PostgreSQL 9.4: NoSQL on ACID

Oleg Bartunov, Teodor Sigaev
• Locale support
• Extendability (indexing)
• GiST (KNN), GIN, SP-GiST
• Full Text Search (FTS)
• Jsonb, VODKA
• Extensions:
• intarray
• pg_trgm
• ltree
• hstore
• plantuner
https://www.facebook.com/oleg.bartunov
obartunov@gmail.com, teodor@sigaev.ru
https://www.facebook.com/groups/postgresql/

Alexander Korotkov
aekorotkov@gmail.com
●
Indexed regexp search
●
GIN compression & fast scan
●
Fast GiST build
●
Range types indexing
●
Split for GiST
●
Indexing for jsonb
●
jsquery
●
Generic WAL + create am (WIP)

Agenda
• The problem
• Hstore
• Introduction to jsonb indexing
• Jsquery - Jsonb Query Language
• Exercises on jsonb GIN opclasses with Jsquery support
• VODKA access method

The problem
• The world of data and applications is changing
• BIG DATA (Volume of data,Velocity of data in-out, Variety of data)
• Web applications are service-oriented
• Service itself can aggregate data, check consistency of data
• High concurrency, simple queries
• Simple database (key-value) is ok
• Eventual consistency is ok, no ACID overhead
• Application needs faster releases
• NoSQL databases match all of these — scalable, efficient, fault-tolerant,
no rigid schema, ready to accept any data.

NoSQL (концептуальные предпосылки)
• Реляционные СУБД — интеграционные
• Все приложения общаются через СУБД
• SQL — универсальный язык работы с данными
• Все изменения в СУБД доступны всем
• Изменения схемы очень затратны, медл. релизы
• Рассчитаны на интерактивную работу
• Интересны агрегаты, а не сами данные, нужен SQL
• SQL отслеживает транзакционность, ограничения целостности... вместо человека

NoSQL (концептуальные предпосылки)
• Сервисная архитектура изменила подход к СУБД
• Приложение состоит из сервисов, SQL->HTTP
• Сервисам не нужна одна монолитная СУБД
• Часто достаточно простых key-value СУБД
• Схема меняется «на ходу», быстрые релизы
• ACID → BASE
• Сервисы — это программы, которые могут сами заниматься агрегированием
• Сервисы могут сами следить за целостностью данных
• Много данных, аналитика, большое кол-во одновременных запросов
• Распределенность - кластеры дешевых shared-nothing машин
• NoSQL —горизонтальная масштабируемость и производительность

NoSQL
• Key-value databases
• Ordered k-v for ranges support
• Column family (column-oriented) stores
• Big Table — value has structure:
• column families, columns, and timestamped versions (maps-of maps-of maps)
• Document databases
• Value has arbitrary structure
• Graph databases — evolution od ordered-kv

NoSQL databases (wikipedia) …+++
Document store
* Lotus Notes
* CouchDB
* MongoDB
* Apache Jackrabbit
* Colayer
* XML databases
o MarkLogic Server
o eXist
Graph
* Neo4j
* AllegroGraph
Tabular
* BigTable
* Mnesia
* Hbase
* Hypertable
Key/value store on disk
* Tuple space
* Memcachedb
* Redis
* SimpleDB
* flare
* Tokyo Cabinet
* BigTable
Key/value cache in RAM
* memcached
* Velocity
* Redis
Eventually-consistent key-value store
* Dynamo
* Cassandra
* Project Voldemort
Ordered key-value store
* NMDB
* Luxio
* Memcachedb
* Berkeley DB
Object database
* Db4o
* InterSystems Caché
* Objectivity/DB
* ZODB

CAP теорема: Наш вариант

The problem
• What if application needs ACID and flexibility of NoSQL ?
• Relational databases work with data with schema known in advance
• One of the major compaints to relational databases is rigid schema.
It's not easy to change schema online (ALTER TABLE … ADD COLUMN...)
• Application should wait for schema changing, infrequent releases
• NoSQL uses json format, why not have it in relational database ?
JSON in PostgreSQL
This is the challenge !

Challenge to PostgreSQL !
• Full support of semi-stuctured data in PostgreSQL
• Storage
• Operators and functions
• Efficiency (fast access to storage, indexes)
• Integration with CORE (planner, optimiser)
• Actually, PostgreSQL is schema-less database since 2003 — hstore, one
of the most popular extension !

Introduction to Hstore
id col1 col2 col3 col4 col5 A lot of columns
key1, …. keyN
●
The problem:
●
Total number of columns may be very large
●
Only several fields are searchable ( used in WHERE)
●
Other columns are used only to output
●
These columns may not known in advance
●
Solution
●
New data type (hstore), which consists of (key,value) pairs (a'la perl hash)

Introduction to Hstore
id col1 col2 col3 col4 col5 Hstore
key1=>val1, key2=>val2,.....
●
Easy to add key=>value pair
●
No need change schema, just change hstore.
●
Schema-less PostgreSQL in 2003 !

Introduction to hstore
• Hstore — key/value binary storage (inspired by perl hash)
'a=>1, b=>2'::hstore
• Key, value — strings
• Get value for a key: hstore -> text
• Operators with indexing support (GiST, GIN)
Check for key: hstore ? text
Contains: hstore @> hstore
• check documentations for more
• Functions for hstore manipulations (akeys, avals, skeys, svals, each,......)
• Hstore provides PostgreSQL schema-less feature !
• Faster releases, no problem with schema upgrade

Hstore binary storage
Npairs:31
Varlena
header
New version flag:1
Key endpos:
31
Val endpos:
31
ISNULL:1
key... val ...
Start End
First key 0 HEntry[0]
i-th key HEntry[i*2 - 1] HEntry[i*2]
i-th value HEntry[i*2] HEntry[i*2 + 1]
String arrayHEntry array
Pairs are lexicographically ordered by key

Hstore limitations
• Levels: unlimited
• Number of elements in array: 2^31
• Number of pairs in hash: 2^31
• Length of string: 2^31 bytes
2^31 bytes = 2 GB

History of hstore development
• May 16, 2003 — first version of hstore

History of hstore development
• May 16, 2003 - first (unpublished) version of hstore for PostgreSQL
7.3
• Dec, 05, 2006 - hstore is a part of PostgreSQL 8.2
(thanks, Hubert Depesz Lubaczewski!)
• May 23, 2007 - GIN index for hstore, PostgreSQL 8.3
• Sep, 20, 2010 - Andrew Gierth improved hstore, PostgreSQL 9.0

Inverted Index
QUERY: compensation accelerometers
INDEX: accelerometers compensation
5,10,25,28,3030,36,58,59,61,73,74 3030,68
RESULT: 3030

GIN improvements
• GIN in 9.4 is greatly improved
• Posting lists compression (varbyte encoding) — smaller indexes
• 9.3: always 6 bytes (4 bytes blockNumber , 2 bytes offset): 90 bytes
(0,8) (0,14) (0,17) (0,22) (0,26) (0,33) (0,34) (0,35) (0,45) (0,47) (0,48) (1,3) (1,4)
(1,6) (1,8)
• 9.4: 1-6 bytes per each item, deltas from previous item: 21 bytes
(0,8) +6 +3 +5 +4 +7 +1 +1 +10 +2 +1 +2051 +1+2 +2
SELECT g % 10 FROM generate_series(1,10000000) g; 11Mb vs 58Mb
• Fast scan of posting lists - «rare & frequent» queries much faster
• 9.3: read posting lists for «rare» and «frequent» and join them
Time(frequent & rare) ~ Time(frequent)
• 9.4: start from posting list for «rare» and skip «frequent» list if no match
Time(frequent & rare) ~ Time(rare)

Hstore is DEAD ? No !
• How hstore benefits by GIN improvement in 9.4 ?
GIN stands for Generalized Inverted Index, so virtually all data types, which use
GIN, get benefit !
• Default hstore GIN opclass considers keys and values separately
• Keys are «frequent», value are «rare»
• Contains query: hstore @> 'key=>value' improved a lot for «rare» values
• Index size is smaller, less io

Hstore 9.3 vs 9.4
Total: 7240858 geo records:
"fcode"=>"RFSU",
"point"=>"(8.85,112.53333)",
"fclass"=>"U",
"asciiname"=>"London Reefs",
"elevation"=>NULL,
"geonameid"=>"1879967",
"population"=>"0"
Query:
SELECT count(*) FROM geo
WHERE geo @> 'fcode=>STM';
gin_hstore_ops: index keys and values
gin_hstore_bytea_ops = gin_hstore_ops, no collation comparison
gin_hstore_hash_ops: index hash(key.value)

Hstore 9.3 vs 9.4
|-------------------------+-------+----------+-------+---------|
| Name | Type | Owner | Table | Size |
|-------------------------+-------+----------+-------+---------|
| geo | table | postgres | | 1352 MB |
| geo_hstore_bytea_ops | index | postgres | geo | 1680 MB |
| geo_hstore_hash_ops_idx | index | postgres | geo | 1073 MB |
|-------------------------+-------+----------+-------+---------|
|-------------------------+-------+----------+-------+---------|
| Name | Type | Owner | Table | Size |
|-------------------------+-------+----------+-------+---------|
| geo | table | postgres | | 1352 MB |
| geo_hstore_bytea_ops | index | postgres | geo | 1296 MB |
| geo_hstore_hash_ops_idx | index | postgres | geo | 925 MB |
|-------------------------+-------+----------+-------+---------|
9.4
9.3
CREATE OPERATOR CLASS gin_hstore_bytea_ops FOR TYPE hstore
….....................................................................................
FUNCTION 1 byteacmp(bytea,bytea),
….....................................................................................
STORAGE bytea;
CREATE INDEX: 239 s Much faster comparison (no collation)
CREATE OPERATOR CLASS gin_hstore_ops FOR TYPE hstore
….....................................................................................
FUNCTION 1 bttextcmp(text,text),,
….....................................................................................
STORAGE text;
CREATE INDEX: 2870 s

Hstore 9.3 vs 9.4
SUMMARY:
●
9.4 GIN posting list compression:
indexes are smaller
●
9.4 GIN is smart regarding 'freq & rare' queries:
time (freq & rare) ~ time (rare) instead of
time (freq & rare) ~ time (freq)
●
gin_hstore_hash_ops is good on 9.3 & 9.4 and
faster default gin opclass
●
Use gin_hstore_bytea_ops instead of default
gin_hstore_ops — much faster create index
Get hstore_ops from:
from https://github.com/akorotkov/hstore_ops

Introduction to hstore
• Hstore benefits
• In provides a flexible model for storing a semi-structured data in relational
database
• hstore has binary storage and rich set of operators and functions, indexes
• Hstore drawbacks
• Too simple model !
Hstore key-value model doesn't supports tree-like structures as json
(introduced in 2006, 3 years after hstore)
• Json — popular and standartized (ECMA-404 The JSON Data
Interchange Standard, JSON RFC-7159)
• Json — PostgreSQL 9.2, textual storage

Hstore vs Json
SELECT sum((v->'a')::text::int) FROM json_test;
851.012 ms
SELECT sum((v->'a')::int) FROM hstore_test;
330.027 ms
• hstore is faster than json even on simple data
CREATE TABLE hstore_test AS (SELECT
'a=>1, b=>2, c=>3, d=>4, e=>5'::hstore AS v
FROM generate_series(1,1000000));
CREATE TABLE json_test AS (SELECT
'{"a":1, "b":2, "c":3, "d":4, "e":5}'::json AS v
FROM generate_series(1,1000000));

Hstore vs Json
• PostgreSQL already has json since 9.2, which supports document-
based model, but
• It's slow, since it has no binary representation and needs to be parsed every
time
• Hstore is fast, thanks to binary representation and index support
• It's possible to convert hstore to json and vice versa, but current hstore is
limited to key-value
• Need hstore with document-based model. Share it's
binary representation with json !

Nested hstore & jsonb
• Nested hstore at PGCon-2013, Ottawa, Canada ( May 24) — thanks
Engine Yard for support !
One step forward true json data type.Nested hstore with arrays support
• Binary storage for nested data at PGCon Europe — 2013, Dublin, Ireland
(Oct 29)
Binary storage for nested data structuresand application to hstore data type
• November, 2013 — binary storage was reworked, nested hstore and
jsonb share the same storage. Andrew Dunstan joined the project.
• January, 2014 - binary storage moved to core

Nested hstore & jsonb
• Feb-Mar, 2014 - Peter Geoghegan joined the project, nested hstore was
cancelled in favour to jsonb (Nested hstore patch for 9.3).
• Mar 23, 2014 Andrew Dunstan committed jsonb to 9.4 branch !
pgsql: Introduce jsonb, a structured format for storing json.
Introduce jsonb, a structured format for storing json.
The new format accepts exactly the same data as the json type. However, it is
stored in a format that does not require reparsing the orgiginal text in order
to process it, making it much more suitable for indexing and other operations.
Insignificant whitespace is discarded, and the order of object keys is not
preserved. Neither are duplicate object keys kept - the later value for a given
key is the only one stored.

Jsonb vs Json
SELECT '{"c":0, "a":2,"a":1}'::json, '{"c":0, "a":2,"a":1}'::jsonb;
json | jsonb
-----------------------+------------------
{"c":0, "a":2,"a":1} | {"a": 1, "c": 0}
(1 row)
• json: textual storage «as is»
• jsonb: no whitespaces
• jsonb: no duplicate keys, last key win
• jsonb: keys are sorted

Jsonb vs Json
• Data
• 1,252,973 Delicious bookmarks
• Server
• MBA, 8 GB RAM, 256 GB SSD
• Test
• Input performance - copy data to table
• Access performance - get value by key
• Search performance contains @> operator

Jsonb vs Json
• Input performance (parser)
Copy data (1,252,973 rows) as text, json,jsonb
copy tt from '/path/to/test.dump'
Text: 34 s - as is
Json: 37 s - json validation
Jsonb: 43 s - json validation, binary storage

Jsonb vs Json (binary storage)
• Access performance — get value by key
• Base: SELECT js FROM js;
• Jsonb: SELECT j->>'updated' FROM jb;
• Json: SELECT j->>'updated' FROM js;
Base: 0.6 s
Jsonb: 1 s 0.4
Json: 9.6 s 9
Jsonb ~ 20X faster Json

Jsonb vs Json
EXPLAIN ANALYZE SELECt count(*) FROM js WHERE js #>>'{tags,0,term}' = 'NYC';
QUERY PLAN
----------------------------------------------------------------------------
Aggregate (cost=187812.38..187812.39 rows=1 width=0)
(actual time=10054.602..10054.602 rows=1 loops=1)
-> Seq Scan on js (cost=0.00..187796.88 rows=6201 width=0)
Filter: ((js #>> '{tags,0,term}'::text[]) = 'NYC'::text)
Rows Removed by Filter: 1252850
Planning time: 0.078 ms
Execution runtime: 10054.635 ms
(6 rows)
Json: no contains @> operator,
search first array element

Jsonb vs Json (binary storage)
EXPLAIN ANALYZE SELECT count(*) FROM jb WHERE jb @> '{"tags":[{"term":"NYC"}]}'::jsonb;
QUERY PLAN
---------------------------------------------------------------------------------------
-> Seq Scan on jb (cost=0.00..191518.16 rows=1253 width=0)
Filter: (jb @> '{"tags": [{"term": "NYC"}]}'::jsonb)
Execution runtime: 1263.225 ms Execution runtime: 10054.635 ms
(6 rows)

Jsonb vs Json (GIN: key && value)
QUERY PLAN
---------------------------------------------------------------------------------------
-> Bitmap Heap Scan on jb (cost=73.71..4769.59 rows=1253 width=0)
Recheck Cond: (jb @> '{"tags": [{"term": "NYC"}]}'::jsonb)
Heap Blocks: exact=285
-> Bitmap Index Scan on gin_jb_idx (cost=0.00..73.40 rows=1253 width=0)
Index Cond: (jb @> '{"tags": [{"term": "NYC"}]}'::jsonb)
(8 rows)
CREATE INDEX gin_jb_idx ON jb USING gin(jb);

Jsonb vs Json (GIN: hash path.value)
QUERY PLAN
---------------------------------------------------------------------------------------
-> Bitmap Heap Scan on jb (cost=33.71..4729.59 rows=1253 width=0)
Recheck Cond: (jb @> '{"tags": [{"term": "NYC"}]}'::jsonb)
-> Bitmap Index Scan on gin_jb_path_idx
(cost=0.00..33.40 rows=1253 width=0) (actual time=0.062..0.062 rows=285 loops=1)
Index Cond: (jb @> '{"tags": [{"term": "NYC"}]}'::jsonb)
(8 rows)
CREATE INDEX gin_jb_path_idx ON jb USING gin(jb jsonb_path_ops);

PostgreSQL 9.4 vs Mongo 2.6.0
• Operator contains @>
• json : 10 s seqscan
• jsonb : 8.5 ms GIN jsonb_ops
• jsonb : 0.7 ms GIN jsonb_path_ops
• mongo : 1.0 ms btree index
• Index size
• jsonb_ops - 636 Mb (no compression, 815Mb)
jsonb_path_ops - 295 Mb
• jsonb_path_ops (tags) - 44 Mb USING gin((jb->'tags') jsonb_path_ops
• mongo (tags) - 387 Mb
mongo (tags.term) - 100 Mb
•Table size
•postgres : 1.3Gb
•mongo : 1.8Gb
•Input performance:
• Text : 34 s
• Json : 37 s
• Jsonb : 43 s
• mongo : 13 m

Что сейчас может Jsonb ?
• Contains operators - jsonb @> jsonb, jsonb <@ jsonb (GIN indexes)
jb @> '{"tags":[{"term":"NYC"}]}'::jsonb
Keys should be specified from root
●
Equivalence operator — jsonb = jsonb (GIN indexes)
• Exists operators — jsonb ? text, jsonb ?! text[], jsonb ?& text[] (GIN indexes)
jb WHERE jb ?| '{tags,links}'
Only root keys supported
• Operators on jsonb parts (functional indexes)
SELECT ('{"a": {"b":5}}'::jsonb -> 'a'->>'b')::int > 2;
CREATE INDEX ….USING BTREE ( (jb->'a'->>'b')::int);
Very cumbersome, too many functional indexes

Найти что-нибудь красное
• Table "public.js_test"
Column | Type | Modifiers
--------+---------+-----------
id | integer | not null
value | jsonb |
select * from js_test;
id | value
----+-----------------------------------------------------------------------
1 | [1, "a", true, {"b": "c", "f": false}]
2 | {"a": "blue", "t": [{"color": "red", "width": 100}]}
3 | [{"color": "red", "width": 100}]
4 | {"color": "red", "width": 100}
5 | {"a": "blue", "t": [{"color": "red", "width": 100}], "color": "red"}
6 | {"a": "blue", "t": [{"color": "blue", "width": 100}], "color": "red"}
7 | {"a": "blue", "t": [{"color": "blue", "width": 100}], "colr": "red"}
8 | {"a": "blue", "t": [{"color": "green", "width": 100}]}
9 | {"color": "green", "value": "red", "width": 100}
(9 rows)

• WITH RECURSIVE t(id, value) AS ( SELECT * FROM
js_test
UNION ALL
(
SELECT
t.id,
COALESCE(kv.value, e.value) AS value
FROM
t
LEFT JOIN LATERAL
jsonb_each(
CASE WHEN jsonb_typeof(t.value) =
'object' THEN t.value
ELSE NULL END) kv ON true
LEFT JOIN LATERAL
jsonb_array_elements(
CASE WHEN
jsonb_typeof(t.value) = 'array' THEN t.value
ELSE NULL END) e ON true
WHERE
kv.value IS NOT NULL OR e.value IS
NOT NULL
)
)
SELECT
js_test.*
FROM
(SELECT id FROM t WHERE value @> '{"color":
"red"}' GROUP BY id) x
JOIN js_test ON js_test.id = x.id;
• Весьма непростое решение !

Что хочется ?
• Need Jsonb query language
• Simple and effective way to search in arrays (and other iterative
searches)
• More comparison operators (сейчас только =)
• Types support
• Schema support (constraints on keys, values)
• Indexes support

Что хочется ?
• Need Jsonb query language
• Simple and effective way to search in arrays (and other iterative
searches)
• More comparison operators (сейчас только =)
• Types support
• Schema support (constraints on keys, values)
• Indexes support
• Introduce Jsquery - textual data type and @@ match operator
jsonb @@ jsquery

PGCon-2014, Май, Оттава

• # - any element array
• % - any key
• * - anything
• $ - current element
• Use "double quotes" for key !
SELECT '{"a": {"b": [1,2,3]}}'::jsonb @@ 'a.b.# = 2';
SELECT '{"a": {"b": [1,2,3]}}'::jsonb @@ '%.b.# = 2';
SELECT '{"a": {"b": [1,2,3]}}'::jsonb @@ '*.# = 2';
select '{"a": {"b": [1,2,3]}}'::jsonb @@ 'a.b.# ($ = 2 OR $ < 3)';
select 'a1."12222" < 111'::jsquery;
path ::= key
| path '.' key_any
| NOT '.' key_any
key ::= '*'
| '#'
| '%'
| '$'
| STRING
….....
key_any ::= key
| NOT

• Scalar
• Test for key existence
• Array overlap
• Array contains
• Array contained
select '{"a": {"b": [1,2,3]}}'::jsonb @@ 'a.b.# IN (1,2,5)';
select '{"a": {"b": [1,2,3]}}'::jsonb @@ 'a.b = *';
select '{"a": {"b": [1,2,3]}}'::jsonb @@ 'a.b && [1,2,5]';
select '{"a": {"b": [1,2,3]}}'::jsonb @@ 'a.b @> [1,2]';
select '{"a": {"b": [1,2,3]}}'::jsonb @@ 'a.b <@ [1,2,3,4,5]';
value_expr
::= '=' scalar_value
| IN '(' value_list ')'
| '=' array
| '=' '*'
| '<' NUMERIC
| '<' '=' NUMERIC
| '>' NUMERIC
| '>' '=' NUMERIC
| '@' '>' array
| '<' '@' array
| '&' '&' array
| IS ARRAY
| IS NUMERIC
| IS OBJECT
| IS STRING
| IS BOOLEAN

• Type checking
select '{"x": true}' @@ 'x IS boolean'::jsquery,
'{"x": 0.1}' @@ 'x IS numeric'::jsquery;
?column? | ?column?
----------+----------
t | t
IS BOOLEAN
IS NUMERIC
IS ARRAY
IS OBJECT
IS STRINGselect '{"a":{"a":1}}' @@ 'a IS object'::jsquery;
?column?
----------
t
select '{"a":["xxx"]}' @@ 'a IS array'::jsquery, '["xxx"]' @@ '$ IS array'::jsquery;
?column? | ?column?
----------+----------
t | t

• How many products are similar to "B000089778" and have
product_sales_rank in range between 10000-20000 ?
• SQL
SELECT count(*) FROM jr WHERE (jr->>'product_sales_rank')::int > 10000
and (jr->> 'product_sales_rank')::int < 20000 and
….boring stuff
• Jsquery
SELECT count(*) FROM jr WHERE jr @@ ' similar_product_ids &&
["B000089778"] AND product_sales_rank( $ > 10000 AND $ < 20000)'
• Mongodb
db.reviews.find( { $and :[ {similar_product_ids: { $in ["B000089778"]}},
{product_sales_rank:{$gt:10000, $lt:20000}}] } ).count()

• WITH RECURSIVE t(id, value) AS ( SELECT * FROM
js_test
UNION ALL
(
SELECT
t.id,
COALESCE(kv.value, e.value) AS value
FROM
t
LEFT JOIN LATERAL
jsonb_each(
CASE WHEN jsonb_typeof(t.value) =
'object' THEN t.value
ELSE NULL END) kv ON true
LEFT JOIN LATERAL
jsonb_array_elements(
CASE WHEN
jsonb_typeof(t.value) = 'array' THEN t.value
ELSE NULL END) e ON true
WHERE
kv.value IS NOT NULL OR e.value IS
NOT NULL
)
)
SELECT
js_test.*
FROM
(SELECT id FROM t WHERE value @> '{"color":
"red"}' GROUP BY id) x
JOIN js_test ON js_test.id = x.id;
• Jsquery
SELECT * FROM js_test
WHERE
value @@ '*.color = "red"';

Еще пример
• SQL
SELECT * FROM js_test2 js
WHERE NOT EXISTS (
SELECT 1
FROM
jsonb_array_elements(js.value) el
WHERE EXISTS (
SELECT 1
FROM jsonb_each(el.value) kv
WHERE NOT
kv.value::text::numeric BETWEEN
0.0 AND 1.0));
• Jsquery
SELECT * FROM js_test2 js
WHERE '#:.%:($ >= 0 AND $ <= 1)';

explain( analyze, buffers) select count(*) from jb where jb @> '{"tags":[{"term":"NYC"}]}'::jsonb;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------
Aggregate (cost=191517.30..191517.31 rows=1 width=0) (actual time=1039.422..1039.423 rows=1 loops=1)
Buffers: shared hit=97841 read=78011
-> Seq Scan on jb (cost=0.00..191514.16 rows=1253 width=0) (actual time=0.006..1039.310 rows=285 loops=1)
Filter: (jb @> '{"tags": [{"term": "NYC"}]}'::jsonb)
Buffers: shared hit=97841 read=78011
Execution time: 1039.444 ms
explain( analyze,costs off) select count(*) from jb where jb @@ 'tags.#.term = "NYC"';
QUERY PLAN
--------------------------------------------------------------------
Aggregate (actual time=891.707..891.707 rows=1 loops=1)
-> Seq Scan on jb (actual time=0.010..891.553 rows=285 loops=1)
Filter: (jb @@ '"tags".#."term" = "NYC"'::jsquery)

Jsquery (indexes)
• GIN opclasses with jsquery support
• jsonb_value_path_ops — use Bloom filtering for key matching
{"a":{"b":{"c":10}}} → 10.( bloom(a) or bloom(b) or bloom(c) )
• Good for key matching (wildcard support) , not good for range query
• jsonb_path_value_ops — hash path (like jsonb_path_ops)
{"a":{"b":{"c":10}}} → hash(a.b.c).10
• No wildcard support, no problem with ranges
List of relations
Schema | Name | Type | Owner | Table | Size | Description
--------+-------------------------+-------+----------+--------------+---------+-------------
public | jb | table | postgres | | 1374 MB |
public | jb_value_path_idx | index | postgres | jb | 306 MB |
public | jb_gin_idx | index | postgres | jb | 544 MB |
public | jb_path_value_idx | index | postgres | jb | 306 MB |
public | jb_path_idx | index | postgres | jb | 251 MB |

Jsquery (indexes)
explain( analyze,costs off) select count(*) from jb where jb @@ 'tags.#.term = "NYC"';
QUERY PLAN
-------------------------------------------------------------------------------------------------
-> Bitmap Heap Scan on jb (actual time=0.115..0.580 rows=285 loops=1)
Recheck Cond: (jb @@ '"tags".#."term" = "NYC"'::jsquery)
-> Bitmap Index Scan on jb_value_path_idx (actual time=0.073..0.073 rows=285 loops=1)
Index Cond: (jb @@ '"tags".#."term" = "NYC"'::jsquery)
(7 rows)

Jsquery (indexes)
explain( analyze,costs off) select count(*) from jb where jb @@ '*.term = "NYC"';
QUERY PLAN
-------------------------------------------------------------------------------------------------
Recheck Cond: (jb @@ '*."term" = "NYC"'::jsquery)
-> Bitmap Index Scan on jb_value_path_idx (actual time=0.113..0.113 rows=285 loops=1)
Index Cond: (jb @@ '*."term" = "NYC"'::jsquery)
(7 rows)

Citus dataset{
"customer_id": "AE22YDHSBFYIP",
"product_category": "Business & Investing",
"product_group": "Book",
"product_id": "1551803542",
"product_sales_rank": 11611,
"product_subcategory": "General",
"product_title": "Start and Run a Coffee Bar (Start & Run a)",
"review_date": {
"$date": 31363200000
},
"review_helpful_votes": 0,
"review_rating": 5,
"review_votes": 10,
"similar_product_ids": [
"0471136174",
"0910627312",
"047112138X",
"0786883561",
"0201570483"
]
}
• 3023162 reviews from Citus
1998-2000 years
• 1573 MB

Jsquery (indexes)
explain (analyze, costs off) select count(*) from jr where
jr @@ ' similar_product_ids && ["B000089778"]';
QUERY PLAN
------------------------------------------------------------------------------------------------
-> Bitmap Heap Scan on jr (actual time=0.084..0.337 rows=185 loops=1)
Recheck Cond: (jr @@ '"similar_product_ids" && ["B000089778"]'::jsquery)
-> Bitmap Index Scan on jr_path_value_idx (actual time=0.057..0.057 rows=185 loops=1)
Index Cond: (jr @@ '"similar_product_ids" && ["B000089778"]'::jsquery)
(7 rows)

Jsquery (indexes)
jr @@ ' similar_product_ids && ["B000089778"]
AND product_sales_rank( $ > 10000 AND $ < 20000)';
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------
Recheck Cond: (jr @@ '("similar_product_ids" && ["B000089778"] &
"product_sales_rank"($ > 10000 & $ < 20000))'::jsquery)
Index Cond: (jr @@ '("similar_product_ids" && ["B000089778"] &
"product_sales_rank"($ > 10000 & $ < 20000))'::jsquery)
Execution time: 129.309 ms !!! No statistics
(7 rows)
• No statistics, no planning :(
Not selective, better not use index!

MongoDB 2.6.0
db.reviews.find( { $and :[ {similar_product_ids: { $in:["B000089778"]}}, {product_sales_rank:{$gt:10000, $lt:20000}}] } )
.explain()
{
"n" : 45,
….................
"millis" : 7,
"indexBounds" : {
"similar_product_ids" : [ index size = 400 MB just for similar_product_ids !!!
[
"B000089778",
"B000089778"
]
]
},
}

Jsquery (indexes)
explain (analyze,costs off) select count(*) from jr where
jr @@ ' similar_product_ids && ["B000089778"]'
and (jr->>'product_sales_rank')::int>10000 and (jr->>'product_sales_rank')::int<20000;
-----------------------------------------------------------------------------------------------------------------------------------------
Filter: ((((jr ->> 'product_sales_rank'::text))::integer > 10000) AND
(((jr ->> 'product_sales_rank'::text))::integer < 20000))
Execution time: 0.506 ms Potentially, query could be faster Mongo !
(9 rows)
• If we rewrite query and use planner

Jsquery (optimiser) — NEW !
• Jsquery now has built-in optimiser for simple queries.
jr @@ 'similar_product_ids && ["B000089778"]
AND product_sales_rank( $ > 10000 AND $ < 20000)'
------------------------------------------------------------------------------------------------------------------------------------------
Recheck Cond: (jr @@ '("similar_product_ids" && ["B000089778"] AND
"product_sales_rank"($ > 10000 AND $ < 20000))'::jsquery)
Rows Removed by Index Recheck: 140
Index Cond: (jr @@ '("similar_product_ids" && ["B000089778"] AND
Execution time: 0.480 ms vs 7 ms MongoDB !

• Jsquery now has built-in optimiser for simple queries.
Analyze query tree and push non-selective parts to recheck (like filter)
Selectivity classes:
1) Equality (x = c)
2) Range (c1 < x < c2)
3) Inequality (c > c1)
4) Is (x is type)
5) Any (x = *)
SELECT gin_debug_query_path_value('similar_product_ids && ["B000089778"]
AND product_sales_rank( $ > 10000 AND $ < 20000)');
gin_debug_query_path_value
-------------------------------------------------
similar_product_ids.# = "B000089778" , entry 0 +

• Jsquery optimiser pushes non-selective operators to recheck
jr @@ 'similar_product_ids && ["B000089778"]
AND product_sales_rank( $ > 10000 AND $ < 20000)'
------------------------------------------------------------------------------------------------------------------------------------------
Recheck Cond: (jr @@ '("similar_product_ids" && ["B000089778"] AND
Rows Removed by Index Recheck: 140
Index Cond: (jr @@ '("similar_product_ids" && ["B000089778"] AND

Jsquery (HINTING) — NEW !
• Jsquery now has HINTING ( if you don't like optimiser)!
explain (analyze, costs off) select count(*) from jr where jr @@ 'product_sales_rank > 10000'
----------------------------------------------------------------------------------------------------------
Recheck Cond: (jr @@ '"product_sales_rank" > 10000'::jsquery)
-> Bitmap Index Scan on jr_path_value_idx (actual time=1052.483..1052.48
rows=2373140 loops=1)
Index Cond: (jr @@ '"product_sales_rank" > 10000'::jsquery)
• Better not to use index — HINT /* --noindex */
explain (analyze, costs off) select count(*) from jr where jr @@ 'product_sales_rank /*-- noindex */ > 10000';
----------------------------------------------------------------------------------
-> Seq Scan on jr (actual time=0.013..1222.123 rows=2373140 loops=1)
Filter: (jr @@ '"product_sales_rank" /*-- noindex */ > 10000'::jsquery)

Contrib/jsquery
• Jsquery index support is quite efficient ( 0.5 ms vs Mongo 7 ms ! )
• Future direction
• Make jsquery planner friendly
• Need statistics for jsonb
• Availability
• Jsquery + opclasses are available as extensions
• Grab it from https://github.com/akorotkov/jsquery (branch master) ,
we need your feedback !

PostgreSQL 9.4+
●
Open-source
●
Relational database
●
Strong support of json

Что дальше ?
• SQL-level jsquery (расширяемость, статистика)
• VODKA access method ! VODKA Optimized Dendriform Keys Array
• Комбинация произвольных методов доступа

Better indexing ...
• Provide interface to change hardcoded B-tree in Entry tree
• Use spgist opclass for storing paths and values as is (strings hashed in values)
• We may go further - provide interface to change hardcoded B-tree in
posting tree
• GIS aware full text search !
• New index access method
CREATE INDEX … USING VODKA

GIN History
• Introduced at PostgreSQL Anniversary Meeting in Toronto, Jul 7-8, 2006
by Oleg Bartunov and Teodor Sigaev

GIN History
• Introduced at PostgreSQL Anniversary Meeting in Toronto, Jul 7-8, 2006
by Oleg Bartunov and Teodor Sigaev
• Supported by JFG Networks (France)
• «Gin stands for Generalized Inverted iNdex and should be considered as
a genie, not a drink.»
• Alexander Korotkov, Heikki Linnakangas have joined GIN++ development
in 2013

GIN History
TODO
----
Nearest future:
* Opclasses for all types (no programming, just many catalog changes).
Distant future:
* Replace B-tree of entries to something like GiST (VODKA ! 2014)
* Add multicolumn support
* Optimize insert operations (background index insertion)
• From GIN Readme, posted in -hackers, 2006-04-26

GIN index structure for jsonb
{
"product_sales_rank": 15000
},
{
"product_group": "Music",
}

Vodka index structure for jsonb
{
},
{
"product_group": "Music",
}

select count(*) from jb where jb @@ 'tags.#.term = "NYC"';
-------------------------------------------------------------------------------------------
Recheck Cond: (jb @@ '"tags".#."term" = "NYC"'::jsquery)
-> Bitmap Index Scan on jb_vodka_idx (actual time=0.108..0.108 rows=285 loops=1)
Index Cond: (jb @@ '"tags".#."term" = "NYC"'::jsquery)
Execution time: 0.456 ms (0.634 ms, GIN jsonb_value_path_ops)
select count(*) from jb where jb @@ '*.term = "NYC"';
-------------------------------------------------------------------------------------------
Recheck Cond: (jb @@ '*."term" = "NYC"'::jsquery)
-> Bitmap Index Scan on jb_vodka_idx (actual time=0.214..0.214 rows=285 loops=1)
Index Cond: (jb @@ '*."term" = "NYC"'::jsquery)
Execution time: 0.526 ms (0.716 ms, GIN jsonb_path_value_ops)

explain (analyze, costs off) select count(*) from jr where jr @@ ' similar_product_ids && ["B000089778"]';
QUERY PLAN
-------------------------------------------------------------------------------------------
-> Bitmap Index Scan on jr_vodka_idx (actual time=0.077..0.077 rows=185 loops=1)
Execution time: 0.237 ms (0.394 ms, GIN jsonb_path_value_idx)
(7 rows)

There are can be different flavors of Vodka

Find twirled spaghetti
Spaghetti indexing ...

R-tree fails here — bounding box of each separate spaghetti is the same
Spaghetti indexing ...

Ottawa downtown: York and George
streets

Summary
• contrib/jsquery for 9.4
• Jsquery - Jsonb Query Language
• Two GIN opclasses with jsquery support
• Grab it from https://github.com/akorotkov/jsquery (branch master)
• Prototype of VODKA access method
• Plans for improving indexing infrastructure
• This work was supported by

Another view on VODKA
• VODKA CONNECTING INDEXES
• composite index, which combines different access methods
• Nested search trees

postgrespro.ru
• Ищем инженеров:
• 24x7 поддержка
• Консалтинг & аудит
• Разработка админских
приложений
• Пакеты
• Ищем си-шников для работы
над постгресом:
• Неубиваемый и масштабируемый
кластер
• Хранилища (in-memory, column-
storage...)

VODKA Optimized Dendriform Keys Array

PostgreSQL 9.4: NoSQL on ACID

More Related Content

What's hot (20)

Similar to PostgreSQL 9.4: NoSQL on ACID (20)

Recently uploaded (20)

PostgreSQL 9.4: NoSQL on ACID