Repository files navigation

JsQuery – is a language to query jsonb data type, introduced in PostgreSQLrelease 9.4.

It's primary goal is to provide an additional functionality to jsonb(currently missing in PostgreSQL), such as a simple and effective wayto search in nested objects and arrays, more comparison operators withindexes support. We hope, that jsquery will be eventually a part ofPostgreSQL.

Jsquery is released as jsquery data type (similar to tsquery) and @@match operator for jsonb.

• Teodor Sigaevteodor@sigaev.ru, Postgres Professional, Moscow, Russia
• Alexander Korotkovaekorotkov@gmail.com, Postgres Professional, Moscow, Russia
• Oleg Bartunovoleg@sai.msu.su, Postgres Professional, Moscow, Russia

JsQuery is realized as an extension and not available in default PostgreSQLinstallation. It is available fromgithubunder the same license asPostgreSQLand supports PostgreSQL 9.4+.

Development is sponsored byWargaming.net.

JsQuery is PostgreSQL extension which requires PostgreSQL 9.4 or higher.Before build and install you should ensure following:

• PostgreSQL version is 9.4 or higher.
• You have development package of PostgreSQL installed or you builtPostgreSQL from source.
• You have flex and bison installed on your system. JsQuery was tested onflex 2.5.37-2.5.39, bison 2.7.12.
• Your PATH variable is configured so that pg_config command available, or set PG_CONFIG variable.

Typical installation procedure may look like this:

$ git clone https://github.com/postgrespro/jsquery.git$ cd jsquery$ make USE_PGXS=1$ sudo make USE_PGXS=1 install$ make USE_PGXS=1 installcheck$ psql DB -c "CREATE EXTENSION jsquery;"

JsQuery extension containsjsquery datatype which represents whole JSON queryas a single value (liketsquery does for fulltext search). The query is anexpression on JSON-document values.

Simple expression is specified aspath binary_operator value orpath unary_operator. See following examples.

• x = "abc" – value of key "x" is equal to "abc";
• $ @> [4, 5, "zzz"] – the JSON document is an array containing values4, 5 and "zzz";
• "abc xyz" >= 10 – value of key "abc xyz" is greater than or equal to 10;
• volume IS NUMERIC – type of key "volume" is numeric.
• $ = true – the whole JSON document is just a true.
• similar_ids.@# > 5 – similar_ids is an array or object of length greaterthan 5;
• similar_product_ids.# = "0684824396" – array "similar_product_ids"contains string "0684824396".
• *.color = "red" – there is object somewhere which key "color" has value"red".
• foo = * – key "foo" exists in object.

Path selects set of JSON values to be checked using given operators. Inthe simplest case path is just an key name. In general path is key names andplaceholders combined by dot signs. Path can use following placeholders:

• # – any index of array;
• % – any key of object;
• * – any sequence of array indexes and object keys;
• @# – length of array or object, could be only used as last component ofpath;
• $ – the whole JSON document as single value, could be only the whole path.

Expression is true when operator is true against at least one value selectedby path.

Key names could be given either with or without double quotes. Key nameswithout double quotes shouldn't contain spaces, start with number or concurwith jsquery keyword.

The supported binary operators are:

• Equality operator:=;
• Numeric comparison operators:>,>=,<,<=;
• Search in the list of scalar values usingIN operator;
• Array comparison operators:&& (overlap),@> (contains),<@ (contained in).

The supported unary operators are:

• Check for existence operator:= *;
• Check for type operators:IS ARRAY,IS NUMERIC,IS OBJECT,IS STRINGandIS BOOLEAN.

Expressions could be complex. Complex expression is a set of expressionscombined by logical operators (AND,OR,NOT) and grouped using braces.

Examples of complex expressions are given below.

• a = 1 AND (b = 2 OR c = 3) AND NOT d = 1
• x.% = true OR x.# = true

Prefix expressions are expressions given in the form path (subexpression).In this case path selects JSON values to be checked using given subexpression.Check results are aggregated in the same way as in simple expressions.

• #(a = 1 AND b = 2) – exists element of array which a key is 1 and b key is 2
• %($ >= 10 AND $ <= 20) – exists object key which values is between 10 and 20

Path also could contain following special placeholders with "every" semantics:

• #: – every indexes of array;
• %: – every key of object;
• *: – every sequence of array indexes and object keys.

Consider following example.

%.#:($ >= 0 AND $ <= 1)

This example could be read as following: there is at least one key which valueis array of numerics between 0 and 1.

We can rewrite this example in the following form with extra braces.

%(#:($ >= 0 AND $ <= 1))

The first placeholder% checks that expression in braces is true for at leastone value in object. The second placeholder#: checks value to be array andall its elements satisfy expressions in braces.

We can rewrite this example without#: placeholder as follows.

%(NOT #(NOT ($ >= 0 AND $ <= 1)) AND $ IS ARRAY)

In this example we transform assertion that every element of array satisfy somecondition to assertion that there is no one element which doesn't satisfy thesame condition.

Some examples of using paths are given below.

• numbers.#: IS NUMERIC – every element of "numbers" array is numeric.
• *:($ IS OBJECT OR $ IS BOOLEAN) – JSON is a structure of nested objectswith booleans as leaf values.
• #:.%:($ >= 0 AND $ <= 1) – each element of array is object containingonly numeric values between 0 and 1.
• documents.#:.% = * – "documents" is array of objects containing at leastone key.
• %.#: ($ IS STRING) – JSON object contains at least one array of strings.
• #.% = true – at least one array element is objects which contains at leastone "true" value.

Usage of path operators and braces need some explanation. When same pathoperators are used multiple times they may refer different values while you canrefer same value multiple time by using braces and$ operator. See followingexamples.

• # < 10 AND # > 20 – exists element less than 10 and exists another elementgreater than 20.
• #($ < 10 AND $ > 20) – exists element which both less than 10 and greaterthan 20 (impossible).
• #($ >= 10 AND $ <= 20) – exists element between 10 and 20.
• # >= 10 AND # <= 20 – exists element great or equal to 10 and existsanother element less or equal to 20. Query can be satisfied by array withno elements between 10 and 20, for instance [0,30].

Same rules apply when you search inside objects and branchy structures.

Type checking operators and "every" placeholders are useful for documentschema validation. JsQuery matchig operator@@ is immutable and can be usedin CHECK constraint. See following example.

CREATETABLEjs (    idserial,    data jsonb,CHECK (data @@'        name IS STRING AND        similar_ids.#: IS NUMERIC AND        points.#:(x IS NUMERIC AND y IS NUMERIC)'::jsquery));

In this example check constraint validates that in "data" jsonb column:value of "name" key is string, value of "similar_ids" key is array of numerics,value of "points" key is array of objects which contain numeric values in"x" and "y" keys.

See ourpgconf.eu presentationfor more examples.

JsQuery extension contains two operator classes (opclasses) for GIN whichprovide different kinds of query optimization.

• jsonb_path_value_ops
• jsonb_value_path_ops

In each of two GIN opclasses jsonb documents are decomposed into entries. Eachentry is associated with particular value and it's path. Difference betweenopclasses is in the entry representation, comparison and usage for searchoptimization.

For example, jsonb document{"a": [{"b": "xyz", "c": true}, 10], "d": {"e": [7, false]}}would be decomposed into following entries:

• "a".#."b"."xyz"
• "a".#."c".true
• "a".#.10
• "d"."e".#.7
• "d"."e".#.false

Since JsQuery doesn't support search in particular array index, we considerall array elements to be equivalent. Thus, each array element is marked withsame# sign in the path.

Major problem in the entries representation is its size. In the given examplekey "a" is presented three times. In the large branchy documents with longkeys size of naive entries representation becomes unreasonable. Both opclassesaddress this issue but in a slightly different way.

jsonb_path_value_ops represents entry as pair of path hash and value.Following pseudocode illustrates it.

(hash(path_item_1.path_item_2. ... .path_item_n); value)

In comparison of entries path hash is the higher part of entry and value isits lower part. This determines the features of this opclass. Since pathis hashed and it is higher part of entry we need to know the full path tothe value in order to use it for search. However, once path is specifiedwe can use both exact and range searches very efficiently.

jsonb_value_path_ops represents entry as pair of value and bloom filterof path.

(value; bloom(path_item_1) | bloom(path_item_2) | ... | bloom(path_item_n))

In comparison of entries value is the higher part of entry and bloom filter ofpath is its lower part. This determines the features of this opclass. Sincevalue is the higher part of entry we can perform only exact value searchefficiently. Range value search is possible as well but we would have tofilter all the the different paths where matching values occur. Bloom filterover path items allows index usage for conditions containing% and* intheir paths.

JsQuery opclasses perform complex query optimization. Thus it's valuable fordeveloper or administrator to see the result of such optimization.Unfortunately, opclasses aren't allowed to do any custom output to theEXPLAIN. That's why JsQuery provides following functions which allows to seehow particular opclass optimizes given query.

• gin_debug_query_path_value(jsquery) – for jsonb_path_value_ops
• gin_debug_query_value_path(jsquery) – for jsonb_value_path_ops

Result of these functions is a textual representation of query tree whichleafs are GIN search entries. Following examples show different results ofquery optimization by different opclasses.

# SELECT gin_debug_query_path_value('x = 1 AND (*.y = 1 OR y = 2)'); gin_debug_query_path_value---------------------------- x = 1 , entry 0           +# SELECT gin_debug_query_value_path('x = 1 AND (*.y = 1 OR y = 2)'); gin_debug_query_value_path---------------------------- AND                       +   x = 1 , entry 0         +   OR                      +     *.y = 1 , entry 1     +     y = 2 , entry 2       +

Unfortunately, jsonb have no statistics yet. That's why JsQuery optimizer hasto do imperative decision while selecting conditions to be evaluated usingindex. This decision is made by assumtion that some condition types are lessselective than others. Optimizer divides conditions into following selectivityclass (listed by descending of selectivity).

1. Equality (x = c)
2. Range (c1 < x < c2)
3. Inequality (x > c)
4. Is (x is type)
5. Any (x = *)

Optimizer evades index evaluation of less selective conditions when possible.For example, in thex = 1 AND y > 0 queryx = 1 is assumed to be moreselective thany > 0. That's why index isn't used for evaluation ofy > 0.

# SELECT gin_debug_query_path_value('x = 1 AND y > 0'); gin_debug_query_path_value---------------------------- x = 1 , entry 0           +

With lack of statistics decisions made by optimizer can be inaccurate. That'swhy JsQuery supports hints. Comments/*-- index */ and/*-- noindex */placed in the conditions forces optimizer to use and not use indexcorrespondingly.

SELECT gin_debug_query_path_value('x = 1 AND y /*-- index */ > 0'); gin_debug_query_path_value---------------------------- AND                       +   x = 1 , entry 0         +   y > 0 , entry 1         +SELECT gin_debug_query_path_value('x /*-- noindex */ = 1 AND y > 0'); gin_debug_query_path_value ----------------------------  y > 0 , entry 0           +

Please, notice, that JsQuery is still under development and while it'sstable and tested, it may contains some bugs. Don't hesitate to raiseissues at github with yourbug reports.

If you're lacking of some functionality in JsQuery and feeling power toimplement it then you're welcome to make pull requests.