Deep dive to PostgreSQL Indexes

© 2019 Percona
Termini PostgreSQL di Deep Dee Intu

© 2019 Percona
Ibrar Ahmed
Deep Dive Into PostgreSQL Indexes
Learn PostgreSQL index
Senior Database Architect - Percona LLC
October – 16, 2019

© 2019 Percona
• Rows / Tuples stored in a table
• Every table in PostgreSQL has physical disk file(s)
postgres=# CREATE TABLE foo(id int, name text);
postgres=# SELECT relfilenode FROM pg_class WHERE relname LIKE 'foo’;
relfilenode
-------------
16384
• The physical files on disk can be seen in the PostgreSQL $PGDATA directory.
$ ls -lrt $PGDATA/base/13680/16384
-rw------- 1 vagrant vagrant 0 Apr 29 11:48 $PGDATA/base/13680/16384
Tables (Heap)
“Tuple” and “Rows” are synonym
$PGDATA is Data Directory
“pg_class” is system table to contain table information“Relfilenode” is table file name of that table
“16384” is table filename
• Tuple stored in a table does not have any order

© 2019 Percona
EXPLAIN SELECT name FROM bar;
QUERY PLAN
----------------------------------------------------------------------------
Seq Scan on bar (cost=0.00..163693.05 rows=9999905 width=11
EXPLAIN SELECT name FROM bar WHERE id = 5432;
QUERY PLAN
----------------------------------------------------------------------------
Gather (cost=1000.00..116776.94 rows=1 width=11)
Workers Planned: 2
-> Parallel Seq Scan on bar (cost=0.00..115776.84 rows=1 width=11)
Filter: (id = 5432)
Tables (Heap)
• Select whole table, must be a sequential scan.
• Select table’s rows where id is 5432, it should not be a sequential scan.
Make sense?
Why?

© 2019 Percona
Selecting Data from HEAP
Page 0/N
Page 1/N
Page 2/N
Page N/N
Tuple - 1
Tuple - 3
Tuple - 2
Tuple - 1
Tuple - n
Tuple - 3
Tuple - 2
Tuple - 1
Tuple - n
Tuple - 3
Tuple - 2
Tuple - 1
Tuple - n
Tuple - 3
Tuple - 2
Tuple - n
CREATE TABLE foo(id INTEGER, name TEXT);
INSERT INTO foo VALUES(1, 'Alex');
INSERT INTO foo VALUES(2, 'Bob');
SELECT ctid, * FROM foo;
ctid | id | name
-------+----+------
(0,1) | 1 | Alex
(0,2) | 2 | Bob
(2 rows)
H
E
A
P
• How to select the data from the HEAP?
• Need to scan each and every page and look for the
tuple in the page
Cost?

© 2019 Percona
PostgreSQL Indexes
https://www.postgresql.org/docs/current/indexes.html
7

© 2019 Percona
postgres=# EXPLAIN SELECT name FROM bar WHERE id = 5432;
QUERY PLAN
----------------------------------------------------------------------------
Bitmap Heap Scan on bar (cost=939.93..64313.02 rows=50000 width=32)
Recheck Cond: (id = 5432)
-> Bitmap Index Scan on bar_idx (cost=0.00..927.43 rows=50000 width=0)
Index Cond: (id = 5432)
postgres=# CREATE INDEX bar_idx ON bar(id);
QUERY PLAN
----------------------------------------------------------------------------
Seq Scan on bar (cost=0.00..159235.00 rows=38216 width=32)
Filter: (id = 5432)
Why Index?
• Indexes are entry points for tables
• Index used to locate the tuples in the table
• The sole reason to have an index is performance
• Index is stored separately from the table’s main storage (PostgreSQL Heap)
• More storage required to store the index along with original table
64313/159235 * 100 = 40%
Cost of the query

© 2019 Percona
postgres=# CREATE INDEX idx_btree ON bar(id);
postgres=# SELECT relfilenode FROM pg_class WHERE relname LIKE ‘idx_btree’;
relfilenode
-------------
16425
Index
Physical file name of the index
PostgreSQL’s Catalog for relations/index
The physical file on disk can be seen in the PostgreSQL $PGDATA directory.
$ ls -lrt $PGDATA/13680/16425
-rw-------1 vagrant vagrant 1073741824 Apr 29 13:05 $PGDATA/base/13680/16425
• PostgreSQL standard way to create a index
(https://www.postgresql.org/docs/current/sql-createindex.html)
• PostgreSQL index has its own file on disk.

© 2019 Percona
postgres=# CREATE INDEX bar_idx ON bar(id);
Creating Index 1/2
”bar” is a table and ”id” is column
QUERY PLAN
----------------------------------------------------------------------
Recheck Cond: (id = 5432)
-> Bitmap Index Scan on bar_idx (cost=0.00..927.43 rows=50000 width=0)
Index Cond: (id = 5432)
• Index based on single column of the table

© 2019 Percona
CREATE INDEX CONCURRENTLY idx_btree ON bar USING BTREE(id);
CREATE INDEX
Time: 23025.372 ms (00:23.025)
CREATE INDEX idx_btree ON bar USING BTREE(id);
CREATE INDEX
Time: 12303.172 ms (00:12.303)
Creating Index 2/2
• PostgreSQL locks the table when creating index
• CONCURRENTLY option creates the index without locking the table

© 2019 Percona
Expression Index 1/2
CREATE INDEX idx_exp ON bar (lower(name));
EXPLAIN SELECT * FROM bar WHERE lower(name) LIKE 'Text1';
QUERY PLAN
-------------------------------------------------------------
Filter: (lower((name)::text) ~~ 'Text1'::text)
EXPLAIN SELECT * FROM bar WHERE lower(name) LIKE 'Text1';
QUERY PLAN
-----------------------------------------------------------------------------
Filter: (lower((name)::text) ~~ 'Text1'::text)
-> Bitmap Index Scan on idx_exp (cost=0.00..1147.43 rows=50000 width=0)
Index Cond: (lower((name)::text) = 'Text1'::text)

© 2019 Percona
Expression Index 2/2
postgres=# EXPLAIN SELECT * FROM bar WHERE (dt + (INTERVAL '2 days')) < now();
QUERY PLAN
-------------------------------------------------------------------------------------
Recheck Cond: ((dt + '2 days'::interval) < now())
-> Bitmap Index Scan on idx_math_exp (cost=0.00..61616.43 rows=3333333 width=0)
Index Cond: ((dt + '2 days'::interval) < now())
postgres=# CREATE INDEX idx_math_exp ON bar((dt + (INTERVAL '2 days')));
postgres=# EXPLAIN SELECT * FROM bar WHERE (dt + (INTERVAL '2 days')) < now();
QUERY PLAN
---------------------------------------------------------------
Filter: ((dt + '2 days'::interval) < now())

© 2019 Percona
Partial Index
14
CREATE INDEX idx_part ON bar(id) where id < 1000;
EXPLAIN SELECT * FROM bar
WHERE id < 1000
AND name LIKE 'text1000’;
QUERY PLAN
-----------------------------------------------------------------------
Recheck Cond: (id < 1000)
Filter: ((name)::text ~~ 'text1000'::text)
-> Bitmap Index Scan on idx_part (cost=0.00..195.28 rows=3333333
width=0)
Index Cond: (id < 1000)
SELECT pg_size_pretty(pg_total_relation_size('idx_part'));
pg_size_pretty
----------------
240 kB
(1 row)
CREATE INDEX idx_full ON bar(id);
EXPLAIN SELECT * FROM bar
WHERE id < 1000
AND name LIKE 'text1000’;
QUERY PLAN
------------------------------------------------------------------------
--
Recheck Cond: (id < 1000)
Filter: ((name)::text ~~ 'text1000'::text)
-> Bitmap Index Scan on idx_full (cost=0.00..61564.43 rows=3333333
width=0)
Index Cond: (id < 1000)
SELECT pg_size_pretty(pg_total_relation_size('idx_full'));
pg_size_pretty
----------------
214 MB
(1 row)
Look at the size of the index
Why create full index if we don’t
need that.
Index Partial Index
Index where id < 1000 only
Q: What will happen when we query where id >1000?
A: Answer is simple, this index won’t selected.

© 2019 Percona
Index Types
https://www.postgresql.org/docs/current/indexes-types.html

© 2019 Percona
B-Tree Index 1/2
16
CREATE INDEX idx_btree ON foo USING BTREE (name);
• What is a B-Tree index?
• Supported Operators
• Less than <
• Less than equal to <=
• Equal =
• Greater than equal to >=
• Greater than >
Wikipedia: (https://en.wikipedia.org/wiki/Self-
balancing_binary_search_tree)
In computer science, a self-balancing (or height-balanced) binary search tree
is any node-based binary search tree that automatically keeps its height
small in the face of arbitrary item insertions and deletions.
postgres=# EXPLAIN ANALYZE SELECT * FROM foo WHERE name = 'text%';
QUERY PLAN
----------------------------------------------------------------------------------------------------------------
Index Scan using idx_btree on foo (cost=0.43..8.45 rows=1 width=19) (actual time=0.015..0.015 rows=0 loops=1)
Index Cond: ((name)::text = 'text%'::text)
Planning Time: 0.105 ms
Execution Time: 0.031 ms
(4 rows)

© 2019 Percona
B-Tree Index 2/2
17
Page 0/N
Page 1/N
Page 2/N
Page N/N
Tuple - 1
Tuple - 3
Tuple - 2
Tuple - 1
Tuple - n
Tuple - 3
Tuple - 2
Tuple - 1
Tuple - n
Tuple - 3
Tuple - 2
Tuple - 1
Tuple - n
Tuple - 3
Tuple - 2
Tuple - n
CREATE TABLE foo(id INTEGER, name TEXT, … );
INSERT INTO foo VALUES(1, 'Alex’, …);
INSERT INTO foo VALUES(2, ‘Bob, …
SELECT ctid, * FROM foo;
ctid | id | name …
-------+----+------
(0,1) | 1 | Alex …
(0,2) | 2 | Bob H
E
A
P
ctid | name
-------+------
(0,1) | Alex
(0,2) | Bob
Index have the key and the location of the tuple.

© 2019 Percona
HASH Index
• What is a Hash index?
• Hash indexes only handles equality operators
• Hash function is used to locate the tuples
18
CREATE INDEX idx_hash ON bar USING HASH (name);
postgres=# d bar
Table "public.bar"
Column | Type | Collation | Nullable | Default
--------+-------------------+-----------+----------+---------
id | integer | | |
name | character varying | | |
dt | date | | |
Indexes:
"idx_btree" btree (name)
"idx_hash" btree (name)
EXPLAIN ANALYZE SELECT * FROM bar WHERE name = 'text%';
QUERY PLAN
Index Scan using idx_hash on bar (cost=0.43..8.45 rows=1 width=19) (actual time=0.023..0.023
rows=0 loops=1)
Index Cond: ((name)::text = 'text%'::text)
(4 rows)

© 2019 Percona
BRIN Index 1/3
• BRIN is a “Block Range Index”
• Used when columns have some correlation with their physical location in the table
• Space optimized because BRIN index contains only three items
• Page number
• Min value of column
• Max value of column
19
Like date, date of city have logical group

© 2019 Percona
BRIN Index 2/3
Sequential Scan
20
postgres=# EXPLAIN ANALYZE SELECT *
FROM bar
WHERE dt > '2022-09-28’
AND dt < '2022-10-28';
QUERY PLAN
------------------------------------------------------
Bitmap Heap Scan on bar (cost=92.03..61271.08 rows=1
width=27) (actual time=1.720..4.186 rows=29 loops=1)
Recheck Cond: ((dt > '2022-09-28 00:00:00’)
AND (dt < '2022-10-28 00:00:00'))
Rows Removed by Index Recheck: 18716
Heap Blocks: lossy=128
-> Bitmap Index Scan on idx_brin
(cost=0.00..92.03 rows=17406 width=0)
(actual time=1.456..1.456 rows=1280 loops=1)
Index Cond: ((dt > '2022-09-28 00:00:00’)
AND (dt < '2022-10-28 00:00:00'))
(8 rows)
postgres=# EXPLAIN ANALYZE SELECT *
FROM bar
WHERE dt > '2022-09-28’
AND dt < '2022-10-28';
QUERY PLAN
-----------------------------------------------------
Seq Scan on bar (cost=0.00..2235285.00 rows=1
width=27)
(actual time=0.139..7397.090 rows=29
loops=1)
Filter: ((dt > '2022-09-28 00:00:00)
AND (dt < '2022-10-28 00:00:00))
Rows Removed by Filter: 99999971
(5 rows)
BRIN Index

© 2019 Percona
BRIN Index 3/3
21
• CREATE INDEX idx_btree ON bar USING BTREE (date);
• CREATE INDEX idx_hash ON bar USING HASH (date);
• CREATE INDEX idx_brin ON bar USING BRIN (date);
21504
48128
48
1
10
100
1000
10000
100000
Size of Index
(KB)
Index Sizes
B-Tree Hash BRIN

© 2019 Percona
GIN Index 1/2
• Generalized Inverted Index
• GIN is to handle where we need to index composite values
• Slow while creating the index because it needs to scan the document up front
22
postgres=# d bar
Table "public.bar"
--------+---------+-----------+----------+---------
id | integer | | |
name | jsonb | | |
dt | date | | |
postgres=# SELECT DISTINCT name, dt FROM bar LIMIT 5;
name | dt
---------------------------------------------------------------------------+------------
{"name": "Alex", "phone": ["333-333-333", "222-222-222", "111-111-111"]} | 2019-05-13
{"name": "Bob", "phone": ["333-333-444", "222-222-444", "111-111-444"]} | 2019-05-14
{"name": "John", "phone": ["333-3333", "777-7777", "555-5555"]} | 2019-05-15
{"name": "David", "phone": ["333-333-555", "222-222-555", "111-111-555"]} | 2019-05-16
(4 rows)

© 2019 Percona
GIN Index 2/2
• Generalized Inverted Index
• GIN is to handle where we need to index composite values
• Slow while creating index because it needs to scan the document up front
23
postgres=# EXPLAIN ANALYZE SELECT * FROM bar
WHERE name @> '{"name": "Alex"}’;
QUERY PLAN
-----------------------------------------------------
Seq Scan on bar (cost=0.00..108309.34 rows=3499
width=96) (actual time=396.019..1050.143 rows=1000000
loops=1)
Filter: (name @> '{"name": "Alex"}'::jsonb)
Rows Removed by Filter: 3000000
CREATE INDEX idx_gin ON bar USING GIN (name);
postgres=# EXPLAIN ANALYZE SELECT * FROM bar
WHERE name @> '{"name": "Alex"}';
QUERY PLAN
-----------------------------------------------------
Bitmap Heap Scan on bar (cost=679.00..13395.57
rows=4000 width=96) (actual time=91.110..445.112
rows=1000000 loops=1)
Recheck Cond: (name @> '{"name": "Alex"}'::jsonb)
Heap Blocks: exact=16394
-> Bitmap Index Scan on
idx_gin (cost=0.00..678.00 rows=4000 width=0)
(actual time=89.033..89.033 rows=1000000 loops=1)
Index Cond: (name @> '{"name":
"Alex"}'::jsonb)

© 2019 Percona
GiST Index
• Generalized Search Tree
• It is Tree-structured access method
• It is a indexing framework used for indexing of complex data types.
• Used to find the point within box
• Used for full text search
• Intarray
24

© 2019 Percona
Where and What?
• B-Tree: Use this index for most of the queries and different data types
• Hash: Used for equality operators
• BRIN: For really large sequentially lineup datasets
• GIN: Used for documents and arrays
• GiST: Used for full text search
25

© 2019 Percona
Index Only Scans
• Index is stored separately from the table’s main storage (PostgreSQL Heap)
• Query needs to scan both the index and the heap
• Index Only Scans only used when all the columns in the query part of the index
• In this case PostgreSQL fetches data from index only
26
In PostgreSQL term

© 2019 Percona
Index Only Scans
27
CREATE INDEX idx_btree_ios ON bar (id, name);
EXPLAIN SELECT id, name FROM bar WHERE id > 100000 AND id <100010;
QUERY PLAN
Index Only Scan using idx_btree_ios on bar (cost=0.56..99.20 rows=25 width=15)
Index Cond: ((id > 100000) AND (id < 100010))
(2 rows)
EXPLAIN SELECT id, name, dt FROM bar WHERE id > 100000 AND id <100010;
QUERY PLAN
Index Scan using idx_btree_ios on bar (cost=0.56..99.20 rows=25 width=19)
Index Cond: ((id > 100000) AND (id < 100010))
(2 rows)

© 2019 Percona
Duplicate Indexes
28
SELECT indrelid::regclass relname, indkey, amname
FROM pg_index i, pg_opclass o, pg_am a
WHERE o.oid = ALL (indclass)
AND a.oid = o.opcmethod
GROUP BY relname, indclass, amname, indkey
HAVING count(*) > 1;
relname | indkey | amname
---------+--------+--------
bar | 2 | btree
(1 row)
SELECT indrelid::regclass relname,
indexrelid::regclass indexname, indkey
FROM pg_index
GROUP BY relname,indexname,indkey;
relname | indexname | indkey
--------------------------+-----------------------------------------------+---------
pg_index | pg_index_indexrelid_index | 1
pg_toast.pg_toast_2615 | pg_toast.pg_toast_2615_index | 1 2
pg_constraint | pg_constraint_conparentid_index | 11

© 2019 Percona
?
“Poor leaders rarely ask questions of themselves or others. Good leaders, on the other
hand, ask many questions. Great leaders ask the great questions.”
Michael Marquardt author of
Leading with Questions

Deep dive to PostgreSQL Indexes

Recommended

More Related Content

What's hot (20)

Similar to Deep dive to PostgreSQL Indexes (20)

Recently uploaded (20)

Deep dive to PostgreSQL Indexes