On Fri, Feb 17, 2017 at 11:05:31PM +0900, Michael Paquier wrote:
> On Fri, Feb 17, 2017 at 10:43 PM, Andreas Karlsson <andreas@proxel.se> wrote:
> > Thinking about this makes me wonder about why you decided to use a
> > transaction per index in many of the steps rather than a transaction per
> > step. Most steps should be quick. The only steps I think the makes sense to
> > have a transaction per table are.
>
> I don't recall all the details to be honest :)
>
> > 1) When building indexes to avoid long running transactions.
> > 2) When validating the new indexes, also to avoid long running transactions.
> >
> > But when swapping the indexes or when dropping the old indexes I do not see
> > any reason to not just use one transaction per step since we do not even
> > have to wait for any locks (other than WaitForLockers which we just want to
> > call once anyway since all indexes relate to the same table).
>
> Perhaps, this really needs a careful lookup.
>
> By the way, as this patch is showing up for the first time in this
> development cycle, would it be allowed in the last commit fest? That's
> not a patch in the easy category, far from that, but it does not
> present a new concept.
FYI, the thread started on 2013-11-15.
-- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB
http://enterprisedb.com
+ As you are, so once was I. As I am, so you will be. +
+ Ancient Roman grave inscription +