On Wed, Oct 14, 2015 at 6:28 PM, dinesh kumar <dineshkumar02@gmail.com> wrote:
> I see this feature as an add on to do the parallel DML operations.
> There won't be any problem, if operations are mutually exclusive.
> I mean, each session operates on unique set of tuples.
>
> In the above case, we don't even need of SKIP LOCKED wait policy.
>
> But, when it comes to mutually depend operations, isn't it nice to provide,
> how much were locked by the other sessions. OR atlest a HINT to the other
> session like,
>
> GET DIAGNOSTICS var = DID_I_MISS_ANYTHING_FROM_OTHER_SESSIONS;
>
> I agree that, adding counter will take a performance hit.
> Rather going to my actual proposal on providing the counter value,
> isn't it good to provide a boolean type HINT, if we miss atleast a single
> tuple.
Suppose there are 5 locked rows and 5 unlocked rows in the heap and you do this:
select * from t1 for share skip locked limit 5
The Boolean you propose will be false if the first 5 rows in physical
order are locked, and otherwise it will be false. But there's no
difference between those two scenarios from the perspective of the
application. Here's another example:
with foo as (select * from t1 for share skip locked) select * from foo
where a = 2;
If foo contains any locked rows at all, this will return true,
regardless of whether a = 2.
It's true that, for a lot of normal-ish queries, LockRows is applied
late enough that your proposed Boolean would return the intended
answer. But there are a bunch of exceptions, like the ones shown
above, and there might be more in the future.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company