ALTER TABLE ... DISABLE TRIGGER vs. AccessExclusiveLock
Hackers,
Experience and a read through backend/commands/tablecmds.c's
AlterTable() indicate that ALTER TABLE ... DISABLE TRIGGER obtains an
exclusive lock on the table (as does any ALTER TABLE).
Blocking other readers from a table when we've, within the body of a
transaction performing a bulk update operation where we don't want /
need triggers to fire, seems at first glance to be over-kill. I can
see how AlterTable()'s complex logic is made less complex through 'get
and keep a big lock', since most of its operational modes really do
need exclusive access, but is it strictly required for ... DISABLE /
REENABLE TRIGGER?
Could, say, RowExclusiveLock hypothetically provide adequate
protection, allowing concurrent reads, but blocking out any other
writers (for ENABLE / DISABLE TRIGGER) -- such as if driven through a
new statement other than ALTER TABLE -- such as "DISABLE TRIGGER foo
ON tbar" ?
Thanks!
----
James Robinson
Socialserve.com
On Tue, Jul 27, 2010 at 3:07 PM, James Robinson
<jlrobins@socialserve.com> wrote:
Experience and a read through backend/commands/tablecmds.c's AlterTable()
indicate that ALTER TABLE ... DISABLE TRIGGER obtains an exclusive lock on
the table (as does any ALTER TABLE).Blocking other readers from a table when we've, within the body of a
transaction performing a bulk update operation where we don't want / need
triggers to fire, seems at first glance to be over-kill. I can see how
AlterTable()'s complex logic is made less complex through 'get and keep a
big lock', since most of its operational modes really do need exclusive
access, but is it strictly required for ... DISABLE / REENABLE TRIGGER?Could, say, RowExclusiveLock hypothetically provide adequate protection,
allowing concurrent reads, but blocking out any other writers (for ENABLE /
DISABLE TRIGGER) -- such as if driven through a new statement other than
ALTER TABLE -- such as "DISABLE TRIGGER foo ON tbar" ?
Funny you should mention this. There is a pending patch to do
something very much along these line.
https://commitfest.postgresql.org/action/patch_view?id=347
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company