Re: jsonpath - Mailing list pgsql-hackers

From Nikita Glukhov
Subject Re: jsonpath
Date
Msg-id c19ee9e4-e832-efc3-d6cf-3cacc6462315@postgrespro.ru
Whole thread Raw
In response to Re: jsonpath  (Alexander Korotkov <a.korotkov@postgrespro.ru>)
Responses Re: jsonpath  (Nikita Glukhov <n.gluhov@postgrespro.ru>)
List pgsql-hackers

On 02.03.2018 00:57, Alexander Korotkov wrote:

On Fri, Mar 2, 2018 at 12:40 AM, Nikita Glukhov <n.gluhov@postgrespro.ru> wrote:
On 28.02.2018 06:55, Robert Haas wrote:

On Mon, Feb 26, 2018 at 10:34 AM, Nikita Glukhov
<n.gluhov@postgrespro.ru> wrote:
Attached 10th version of the jsonpath patches.

1. Fixed error handling in arithmetic operators.

    Now run-time errors in arithmetic operators are catched (added
    PG_TRY/PG_CATCH around operator's functions calls) and converted into
    Unknown values in predicates as it is required by the standard:
I think we really need to rename PG_TRY and PG_CATCH or rethink this
whole interface so that people stop thinking they can use it to
prevent errors from being thrown.

I understand that it is unsafe to call arbitrary function inside PG_TRY without
rethrowing of caught errors in PG_CATCH, but in jsonpath only the following
numeric and datetime functions with known behavior are called inside PG_TRY
and only errors of category ERRCODE_DATA_EXCEPTION are caught:

numeric_add()
numeric_mul()
numeric_div()
numeric_mod()
numeric_float8()
float8in()
float8_numeric()
to_datetime()

That seems like a quite limited list of functions.  What about reworking them
providing a way of calling them without risk of exception?  For example, we can
have numeric_add_internal() function which fill given data structure with
error information instead of throwing the error.  numeric_add() would be a
wrapper over numeric_add_internal(), which throws an error if corresponding
data structure is filled.  In jsonpath we can call numeric_add_internal() and
interpret errors in another way.  That seems to be better than use of PG_TRY
and PG_CATCH.
Attached 12th version of jsonpath patches.

I added the 7th patch where the following functions were extracted for 
safe error handling in jsonpath:

numeric_add_internal()
numeric_sub_internal()
numeric_mul_internal()
numeric_div_internal()
numeric_mod_internal()
float8_numeric_internal()
numeric_float8_internal()
float8in_internal()

Errors are passed to caller with new ereport_safe() macro when
ErrorData **edata is not NULL:

+#define ereport_safe(edata, elevel, rest) \
+    do { \
+        if (edata) { \
+            errstart(elevel, __FILE__, __LINE__, PG_FUNCNAME_MACRO, TEXTDOMAIN); \
+            (rest); \
+            *(edata) = CopyErrorData(); \
+            FlushErrorState(); \
+        } else { \
+            ereport(elevel, rest); \
+        } \
+    } while (0)


But to_datetime() is still called in jsonpath inside PG_TRY/PG_CATCH block
because it needs too deep error propagation.

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

Attachment

pgsql-hackers by date:

Previous
From: Jeevan Chalke
Date:
Subject: Re: [HACKERS] Partition-wise aggregation/grouping
Next
From: David Steele
Date:
Subject: Re: Re: PATCH: pgbench - break out timing data for initializationphases