On Sat, Sep 10, 2011 at 12:32 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Thu, Sep 8, 2011 at 8:42 PM, Fujii Masao <masao.fujii@gmail.com> wrote:
>> Another idea to avoid spinlock contention is save the timestamp in
>> PgBackendStatus (which contains information for pg_stat_activity).
>> This enables us to write and read the timestamp without spinlock.
>> Comments?
>
> That seems like a possibly promising approach, in that each backend
> could update the information separately, and it's the reader's job to
> go find the maximum of all those values when needed. So the overhead
> is (properly, in this case) placed on the reader instead of the
> writer. But it's a bit tricky, because when the reader wants that
> maximum, it has to take into account inactive backends that may have
> committed transactions before exiting, not just the ones that are
> still connected.
Yes, that's what I was thinking. The attached patch has adopted that
approach.
Regards,
--
Fujii Masao
NIPPON TELEGRAPH AND TELEPHONE CORPORATION
NTT Open Source Software Center
