56.3. Extensibility

Traditionally, implementing a new index access method meant a lot of difficult work. It was necessary to understand the inner workings of the database, such as the lock manager and Write-Ahead Log. The GiST interface has a high level of abstraction, requiring the access method implementer only to implement the semantics of the data type being accessed. The GiST layer itself takes care of concurrency, logging and searching the tree structure.

This extensibility should not be confused with the extensibility of the other standard search trees in terms of the data they can handle. For example, PostgreSQL supports extensible B-trees and hash indexes. That means that you can use PostgreSQL to build a B-tree or hash over any data type you want. But B-trees only support range predicates (<, =, >), and hash indexes only support equality queries.

So if you index, say, an image collection with a PostgreSQL B-tree, you can only issue queries such as "is imagex equal to imagey", "is imagex less than imagey" and "is imagex greater than imagey". Depending on how you define "equals", "less than" and "greater than" in this context, this could be useful. However, by using a GiST based index, you could create ways to ask domain-specific questions, perhaps "find all images of horses" or "find all over-exposed images".

All it takes to get a GiST access method up and running is to implement several user-defined methods, which define the behavior of keys in the tree. Of course these methods have to be pretty fancy to support fancy queries, but for all the standard queries (B-trees, R-trees, etc.) they're relatively straightforward. In short, GiST combines extensibility along with generality, code reuse, and a clean interface.

There are seven methods that an index operator class for GiST must provide, and an eighth that is optional. Correctness of the index is ensured by proper implementation of the same, consistent and union methods, while efficiency (size and speed) of the index will depend on the penalty and picksplit methods. The remaining two basic methods are compress and decompress, which allow an index to have internal tree data of a different type than the data it indexes. The leaves are to be of the indexed data type, while the other tree nodes can be of any C struct (but you still have to follow PostgreSQL data type rules here, see about varlena for variable sized data). If the tree's internal data type exists at the SQL level, the STORAGE option of the CREATE OPERATOR CLASS command can be used. The optional eighth method is distance, which is needed if the operator class wishes to support ordered scans (nearest-neighbor searches).

consistent

Given an index entry p and a query value q, this function determines whether the index entry is "consistent" with the query; that is, could the predicate "indexed_column indexable_operator q" be true for any row represented by the index entry? For a leaf index entry this is equivalent to testing the indexable condition, while for an internal tree node this determines whether it is necessary to scan the subtree of the index represented by the tree node. When the result is true, a recheck flag must also be returned. This indicates whether the predicate is certainly true or only possibly true. If recheck = false then the index has tested the predicate condition exactly, whereas if recheck = true the row is only a candidate match. In that case the system will automatically evaluate the indexable_operator against the actual row value to see if it is really a match. This convention allows GiST to support both lossless and lossy index structures.

The SQL declaration of the function must look like this:

CREATE OR REPLACE FUNCTION my_consistent(internal, data_type, smallint, oid, internal)
RETURNS bool
AS 'MODULE_PATHNAME'
LANGUAGE C STRICT;

And the matching code in the C module could then follow this skeleton:

Datum       my_consistent(PG_FUNCTION_ARGS);
PG_FUNCTION_INFO_V1(my_consistent);

Datum
my_consistent(PG_FUNCTION_ARGS)
{
    GISTENTRY  *entry = (GISTENTRY *) PG_GETARG_POINTER(0);
    data_type  *query = PG_GETARG_DATA_TYPE_P(1);
    StrategyNumber strategy = (StrategyNumber) PG_GETARG_UINT16(2);
    /* Oid subtype = PG_GETARG_OID(3); */
    bool       *recheck = (bool *) PG_GETARG_POINTER(4);
    data_type  *key = DatumGetDataType(entry->key);
    bool        retval;

    /*
     * determine return value as a function of strategy, key and query.
     *
     * Use GIST_LEAF(entry) to know where you're called in the index tree,
     * which comes handy when supporting the = operator for example (you could
     * check for non empty union() in non-leaf nodes and equality in leaf
     * nodes).
     */

    *recheck = true;        /* or false if check is exact */

    PG_RETURN_BOOL(retval);
}

Here, key is an element in the index and query the value being looked up in the index. The StrategyNumber parameter indicates which operator of your operator class is being applied — it matches one of the operator numbers in the CREATE OPERATOR CLASS command. Depending on what operators you have included in the class, the data type of query could vary with the operator, but the above skeleton assumes it doesn't.

union

This method consolidates information in the tree. Given a set of entries, this function generates a new index entry that represents all the given entries.

The SQL declaration of the function must look like this:

CREATE OR REPLACE FUNCTION my_union(internal, internal)
RETURNS internal
AS 'MODULE_PATHNAME'
LANGUAGE C STRICT;

And the matching code in the C module could then follow this skeleton:

Datum       my_union(PG_FUNCTION_ARGS);
PG_FUNCTION_INFO_V1(my_union);

Datum
my_union(PG_FUNCTION_ARGS)
{
    GistEntryVector *entryvec = (GistEntryVector *) PG_GETARG_POINTER(0);
    GISTENTRY  *ent = entryvec->vector;
    data_type  *out,
               *tmp,
               *old;
    int         numranges,
                i = 0;

    numranges = entryvec->n;
    tmp = DatumGetDataType(ent[0].key);
    out = tmp;

    if (numranges == 1)
    {
        out = data_type_deep_copy(tmp);

        PG_RETURN_DATA_TYPE_P(out);
    }

    for (i = 1; i < numranges; i++)
    {
        old = out;
        tmp = DatumGetDataType(ent[i].key);
        out = my_union_implementation(out, tmp);
    }

    PG_RETURN_DATA_TYPE_P(out);
}

As you can see, in this skeleton we're dealing with a data type where union(X, Y, Z) = union(union(X, Y), Z). It's easy enough to support data types where this is not the case, by implementing the proper union algorithm in this GiST support method.

The union implementation function should return a pointer to newly palloc()ed memory. You can't just return whatever the input is.

compress

Converts the data item into a format suitable for physical storage in an index page.

The SQL declaration of the function must look like this:

CREATE OR REPLACE FUNCTION my_compress(internal)
RETURNS internal
AS 'MODULE_PATHNAME'
LANGUAGE C STRICT;

And the matching code in the C module could then follow this skeleton:

Datum       my_compress(PG_FUNCTION_ARGS);
PG_FUNCTION_INFO_V1(my_compress);

Datum
my_compress(PG_FUNCTION_ARGS)
{
    GISTENTRY  *entry = (GISTENTRY *) PG_GETARG_POINTER(0);
    GISTENTRY  *retval;

    if (entry->leafkey)
    {
        /* replace entry->key with a compressed version */
        compressed_data_type *compressed_data = palloc(sizeof(compressed_data_type));

        /* fill *compressed_data from entry->key ... */

        retval = palloc(sizeof(GISTENTRY));
        gistentryinit(*retval, PointerGetDatum(compressed_data),
                      entry->rel, entry->page, entry->offset, FALSE);
    }
    else
    {
        /* typically we needn't do anything with non-leaf entries */
        retval = entry;
    }

    PG_RETURN_POINTER(retval);
}

You have to adapt compressed_data_type to the specific type you're converting to in order to compress your leaf nodes, of course.

Depending on your needs, you could also need to care about compressing NULL values in there, storing for example (Datum) 0 like gist_circle_compress does.

decompress

The reverse of the compress method. Converts the index representation of the data item into a format that can be manipulated by the database.

The SQL declaration of the function must look like this:

CREATE OR REPLACE FUNCTION my_decompress(internal)
RETURNS internal
AS 'MODULE_PATHNAME'
LANGUAGE C STRICT;

And the matching code in the C module could then follow this skeleton:

Datum       my_decompress(PG_FUNCTION_ARGS);
PG_FUNCTION_INFO_V1(my_decompress);

Datum
my_decompress(PG_FUNCTION_ARGS)
{
    PG_RETURN_POINTER(PG_GETARG_POINTER(0));
}

The above skeleton is suitable for the case where no decompression is needed.

penalty

Returns a value indicating the "cost" of inserting the new entry into a particular branch of the tree. Items will be inserted down the path of least penalty in the tree. Values returned by penalty should be non-negative. If a negative value is returned, it will be treated as zero.

The SQL declaration of the function must look like this:

CREATE OR REPLACE FUNCTION my_penalty(internal, internal, internal)
RETURNS internal
AS 'MODULE_PATHNAME'
LANGUAGE C STRICT;  -- in some cases penalty functions need not be strict

And the matching code in the C module could then follow this skeleton:

Datum       my_penalty(PG_FUNCTION_ARGS);
PG_FUNCTION_INFO_V1(my_penalty);

Datum
my_penalty(PG_FUNCTION_ARGS)
{
    GISTENTRY  *origentry = (GISTENTRY *) PG_GETARG_POINTER(0);
    GISTENTRY  *newentry = (GISTENTRY *) PG_GETARG_POINTER(1);
    float      *penalty = (float *) PG_GETARG_POINTER(2);
    data_type  *orig = DatumGetDataType(origentry->key);
    data_type  *new = DatumGetDataType(newentry->key);

    *penalty = my_penalty_implementation(orig, new);
    PG_RETURN_POINTER(penalty);
}

The penalty function is crucial to good performance of the index. It'll get used at insertion time to determine which branch to follow when choosing where to add the new entry in the tree. At query time, the more balanced the index, the quicker the lookup.

picksplit

When an index page split is necessary, this function decides which entries on the page are to stay on the old page, and which are to move to the new page.

The SQL declaration of the function must look like this:

CREATE OR REPLACE FUNCTION my_picksplit(internal, internal)
RETURNS internal
AS 'MODULE_PATHNAME'
LANGUAGE C STRICT;

And the matching code in the C module could then follow this skeleton:

Datum       my_picksplit(PG_FUNCTION_ARGS);
PG_FUNCTION_INFO_V1(my_picksplit);

Datum
my_picksplit(PG_FUNCTION_ARGS)
{
    GistEntryVector *entryvec = (GistEntryVector *) PG_GETARG_POINTER(0);
    OffsetNumber maxoff = entryvec->n - 1;
    GISTENTRY  *ent = entryvec->vector;
    GIST_SPLITVEC *v = (GIST_SPLITVEC *) PG_GETARG_POINTER(1);
    int         i,
                nbytes;
    OffsetNumber *left,
               *right;
    data_type  *tmp_union;
    data_type  *unionL;
    data_type  *unionR;
    GISTENTRY **raw_entryvec;

    maxoff = entryvec->n - 1;
    nbytes = (maxoff + 1) * sizeof(OffsetNumber);

    v->spl_left = (OffsetNumber *) palloc(nbytes);
    left = v->spl_left;
    v->spl_nleft = 0;

    v->spl_right = (OffsetNumber *) palloc(nbytes);
    right = v->spl_right;
    v->spl_nright = 0;

    unionL = NULL;
    unionR = NULL;

    /* Initialize the raw entry vector. */
    raw_entryvec = (GISTENTRY **) malloc(entryvec->n * sizeof(void *));
    for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
        raw_entryvec[i] = &(entryvec->vector[i]);

    for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
    {
        int         real_index = raw_entryvec[i] - entryvec->vector;

        tmp_union = DatumGetDataType(entryvec->vector[real_index].key);
        Assert(tmp_union != NULL);

        /*
         * Choose where to put the index entries and update unionL and unionR
         * accordingly. Append the entries to either v_spl_left or
         * v_spl_right, and care about the counters.
         */

        if (my_choice_is_left(unionL, curl, unionR, curr))
        {
            if (unionL == NULL)
                unionL = tmp_union;
            else
                unionL = my_union_implementation(unionL, tmp_union);

            *left = real_index;
            ++left;
            ++(v->spl_nleft);
        }
        else
        {
            /*
             * Same on the right
             */
        }
    }

    v->spl_ldatum = DataTypeGetDatum(unionL);
    v->spl_rdatum = DataTypeGetDatum(unionR);
    PG_RETURN_POINTER(v);
}

Like penalty, the picksplit function is crucial to good performance of the index. Designing suitable penalty and picksplit implementations is where the challenge of implementing well-performing GiST indexes lies.

same

Returns true if two index entries are identical, false otherwise.

The SQL declaration of the function must look like this:

CREATE OR REPLACE FUNCTION my_same(internal, internal, internal)
RETURNS internal
AS 'MODULE_PATHNAME'
LANGUAGE C STRICT;

And the matching code in the C module could then follow this skeleton:

Datum       my_same(PG_FUNCTION_ARGS);
PG_FUNCTION_INFO_V1(my_same);

Datum
my_same(PG_FUNCTION_ARGS)
{
    prefix_range *v1 = PG_GETARG_PREFIX_RANGE_P(0);
    prefix_range *v2 = PG_GETARG_PREFIX_RANGE_P(1);
    bool       *result = (bool *) PG_GETARG_POINTER(2);

    *result = my_eq(v1, v2);
    PG_RETURN_POINTER(result);
}

For historical reasons, the same function doesn't just return a Boolean result; instead it has to store the flag at the location indicated by the third argument.

distance

Given an index entry p and a query value q, this function determines the index entry's "distance" from the query value. This function must be supplied if the operator class contains any ordering operators. A query using the ordering operator will be implemented by returning index entries with the smallest "distance" values first, so the results must be consistent with the operator's semantics. For a leaf index entry the result just represents the distance to the index entry; for an internal tree node, the result must be the smallest distance that any child entry could have.

The SQL declaration of the function must look like this:

CREATE OR REPLACE FUNCTION my_distance(internal, data_type, smallint, oid)
RETURNS float8
AS 'MODULE_PATHNAME'
LANGUAGE C STRICT;

And the matching code in the C module could then follow this skeleton:

Datum       my_distance(PG_FUNCTION_ARGS);
PG_FUNCTION_INFO_V1(my_distance);

Datum
my_distance(PG_FUNCTION_ARGS)
{
    GISTENTRY  *entry = (GISTENTRY *) PG_GETARG_POINTER(0);
    data_type  *query = PG_GETARG_DATA_TYPE_P(1);
    StrategyNumber strategy = (StrategyNumber) PG_GETARG_UINT16(2);
    /* Oid subtype = PG_GETARG_OID(3); */
    data_type  *key = DatumGetDataType(entry->key);
    double      retval;

    /*
     * determine return value as a function of strategy, key and query.
     */

    PG_RETURN_FLOAT8(retval);
}

The arguments to the distance function are identical to the arguments of the consistent function, except that no recheck flag is used. The distance to a leaf index entry must always be determined exactly, since there is no way to re-order the tuples once they are returned. Some approximation is allowed when determining the distance to an internal tree node, so long as the result is never greater than any child's actual distance. Thus, for example, distance to a bounding box is usually sufficient in geometric applications. The result value can be any finite float8 value. (Infinity and minus infinity are used internally to handle cases such as nulls, so it is not recommended that distance functions return these values.)

All the GiST support methods are normally called in short-lived memory contexts; that is, CurrentMemoryContext will get reset after each tuple is processed. It is therefore not very important to worry about pfree'ing everything you palloc. However, in some cases it's useful for a support method to cache data across repeated calls. To do that, allocate the longer-lived data in fcinfo->flinfo->fn_mcxt, and keep a pointer to it in fcinfo->flinfo->fn_extra. Such data will survive for the life of the index operation (e.g., a single GiST index scan, index build, or index tuple insertion). Be careful to pfree the previous value when replacing a fn_extra value, or the leak will accumulate for the duration of the operation.