On Mon, Oct 12, 2020 at 7:59 AM Andy Fan <zhihui.fan1213@gmail.com> wrote:
>
> Sorry for the late reply! Suppose we have partition defined like this:
> p
> - p1
> - p2
>
> When you talk about "the statistics from the partitioned table", do you
> mean the statistics from p or p1/p2? I just confirmed there is no statistics
> for p (at least pg_class.reltuples = -1), so I think you are talking about
> p1/p2.
I am talking about p when I say statistics from the partitioned table.
I see that pg_statistic row from p is well populated.
pg_class.reltuples = -1 indicates that the heap doesn't have any rows.
set_rel_size() sets the number of rows in the partitioned table based
on the rows in individual unpruned partitions.
Glad to know that, Thanks for this info!
>
> Here we are talking about partkey = $1 or partkey = RunTimeValue.
> so even the value can hit 1 partition only, but since we don't know
> the exact value, so we treat all the partition equally. so looks
> nothing wrong with partition level estimation. However when we cost
> the Append path, we need know just one of them can be hit, then
> we need do something there. Both AppendPath->rows/total_cost
> should be adjusted (That doesn't happen now).
I think in this case we can safely assume that only one partition will
remain so normalize costs considering that only one partition will
survive.
Exactly. What I am trying to do is fix this at create_append_path,
do you have different suggestions? about the pkey > $1 case, I think
even if we use the statistics from partition level, it would be
hard-code as well since we don't know what value $1 is.
I have gone through the main part of the RunTime partition prune, hope
I can update a runnable patch soon. The main idea is fix the rows/
costs at create_append_path stage. So any suggestion in a different
direction will be very useful.