Atri Sharma <atri.jiit@gmail.com> writes:
> One of the factors that leads to bad estimates is that the histogram of the
> values of a column maintained by the planner gets old by time and the data
> in the column changes. So, the histogram is no longer a quite accurate view
> of the data and it leads to bad selectivity.
TBH, this is so far down the list of problems that it'll be a long time
before we need to worry about it. It's certainly not the number one
priority for any project to model risk in the planner.
The thing that I think is probably the number one problem is estimates
that depend on an assumption of uniform distribution of sought-after rows
among those encountered by a scan. This is usually where bad plans for
LIMIT queries are coming from. We could certainly add some sort of fudge
factor to those costs, but I'd like to have a more-or-less principled
framework for doing so.
regards, tom lane