Re: How hard would a "path" operator be to implement in PostgreSQL - Mailing list pgsql-general

On Aug 19, 2012, at 21:28, Chris Travers <chris.travers@gmail.com> wrote:

> Hi;
>
> I have been reading up on object-relational features of Oracle and DB2 and found that one of the big things they have
thatwe don't is a path operator.  The idea is that you can use the path operator to follow some subset of foreign keys
calledrefs.   

Why do you feel this is a "big thing".  Sure, you can save a few keystrokes when writing multi-table queries but that
doesn'tseem all that great and now you are introducing ambiguity into the planner/query when trying to resolve these
implicitjoins.  I concur that introducing an explicit REF is a poor choice taken at face value since now you have to
rememberwhat references are present.  With FOREIGN KEYS you introduce a logical constraint but you are able to perform
anjoin between two tables independent of the presence of an FK. 

>
> Suppose we have a table (Oracle/DB2 styles here):
>
> CREATE TABLE country (
>     REF IS OID USER GENERATED,
>     id int serial not null unique,
>     name text primary key,
>     short_name text not null unique
> );
>
> CREATE TABLE address (
>     street_text text not null,
>     city text not null,
>     state_province text.
>     country REF(country)
> );
>
> In oracle this might allow you to do something like:
>
> SELECT * from address where address.country.short_name = 'US';
>
> In DB2 this might be done like:
>
> SELECT * FROM address WHERE address->country->short_name = 'US';
>
> I like DB2's approach better because there is no ambiguity between namespace resolution but I don't entirely like the
waythe refs work as separate from standard foreign keys. 
>
> What I am thinking about is a function which accepts a row and a destination table name, looks up foreign keys in and
retrievesthe row from the other table, returning it, if the table name exists.   This could then be mapped to an
operatorwhich would avoid some problems.  This could then be: 
>
> SELECT * FROM address where (address->'country').short_name = 'US'

short for:

SELECT address.*
FROM address
JOIN country ON (country_id = country.id)
WHERE short_name = 'US';

depending on naming scheme a NATURAL JOIN or USING can shorten this a bit.

I guess there is some value in not making the columns from country available to the SELECT LIST...

Another form equates to...

SELECT * FROM address WHERE EXISTS (SELECT 1 FROM country WHERE id = country_id AND short_name = 'US';

This second form seems a more direct translation of the query.  At first glance this would seem to be a bad query to
writefor the given goal but maybe the planner does the right thing and so the two pure forms perform equivalently? 

>
> Or if we also have a continents table:
>
> SELECT * FROM address where (address->'country'->'continent').name = 'North America';
>
> Obviously these examples assume a very small number of address records and are largely contrived from the IBM
examples. However you could also do this on a small return set: 

As implied above if this only is going to generate meaningful plans on small datasets its value seems even more
limited...

>
> select (a->'country').short_name from address where ....;
>
> Or even:
>
> select (a->'country').* from address where......
>
> The next question is whether there is a way to pass country in as an identifier so there is no need to use single
quotes. This would make things a little more transparent if possible but I would be happy without this.  Eventually it
mightbe kinda useful (for those porting O-R stuff from Oracle or DB2) to have a path operator built in with a concept
ofa default, implicit join. 
>
> Any thoughts?  If it can be done in plain SQL and inlined that would be ideal but in the prototyping state, that
isn'tso important and I expect that it is not. 

I am presuming this syntax and capability is not specified in the SQL standard anywhere?

I don't have a dog in the "make things easier to port" fight but this seems like something that would be fairly easy to
identifyin Oracle/DB2 SQL code and to rewrite using explicit joins without losing any actual functionality. 

Given that you come across this via "reading" it doesn't sound like you have actually been bitten by its lack nor are
havingpeople complain about its lack so I suppose you are arguing for its inclusion based upon its own merits in making
codingeasier. 

>
> Best Wishes,
> Chris Traves