On 2015-07-28 22:51:55 +0300, Heikki Linnakangas wrote:
> >checkpoint continuous flushing
>
> This does a big memory allocation at checkpoint, which Tom vehemently
> objects to.
Uh. Didn't he just object to failing in that case? IIRC he even
indicated tentative assent, a year or so back, with my idea of just
pre-allocating all the required memory in shared memory, so that we
don't need to allocate anything at some point.
I've not read the last version of the patch, but in my old version the
allocation wasn't actually that large in comparison to the size of
shared buffers itself.
> I don't much like it either, although I would be OK with a more
> moderately-sized allocation.
That'll disallow some mighty fine optimizations like e.g. being able to
do the fsyncs of files early, directly after we wrote out all their
buffer, thereby reducing how much dirty data (written out by backend)
that needs to be flushed to disk.
Greetings,
Andres Freund