On 5 April 2016 at 17:43, Rod Taylor <rod.tay...@gmail.com> wrote:
> The intention of this feature is to give the ability to slip into a normal
> workload for non-urgent maintenance work. In essence, instead of lock
> waiters being in a Queue, DEFERRABLE causes the current lock statement to
> always be last.
Good idea; this was on my list of things to implement. I was going to call
it WAIT PATIENTLY option.
> It was discussed at last years pgCon as useful for replication tools
> adding/removing triggers. I've also seen more than one plpgsql loop using
> subtransactions and LOCK TABLE .. NOWAIT to achieve a similar effect. IMO,
> it's much cleaner built in.
Agreed, but your implementation is essentially just the same looping
concept, which I don't much like.
> If a lock is successfully obtained on one table, but not on all tables, it
> releases that lock and will retry to get them as a group in the future.
> Since inheritance acts as a group of tables (top + recursive cascade to
> children), this implementation is necessary even if only a single table is
> specified in the command.
I'd prefer to see this as a lock wait mode where it sits in the normal lock
queue BUT other lock requestors are allowed to queue jump past it. That
should be just a few lines changed in the lock conflict checker and some
sleight of hand in the lock queue code.
That way we avoid the busy-wait loop and multiple DEFERRABLE lock waiters
queue up normally.
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services