Deep Dive Into PostgreSQL Indexes

Ibrar Ahmed
Senior Database Architect - Percona LLC
May 2019
Table Characteristics

• 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

• Tuple stored in a table does not have any order

2
Selecting Data 1/2
• Select whole table, must be a sequential scan.

• Select table’s rows where id is 5432, it should not be a sequential scan.

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)

3
Selecting Data 2/2
CREATE TABLE foo(id INTEGER, name TEXT); Page 0/N Tuple - 1
Tuple - 2

INSERT INTO foo VALUES(1, 'Alex'); Tuple - 3

INSERT INTO foo VALUES(2, 'Bob');

Tuple - n

Tuple - 1
Page 1/N Tuple - 2
Tuple - 3
SELECT ctid, * FROM foo;
ctid | id | name
-------+----+------ Tuple - n H
Tuple - 1
Page 2/N E
(0,1) | 1 | Alex Tuple - 2
A
Tuple - 3
(0,2) | 2 | Bob P
(2 rows)
Tuple - n

• How to select the data from the HEAP?

• Need to scan each and every page and look for the Page N/N
Tuple - 1
Tuple - 2
tuple in the page Tuple - 3

Cost? Tuple - n

4
PostgreSQL Indexes
https://www.postgresql.org/docs/current/indexes.html

5
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
postgres=# EXPLAIN SELECT name FROM bar WHERE id = 5432;
QUERY PLAN
----------------------------------------------------------------------------
Seq Scan on bar (cost=0.00..159235.00 rows=38216 width=32)
Filter: (id = 5432)

postgres=# CREATE INDEX bar_idx ON bar(id);

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)

6
Index

• PostgreSQL standard way to create a index

(https://www.postgresql.org/docs/current/sql-createindex.html)

postgres=# CREATE INDEX idx_btree ON bar(id);

postgres=# SELECT relfilenode FROM pg_class WHERE relname LIKE ‘idx_btree’;

relfilenode
-------------
16425

• PostgreSQL index has its own file on disk.

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

7
Creating Index 1/2

• Index based on single column of the table

postgres=# CREATE INDEX bar_idx ON bar(id);

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)

8
Creating Index 2/2
PostgreSQL locks the table when creating index

CREATE INDEX idx_btree ON bar USING BTREE(id);

CREATE INDEX
Time: 12303.172 ms (00:12.303)

CONCURRENTLY option creates the index without locking the table

CREATE INDEX CONCURRENTLY idx_btree ON bar USING BTREE(id);

CREATE INDEX
Time: 23025.372 ms (00:23.025)

9
Expression Index 1/2
EXPLAIN SELECT * FROM bar WHERE lower(name) LIKE 'Text1';
QUERY PLAN
-------------------------------------------------------------
Seq Scan on bar (cost=0.00..213694.00 rows=50000 width=40)
Filter: (lower((name)::text) ~~ 'Text1'::text)

CREATE INDEX idx_exp ON bar (lower(name));

EXPLAIN SELECT * FROM bar WHERE lower(name) LIKE 'Text1';

QUERY PLAN
-----------------------------------------------------------------------------
Bitmap Heap Scan on bar (cost=1159.93..64658.02 rows=50000 width=40)
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)

1
0
Expression Index 2/2
postgres=# EXPLAIN SELECT * FROM bar WHERE (dt + (INTERVAL '2 days')) < now();
QUERY PLAN
---------------------------------------------------------------
Seq Scan on bar (cost=0.00..238694.00 rows=3333333 width=40)
Filter: ((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
-------------------------------------------------------------------------------------
Bitmap Heap Scan on bar (cost=62449.77..184477.10 rows=3333333 width=40)
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())

1
1
Partial Index
Index Partial Index
CREATE INDEX idx_full ON bar(id); CREATE INDEX idx_part ON bar(id) where id < 10000;

EXPLAIN SELECT * FROM bar EXPLAIN SELECT * FROM bar

WHERE id < 1000 WHERE id < 1000

AND name LIKE 'text1000’; AND name LIKE 'text1000’;

QUERY PLAN QUERY PLAN

------------------------------------------------------------------------ -----------------------------------------------------------------------
--
Bitmap Heap Scan on bar (cost=199.44..113893.44 rows=16667 width=40)
Bitmap Heap Scan on bar (cost=61568.60..175262.59 rows=16667 width=40)
Recheck Cond: (id < 1000)
Recheck Cond: (id < 1000)
Filter:
Q: What will happen when we query ((name)::text
where id >1000? ~~ 'text1000'::text)
Filter: ((name)::text ~~ 'text1000'::text)
-> Bitmap Index Scan on idx_part (cost=0.00..195.28 rows=3333333
-> Bitmap Index Scan on idx_full (cost=0.00..61564.43 rows=3333333 width=0)
width=0)
A: Answer is simple, this index won’t selected.
Index Cond: (id < 1000)
Index Cond: (id < 1000)

SELECT pg_size_pretty(pg_total_relation_size('idx_part'));
SELECT pg_size_pretty(pg_total_relation_size('idx_full'));
pg_size_pretty
pg_size_pretty
----------------
----------------
240 kB
214 MB
(1 row)
(1 row)

12
 Index Types
https://www.postgresql.org/docs/current/indexes-types.html

13
B-Tree Index 1/2
• What is a B-Tree index? Wikipedia: (https://en.wikipedia.org/wiki/Self-
• Supported Operators balancing_binary_search_tree)
• Less than < In computer science, a self-balancing (or height-balanced) binary search tree
• Less than equal to <=
• Equal = is any node-based binary search tree that automatically keeps its height
• Greater than equal to >= small in the face of arbitrary item insertions and deletions.
• Greater than >

CREATE INDEX idx_btree ON foo USING BTREE (name);

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)

14
B-Tree Index 2/2
CREATE TABLE foo(id INTEGER, name TEXT); Page 0/N Tuple - 1
Tuple - 2
INSERT INTO foo VALUES(1, 'Alex'); Tuple - 3

INSERT INTO foo VALUES(2, 'Bob');

Tuple - n
SELECT ctid, * FROM foo; Tuple - 1
Page 1/N
ctid | id | name Tuple - 2
Tuple - 3
-------+----+------
(0,1) | 1 | Alex
(0,2) | 2 | Bob
Tuple - n H
Tuple - 1
Page 2/N E
Tuple - 2
Tuple - 3
A
P
Index have the key and the location of the tuple.
ctid | name Tuple - n

-------+------
(0,1) | Alex Tuple - 1
(0,2) | Bob Page N/N Tuple - 2

(2,2) | Alex Tuple - 3

Tuple - n

15
HASH Index
• What is a Hash index? postgres=# \d bar
Table "public.bar"
• Hash indexes only handles equality operators Column | Type | Collation | Nullable | Default
--------+-------------------+-----------+----------+---------
• Hash function is used to locate the tuples id
name
| integer |
| character varying |
|
|
|
|
dt | date | | |
Indexes:
CREATE INDEX idx_hash ON bar USING HASH (name); "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)
Planning Time: 0.080 ms
Execution Time: 0.041 ms
(4 rows)

16
BRIN Index 1/2

• 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

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);

17
BRIN Index 2/2
Sequential Scan BRIN Index

postgres=# EXPLAIN ANALYZE SELECT * postgres=# EXPLAIN ANALYZE SELECT *

FROM bar FROM bar
WHERE dt > '2022-09-28’ WHERE dt > '2022-09-28’
AND dt < '2022-10-28'; AND dt < '2022-10-28';
QUERY PLAN QUERY PLAN
----------------------------------------------------- ------------------------------------------------------
Seq Scan on bar (cost=0.00..2235285.00 rows=1 Bitmap Heap Scan on bar (cost=92.03..61271.08 rows=1
width=27) width=27) (actual time=1.720..4.186 rows=29 loops=1)
(actual time=0.139..7397.090 rows=29 Recheck Cond: ((dt > '2022-09-28 00:00:00’)
loops=1) AND (dt < '2022-10-28 00:00:00'))
Filter: ((dt > '2022-09-28 00:00:00) Rows Removed by Index Recheck: 18716
AND (dt < '2022-10-28 00:00:00)) Heap Blocks: lossy=128
Rows Removed by Filter: 99999971 -> Bitmap Index Scan on idx_brin
Planning Time: 0.114 ms (cost=0.00..92.03 rows=17406 width=0)
Execution Time: 7397.107 ms (actual time=1.456..1.456 rows=1280 loops=1)
(5 rows) Index Cond: ((dt > '2022-09-28 00:00:00’)
AND (dt < '2022-10-28 00:00:00'))
Planning Time: 0.130 ms
Execution Time: 4.233 ms
(8 rows)

18
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

postgres=# \d bar
Table "public.bar"
Column | Type | Collation | Nullable | Default
--------+---------+-----------+----------+---------
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)

19
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
CREATE INDEX idx_gin ON bar USING GIN (name);

postgres=# EXPLAIN ANALYZE SELECT * FROM bar postgres=# EXPLAIN ANALYZE SELECT * FROM bar
WHERE name @> '{"name": "Alex"}’; WHERE name @> '{"name": "Alex"}';
QUERY PLAN QUERY PLAN
----------------------------------------------------- -----------------------------------------------------
Seq Scan on bar (cost=0.00..108309.34 rows=3499 Bitmap Heap Scan on bar (cost=679.00..13395.57
width=96) (actual time=396.019..1050.143 rows=1000000 rows=4000 width=96) (actual time=91.110..445.112
loops=1) rows=1000000
Even if you create a BTREE index, it won’t be loops=1)
considered.
Filter: (name @> '{"name": Because
"Alex"}'::jsonb)
it does not know the individual element in (name
Recheck Cond: value. @> '{"name": "Alex"}'::jsonb)
Rows Removed by Filter: 3000000 Heap Blocks: exact=16394
Planning Time: 0.107 ms -> Bitmap Index Scan on
Execution Time: 1079.861 ms 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)
Planning Time: 0.168 ms
Execution Time: 475.447 ms
20
GiST Index

• Generalized Search Tree

• A GiST index is lossy

• Tree-structured access method

21
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

22
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

23
Index Only Scans
CREATE INDEX idx_btree_ios ON bar (id,name);

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)

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)

24
Duplicate Indexes
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

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)

25
Unused Indexes
SELECT relname, indexrelname, idx_scan
FROM pg_catalog.pg_stat_user_indexes;

relname | indexrelname | idx_scan

---------+---------------+----------
foo | idx_foo_date | 0
bar | idx_btree | 0
bar | idx_btree_id | 0
bar | idx_btree_name| 6
bar | idx_brin_brin | 4
(7 rows)

26
 ?
“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

27
Thank You to Our Sponsors
Rate My Session

29