Alvaro Herrera <alvherre@commandprompt.com> wrote:
> Please put this on hold until we decide what to do with Heikki's patch
> to update OldestXmin during vacuum.
Yes, I think his patch is very useful, but it has a little conflict
with my patch.
> One idea is to count the tuples actually cleared during vacuum. Another
> idea is to reread pgstat data after each OldestXmin recalculation to get
> accurate dead tuple counting. Neither of these seem very satisfying.
I'm thinking to use the larger value of follows:
- dead tuples statistic count at the beginning of vacuum
- count the tuples actually cleared during vacuum
The latter is usually larger with Heikki's patch. In the other hand,
the formar is larger in only cases when many DELETEs were rollbacked.
Regards,
---
ITAGAKI Takahiro
NTT Open Source Software Center