mirror of
https://git.postgresql.org/git/postgresql.git
synced 2024-09-12 21:38:18 +02:00
Discussion: https://postgr.es/m/CAEepm%3D3sijpGr8tXdyz-7EJJZfhQHABPKEQ29gpnb7-XSy%2B%3D5A%40mail.gmail.com Reported-by: Thomas Munro Backpatch-through: 9.6
284 lines
9.8 KiB
Plaintext
284 lines
9.8 KiB
Plaintext
<!-- doc/src/sgml/bloom.sgml -->
|
|
|
|
<sect1 id="bloom" xreflabel="bloom">
|
|
<title>bloom</title>
|
|
|
|
<indexterm zone="bloom">
|
|
<primary>bloom</primary>
|
|
</indexterm>
|
|
|
|
<para>
|
|
<literal>bloom</literal> provides an index access method based on
|
|
<ulink url="https://en.wikipedia.org/wiki/Bloom_filter">Bloom filters</ulink>.
|
|
</para>
|
|
|
|
<para>
|
|
A Bloom filter is a space-efficient data structure that is used to test
|
|
whether an element is a member of a set. In the case of an index access
|
|
method, it allows fast exclusion of non-matching tuples via signatures
|
|
whose size is determined at index creation.
|
|
</para>
|
|
|
|
<para>
|
|
A signature is a lossy representation of the indexed attribute(s), and as
|
|
such is prone to reporting false positives; that is, it may be reported
|
|
that an element is in the set, when it is not. So index search results
|
|
must always be rechecked using the actual attribute values from the heap
|
|
entry. Larger signatures reduce the odds of a false positive and thus
|
|
reduce the number of useless heap visits, but of course also make the index
|
|
larger and hence slower to scan.
|
|
</para>
|
|
|
|
<para>
|
|
This type of index is most useful when a table has many attributes and
|
|
queries test arbitrary combinations of them. A traditional btree index is
|
|
faster than a bloom index, but it can require many btree indexes to support
|
|
all possible queries where one needs only a single bloom index. Note
|
|
however that bloom indexes only support equality queries, whereas btree
|
|
indexes can also perform inequality and range searches.
|
|
</para>
|
|
|
|
<sect2>
|
|
<title>Parameters</title>
|
|
|
|
<para>
|
|
A <literal>bloom</literal> index accepts the following parameters in its
|
|
<literal>WITH</literal> clause:
|
|
</para>
|
|
|
|
<variablelist>
|
|
<varlistentry>
|
|
<term><literal>length</literal></term>
|
|
<listitem>
|
|
<para>
|
|
Length of each signature (index entry) in bits. It is rounded up to the
|
|
nearest multiple of <literal>16</literal>. The default is
|
|
<literal>80</literal> bits and the maximum is <literal>4096</literal>.
|
|
</para>
|
|
</listitem>
|
|
</varlistentry>
|
|
</variablelist>
|
|
<variablelist>
|
|
<varlistentry>
|
|
<term><literal>col1 — col32</literal></term>
|
|
<listitem>
|
|
<para>
|
|
Number of bits generated for each index column. Each parameter's name
|
|
refers to the number of the index column that it controls. The default
|
|
is <literal>2</literal> bits and maximum is <literal>4095</literal>. Parameters for
|
|
index columns not actually used are ignored.
|
|
</para>
|
|
</listitem>
|
|
</varlistentry>
|
|
</variablelist>
|
|
</sect2>
|
|
|
|
<sect2>
|
|
<title>Examples</title>
|
|
|
|
<para>
|
|
This is an example of creating a bloom index:
|
|
</para>
|
|
|
|
<programlisting>
|
|
CREATE INDEX bloomidx ON tbloom USING bloom (i1,i2,i3)
|
|
WITH (length=80, col1=2, col2=2, col3=4);
|
|
</programlisting>
|
|
|
|
<para>
|
|
The index is created with a signature length of 80 bits, with attributes
|
|
i1 and i2 mapped to 2 bits, and attribute i3 mapped to 4 bits. We could
|
|
have omitted the <literal>length</literal>, <literal>col1</literal>,
|
|
and <literal>col2</literal> specifications since those have the default values.
|
|
</para>
|
|
|
|
<para>
|
|
Here is a more complete example of bloom index definition and usage, as
|
|
well as a comparison with equivalent btree indexes. The bloom index is
|
|
considerably smaller than the btree index, and can perform better.
|
|
</para>
|
|
|
|
<programlisting>
|
|
=# CREATE TABLE tbloom AS
|
|
SELECT
|
|
(random() * 1000000)::int as i1,
|
|
(random() * 1000000)::int as i2,
|
|
(random() * 1000000)::int as i3,
|
|
(random() * 1000000)::int as i4,
|
|
(random() * 1000000)::int as i5,
|
|
(random() * 1000000)::int as i6
|
|
FROM
|
|
generate_series(1,10000000);
|
|
SELECT 10000000
|
|
=# CREATE INDEX bloomidx ON tbloom USING bloom (i1, i2, i3, i4, i5, i6);
|
|
CREATE INDEX
|
|
=# SELECT pg_size_pretty(pg_relation_size('bloomidx'));
|
|
pg_size_pretty
|
|
----------------
|
|
153 MB
|
|
(1 row)
|
|
=# CREATE index btreeidx ON tbloom (i1, i2, i3, i4, i5, i6);
|
|
CREATE INDEX
|
|
=# SELECT pg_size_pretty(pg_relation_size('btreeidx'));
|
|
pg_size_pretty
|
|
----------------
|
|
387 MB
|
|
(1 row)
|
|
</programlisting>
|
|
|
|
<para>
|
|
A sequential scan over this large table takes a long time:
|
|
<programlisting>
|
|
=# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 898732 AND i5 = 123451;
|
|
QUERY PLAN
|
|
------------------------------------------------------------------------------------------------------------
|
|
Seq Scan on tbloom (cost=0.00..213694.08 rows=1 width=24) (actual time=1445.438..1445.438 rows=0 loops=1)
|
|
Filter: ((i2 = 898732) AND (i5 = 123451))
|
|
Rows Removed by Filter: 10000000
|
|
Planning time: 0.177 ms
|
|
Execution time: 1445.473 ms
|
|
(5 rows)
|
|
</programlisting>
|
|
</para>
|
|
|
|
<para>
|
|
So the planner will usually select an index scan if possible.
|
|
With a btree index, we get results like this:
|
|
<programlisting>
|
|
=# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 898732 AND i5 = 123451;
|
|
QUERY PLAN
|
|
--------------------------------------------------------------------------------------------------------------------------------
|
|
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)
|
|
Index Cond: ((i2 = 898732) AND (i5 = 123451))
|
|
Heap Fetches: 0
|
|
Planning time: 0.193 ms
|
|
Execution time: 445.770 ms
|
|
(5 rows)
|
|
</programlisting>
|
|
</para>
|
|
|
|
<para>
|
|
Bloom is better than btree in handling this type of search:
|
|
<programlisting>
|
|
=# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 898732 AND i5 = 123451;
|
|
QUERY PLAN
|
|
---------------------------------------------------------------------------------------------------------------------------
|
|
Bitmap Heap Scan on tbloom (cost=178435.39..178439.41 rows=1 width=24) (actual time=76.698..76.698 rows=0 loops=1)
|
|
Recheck Cond: ((i2 = 898732) AND (i5 = 123451))
|
|
Rows Removed by Index Recheck: 2439
|
|
Heap Blocks: exact=2408
|
|
-> Bitmap Index Scan on bloomidx (cost=0.00..178435.39 rows=1 width=0) (actual time=72.455..72.455 rows=2439 loops=1)
|
|
Index Cond: ((i2 = 898732) AND (i5 = 123451))
|
|
Planning time: 0.475 ms
|
|
Execution time: 76.778 ms
|
|
(8 rows)
|
|
</programlisting>
|
|
Note the relatively large number of false positives: 2439 rows were
|
|
selected to be visited in the heap, but none actually matched the
|
|
query. We could reduce that by specifying a larger signature length.
|
|
In this example, creating the index with <literal>length=200</literal>
|
|
reduced the number of false positives to 55; but it doubled the index size
|
|
(to 306 MB) and ended up being slower for this query (125 ms overall).
|
|
</para>
|
|
|
|
<para>
|
|
Now, the main problem with the btree search is that btree is inefficient
|
|
when the search conditions do not constrain the leading index column(s).
|
|
A better strategy for btree is to create a separate index on each column.
|
|
Then the planner will choose something like this:
|
|
<programlisting>
|
|
=# EXPLAIN ANALYZE SELECT * FROM tbloom WHERE i2 = 898732 AND i5 = 123451;
|
|
QUERY PLAN
|
|
------------------------------------------------------------------------------------------------------------------------------
|
|
Bitmap Heap Scan on tbloom (cost=9.29..13.30 rows=1 width=24) (actual time=0.148..0.148 rows=0 loops=1)
|
|
Recheck Cond: ((i5 = 123451) AND (i2 = 898732))
|
|
-> BitmapAnd (cost=9.29..9.29 rows=1 width=0) (actual time=0.145..0.145 rows=0 loops=1)
|
|
-> 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)
|
|
Index Cond: (i5 = 123451)
|
|
-> 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)
|
|
Index Cond: (i2 = 898732)
|
|
Planning time: 2.049 ms
|
|
Execution time: 0.280 ms
|
|
(9 rows)
|
|
</programlisting>
|
|
Although this query runs much faster than with either of the single
|
|
indexes, we pay a large penalty in index size. Each of the single-column
|
|
btree indexes occupies 214 MB, so the total space needed is over 1.2GB,
|
|
more than 8 times the space used by the bloom index.
|
|
</para>
|
|
</sect2>
|
|
|
|
<sect2>
|
|
<title>Operator Class Interface</title>
|
|
|
|
<para>
|
|
An operator class for bloom indexes requires only a hash function for the
|
|
indexed data type and an equality operator for searching. This example
|
|
shows the operator class definition for the <type>text</type> data type:
|
|
</para>
|
|
|
|
<programlisting>
|
|
CREATE OPERATOR CLASS text_ops
|
|
DEFAULT FOR TYPE text USING bloom AS
|
|
OPERATOR 1 =(text, text),
|
|
FUNCTION 1 hashtext(text);
|
|
</programlisting>
|
|
</sect2>
|
|
|
|
<sect2>
|
|
<title>Limitations</title>
|
|
<para>
|
|
<itemizedlist>
|
|
<listitem>
|
|
<para>
|
|
Only operator classes for <type>int4</type> and <type>text</type> are
|
|
included with the module.
|
|
</para>
|
|
</listitem>
|
|
|
|
<listitem>
|
|
<para>
|
|
Only the <literal>=</literal> operator is supported for search. But
|
|
it is possible to add support for arrays with union and intersection
|
|
operations in the future.
|
|
</para>
|
|
</listitem>
|
|
|
|
<listitem>
|
|
<para>
|
|
<literal>bloom</literal> access method doesn't support
|
|
<literal>UNIQUE</literal> indexes.
|
|
</para>
|
|
</listitem>
|
|
|
|
<listitem>
|
|
<para>
|
|
<literal>bloom</literal> access method doesn't support searching for
|
|
<literal>NULL</literal> values.
|
|
</para>
|
|
</listitem>
|
|
</itemizedlist>
|
|
</para>
|
|
</sect2>
|
|
|
|
<sect2>
|
|
<title>Authors</title>
|
|
|
|
<para>
|
|
Teodor Sigaev <email>teodor@postgrespro.ru</email>,
|
|
Postgres Professional, Moscow, Russia
|
|
</para>
|
|
|
|
<para>
|
|
Alexander Korotkov <email>a.korotkov@postgrespro.ru</email>,
|
|
Postgres Professional, Moscow, Russia
|
|
</para>
|
|
|
|
<para>
|
|
Oleg Bartunov <email>obartunov@postgrespro.ru</email>,
|
|
Postgres Professional, Moscow, Russia
|
|
</para>
|
|
</sect2>
|
|
|
|
</sect1>
|