in this pr, we add a new algorithm to estimate semi/anti join row count.
In original alg., we reduce row count from cross join. usually, this is not good.
for example, L left semi join R on L.a=R.a
suppose L is larger than R, and ndv(L.a) < ndv(R.a)
the estimated row count is rowcount(R) * rowcount(L) / ndv(R.a). in most cases, the estimated row count is larger than rowcount(L).
in new alg, we use ndv(R.a)/originalNdv(R.a) to estimate result rowCount. the basic idea is as following:
1. Suppose ndv(R.a) reduced from m to n.
2. Assume that the value space of L.a is the same as R.a if R.a is not filtered.(this assumption is also hold in original alg.)
regard `L left join R` as a filter applied on L, that is, if L.a is in R.a, then this tuple stays in result.
R.a shrinks to m/n, so L.a also shrinks to m/n
b9381570d6