH HOLD), but 
> there is the downside (recently discussed on one of the lists) that 
> when the transaction is ended, the cursor is copied to a tuplestore. 
> As long as the tuplestore fits within sort_mem (work_mem in 7.5+), it 
> will be held entirely within memory. If not, it will spill to disk.

No, I think that's an overhead I can't afford. Let's not forget that I'm 
a driver, and I'm expected to be relatively transparent, performance 
wise. At the moment, I'm not using cursors (lazy fetching), so libpq is 
allocating quite a bit of memory on the client side. Reserving such an 
amount on the server side seems a little too much.

Just out of curiosity, how do I manually destroy the cursor when it's no 
longer needed? Just do "close refcursor"?

Also, does this copy take place when the table is changed, or as soon as 
the transaction ends? If the former, it may not matter.

In any case, my original concern was that I'll be changing semantics by 
opening another transaction. Upon retrospect, however, that's precisely 
what PostgreSQL is doing if one is not active. Assuming the OLE DB user 
uses the Transaction interface, and does not send a "begin" command, I'm 
capable of knowing whether I'm inside a transaction or not.

> Joe
         Shachar

-- 
Shachar Shemesh
Lingnu Open Source Consulting
http://www.lingnu.com/