Re: Fix checkpoint skip logic on idle systems by tracking LSN progress

On Thu, Dec 22, 2016 at 6:41 AM, David Steele <david(at)pgmasters(dot)net> wrote:
> On 12/21/16 4:28 PM, Andres Freund wrote:
>
>> Working on committing this (tomorrow morning, not tonight). There's
>> some relatively minor things I want to change:

Thanks for looking at this patch.

>> - I don't like the name XLogSetFlags() - it's completely unclear what
>> that those flags refer to - it could just as well be replay
>> related. XLogSetRecordFlags()?
>
> That sounds a bit more clear.

Fine for me.

>> - Similarly I don't like the name "progress LSN" much. What does
>> "progress" really mean in that". Maybe "consistency LSN"?
>
> Yes, please. I think that really cuts to the core of what the patch is
> about. Progress made perfect sense to me, but consistency is always the
> goal, and what we are saying here is that this is the last xlog record that
> is required to achieve consistency. Anything that happens to be after it is
> informational only.

Fine as well.

>> - It's currently required to avoid triggering archive timeouts and
>> checkpoints triggering each other, but I'm nervous marking all xlog
>> switches as unimportant. I think it'd be better to only mark timeout
>> triggered switches as such.
>
> That seems fine to me. If the system is truly idle that might trigger one
> more xlog switch that is needed, but it seems like a reasonable compromise.

On a long-running embedded system the difference won't matter much. So
I guess I'm fine with this bit as well.
--
Michael