Chapter 43. Interfacing Extensions To Indices

The procedures described thus far let you define a new type, new functions and new operators. However, we cannot yet define a secondary index (such as a B-tree, R-tree or hash access method) over a new type or its operators.

Look back at The major Postgres system catalogs. The right half shows the catalogs that we must modify in order to tell Postgres how to use a user-defined type and/or user-defined operators with an index (i.e., pg_am, pg_amop, pg_amproc, pg_operator and pg_opclass). Unfortunately, there is no simple command to do this. We will demonstrate how to modify these catalogs through a running example: a new operator class for the B-tree access method that stores and sorts complex numbers in ascending absolute value order.

The pg_am class contains one instance for every user defined access method. Support for the heap access method is built into Postgres, but every other access method is described here. The schema is

The object ID of the instance in pg_am is used as a foreign key in lots of other classes. You don't need to add a new instance to this class; all you're interested in is the object ID of the access method instance you want to extend:

SELECT oid FROM pg_am WHERE amname = 'btree';

 oid
-----
 403
(1 row)
   

The amstrategies attribute exists to standardize comparisons across data types. For example, B-trees impose a strict ordering on keys, lesser to greater. Since Postgres allows the user to define operators, Postgres cannot look at the name of an operator (eg, ">" or "<") and tell what kind of comparison it is. In fact, some access methods don't impose any ordering at all. For example, R-trees express a rectangle-containment relationship, whereas a hashed data structure expresses only bitwise similarity based on the value of a hash function. Postgres needs some consistent way of taking a qualification in your query, looking at the operator and then deciding if a usable index exists. This implies that Postgres needs to know, for example, that the "<=" and ">" operators partition a B-tree. Postgres uses strategies to express these relationships between operators and the way they can be used to scan indices.

Defining a new set of strategies is beyond the scope of this discussion, but we'll explain how B-tree strategies work because you'll need to know that to add a new operator class. In the pg_am class, the amstrategies attribute is the number of strategies defined for this access method. For B-trees, this number is 5. These strategies correspond to

The idea is that you'll need to add procedures corresponding to the comparisons above to the pg_amop relation (see below). The access method code can use these strategy numbers, regardless of data type, to figure out how to partition the B-tree, compute selectivity, and so on. Don't worry about the details of adding procedures yet; just understand that there must be a set of these procedures for int2, int4, oid, and every other data type on which a B-tree can operate.

Sometimes, strategies aren't enough information for the system to figure out how to use an index. Some access methods require other support routines in order to work. For example, the B-tree access method must be able to compare two keys and determine whether one is greater than, equal to, or less than the other. Similarly, the R-tree access method must be able to compute intersections, unions, and sizes of rectangles. These operations do not correspond to user qualifications in SQL queries; they are administrative routines used by the access methods, internally.

In order to manage diverse support routines consistently across all Postgres access methods, pg_am includes an attribute called amsupport. This attribute records the number of support routines used by an access method. For B-trees, this number is one -- the routine to take two keys and return -1, 0, or +1, depending on whether the first key is less than, equal to, or greater than the second.

Note: Strictly speaking, this routine can return a negative number (< 0), 0, or a non-zero positive number (> 0).

The amstrategies entry in pg_am is just the number of strategies defined for the access method in question. The procedures for less than, less equal, and so on don't appear in pg_am. Similarly, amsupport is just the number of support routines required by the access method. The actual routines are listed elsewhere.

By the way, the amorderstrategy entry tells whether the access method supports ordered scan. Zero means it doesn't; if it does, amorderstrategy is the number of the strategy routine that corresponds to the ordering operator. For example, btree has amorderstrategy = 1 which is its "less than" strategy number.

The next class of interest is pg_opclass. This class exists only to associate an operator class name and perhaps a default type with an operator class oid. Some existing opclasses are int2_ops, int4_ops, and oid_ops. You need to add an instance with your opclass name (for example, complex_abs_ops) to pg_opclass. The oid of this instance will be a foreign key in other classes, notably pg_amop.

INSERT INTO pg_opclass (opcname, opcdeftype)
    SELECT 'complex_abs_ops', oid FROM pg_type WHERE typname = 'complex';

SELECT oid, opcname, opcdeftype
    FROM pg_opclass
    WHERE opcname = 'complex_abs_ops';

  oid   |     opcname     | opcdeftype
--------+-----------------+------------
 277975 | complex_abs_ops |     277946
(1 row)
   

The above example assumes that you want to make this new opclass the default index opclass for the complex datatype. If you don't, just insert zero into opcdeftype, rather than inserting the datatype's oid:

INSERT INTO pg_opclass (opcname, opcdeftype) VALUES ('complex_abs_ops', 0);
   

So now we have an access method and an operator class. We still need a set of operators. The procedure for defining operators was discussed earlier in this manual. For the complex_abs_ops operator class on Btrees, the operators we require are:

        absolute value less-than
        absolute value less-than-or-equal
        absolute value equal
        absolute value greater-than-or-equal
        absolute value greater-than
   

Suppose the code that implements the functions defined is stored in the file PGROOT/src/tutorial/complex.c

Part of the C code looks like this: (note that we will only show the equality operator for the rest of the examples. The other four operators are very similar. Refer to complex.c or complex.source for the details.)

#define Mag(c) ((c)->x*(c)->x + (c)->y*(c)->y)

         bool
         complex_abs_eq(Complex *a, Complex *b)
         {
             double amag = Mag(a), bmag = Mag(b);
             return (amag==bmag);
         }
   

We make the function known to Postgres like this:

CREATE FUNCTION complex_abs_eq(complex, complex)
              RETURNS bool
              AS 'PGROOT/tutorial/obj/complex.so'
              LANGUAGE 'c';
   

There are some important things that are happening here.

First, note that operators for less-than, less-than-or-equal, equal, greater-than-or-equal, and greater-than for complex are being defined. We can only have one operator named, say, = and taking type complex for both operands. In this case we don't have any other operator = for complex, but if we were building a practical datatype we'd probably want = to be the ordinary equality operation for complex numbers. In that case, we'd need to use some other operator name for complex_abs_eq.

Second, although Postgres can cope with operators having the same name as long as they have different input datatypes, C can only cope with one global routine having a given name, period. So we shouldn't name the C function something simple like abs_eq. Usually it's a good practice to include the datatype name in the C function name, so as not to conflict with functions for other datatypes.

Third, we could have made the Postgres name of the function abs_eq, relying on Postgres to distinguish it by input datatypes from any other Postgres function of the same name. To keep the example simple, we make the function have the same names at the C level and Postgres level.

Finally, note that these operator functions return Boolean values. The access methods rely on this fact. (On the other hand, the support function returns whatever the particular access method expects -- in this case, a signed integer.) The final routine in the file is the "support routine" mentioned when we discussed the amsupport attribute of the pg_am class. We will use this later on. For now, ignore it.

Now we are ready to define the operators:

CREATE OPERATOR = (
     leftarg = complex, rightarg = complex,
     procedure = complex_abs_eq,
     restrict = eqsel, join = eqjoinsel
         )
   

The next step is to add entries for these operators to the pg_amop relation. To do this, we'll need the oids of the operators we just defined. We'll look up the names of all the operators that take two complexes, and pick ours out:

    SELECT o.oid AS opoid, o.oprname
     INTO TABLE complex_ops_tmp
     FROM pg_operator o, pg_type t
     WHERE o.oprleft = t.oid and o.oprright = t.oid
      and t.typname = 'complex';

 opoid  | oprname
--------+---------
 277963 | +
 277970 | <
 277971 | <=
 277972 | =
 277973 | >=
 277974 | >
(6 rows)
   

Now we're ready to update pg_amop with our new operator class. The most important thing in this entire discussion is that the operators are ordered, from less than through greater than, in pg_amop. We add the instances we need:

    INSERT INTO pg_amop (amopid, amopclaid, amopopr, amopstrategy)
        SELECT am.oid, opcl.oid, c.opoid, 1
        FROM pg_am am, pg_opclass opcl, complex_ops_tmp c
        WHERE amname = 'btree' AND
            opcname = 'complex_abs_ops' AND
            c.oprname = '<';
   

The next step is registration of the "support routine" previously described in our discussion of pg_am. The oid of this support routine is stored in the pg_amproc class, keyed by the access method oid and the operator class oid. First, we need to register the function in Postgres (recall that we put the C code that implements this routine in the bottom of the file in which we implemented the operator routines):

    CREATE FUNCTION complex_abs_cmp(complex, complex)
     RETURNS int4
     AS 'PGROOT/tutorial/obj/complex.so'
     LANGUAGE 'c';

    SELECT oid, proname FROM pg_proc
     WHERE proname = 'complex_abs_cmp';

  oid   |     proname
--------+-----------------
 277997 | complex_abs_cmp
(1 row)
   

    INSERT INTO pg_amproc (amid, amopclaid, amproc, amprocnum)
        SELECT a.oid, b.oid, c.oid, 1
            FROM pg_am a, pg_opclass b, pg_proc c
            WHERE a.amname = 'btree' AND
                b.opcname = 'complex_abs_ops' AND
                c.proname = 'complex_abs_cmp';
   

And we're done! (Whew.) It should now be possible to create and use btree indexes on complex columns.