How do I keep FOR JSON PATH from escaping query results?

Question

I\'m trying to write a fairly complicated SQL Query that produces JSON as the result. All is working great except for some hardcoded arrays I need to have deeper in the hier

Answer 1

I have tested peformance of both solutions:

first - via JSON_QUERY:

declare @i int = 1;
SELECT
    JSON_QUERY(
        CASE when @i = 1 THEN
        (
            SELECT * FROM
            (
                select textCol AS Stuff from table1 where id % 2 = 0 
                UNION ALL
                SELECT textCol AS Stuff from table1 where id % 2 <> 0
            ) AS SubSelect
            FOR JSON PATH
        )
        ELSE
        (
            SELECT textCol AS Stuff from table1
            FOR JSON PATH
        )
        END
    ) AS WhyItMatters
FOR JSON path

gives me average execution time 91ms.

Second:

declare @i int = 1;
SELECT
    (SELECT * FROM
        (
            select textCol AS Stuff from table1 where id % 2 = 0 and @i = 1
            UNION ALL
            SELECT textCol AS Stuff from table1 where id % 2 <> 0 and @i = 1
            union all
            SELECT textCol AS Stuff from table1 where @i <> 1
        ) AS SubSelect
        FOR JSON PATH
    ) AS WhyItMatters
FOR JSON path

gives me average execution time 45ms.

table1 contains 12727 rows. Resulting JSON has length about 1500000 characters.

Answer 2

There's some fascinating behavior going on in the optimizer for this query, and I'm not sure if it's a bug. The following query will not add escaping:

SELECT
    'Hi' AS Greeting,
    (
        CASE WHEN 1 = 1 THEN (
            SELECT * FROM (
                SELECT 'qwerty' AS [Stuff]
                UNION ALL
                SELECT 'zxcvb' AS [Stuff]
            ) _
            FOR JSON PATH
        ) ELSE (
            SELECT 'asdf' AS [Stuff]
            FOR JSON PATH
        )
        END
    ) AS WhyItMatters
FOR JSON PATH

The CASE can be optimized away, and it is optimized away, and the end result is nicely nested JSON. But if we remove the ability to optimize things away, it degenerates into pasting in an escaped string:

SELECT
    'Hi' AS Greeting,
    (
        CASE WHEN RAND() = 1 THEN (
            SELECT * FROM (
                SELECT 'qwerty' AS [Stuff]
                UNION ALL
                SELECT 'zxcvb' AS [Stuff]
            ) _
            FOR JSON PATH
        ) ELSE (
            SELECT 'asdf' AS [Stuff]
            FOR JSON PATH
        )
        END
    ) AS WhyItMatters
FOR JSON PATH

It seems illogical that one query would result in processing typed JSON and the other would not, but there you go. JSON is not an actual type in T-SQL (unlike XML), so we can't CAST or CONVERT, but JSON_QUERY will do roughly the same thing:

SELECT
    'Hi' AS Greeting,
    JSON_QUERY(
        CASE WHEN RAND() = 1 THEN (
            SELECT * FROM (
                SELECT 'qwerty' AS [Stuff]
                UNION ALL
                SELECT 'zxcvb' AS [Stuff]
            ) _
            FOR JSON PATH
        ) ELSE (
            SELECT 'asdf' AS [Stuff]
            FOR JSON PATH
        )
        END
    ) AS WhyItMatters
FOR JSON PATH

Note that this also works if the argument is already JSON (in the constant case), so it's safe to add regardless.

Answer 3

I have found one possible solution but I really don't like it. I'm posting what I have in hopes that somebody has a better solution.

Using a WHERE statement on every branch of my UNION with either the affirmative or exact negative of my CASE statement can prevent the "strigifying" of my results.

For example, this query:

SELECT
    'Hi' AS Greeting,
    (
        SELECT * FROM
        (
            SELECT 'asdf' AS Stuff WHERE DatePart(second, GetDate()) % 2 = 0
            UNION ALL
            SELECT 'qwerty' AS Stuff WHERE DatePart(second, GetDate()) % 2 = 1
            UNION ALL
            SELECT 'zxcvb' AS Stuff WHERE DatePart(second, GetDate()) % 2 = 1
        ) AS SubSelect
        FOR JSON PATH
    ) AS Try1
FOR JSON PATH

provides these results:

[
    {
        "Greeting": "Hi",
        "Try1": [
            {
                "Stuff": "qwerty"
            },
            {
                "Stuff": "zxcvb"
            }
        ]
    }
]

If nothing better can be found, I can move forward with this. But this seems like a hacky way to control this.