On 2020/02/18 12:39, Michael Paquier wrote:
> On Mon, Feb 17, 2020 at 10:21:23PM +0900, Fujii Masao wrote:
>> On 2020/02/17 18:48, Michael Paquier wrote:
>>> Actually, I have some questions:
>>> 1) Should a new wait event be added in recoveryPausesHere()? That
>>> would be IMO useful.
>>
>> Yes, it's useful, I think. But it's better to implement that
>> as a separate patch.
>
> No problem for me.
On second thought, it's OK to add that event into the patch.
Attached is the updated version of the patch. This patch adds
two wait events for WAL archiving and recovery pause.
>>> 2) Perhaps those two points should be replaced with WaitLatch(), where
>>> we would use the new wait events introduced?
>>
>> For what? Maybe it should, but I'm not sure it's worth the trouble.
>
> I don't have more to offer than signal handling consistency for both
> without relying on pg_usleep()'s behavior depending on the platform,
> and power consumption. For the recovery pause, the second argument
> may not be worth carrying, but we never had this argument for the
> archiving wait, did we?
I have no idea about this. But I wonder how much that change
is helpful to reduce the power consumption because waiting
for WAL archive during the backup basically not so frequently
happens.
Regards,
--
Fujii Masao
NTT DATA CORPORATION
Advanced Platform Technology Group
Research and Development Headquarters