On Tue, Aug 28, 2018 at 2:40 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Thomas Munro <thomas.munro@enterprisedb.com> writes:
> > We could probably improve that situation by making syscache lookups
> > (and probably other things too) fail when called from _PG_init() in
> > regular backends so that extension authors are made aware of this
> > hazard, or perhaps go the other way and change the order we do things
> > in parallel workers.
>
> Hmm. There's an argument to be made for the latter: we don't really
> want stuff failing in parallel workers if it works fine normally.
>
> On the other hand, it seems clear to me that we *don't* want extensions to
> be doing stuff like syscache lookups in _PG_init(), because that would
> prevent them from working as shared_preload_libraries entries.
>
> And on the third hand, intentionally breaking code that used to work
> isn't likely to win us many friends either. So I'm not sure that your
> first option is really tenable. Perhaps we could get away with doing
> it in HEAD and not back-patching ... but that does little for existing
> problems.
Ok, so I guess we should just make it work then. Here's throw-away
test patch to make "unaccent" blow up in this way, and a draft patch
to reorder things so that it works.
CREATE EXTENSION unaccent;
CREATE TABLE t AS SELECT 'hello world' AS v FROM generate_series(1, 1000000);
ANALYZE t;
EXPLAIN ANALYZE SELECT COUNT(unaccent(v)) FROM t;
--
Thomas Munro
http://www.enterprisedb.com
