Hi,
On 24.08.2022 01:12, Andrew Dunstan wrote:
On 2022-08-23 Tu 17:29, Nikita Glukhov wrote:
Here is my plan:
0. Take my last v7-0001 patch as a base. It already contains refactoring
of JsonCoercion code. (Fix 0002 is not needed anymore, because it is for
json[b] domains, which simply will not be supported.)
1. Replace JSON_COERCION_VIA_EXPR in JsonCoercion with new
JsonCoercionType(s) for hardcoded coercions.
JsonCoerion node were completely removed because they are not needed
anymore (see p. 4).
2. Disable all non-JSON-compatible output types in coerceJsonFuncExpr().
3. Add missing safe type input functions for integers, numerics, and
maybe others.
We need to write much more functions, than I expected. And I still didn't
implemented safe input functions for numeric and datetime types.
I will start to do it tomorrow.
4. Implement hardcoded coercions using these functions in
ExecEvalJsonExprCoercion().
That was done using simple `switch (returning_typid) { .. }`,
which can be nested into `switch (jbv->type)`.
5. Try to allow only constants (and also maybe column/parameter
references) in JSON_VALUE's DEFAULT expressions. This should be enough
for the most of practical cases. JSON_QUERY even does not have DEFAULT
expressions -- it has only EMPTY ARRAY and EMPTY OBJECT, which can be
treated as simple JSON constants.
I have not tried to implement this yet.
er, really? This is from the regression output:
SELECT JSON_QUERY(jsonb '[]', '$[*]' DEFAULT '"empty"' ON EMPTY); json_query
------------ "empty"
(1 row)
SELECT JSON_QUERY(jsonb '[1,2]', '$[*]' DEFAULT '"empty"' ON ERROR); json_query
------------ "empty"
(1 row)
This is another extension. SQL standard defines only
EMPTY ARRAY an EMPTY OBJECT behavior for JSON_QUERY:
<JSON query empty behavior> ::=
ERROR
| NULL
| EMPTY ARRAY
| EMPTY OBJECT
<JSON query error behavior> ::=
ERROR
| NULL
| EMPTY ARRAY
| EMPTY OBJECT
But it is possible to allow all other expressions in ERROR ON ERROR
case, and I don't know if it will be consistent enough to allow some
expressions in one case and deny in other.
And there is another problem: expressions can be only checked for
Const-ness only after expression simplification. AFAIU, at the
parsing stage they look like 'string'::type. So, it's unclear if it
is correct to check expressions in ExecInitExpr().
6. Remove subtransactions.
They were completely removed. Only DEFAULT expression needs to be fixed now.
Sounds like a good plan, modulo the issues in item 5. I would rather
lose some features temporarily than try to turn handsprings to make them
work and jeopardize the rest.
I'll look forward to seeing your patch in the morning :-)
v8 - is a highly WIP patch, which I failed to finish today.
Even some test cases fail now, and they simply show unfinished
things like casts to bytea (they can be simply removed) and missing
safe input functions.
--
Nikita Glukhov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company
