On Fri, 2009-08-28 at 17:55 +0300, Martin Pihlak wrote:
> This is weird, indeed it seems that for some reason the recovery restartpoints
> are not created.
>
> Looking quickly at RecoveryRestartPoint() in xlog.c, there are two cases when it
> doesn't do a checkpoint. For one thing, it checks if if enough time has elapsed
> since last controlfile modification. If not, the checkpoint is skipped. I'm wondering
> if it does the correct thing if the clocks of two machines are too far apart.
Probably not that.
> Another check is "is it safe to do a checkpoint". This is logged with DEBUG2, so
> it should be visible if you set the logging level accordingly.
This seems like the most likely cause. I would guess that one of your
GiST indexes has a corruption in it and is preventing a restartpoint
from taking place. That has happened previously.
> Alternatively, you could attach a debugger to the recovery process and see if the
> RecoveryRestartPoint() and CheckPointGuts() functions are called at all.
I would say no need for that, but you can check the Gist pending actions
table.
--
Simon Riggs www.2ndQuadrant.com