The nested
function when used in the WHERE
clause of an SQL statement filters documents of nested
object type based on field properties. Two syntax options are available to users, see section 1.1 Syntax. Both syntax options achieve the goal of filtering a nested
field with an Operator
and a Literal
value. Although semantically the syntax options can produce the same query, they will produce different DSL to push to OpenSearch. This difference is based on the AND/OR/NOT
boolean queries residing inside a nested
query or vice versa. See section 2.1 and 2.2 for examples depicting this alternate DSL creation. Testing large groups of push down execution times for identical queries using the different syntax options showed a nominal difference of approximately 3% for Syntax option 2.
- Overview
- Syntax
- Changes To Core
- Example Data And Usage
- Example Queries
- Architecture Diagrams
- Syntax Option 1 Object Tree to DSL
- Syntax Option 2 Object Tree to DSL
- Additional Info
- Release Schedule
The nested function has two syntax options when used in the WHERE
clause of an SQL statement. Both syntax options can produce the same OpenSearch response, but will form different DSL used in the query.
nested(path, condition)
nested(field | field, path) OPERATOR LITERAL
- FilterQueryBuilder: Added logic to handle
nested
functions inWHERE
clause as predicate expression andFunctionExpression
. - LuceneQuery: Added logic to handle
nested
functions andFunctionExpression
's on both sides ofOPERATOR
inWHERE
clause.
A basic example from data mapping to response from SQL plugin.
Mapping:
{
"mappings": {
"properties": {
"message": {
"type": "nested",
"properties": {
"info": {
"type": "keyword",
"index": "true"
}
}
}
}
}
}
Dataset:
{"index":{"_id":"1"}}
{"message":{"info":"a"}}
{"index":{"_id":"2"}}
{"message":{"info":"c"}}
Query:
SELECT * FROM nested_objects WHERE nested(message, message.info = 'a');
SELECT * FROM nested_objects WHERE nested(message.info) = 'a';
Both queries produce same response from OpenSearch.
Response:
{
"schema": [
{
"name": "message",
"type": "nested"
}
],
"datarows": [
[
[
{
"info": "a"
}
]
]
],
"total": 1,
"size": 1,
"status": 200
}
A basic nested function in the SELECT
clause. This example filters the nested
object message
and the inner field info
with the value of a
.
SELECT * FROM nested_objects WHERE nested(message.info) = 'a' OR nested(message.author) = 'elm';
Using the same query as above but with syntax option 1, the response from OpenSearch will be the same, but the DSL used in the query will differ.
SELECT * FROM nested_objects WHERE nested(message, message.info = 'a' OR messate.author = 'elm');
This example has a nested function in the SELECT
clause and the WHERE
clause of the query. Using the nested
function in the SELECT
clause will flatten the response nested objects from OpenSearch. See Nested In Select for more details on flattening.
SELECT nested(message.info) FROM nested_objects WHERE nested(message.info) = 'a';
When using syntax option 1 we need separate function calls when the path
is not identical. This query exemplifies how a user can use multiple queries with differing path
values.
SELECT * FROM nested_objects WHERE nested(message, message.info = 'a') OR nested(comment, comment.data = '123');
The following example illustrates the object tree that is built from the example query. This query's object tree will impact the structure of the DSL to be pushed to OpenSearch.
Example Query: SELECT * FROM nested_objects WHERE nested(message, message.info = 'a' OR message.info = 'b' AND message.dayOfWeek > 4);
graph TB;
A[Function: NESTED\n<hr>arguments]-->B1[ReferenceExpression:\nmessage]
A-->B2[Function: OR\n<hr>arguments]
B2-->C1[Function: =\n<hr>arguments]
B2-->C2[Function: AND\n<hr>arguments]
C1-->D1[ReferenceExpression:\nmessage.info]
C1-->D2[LiteralExpression:\na]
C2-->D3[Function: =\n<hr>arguments]
C2-->D4[Function: >\n<hr>arguments]
D3-->E1[ReferenceExpression:\nmessage.info]
D3-->E2[LiteralExpression:\nb]
D4-->E3[ReferenceExpression:\nmessage.dayOfWeek]
D4-->E4[LiteralExpression:\n4]
This diagram illustrates the steps the SQL plugin does to translate the object tree to DSL for execution in OpenSearch. Notice that the nested
function forms the outer portion of the DSL query.
stateDiagram-v2
direction LR
NestedNode --> OrFuncNode
OrFuncNode --> OrFuncNode1
OrFuncNode --> OrFuncNode2
OrFuncNode2 --> AndFuncNode1
OrFuncNode2 --> AndFuncNode2
state "Function: NESTED" as NestedNode {
state "visitFunction(NESTED)" as NestedNodeVisitFunc
state "'nested': {\n 'query': {\n...\n}}" as NestedNodeDSL
state "visitFunction(condition)" as NestedNodeVisitCondition
NestedNodeVisitFunc --> NestedNodeDSL
NestedNodeVisitFunc --> NestedNodeVisitCondition
}
state "Function: OR" as OrFuncNode {
state "visitFunction(OR)" as OrFuncNodeVisitFunc
state "'bool': {\n 'should': {\n...\n}}" as OrFuncNodeDSL
state "arguments.accept()" as OrFuncNodeArgs
OrFuncNodeVisitFunc --> OrFuncNodeDSL
OrFuncNodeDSL --> OrFuncNodeArgs
}
state "Function: =" as OrFuncNode1 {
state "visitFunction(=)" as OrFuncNode1VisitFunc
state "'term': {\n 'message.info': {\n  'value':'a'\n}}" as OrFuncNode1DSL
OrFuncNode1VisitFunc --> OrFuncNode1DSL
}
state "Function: AND" as OrFuncNode2 {
state "visitFunction(AND)" as OrFuncNode2VisitFunc
state "'bool': {\n 'filter': {\n...\n}}" as OrFuncNode2DSL
state "arguments.accept()" as OrFuncNode2Args
OrFuncNode2VisitFunc --> OrFuncNode2DSL
OrFuncNode2DSL --> OrFuncNode2Args
}
state "Function: =" as AndFuncNode1 {
state "visitFunction(=)" as AndFuncNode1VisitFunc
state "'term': {\n 'message.info': {\n  'value':'b'\n}}" as AndFuncNode1DSL
AndFuncNode1VisitFunc --> AndFuncNode1DSL
}
state "Function: >" as AndFuncNode2 {
state "visitFunction(>)" as AndFuncNode2VisitFunc
state "'range': {\n 'message.dayOfWeek': {\n  'from':4\n}}" as AndFuncNode2DSL
AndFuncNode2VisitFunc --> AndFuncNode2DSL
}
After pushing down the filter object tree we have a DSL query to push to OpenSearch. This query forms the boolean logic inside the nested
query. This mirrors the syntax option chosen which has the condition specified inside the nested
function.
{
"nested" : {
"query" : {
"bool" : {
"should" : [
{
"term" : {
"message.info" : {
"value" : "a",
"boost" : 1.0
}
}
},
{
"bool" : {
"filter" : [
{
"term" : {
"message.info" : {
"value" : "b",
"boost" : 1.0
}
}
},
{
"range" : {
"message.dayOfWeek" : {
"from" : 4,
...
}
}
}
]
}
}
]
}
}
}
}
The following example illustrates the object tree that is built from the example query. Rather than have the boolean logic specified within the nested
function, we have the nested
function used in a predicate expression.
Example Query: SELECT * FROM nested_objects WHERE nested(message.info) = 'a' OR nested(message.info) = 'b' AND nested(message.dayOfWeek) > 4;
graph TB;
A[Function: OR\n<hr>arguments]-->B1[Function: =\n<hr>arguments]
A-->B2[Function: AND\n<hr>arguments]
B1-->C1[Function: NESTED\n<hr>arguments]
B1-->C2[LiteralExpression:\na]
B2-->C3[Function: =\n<hr>arguments]
B2-->C4[Function: >\n<hr>arguments]
C1-->D1[ReferenceExpression:\nmessage.info]
C3-->D2[Function: NESTED\n<hr>arguments]
C3-->D3[LiteralExpression:\nb]
C4-->D4[Function: NESTED\n<hr>arguments]
C4-->D5[LiteralExpression:\n4]
D2-->E1[ReferenceExpression:\nmessage.info]
D4-->E2[ReferenceExpression:\nmessage.dayOfWeek]
This diagram illustrates the steps the SQL plugin goes through to translate the object tree to DSL for execution in OpenSearch. Notice that the boolean operations(should/filter
) form the outer potion of the DSL query.
stateDiagram-v2
direction LR
OrFuncNode --> OrFuncNode1
OrFuncNode --> OrFuncNode2
OrFuncNode2 --> AndFuncNode1
OrFuncNode2 --> AndFuncNode2
state "Function: OR" as OrFuncNode {
state "visitFunction(OR)" as NestedNodeVisitFunc
state "'bool': {\n 'should': {\n...\n}}" as NestedNodeDSL
state "arguments.accept()" as NestedNodeVisitCondition
NestedNodeVisitFunc --> NestedNodeDSL
NestedNodeVisitFunc --> NestedNodeVisitCondition
}
state "Function: =" as OrFuncNode1 {
state "visitFunction(=)" as OrFuncNode1VisitFunc
state "'nested': {\n 'query': {\n   'term'\n    'message.info': {\n     'value': 'a',\n      }}}" as OrFuncNode1DSL
OrFuncNode1VisitFunc --> OrFuncNode1DSL
}
state "Function: AND" as OrFuncNode2 {
state "visitFunction(AND)" as OrFuncNode2VisitFunc
state "'bool': {\n 'filter': {\n...\n}}" as OrFuncNode2DSL
state "arguments.accept()" as OrFuncNode2Args
OrFuncNode2VisitFunc --> OrFuncNode2DSL
OrFuncNode2DSL --> OrFuncNode2Args
}
state "Function: =" as AndFuncNode1 {
state "visitFunction(=)" as AndFuncNode1VisitFunc
state "'nested': {\n 'query': {\n   'term'\n    'message.info': {\n     'value': 'b',\n      }}}" as AndFuncNode1DSL
AndFuncNode1VisitFunc --> AndFuncNode1DSL
}
state "Function: >" as AndFuncNode2 {
state "visitFunction(>)" as AndFuncNode2VisitFunc
state "'nested': {\n 'query': {\n   'range'\n    'message.dayOfWeek': {\n     'from': 4,\n      }}}" as AndFuncNode2DSL
AndFuncNode2VisitFunc --> AndFuncNode2DSL
}
After pushing down the filter object tree we have a DSL query to push to OpenSearch. This query forms the boolean logic as the outer portion of the query and has nested queries used inside. This mirrors the syntax option chosen which has the boolean operators forming a predicate expression with the nested
functions in the SQL query.
{
"bool" : {
"should" : [
{
"nested" : {
"query" : {
"term" : {
"message.info" : {
"value" : "a",
"boost" : 1.0
}
}
},
...
}
},
{
"bool" : {
"filter" : [
{
"nested" : {
"query" : {
"term" : {
"message.info" : {
"value" : "b",
"boost" : 1.0
}
}
},
...
}
},
{
"nested" : {
"query" : {
"range" : {
"message.dayOfWeek" : {
"from" : 4,
...
}
}
},
...
}
}
]
}
}
]
}
}
TBD
A normalization strategy could produce identical DSL independent on the syntax option used in the query. Testing so far hasn't highlighted any major improvements with query execution that necessitate identical DSL. Additional benchmarking and planning will be done to determine a viable normalization strategy and if implementation is required.
See Issues Tracked under Issue 1111 for related PR's and information.