Re: [HACKERS] SQL/JSON in PostgreSQL

On 29.11.2017 05:24, Michael Paquier wrote:

> On Wed, Nov 15, 2017 at 10:17 AM, Nikita Glukhov
> <n(dot)gluhov(at)postgrespro(dot)ru> wrote:
>> Attached the new version of the patches where displaying of SQL/JSON
>> constructor nodes was fixed. I decided not to invent new nodes but to
>> extend
>> existing FuncExpr, Aggref, WindowFunc nodes with new formatting fields that
>> give us ability to override default displaying in ruleutils.c. Also new
>> invisible CoercionForm was added for hiding casts in which FORMAT and
>> RETUNING
>> clauses are transformed.
> Okay, I can see that the patch is still in the same situation two
> weeks after I looked at it first, so I am marking it as returned with
> feedback. This needs to be broken down, and get documentation. At this
> point getting a review out of this patch is not something I'd
> recommend until it is put in a shape that makes it easier. Please also
> help in reviewing other's patches, yours here is very large.

Attached new version of patches:

1. Added some comments for the jsonpath code and some documentation for
jsonpath operators and opclasses.  Sorry that complete documentation for
jsonpath itself is still missing.

2. Added support for custom user-defined item methods and functions in
jsonpath.
This feature allows us to move our non-standard methods (map, reduce,
fold, min,
max) to the extension contrib/jsnopathx. Now user can implement all standard
JavaScript array methods by himself.

3. Added variable jsonpath specifications in SQL/JSON functions.

4. Added subtransactions inside PG_TRY/PG_CATCH block that is used for
ON ERROR
clause support in JSON_EXISTS, JSON_VALUE, JSON_QUERY (see ExecEvalJson() in
src/backend/executor/execExpeInterp.c).

The last addition is the most questionable.  By using standard ON ERROR
сlause
we can specify default value for expression when an error occurs during
casting
JSON data to the target SQL type.  Cast expressions can contain
arbitrary user
functions, so they should be executed inside own subtransaction like it
is done
in plpgsql (see exec_stmt_block()).  Subtransaction start obviously
introduces
significant performance overhead (except the case when ERROR ON ERROR is
used and error handling is unnecessary):

=# EXPLAIN (ANALYZE)
   SELECT JSON_VALUE(jsonb '1', '$' RETURNING int ERROR ON ERROR)
   FROM generate_series(1, 1000000);
 ...
 Execution time: 395.238 ms

=# EXPLAIN (ANALYZE)
   SELECT JSON_VALUE(jsonb '1', '$' RETURNING int)
   FROM generate_series(1, 1000000);
 ...
 Execution time: 914.850 ms

To partially eliminate this overhead, I'm trying to examine cast-expression
volatility:
 * mutable -- subtransaction is started
 * stable -- subtransaction is not started, only new ResourceOwner is
created
 * immutable -- ResourceOwner is not even created
But don't know if it is correct to rely on volatility here.  And also I
doubt
that we can start multiple subtransactions (for each SQL/JSON function
 evaluation) within a single SELECT statement.

--
Nikita Glukhov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company