Arthur Zakirov <a.zakirov@postgrespro.ru> writes:
> On Sat, Mar 24, 2018 at 04:56:36PM -0400, Tom Lane wrote:
>> * And that leads us to not particularly need a view telling which
>> dictionaries are loaded, either. It's just an implementation detail
>> that users don't need to worry about.
> If all dictionaries will be shareable then this view could be removed.
> Unfortunately I think it can't help with leaked segments, I didn't find
> a way to iterate dshash entries. That's why pg_ts_shared_dictionaries()
> scans pg_ts_dict table instead of scanning dshash table.
If you're scanning pg_ts_dict, what happens with dictionaries belonging
to other databases? They won't be visible in your local copy of
pg_ts_dict. Between that and the inability to find leaked segments,
I'm not seeing that this has much use-case.
>> (It might work to use
>> the combination of dictionary OID and TID of the dictionary's pg_ts_dict
>> tuple as the lookup key for shared dictionaries. Oh, and have you
>> thought about the possibility of conflicting OIDs in different DBs?
>> Probably the database OID has to be part of the key, as well.)
> Yes unfortunately ALTER TEXT SEARCH DICTIONARY doesn't reload a
> dictionary. TID can help here. I thought about using XID too when I
> started to work on RELOAD command. But I'm not sure that it is a good
> idea, anyway XID isn't needed in current version.
Actually, existing practice is to check both xmin and tid; see for example
where plpgsql checks if a cached function data structure still matches the
pg_proc row, pl_comp.c around line 175 in HEAD. The other PLs do it
similarly I think. I'm not sure offhand just how much that changes the
risks of a false match compared to testing only one of these fields, but
I'd recommend conforming to the way it's done elsewhere.
regards, tom lane