4
<title>GIN Indexes</title>
7
<primary>index</primary>
8
<secondary>GIN</secondary>
11
<sect1 id="gin-intro">
12
<title>Introduction</title>
15
<acronym>GIN</acronym> stands for Generalized Inverted Index. It is
16
an index structure storing a set of (key, posting list) pairs, where
17
a <quote>posting list</> is a set of rows in which the key occurs. Each
18
indexed value can contain many keys, so the same row ID can appear in
19
multiple posting lists.
23
It is generalized in the sense that a <acronym>GIN</acronym> index
24
does not need to be aware of the operation that it accelerates.
25
Instead, it uses custom strategies defined for particular data types.
29
One advantage of <acronym>GIN</acronym> is that it allows the development
30
of custom data types with the appropriate access methods, by
31
an expert in the domain of the data type, rather than a database expert.
32
This is much the same advantage as using <acronym>GiST</acronym>.
36
The <acronym>GIN</acronym>
37
implementation in <productname>PostgreSQL</productname> is primarily
38
maintained by Teodor Sigaev and Oleg Bartunov. There is more
39
information about <acronym>GIN</acronym> on their
40
<ulink url="http://www.sai.msu.su/~megera/wiki/Gin">website</ulink>.
44
<sect1 id="gin-extensibility">
45
<title>Extensibility</title>
48
The <acronym>GIN</acronym> interface has a high level of abstraction,
49
requiring the access method implementer only to implement the semantics of
50
the data type being accessed. The <acronym>GIN</acronym> layer itself
51
takes care of concurrency, logging and searching the tree structure.
55
All it takes to get a <acronym>GIN</acronym> access method working is to
56
implement four (or five) user-defined methods, which define the behavior of
57
keys in the tree and the relationships between keys, indexed values,
58
and indexable queries. In short, <acronym>GIN</acronym> combines
59
extensibility with generality, code reuse, and a clean interface.
63
The four methods that an operator class for
64
<acronym>GIN</acronym> must provide are:
69
<term>int compare(Datum a, Datum b)</term>
72
Compares keys (not indexed values!) and returns an integer less than
73
zero, zero, or greater than zero, indicating whether the first key is
74
less than, equal to, or greater than the second.
80
<term>Datum *extractValue(Datum inputValue, int32 *nkeys)</term>
83
Returns an array of keys given a value to be indexed. The
84
number of returned keys must be stored into <literal>*nkeys</>.
90
<term>Datum *extractQuery(Datum query, int32 *nkeys,
91
StrategyNumber n, bool **pmatch, Pointer **extra_data)</term>
94
Returns an array of keys given a value to be queried; that is,
95
<literal>query</> is the value on the right-hand side of an
96
indexable operator whose left-hand side is the indexed column.
97
<literal>n</> is the strategy number of the operator within the
98
operator class (see <xref linkend="xindex-strategies">).
99
Often, <function>extractQuery</> will need
100
to consult <literal>n</> to determine the data type of
101
<literal>query</> and the key values that need to be extracted.
102
The number of returned keys must be stored into <literal>*nkeys</>.
103
If the query contains no keys then <function>extractQuery</>
104
should store 0 or -1 into <literal>*nkeys</>, depending on the
105
semantics of the operator. 0 means that every
106
value matches the <literal>query</> and a sequential scan should be
107
performed. -1 means nothing can match the <literal>query</>.
108
<literal>pmatch</> is an output argument for use when partial match
109
is supported. To use it, <function>extractQuery</> must allocate
110
an array of <literal>*nkeys</> booleans and store its address at
111
<literal>*pmatch</>. Each element of the array should be set to TRUE
112
if the corresponding key requires partial match, FALSE if not.
113
If <literal>*pmatch</> is set to NULL then GIN assumes partial match
114
is not required. The variable is initialized to NULL before call,
115
so this argument can simply be ignored by operator classes that do
116
not support partial match.
117
<literal>extra_data</> is an output argument that allows
118
<function>extractQuery</> to pass additional data to the
119
<function>consistent</> and <function>comparePartial</> methods.
120
To use it, <function>extractQuery</> must allocate
121
an array of <literal>*nkeys</> Pointers and store its address at
122
<literal>*extra_data</>, then store whatever it wants to into the
123
individual pointers. The variable is initialized to NULL before
124
call, so this argument can simply be ignored by operator classes that
125
do not require extra data. If <literal>*extra_data</> is set, the
126
whole array is passed to the <function>consistent</> method, and
127
the appropriate element to the <function>comparePartial</> method.
134
<term>bool consistent(bool check[], StrategyNumber n, Datum query,
135
int32 nkeys, Pointer extra_data[], bool *recheck)</term>
138
Returns TRUE if the indexed value satisfies the query operator with
139
strategy number <literal>n</> (or might satisfy, if the recheck
140
indication is returned). The <literal>check</> array has length
141
<literal>nkeys</>, which is the same as the number of keys previously
142
returned by <function>extractQuery</> for this <literal>query</> datum.
144
<literal>check</> array is TRUE if the indexed value contains the
145
corresponding query key, ie, if (check[i] == TRUE) the i-th key of the
146
<function>extractQuery</> result array is present in the indexed value.
147
The original <literal>query</> datum (not the extracted key array!) is
148
passed in case the <function>consistent</> method needs to consult it.
149
<literal>extra_data</> is the extra-data array returned by
150
<function>extractQuery</>, or NULL if none.
151
On success, <literal>*recheck</> should be set to TRUE if the heap
152
tuple needs to be rechecked against the query operator, or FALSE if
153
the index test is exact.
161
Optionally, an operator class for
162
<acronym>GIN</acronym> can supply a fifth method:
168
<term>int comparePartial(Datum partial_key, Datum key, StrategyNumber n,
169
Pointer extra_data)</term>
172
Compare a partial-match query to an index key. Returns an integer
173
whose sign indicates the result: less than zero means the index key
174
does not match the query, but the index scan should continue; zero
175
means that the index key does match the query; greater than zero
176
indicates that the index scan should stop because no more matches
177
are possible. The strategy number <literal>n</> of the operator
178
that generated the partial match query is provided, in case its
179
semantics are needed to determine when to end the scan. Also,
180
<literal>extra_data</> is the corresponding element of the extra-data
181
array made by <function>extractQuery</>, or NULL if none.
189
To support <quote>partial match</> queries, an operator class must
190
provide the <function>comparePartial</> method, and its
191
<function>extractQuery</> method must set the <literal>pmatch</>
192
parameter when a partial-match query is encountered. See
193
<xref linkend="gin-partial-match"> for details.
198
<sect1 id="gin-implementation">
199
<title>Implementation</title>
202
Internally, a <acronym>GIN</acronym> index contains a B-tree index
203
constructed over keys, where each key is an element of the indexed value
204
(a member of an array, for example) and where each tuple in a leaf page is
205
either a pointer to a B-tree over heap pointers (PT, posting tree), or a
206
list of heap pointers (PL, posting list) if the list is small enough.
209
<sect2 id="gin-fast-update">
210
<title>GIN fast update technique</title>
213
Updating a <acronym>GIN</acronym> index tends to be slow because of the
214
intrinsic nature of inverted indexes: inserting or updating one heap row
215
can cause many inserts into the index (one for each key extracted
216
from the indexed value). As of <productname>PostgreSQL</productname> 8.4,
217
<acronym>GIN</> is capable of postponing much of this work by inserting
218
new tuples into a temporary, unsorted list of pending entries.
219
When the table is vacuumed, or if the pending list becomes too large
220
(larger than <xref linkend="guc-work-mem">), the entries are moved to the
221
main <acronym>GIN</acronym> data structure using the same bulk insert
222
techniques used during initial index creation. This greatly improves
223
<acronym>GIN</acronym> index update speed, even counting the additional
224
vacuum overhead. Moreover the overhead can be done by a background
225
process instead of in foreground query processing.
229
The main disadvantage of this approach is that searches must scan the list
230
of pending entries in addition to searching the regular index, and so
231
a large list of pending entries will slow searches significantly.
232
Another disadvantage is that, while most updates are fast, an update
233
that causes the pending list to become <quote>too large</> will incur an
234
immediate cleanup cycle and thus be much slower than other updates.
235
Proper use of autovacuum can minimize both of these problems.
239
If consistent response time is more important than update speed,
240
use of pending entries can be disabled by turning off the
241
<literal>FASTUPDATE</literal> storage parameter for a
242
<acronym>GIN</acronym> index. See <xref linkend="sql-createindex"
243
endterm="sql-createindex-title"> for details.
247
<sect2 id="gin-partial-match">
248
<title>Partial match algorithm</title>
251
GIN can support <quote>partial match</> queries, in which the query
252
does not determine an exact match for one or more keys, but the possible
253
matches fall within a reasonably narrow range of key values (within the
254
key sorting order determined by the <function>compare</> support method).
255
The <function>extractQuery</> method, instead of returning a key value
256
to be matched exactly, returns a key value that is the lower bound of
257
the range to be searched, and sets the <literal>pmatch</> flag true.
258
The key range is then searched using the <function>comparePartial</>
259
method. <function>comparePartial</> must return zero for an actual
260
match, less than zero for a non-match that is still within the range
261
to be searched, or greater than zero if the index key is past the range
268
<sect1 id="gin-tips">
269
<title>GIN tips and tricks</title>
273
<term>Create vs insert</term>
276
Insertion into a <acronym>GIN</acronym> index can be slow
277
due to the likelihood of many keys being inserted for each value.
278
So, for bulk insertions into a table it is advisable to drop the GIN
279
index and recreate it after finishing bulk insertion.
283
As of <productname>PostgreSQL</productname> 8.4, this advice is less
284
necessary since delayed indexing is used (see <xref
285
linkend="gin-fast-update"> for details). But for very large updates
286
it may still be best to drop and recreate the index.
292
<term><xref linkend="guc-maintenance-work-mem"></term>
295
Build time for a <acronym>GIN</acronym> index is very sensitive to
296
the <varname>maintenance_work_mem</> setting; it doesn't pay to
297
skimp on work memory during index creation.
303
<term><xref linkend="guc-work-mem"></term>
306
During a series of insertions into an existing <acronym>GIN</acronym>
307
index that has <literal>FASTUPDATE</> enabled, the system will clean up
308
the pending-entry list whenever it grows larger than
309
<varname>work_mem</>. To avoid fluctuations in observed response time,
310
it's desirable to have pending-list cleanup occur in the background
311
(i.e., via autovacuum). Foreground cleanup operations can be avoided by
312
increasing <varname>work_mem</> or making autovacuum more aggressive.
313
However, enlarging <varname>work_mem</> means that if a foreground
314
cleanup does occur, it will take even longer.
320
<term><xref linkend="guc-gin-fuzzy-search-limit"></term>
323
The primary goal of developing <acronym>GIN</acronym> indexes was
324
to create support for highly scalable, full-text search in
325
<productname>PostgreSQL</productname>, and there are often situations when
326
a full-text search returns a very large set of results. Moreover, this
327
often happens when the query contains very frequent words, so that the
328
large result set is not even useful. Since reading many
329
tuples from the disk and sorting them could take a lot of time, this is
330
unacceptable for production. (Note that the index search itself is very
334
To facilitate controlled execution of such queries
335
<acronym>GIN</acronym> has a configurable soft upper limit on the
336
number of rows returned, the
337
<varname>gin_fuzzy_search_limit</varname> configuration parameter.
338
It is set to 0 (meaning no limit) by default.
339
If a non-zero limit is set, then the returned set is a subset of
340
the whole result set, chosen at random.
343
<quote>Soft</quote> means that the actual number of returned results
344
could differ slightly from the specified limit, depending on the query
345
and the quality of the system's random number generator.
353
<sect1 id="gin-limit">
354
<title>Limitations</title>
357
<acronym>GIN</acronym> doesn't support full index scans: because there are
358
often many keys per value, each heap pointer would be returned many times,
359
and there is no easy way to prevent this.
363
When <function>extractQuery</function> returns zero keys,
364
<acronym>GIN</acronym> will emit an error. Depending on the operator,
365
a void query might match all, some, or none of the indexed values (for
366
example, every array contains the empty array, but does not overlap the
367
empty array), and <acronym>GIN</acronym> cannot determine the correct
368
answer, nor produce a full-index-scan result if it could determine that
373
It is not an error for <function>extractValue</> to return zero keys,
374
but in this case the indexed value will be unrepresented in the index.
375
This is another reason why full index scan is not useful — it would
380
It is possible for an operator class to circumvent the restriction against
381
full index scan. To do that, <function>extractValue</> must return at least
382
one (possibly dummy) key for every indexed value, and
383
<function>extractQuery</function> must convert an unrestricted search into
384
a partial-match query that will scan the whole index. This is inefficient
385
but might be necessary to avoid corner-case failures with operators such
390
<sect1 id="gin-examples">
391
<title>Examples</title>
394
The <productname>PostgreSQL</productname> source distribution includes
395
<acronym>GIN</acronym> operator classes for <type>tsvector</> and
396
for one-dimensional arrays of all internal types. Prefix searching in
397
<type>tsvector</> is implemented using the <acronym>GIN</> partial match
399
The following <filename>contrib</> modules also contain
400
<acronym>GIN</acronym> operator classes:
405
<term>btree-gin</term>
407
<para>B-Tree equivalent functionality for several data types</para>
414
<para>Module for storing (key, value) pairs</para>
419
<term>intarray</term>
421
<para>Enhanced support for int4[]</para>
428
<para>Text similarity using trigram matching</para>