Create nested JSON arrays using FOR JSON PATH

With deeply nested arrays the subqueries get unmanageable quickly:

select id,foo, (select id, bar, (select ... for json path) things, 
(select...) more_things) yet_more, select(...) blarg

I create a relational (non-json) view that joins all my tables and has the json structure embedded in the column aliases, just like for json path does. But I also have [] to indicate that the json node is an array. Like this:

select p.id [id], p.foo [foo], c.name [children[].name], c.id [children[].id],
gp.name [grandparent.name], gc.name [children[].grandchildren[].name]
from parent p
join children c on c.parent_id = p.id .....

I wrote a stored procedure that creates a json view into the non-json view that parses the column names of the relational view and makes the json pretty. See below. Call it with the name of your relational view and it creates a view. It's not thoroughly tested but it works for me. Only caveat is that tables need to have id columns called id. It uses string_agg() and json_array() to the version of sql needs to be pretty new. It's also set up to return an array in the root. It will need tweaking to return an object.


create procedure create_json_from_view
@view_name varchar(max)
as

create table #doc_schema (
    node_level int,             -- nesting level starting with 0
    node_name varchar(max),     -- alias used for this nodes query
    node_path varchar(max),     -- full path to this node
    parent_path varchar(max),   -- full path to it's parents 
    is_array bit,               -- is node marked as array by ending with []
    select_columns varchar(max),-- comma separated path/alias pairs for selected columns on node
    group_by_columns varchar(max), -- comma separated paths for selected columns on node. group by is necessary to prevent duplicates
    node_parent_id varchar(max),   -- the id column path to join subquery to parent. NOTE: ID COLUMN MUST BE CALLED ID
    from_clause varchar(max),   -- from clause built from above fields
    node_query varchar(max)     -- complete query built from above fields
)

/* get each node path from view schema
*/
INSERT INTO #doc_schema (node_path)
select distinct LEFT(COLUMN_NAME,CHARINDEX('.'+ VALUE + '.',COLUMN_NAME) + LEN(VALUE)) node_path 
FROM INFORMATION_SCHEMA.COLUMNS 
CROSS APPLY STRING_SPLIT(COLUMN_NAME, '.') 
WHERE CHARINDEX('.',COLUMN_NAME) > 0
AND RIGHT(COLUMN_NAME,LEN(VALUE)) <> VALUE
and table_name = @view_name

/* node_name past rightmost period or the same as node_path if there is no period
also remove [] from arrays
*/
update #doc_schema set node_name = 
case when charindex('.',node_path) = 0 then replace(node_path,'[]','')
else REPLACE(right(node_path,charindex('.',reverse(node_path)) - 1),'[]','') end

/* if path ends with [] node is array
    escapes are necessary because [] have meaning for like
*/
update #doc_schema set is_array =
case when node_path like '%\[\]' escape '\' then 1 else 0 end --\

/* parent path is everything before last . in node path
    except when the parent is the root, in which case parent is empty string
*/
update #doc_schema set parent_path = 
case when charindex('.',node_path) = 0 then ''
else left(node_path,len(node_path) - charindex('.',reverse(node_path))) end

/* level is how many . in path. an ugly way to count.
*/
update #doc_schema set node_level = len(node_path) - len(replace(node_path,'.','')) + 1

/* set up root node
*/
insert into #doc_schema (node_path,node_name,parent_path,node_level,is_array)
select '','',null,0,1

/* I'm sorry this is so ugly. I just gave up on explaining
    all paths need to be wrapped in [] and internal ] need to be escaped as ]]
*/
update #doc_schema set select_columns = sub2.select_columns, group_by_columns = sub2.group_by_columns
from (
    select node_path,string_agg(column_path + ' ' + column_name,',') select_columns,
    string_agg(column_path,',') group_by_columns
    from (
        select ds.node_path,'['+replace(c.COLUMN_NAME,']',']]')+']' column_path,replace(c.column_name,ds.node_path + '.','') column_name
        from INFORMATION_SCHEMA.COLUMNS c
        join #doc_schema ds
        on (charindex(ds.node_path + '.', c.COLUMN_NAME) = 1
        and charindex('.',replace(c.COLUMN_NAME,ds.node_path + '.','')) = 0)
        or (ds.node_level = 0 and charindex('.',c.COLUMN_NAME) =  0)
        where table_name = @view_name
    ) sub
    group by node_path
) sub2
where #doc_schema.node_path = sub2.node_path

/* id paths for joining subqueries to parents
    Again, the need to be wrapped in [] and and internal ] need to be escaped as ]]
*/
update #doc_schema set node_parent_id  = 
    case when parent_path = '' then '[id]' 
    else '[' + replace(parent_path,']',']]')+'.id]'
    end

/* table aliases for joining subqueries to parents need to be unique
    just use L0 L1 etc based on nesting level
*/
update #doc_schema set from_clause =
    case when node_level = 0 then ' from ' + @view_name + ' L'+cast(node_level as varchar(4)) + ' '
    else ' from ' + @view_name + ' L'+cast(node_level as varchar(4))+' where L'+cast(node_level - 1 as varchar(4))+'.'+ node_parent_id + 
        '  = L'+cast(node_level as varchar(4))+'.'+ node_parent_id 
    end

/* Assemble node query from all parts
    ###subqueries### is a place to put subqueries for node
*/
update #doc_schema set  node_query = 
        ' (select ' + select_columns + ', ###subqueries###' + from_clause 
        + ' group by '+ group_by_columns
        +' for json path) '

/* json path will treat all objects as arrays so select first explicitly
    to prevent [] in json
*/  
update #doc_schema set  node_query =    
    case when is_array = 0
    then '(select JSON_query(' + node_query + ',''$[0]'')) ' + node_name
    else node_query +  + node_name end

/* starting with highest nesting level substitute child subqueries ino
    subquery hold in their parents
*/
declare @counter int = (select max(node_level) from #doc_schema)

while(@counter >= 0)
begin
    update #doc_schema set node_query = replace(node_query,'###subqueries###', subs.subqueries)
    from
    (select parent_path, string_agg(node_query,',') subqueries, node_level from #doc_schema
    group by parent_path, node_level ) subs
    where subs.node_level = @counter and 
    #doc_schema.node_path = subs.parent_path

    set @counter -= 1
end

/* objects and arrays with no subobjects or subarrays still have subquery holder so remove them
*/
update #doc_schema set node_query = replace(node_query,', ###subqueries###', '') where node_level = 0

declare @query nvarchar(max) = (select node_query from #doc_schema where node_level = 0)

/* add wrapper to query to specify column nave otherwise create view will fail
*/
set @query = 
    case when OBJECT_ID(@view_name + '_JSON', 'V') is NULL then 'create' else 'alter' end +
    ' view ' + @view_name + '_json as select' + @query + ' json'

exec sp_executesql @query

You can use the following query:

SELECT pr.person_id AS [person.id], pr.person_name AS [person.name],
    (
        SELECT pt.pet_id AS id, pt.pet_name AS name 
        FROM @Pets pt WHERE pt.pet_owner=pr.person_id 
        FOR JSON PATH
    ) AS [person.pet]
FROM @Persons pr 
FOR JSON PATH, ROOT('pet owners')

For more information, see https://blogs.msdn.microsoft.com/sqlserverstorageengine/2015/10/09/returning-child-rows-formatted-as-json-in-sql-server-queries/