Re: Speedup twophase transactions

On Fri, Mar 17, 2017 at 5:15 PM, Michael Paquier
<michael(dot)paquier(at)gmail(dot)com> wrote:
> On Fri, Mar 17, 2017 at 5:00 PM, Nikhil Sontakke
> <nikhils(at)2ndquadrant(dot)com> wrote:
>> Micheal, it looks like you are working on a final version of this patch? I
>> will wait to review it from my end, then.
>
> I have to admit that I am beginning to get drawn into it...

And here is what I got. I have found a couple of inconsistencies in
the patch, roughly:
- During recovery entries marked with ondisk = true should have their
start and end LSN reset to InvalidXLogRecPtr. This was actually
leading to some inconsistencies in MarkAsPreparing() for 2PC
transactions staying around for more than 2 checkpoints.
- RecoverPreparedTransactions(), StandbyRecoverPreparedTransactions()
and PrescanPreparedTransactions() doing both a scan of pg_twophase and
the shared memory entries was way too complicated. I have changed
things so as only memory entries are scanned by those routines, but an
initial scan of pg_twophase is done before recovery.
- Some inconsistencies in the comments and some typos found on the way.
- Simplification of some routines used in redo, as well as simplified
the set of routines made available to users.

Tests are passing for me, an extra lookup would be nice.
--
Michael