GiST stands for Generalized Search Tree. It is a balanced, tree-structured access method, that acts as a base template in which to implement arbitrary indexing schemes. B-trees, R-trees and many other indexing schemes can be implemented in GiST.
One advantage of GiST is that it allows the development of custom data types with the appropriate access methods, by an expert in the domain of the data type, rather than a database expert.
Some of the information here is derived from the University of California at Berkeley's GiST Indexing Project web site and Marcel Kornacker's thesis, Access Methods for Next-Generation Database Systems. The GiST implementation in PostgreSQL is primarily maintained by Teodor Sigaev and Oleg Bartunov, and there is more information on their web site.
The PostgreSQL source distribution includes several examples of index methods implemented using GiST. The core system currently provides text search support (indexing for tsvector and tsquery) as well as R-Tree equivalent functionality for some of the built-in geometric data types (see src/backend/access/gist/gistproc.c). The following contrib modules also contain GiST operator classes:
btree_gist
B-tree equivalent functionality for several data types
cube
Indexing for multidimensional cubes
hstore
Module for storing (key, value) pairs
intarray
RD-Tree for one-dimensional array of int4 values
ltree
Indexing for tree-like structures
pg_trgm
Text similarity using trigram matching
seg
Indexing for “float ranges”
Building large GiST indexes by simply inserting all the tuples tends to be slow, because if the index tuples are scattered across the index and the index is large enough to not fit in cache, the insertions need to perform a lot of random I/O. Beginning in version 9.2, PostgreSQL supports a more efficient method to build GiST indexes based on buffering, which can dramatically reduce the number of random I/Os needed for non-ordered data sets. For well-ordered data sets the benefit is smaller or non-existent, because only a small number of pages receive new tuples at a time, and those pages fit in cache even if the index as whole does not.
However, buffering index build needs to call the penalty function more often, which consumes some extra CPU resources. Also, the buffers used in the buffering build need temporary disk space, up to the size of the resulting index. Buffering can also influence the quality of the resulting index, in both positive and negative directions. That influence depends on various factors, like the distribution of the input data and the operator class implementation.
By default, a GiST index build switches to the buffering method when the index size reaches effective_cache_size. It can be manually turned on or off by the buffering parameter to the CREATE INDEX command. The default behavior is good for most cases, but turning buffering off might speed up the build somewhat if the input data is ordered.
The core PostgreSQL distribution includes the GiST operator classes shown in Table 64.1. (Some of the optional modules described in Appendix F provide additional GiST operator classes.)
For historical reasons, the inet_ops operator class is not the default class for types inet and cidr. To use it, mention the class name in CREATE INDEX, for example
Name
Indexed Data Type
Indexable Operators
Ordering Operators
box_ops
box
&& &> &< &<| >> << <<| <@ @> @ |&> |>> ~ ~=
circle_ops
circle
&& &> &< &<| >> << <<| <@ @> @ |&> |>> ~ ~=
<->
inet_ops
inet, cidr
&& >> >>= > >= <> << <<= < <= =
point_ops
point
>> >^ << <@ <@ <@ <^ ~=
<->
poly_ops
polygon
&& &> &< &<| >> << <<| <@ @> @ |&> |>> ~ ~=
<->
range_ops
any range type
&& &> &< >> << <@ -|- = @> @>
tsquery_ops
tsquery
<@ @>
tsvector_ops
tsvector
@@
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 five methods that an index operator class for GiST must provide, and four that are 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. Two optional 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). The optional ninth method fetch is needed if the operator class wishes to support index-only scans, except when the compress method is omitted.
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:
And the matching code in the C module could then follow this skeleton:
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 which operators you have included in the class, the data type of query could vary with the operator, since it will be whatever type is on the righthand side of the operator, which might be different from the indexed data type appearing on the lefthand side. (The above code skeleton assumes that only one type is possible; if not, fetching the query argument value would have to depend on the operator.) It is recommended that the SQL declaration of the consistent function use the opclass's indexed data type for the query argument, even though the actual type might be something else depending on the operator.
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:
And the matching code in the C module could then follow this skeleton:
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 result of the union function must be a value of the index's storage type, whatever that is (it might or might not be different from the indexed column's type). The union function should return a pointer to newly palloc()ed memory. You can't just return the input value as-is, even if there is no type change.
As shown above, the union function's first internal argument is actually a GistEntryVector pointer. The second argument is a pointer to an integer variable, which can be ignored. (It used to be required that the union function store the size of its result value into that variable, but this is no longer necessary.)
compress
Converts a data item into a format suitable for physical storage in an index page. If the compress method is omitted, data items are stored in the index without modification.
The SQL declaration of the function must look like this:
And the matching code in the C module could then follow this skeleton:
You have to adapt compressed_data_type to the specific type you're converting to in order to compress your leaf nodes, of course.
decompress
Converts the stored representation of a data item into a format that can be manipulated by the other GiST methods in the operator class. If the decompress method is omitted, it is assumed that the other GiST methods can work directly on the stored data format. (decompress is not necessarily the reverse of the compress method; in particular, if compress is lossy then it's impossible for decompress to exactly reconstruct the original data. decompress is not necessarily equivalent to fetch, either, since the other GiST methods might not require full reconstruction of the data.)
The SQL declaration of the function must look like this:
And the matching code in the C module could then follow this skeleton:
The above skeleton is suitable for the case where no decompression is needed. (But, of course, omitting the method altogether is even easier, and is recommended in such cases.)
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:
And the matching code in the C module could then follow this skeleton:
For historical reasons, the penalty function doesn't just return a float result; instead it has to store the value at the location indicated by the third argument. The return value per se is ignored, though it's conventional to pass back the address of that argument.
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:
And the matching code in the C module could then follow this skeleton:
Notice that the picksplit function's result is delivered by modifying the passed-in v structure. The return value per se is ignored, though it's conventional to pass back the address of 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. (An “index entry” is a value of the index's storage type, not necessarily the original indexed column's type.)
The SQL declaration of the function must look like this:
And the matching code in the C module could then follow this skeleton:
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. The return value per se is ignored, though it's conventional to pass back the address of that 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:
And the matching code in the C module could then follow this skeleton:
The arguments to the distance function are identical to the arguments of the consistent function.
Some approximation is allowed when determining the distance, so long as the result is never greater than the entry's actual distance. Thus, for example, distance to a bounding box is usually sufficient in geometric applications. For an internal tree node, the distance returned must not be greater than the distance to any of the child nodes. If the returned distance is not exact, the function must set *recheck to true. (This is not necessary for internal tree nodes; for them, the calculation is always assumed to be inexact.) In this case the executor will calculate the accurate distance after fetching the tuple from the heap, and reorder the tuples if necessary.
If the distance function returns *recheck = true for any leaf node, the original ordering operator's return type must be float8 or float4, and the distance function's result values must be comparable to those of the original ordering operator, since the executor will sort using both distance function results and recalculated ordering-operator results. Otherwise, the distance function's result values can be any finite float8 values, so long as the relative order of the result values matches the order returned by the ordering operator. (Infinity and minus infinity are used internally to handle cases such as nulls, so it is not recommended that distance functions return these values.)
fetch
Converts the compressed index representation of a data item into the original data type, for index-only scans. The returned data must be an exact, non-lossy copy of the originally indexed value.
The SQL declaration of the function must look like this:
The argument is a pointer to a GISTENTRY struct. On entry, its key field contains a non-NULL leaf datum in compressed form. The return value is another GISTENTRY struct, whose key field contains the same datum in its original, uncompressed form. If the opclass's compress function does nothing for leaf entries, the fetch method can return the argument as-is. Or, if the opclass does not have a compress function, the fetch method can be omitted as well, since it would necessarily be a no-op.
The matching code in the C module could then follow this skeleton:
If the compress method is lossy for leaf entries, the operator class cannot support index-only scans, and must not define a fetch function.
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.