Min Xid problem proposal

Поиск
Список
Период
Сортировка
От Alvaro Herrera
Тема Min Xid problem proposal
Дата
Msg-id 20051209163305.GA25980@surnet.cl
обсуждение исходный текст
Ответы Re: Min Xid problem proposal  (Tom Lane <tgl@sss.pgh.pa.us>)
Список pgsql-hackers
Anybody remembers my patch to allow tracking the minimum Xid present in
a table, allowing to update the freeze xid on a per-table basis?  The
motivation behind it was to remove the requirement of database-wide
vacuums.

The problem I found with it was that it required all tables to be
vacuumed at least once every billion transactions, even frozen tables,
because there was the danger that somebody may insert new tuples into
the table without marking that fact in the catalogs (thus minxid would
remain FrozenTransactionId but reality would be different.)

My proposal to solve that problem, is to make any transaction that
inserts or modifies tuples in a table that is marked as frozen, unfreeze
it first.  The problem I had last time was finding a good spot in the
code for doing so.  I'm now proposing to do it in the parser, in
setTargetTable().  This routine currently opens the target relation and
acquires RowExclusiveLock on it.  At this point we can check its
relminxid, and if it's FreezeTransactionId, open pg_class and change the
value there.

The problem with this is that it seems to turn a possibly innocuous
insert into an operation that needs to open pg_class.  But in the case
of a relation not in the relcache, it will happen anyway, so it's not
really all that serious.  (The assumption here is that an unfreeze event
is at least as unlikely as a cache miss or a cache invalidation.)

The attached patch implements this idea.  (Of course it doesn't work
standalone; it requires the rest of my min-xid patch.)  I attach it here
separately because it's small and the proposal can be reviewed
independently of the rest of the patch, which is quite bigger.

Note that there's a FIXME on heap_unfreeze() saying that the shared
invalidation would not occur if the transaction aborts.  This comment
comes verbatim from vacuum.c's vac_update_relstats(); however I made a
small experiment and it seems to be false.  I'm not sure about it, but
it seems to me to be critical to send a invalidation message so that
other backends will notice immediately when we unfreeze a relation.

Comments?

--
Alvaro Herrera                 http://www.amazon.com/gp/registry/CTMLCN8V17R4
"La victoria es para quien se atreve a estar solo"

Вложения

В списке pgsql-hackers по дате отправления:

Предыдущее
От: Martijn van Oosterhout
Дата:
Сообщение: Re: psql & UTF8
Следующее
От: Bruce Momjian
Дата:
Сообщение: Re: Upcoming PG re-releases