Re: Ambiguity in IS JSON description and logic - Mailing list pgsql-docs

On Thu, Mar 20, 2025 at 7:08 AM Kirk Parker <khp@equatoria.us> wrote:

On Thu, Mar 20, 2025 at 2:46 AM PG Doc comments form <noreply@postgresql.org> wrote:

The following documentation comment has been logged on the website:

Page: https://www.postgresql.org/docs/17/functions-json.html
Description:

On the manual page
https://www.postgresql.org/docs/current/functions-json.html, in the Table
9.48. "SQL/JSON Testing Functions" there is a description of IS JSON. It
includes the next sentence: "If WITH UNIQUE KEYS is specified, then any
object in the expression is also tested to see if it has duplicate keys."
And such text is ambiguous, because the term "object" has certain meaning
regarding json format. In reality the option WITH UNIQUE KEYS allows to
check for duplicated keys any array element not object. For objects, both
WITH UNIQUE KEYS and WITHOUT UNIQUE KEYS return false, and both IS JSON
ARRAY WITH UNIQUE KEY and IS JSON ARRAY WITHOUT UNIQUE KEY return true (it
is at the same time with and without unique values, how it is possible?),
i.e. it works the same as just IS JSON ARRAY. The example code that confirms
my reasoning: 
SELECT
    js.vl AS                                      "tested str",             

    js.vl IS JSON OBJECT WITH UNIQUE KEYS AS      ".. object w. UQ keys",   

    js.vl IS JSON OBJECT WITHOUT UNIQUE KEYS AS   ".. object w/o UQ keys",
    js.vl IS JSON ARRAY WITH UNIQUE KEYS AS       ".. array w. UQ keys",   

    js.vl IS JSON ARRAY WITHOUT UNIQUE KEYS AS    ".. array w/o UQ keys",
    js.vl IS JSON ARRAY ".. array"
FROM (VALUES ('{{"a": "a1"}, {"a": "a2"}}'), ('[{"a": "a1"}, {"a": "a2"}]'),
('["a", "a"]')) AS js(vl);

I'm not sure what should be the right logic for this option, for me it looks
now the same as simple IS JSON ARRAY without any UNIQUE KEY option, but if
we use an option it should be either true for WITH UNIQUE KEYS or WITHOUT
UNIQUE KEYS but not for both at the same time. But anyway the sentence I
showed above should contain "array" instead of "object" because for objects
it returns false independently of applied option.  I tested it on
"PostgreSQL 17.0 on x86_64-windows, compiled by msvc-19.41.34120, 64-bit".

First, WITHOUT UNIQUE KEYS does not mean "confirm that there are duplicate keys", it's just a way of stating the default explicitly. In other words it means "w/o testing for duplicate keys". Thus IS JSON OBJECT and IS JSON OBJECT WITHOUT UNIQUE KEYS will both always return identical results on the same JSON expression. 

Secondly, the UNIQUE test is recursive; for objects maybe the meaning is intuitive, but for JSON arrays -- which don't have any concept of keys; JSON arrays are just ordered lists -- it means "does this array contain any embedded objects with duplicate keys".

See:

 SELECT js,

  js IS JSON "json?",

  js IS JSON OBJECT "object?",

  js IS JSON OBJECT WITH UNIQUE KEYS "object w. UK?",

  js IS JSON OBJECT WITHOUT UNIQUE KEYS "object w/o UK?",

  js IS JSON ARRAY "array?",

  js IS JSON ARRAY WITH UNIQUE KEYS "array w. UK?",

  js IS JSON ARRAY WITHOUT UNIQUE KEYS "array w/o UK?"

FROM (VALUES

  ('[{"a":1},{"b":2,"b":3}]'),       -- expect t for array, array w/o UK

  ('[{"a":1},{"b":2,"c":3}]'),       -- expect t for ALL array tests

  ('{"b":2,"b":3}'),                 -- expect t for object, object w/o UK

  ('{"c":2,"d":3}'),                 -- expect t for ALL object tests

  ('{"c":2,"d":{ "e": 0, "e": 1}}'), -- WITH UNIQUE is recursive for nested objects

  ('{"c":2,"d":{ "e": 0, "f": {"g":1,"g":2}}}'), -- no matter how deep

  ('[{"a":1},{"b":2,"c":{"d":1, "d":2}}]') -- and also tests arrays recursively for embedded objecs

) foo(js);

A couple of side notes:

1. Your first data example is not JSON at all.  It's helpful for this kind of test to include a plain IS JSON column, since any of the IS JSON X tests can fail for two reasons: (a) it's not JSON, or (b) it is JSON but it's not an X.

2. Curiously, the JSON spec itself is completely silent on the meaning of objects with duplicate keys.  PostgreSQL is more helpful in this regard--the docs explicitly state that the last value is the one that is retained by JSONB and used in processing functions.

 

To improve the documentation here, I would suggest simply adding the word "recursively" after "tested":

    If WITH UNIQUE KEYS is specified, then any object in the expression is also tested recursively to see if it has duplicate keys