Alvaro Herrera wrote:
> Matthew T. O'Connor wrote:
> Well, if a table has 10 rows, and we keep the current threshold of 1000
> rows, then this table must have 1002 dead tuples (99% dead tuples, 1002
> dead + 10 live) before being vacuumed. This seems wasteful because
> there are 500 dead tuples on it and only 10 live tuples. So each scan
> must wade through all the dead tuples.
>
> Another small table with 100 tuples will be vacuumed on every iteration
> as well, even if there are just two dead tuples. So you are right --
> maybe dropping it all the way to 0 is too much. But a small value of 10
> is reasonable? That will make the 10 tuple table be vacuumed when there
> are 10 dead tuples (50% of dead tuples), and the 100 tuple table when
> there are 11 (11% of dead tuples). It decreases quickly to the scale
> factor (2%, or do we want to decrease it to 1%?)
I think it's probably fine. I think, that the optimal number for the
base_threhold is probably dependant on the width of the row, for a very
narrow row where you might have many on the same page, 20 or 50 might be
right, but for a very wide table a smaller number might be optimal,
however I think it probably doesn't matter much anyway.
Reducing the default to 10 seems fine, and perhaps even removing it as a
tuning knob. I think there are too many autovacuum knobs and it
confuses people. Is it too late to possibly remove this GUC altogether?
