On 31.08.2022 20:14, Tom Lane wrote:
Robert Haas <robertmhaas@gmail.com> writes:
On Wed, Aug 31, 2022 at 1:06 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
The currently proposed patchset hacks up a relatively small number
of core datatypes to be able to do that. But it's just a hack
and there's no prospect of extension types being able to join
in the fun. I think where we need to start, for v16, is making
an API design that will let any datatype have this functionality.
(I don't say that we'd convert every datatype to do so right away;
in the long run we should, but I'm content to start with just the
same core types touched here.) Beside the JSON stuff, there is
another even more pressing application for such behavior, namely
the often-requested COPY functionality to be able to shunt bad data
off somewhere without losing the entire transfer. In the COPY case
I think we'd want to be able to capture the error message that
would have been issued, which means the current patches are not
at all appropriate as a basis for that API design: they're just
returning a bool without any details.
I would be in favor of making more of an effort than just a few token
data types. The initial patch could just touch a few, but once the
infrastructure is in place we should really make a sweep through the
tree and tidy up.
Sure, but my point is that we can do that in a time-extended fashion
rather than having a flag day where everything must be updated.
The initial patch just needs to update a few types as proof of concept.
And here is a quick POC patch with an example for COPY and float4:
=# CREATE TABLE test (i int, f float4);
CREATE TABLE
=# COPY test (f) FROM stdin WITH (null_on_error (f));
1
err
2
\.
COPY 3
=# SELECT f FROM test; f
--- 1 2
(3 rows)
=# COPY test (i) FROM stdin WITH (null_on_error (i));
ERROR: input function for datatype "integer" does not support error handling
PG_RETURN_ERROR() is a reincarnation of ereport_safe() macro for returning
ErrorData, which was present in older versions (~v18) of SQL/JSON patches.
Later it was replaced with `bool *have_error` and less magical
`if (have_error) ... else ereport(...)`.
Obviously, this needs a separate thread.
--
Nikita Glukhov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company
