Re: One long transaction or multiple short transactions?

On Thu, Oct 08, 2015 at 05:43:11PM -0400, Carlo wrote:
> -----Original Message-----
> From: ktm(at)rice(dot)edu [mailto:ktm(at)rice(dot)edu]
> Sent: October 8, 2015 1:00 PM
> To: Carlo
> Cc: pgsql-performance(at)postgresql(dot)org
> Subject: Re: [PERFORM] One long transaction or multiple short transactions?
>
> On Thu, Oct 08, 2015 at 11:08:55AM -0400, Carlo wrote:
> > >> Sounds like a locking problem
> >
> > This is what I am trying to get at. The reason that I am not
> > addressing hardware or OS configuration concerns is that this is not
> > my environment, but my client's. The client is running my import
> > software and has a choice of how long the transactions can be. They
> > are going for long transactions, and I am trying to determine whether
> > there is a penalty for single long transactions over a configuration
> > which would allow for more successive short transactions. (keep in mind
> all reads and writes are single-row).
> >
> > There are other people working on hardware and OS configuration, and
> > that's why I can't want to get into a general optimization discussion
> > because the client is concerned with just this question.
> >
>
> On October 8, 2015 1:00 PM Ken wrote:
> > Hi Carlo,
>
> > Since the read/writes are basically independent, which is what I take your
> "single-row" comment to mean, by batching them you are balancing two
> > opposing factors. First, larger batches allow you to consolodate I/O and
> other resource requests to make them more efficient per row. Second, larger
> > batches require more locking as the number of rows updated grows. It may
> very well be the case that by halving your batch size that the system can
> > process them more quickly than a single batch that is twice the size.
>
> Just to clarify, one transaction of this type may take longer to commit than
> two successive transactions of half the size?
>

Yes, but where the optimum count is located should be determined by testing.
Just varying the batch size and note where the performance is at a maximum.

Regards,
Ken