8
8
</indexterm>
9
9
10
10
<para>
11
- <literal>bloom</> is a module that implements an index access method. It comes
12
- as an example of custom access methods and generic WAL record usage. But it
13
- is also useful in itself.
11
+ <literal>bloom</> provides an index access method based on
12
+ <ulink url="http://en.wikipedia.org/wiki/Bloom_filter">Bloom filters</ulink>.
14
13
</para>
15
14
16
- <sect2>
17
- <title>Introduction</title>
15
+ <para>
16
+ A Bloom filter is a space-efficient data structure that is used to test
17
+ whether an element is a member of a set. In the case of an index access
18
+ method, it allows fast exclusion of non-matching tuples via signatures
19
+ whose size is determined at index creation.
20
+ </para>
18
21
19
- <para>
20
- The implementation of a
21
- <ulink url="http://en.wikipedia.org/wiki/Bloom_filter">Bloom filter</ulink>
22
- allows fast exclusion of non-candidate tuples via signatures.
23
- Since a signature is a lossy representation of all indexed attributes,
24
- search results must be rechecked using heap information.
25
- The user can specify signature length in bits (default 80, maximum 4096)
26
- and the number of bits generated for each index column (default 2,
27
- maximum 4095).
28
- </para>
22
+ <para>
23
+ A signature is a lossy representation of the indexed attribute(s), and as
24
+ such is prone to reporting false positives; that is, it may be reported
25
+ that an element is in the set, when it is not. So index search results
26
+ must always be rechecked using the actual attribute values from the heap
27
+ entry. Larger signatures reduce the odds of a false positive and thus
28
+ reduce the number of useless heap visits, but of course also make the index
29
+ larger and hence slower to scan.
30
+ </para>
29
31
30
- <para>
31
- This index is useful if a table has many attributes and queries include
32
- arbitrary combinations of them. A traditional <literal>btree</> index is
33
- faster than a bloom index, but it can require many indexes to support all
34
- possible queries where one needs only a single bloom index. A Bloom index
35
- supports only equality comparison. Since it's a signature file, and not a
36
- tree, it always must be read fully, but sequentially, so that index search
37
- performance is constant and doesn't depend on a query.
38
- </para>
39
- </sect2>
32
+ <para>
33
+ This type of index is most useful when a table has many attributes and
34
+ queries test arbitrary combinations of them. A traditional btree index is
35
+ faster than a bloom index, but it can require many btree indexes to support
36
+ all possible queries where one needs only a single bloom index. Note
37
+ however that bloom indexes only support equality queries, whereas btree
38
+ indexes can also perform inequality and range searches.
39
+ </para>
40
40
41
41
<sect2>
42
42
<title>Parameters</title>
43
43
44
44
<para>
45
- <literal>bloom</> indexes accept the following parameters in the
46
- <literal>WITH</>
47
- clause.
45
+ A <literal>bloom</> index accepts the following parameters in its
46
+ <literal>WITH</> clause:
48
47
</para>
49
48
50
49
<variablelist>
51
50
<varlistentry>
52
51
<term><literal>length</></term>
53
52
<listitem>
54
53
<para>
55
- Length of signature in bits
54
+ Length of each signature (index entry) in bits. The default
55
+ is <literal>80</> bits and maximum is <literal>4096</>.
56
56
</para>
57
57
</listitem>
58
58
</varlistentry>
62
62
<term><literal>col1 — col32</></term>
63
63
<listitem>
64
64
<para>
65
- Number of bits generated for each index column
65
+ Number of bits generated for each index column. Each parameter's name
66
+ refers to the number of the index column that it controls. The default
67
+ is <literal>2</> bits and maximum is <literal>4095</>. Parameters for
68
+ index columns not actually used are ignored.
66
69
</para>
67
70
</listitem>
68
71
</varlistentry>
73
76
<title>Examples</title>
74
77
75
78
<para>
76
- An example of an index definition is given below.
79
+ This is an example of creating a bloom index:
77
80
</para>
78
81
79
82
<programlisting>
@@ -82,92 +85,135 @@ CREATE INDEX bloomidx ON tbloom USING bloom (i1,i2,i3)
82
85
</programlisting>
83
86
84
87
<para>
85
- Here, we created a bloom index with a signature length of 80 bits,
86
- and attributes i1 and i2 mapped to 2 bits, and attribute i3 mapped to 4 bits.
88
+ The index is created with a signature length of 80 bits, with attributes
89
+ i1 and i2 mapped to 2 bits, and attribute i3 mapped to 4 bits. We could
90
+ have omitted the <literal>length</>, <literal>col1</>,
91
+ and <literal>col2</> specifications since those have the default values.
87
92
</para>
88
93
89
94
<para>
90
- Here is a fuller example of index definition and usage:
95
+ Here is a more complete example of bloom index definition and usage, as
96
+ well as a comparison with equivalent btree indexes. The bloom index is
97
+ considerably smaller than the btree index, and can perform better.
91
98
</para>
92
99
93
100
<programlisting>
94
- CREATE TABLE tbloom AS
95
- SELECT
96
- random()::int as i1,
97
- random()::int as i2,
98
- random()::int as i3,
99
- random()::int as i4,
100
- random()::int as i5,
101
- random()::int as i6,
102
- random()::int as i7,
103
- random()::int as i8,
104
- random()::int as i9,
105
- random()::int as i10,
106
- random()::int as i11,
107
- random()::int as i12,
108
- random()::int as i13
109
- FROM
110
- generate_series(1,1000);
111
- CREATE INDEX bloomidx ON tbloom USING
112
- bloom (i1, i2, i3, i4, i5, i6, i7, i8, i9, i10, i11, i12);
113
- SELECT pg_relation_size('bloomidx');
114
- CREATE index btree_idx ON tbloom(i1,i2,i3,i4,i5,i6,i7,i8,i9,i10,i11,i12);
115
- SELECT pg_relation_size('btree_idx');
101
+ =# CREATE TABLE tbloom AS
102
+ SELECT
103
+ (random() * 1000000)::int as i1,
104
+ (random() * 1000000)::int as i2,
105
+ (random() * 1000000)::int as i3,
106
+ (random() * 1000000)::int as i4,
107
+ (random() * 1000000)::int as i5,
108
+ (random() * 1000000)::int as i6
109
+ FROM
110
+ generate_series(1,10000000);
111
+ SELECT 10000000
112
+ =# CREATE INDEX bloomidx ON tbloom USING bloom (i1, i2, i3, i4, i5, i6);
113
+ CREATE INDEX
114
+ =# SELECT pg_size_pretty(pg_relation_size('bloomidx'));
115
+ pg_size_pretty
116
+ ----------------
117
+ 153 MB
118
+ (1 row)
119
+ =# CREATE index btreeidx ON tbloom (i1, i2, i3, i4, i5, i6);
120
+ CREATE INDEX
121
+ =# SELECT pg_size_pretty(pg_relation_size('btreeidx'));
122
+ pg_size_pretty
123
+ ----------------
124
+ 387 MB
125
+ (1 row)
116
126
</programlisting>
117
127
128
+ <para>
129
+ A sequential scan over this large table takes a long time:
118
130
<programlisting>
119
- =# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 20 AND i10 = 15 ;
120
- QUERY PLAN
121
- -----------------------------------------------------------------------------------------------------------------
122
- Bitmap Heap Scan on tbloom (cost=1.50..5.52 rows=1 width=52 ) (actual time=0.057..0.057 rows=0 loops=1)
123
- Recheck Cond : ((i2 = 20 ) AND (i10 = 15 ))
124
- -> Bitmap Index Scan on bloomidx (cost=0.00..1.50 rows=1 width=0) (actual time=0.041..0.041 rows=9 loops=1)
125
- Index Cond: ((i2 = 20) AND (i10 = 15))
126
- Total runtime: 0.081 ms
131
+ =# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 898732 AND i5 = 123451 ;
132
+ QUERY PLAN
133
+ ------------------------------------------------------------------------------------------------------------
134
+ Seq Scan on tbloom (cost=0.00..213694.08 rows=1 width=24 ) (actual time=1445.438..1445.438 rows=0 loops=1)
135
+ Filter : ((i2 = 898732 ) AND (i5 = 123451 ))
136
+ Rows Removed by Filter: 10000000
137
+ Planning time: 0.177 ms
138
+ Execution time: 1445.473 ms
127
139
(5 rows)
128
140
</programlisting>
141
+ </para>
129
142
130
143
<para>
131
- Seqscan is slow.
144
+ So the planner will usually select an index scan if possible.
145
+ With a btree index, we get results like this:
146
+ <programlisting>
147
+ =# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 898732 AND i5 = 123451;
148
+ QUERY PLAN
149
+ --------------------------------------------------------------------------------------------------------------------------------
150
+ Index Only Scan using btreeidx on tbloom (cost=0.56..298311.96 rows=1 width=24) (actual time=445.709..445.709 rows=0 loops=1)
151
+ Index Cond: ((i2 = 898732) AND (i5 = 123451))
152
+ Heap Fetches: 0
153
+ Planning time: 0.193 ms
154
+ Execution time: 445.770 ms
155
+ (5 rows)
156
+ </programlisting>
132
157
</para>
133
158
159
+ <para>
160
+ Bloom is better than btree in handling this type of search:
134
161
<programlisting>
135
- =# SET enable_bitmapscan = off;
136
- =# SET enable_indexscan = off;
137
- =# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 20 AND i10 = 15;
138
- QUERY PLAN
139
- --------------------------------------------------------------------------------------------------
140
- Seq Scan on tbloom (cost=0.00..25.00 rows=1 width=52) (actual time=0.162..0.162 rows=0 loops=1)
141
- Filter: ((i2 = 20) AND (i10 = 15))
142
- Total runtime: 0.181 ms
143
- (3 rows)
162
+ =# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 898732 AND i5 = 123451;
163
+ QUERY PLAN
164
+ ---------------------------------------------------------------------------------------------------------------------------
165
+ Bitmap Heap Scan on tbloom (cost=178435.39..178439.41 rows=1 width=24) (actual time=76.698..76.698 rows=0 loops=1)
166
+ Recheck Cond: ((i2 = 898732) AND (i5 = 123451))
167
+ Rows Removed by Index Recheck: 2439
168
+ Heap Blocks: exact=2408
169
+ -> Bitmap Index Scan on bloomidx (cost=0.00..178435.39 rows=1 width=0) (actual time=72.455..72.455 rows=2439 loops=1)
170
+ Index Cond: ((i2 = 898732) AND (i5 = 123451))
171
+ Planning time: 0.475 ms
172
+ Execution time: 76.778 ms
173
+ (8 rows)
144
174
</programlisting>
175
+ Note the relatively large number of false positives: 2439 rows were
176
+ selected to be visited in the heap, but none actually matched the
177
+ query. We could reduce that by specifying a larger signature length.
178
+ In this example, creating the index with <literal>length=200</>
179
+ reduced the number of false positives to 55; but it doubled the index size
180
+ (to 306 MB) and ended up being slower for this query (125 ms overall).
181
+ </para>
145
182
146
- <para>
147
- A btree index will be not used for this query.
148
- </para>
149
-
183
+ <para>
184
+ Now, the main problem with the btree search is that btree is inefficient
185
+ when the search conditions do not constrain the leading index column(s).
186
+ A better strategy for btree is to create a separate index on each column.
187
+ Then the planner will choose something like this:
150
188
<programlisting>
151
- =# DROP INDEX bloomidx;
152
- =# CREATE INDEX btree_idx ON tbloom(i1, i2, i3, i4, i5, i6, i7, i8, i9, i10, i11, i12);
153
- =# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 20 AND i10 = 15;
154
- QUERY PLAN
155
- --------------------------------------------------------------------------------------------------
156
- Seq Scan on tbloom (cost=0.00..25.00 rows=1 width=52) (actual time=0.210..0.210 rows=0 loops=1)
157
- Filter: ((i2 = 20) AND (i10 = 15))
158
- Total runtime: 0.250 ms
159
- (3 rows)
189
+ =# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 898732 AND i5 = 123451;
190
+ QUERY PLAN
191
+ ------------------------------------------------------------------------------------------------------------------------------
192
+ Bitmap Heap Scan on tbloom (cost=9.29..13.30 rows=1 width=24) (actual time=0.148..0.148 rows=0 loops=1)
193
+ Recheck Cond: ((i5 = 123451) AND (i2 = 898732))
194
+ -> BitmapAnd (cost=9.29..9.29 rows=1 width=0) (actual time=0.145..0.145 rows=0 loops=1)
195
+ -> Bitmap Index Scan on tbloom_i5_idx (cost=0.00..4.52 rows=11 width=0) (actual time=0.089..0.089 rows=10 loops=1)
196
+ Index Cond: (i5 = 123451)
197
+ -> Bitmap Index Scan on tbloom_i2_idx (cost=0.00..4.52 rows=11 width=0) (actual time=0.048..0.048 rows=8 loops=1)
198
+ Index Cond: (i2 = 898732)
199
+ Planning time: 2.049 ms
200
+ Execution time: 0.280 ms
201
+ (9 rows)
160
202
</programlisting>
203
+ Although this query runs much faster than with either of the single
204
+ indexes, we pay a large penalty in index size. Each of the single-column
205
+ btree indexes occupies 214 MB, so the total space needed is over 1.2GB,
206
+ more than 8 times the space used by the bloom index.
207
+ </para>
161
208
</sect2>
162
209
163
210
<sect2>
164
- <title>Opclass interface </title>
211
+ <title>Operator Class Interface </title>
165
212
166
213
<para>
167
- The Bloom opclass interface is simple. It requires 1 supporting function:
168
- a hash function for the indexing datatype. It provides 1 search operator:
169
- the equality operator. The example below shows <literal>opclass</>
170
- definition for <literal>text</> datatype.
214
+ An operator class for bloom indexes requires only a hash function for the
215
+ indexed datatype and an equality operator for searching. This example
216
+ shows the opclass definition for the <type>text</> data type:
171
217
</para>
172
218
173
219
<programlisting>
@@ -179,22 +225,21 @@ DEFAULT FOR TYPE text USING bloom AS
179
225
</sect2>
180
226
181
227
<sect2>
182
- <title>Limitation </title>
228
+ <title>Limitations </title>
183
229
<para>
184
-
185
230
<itemizedlist>
186
231
<listitem>
187
232
<para>
188
- For now, only opclasses for <literal >int4</>, <literal >text</> come
189
- with the module. However, users may define more of them .
233
+ Only operator classes for <type >int4</> and <type >text</> are
234
+ included with the module.
190
235
</para>
191
236
</listitem>
192
237
193
238
<listitem>
194
239
<para>
195
- Only the <literal>=</literal> operator is supported for search at the
196
- moment. But it's possible to add support for arrays with contains and
197
- intersection operations in the future.
240
+ Only the <literal>=</literal> operator is supported for search. But
241
+ it is possible to add support for arrays with union and intersection
242
+ operations in the future.
198
243
</para>
199
244
</listitem>
200
245
</itemizedlist>
0 commit comments