// optimizeOrder(l) takes a list and returns a list, containing the same contents
// but with a new ordering, however it wishes.
func (it *And) optimizeOrder(its []graph.Iterator) []graph.Iterator {
var (
// bad contains iterators that can't be (efficiently) nexted, such as
// graph.Optional or graph.Not. Separate them out and tack them on at the end.
out, bad []graph.Iterator
best graph.Iterator
bestCost = int64(1 << 62)
)
// Find the iterator with the projected "best" total cost.
// Total cost is defined as The Next()ed iterator's cost to Next() out
// all of it's contents, and to Contains() each of those against everyone
// else.
for _, root := range its {
if _, canNext := root.(graph.Nexter); !canNext {
bad = append(bad, root)
continue
}
rootStats := root.Stats()
cost := rootStats.NextCost
for _, f := range its {
if _, canNext := f.(graph.Nexter); !canNext {
continue
}
if f == root {
continue
}
stats := f.Stats()
cost += stats.ContainsCost * (1 + (rootStats.Size / (stats.Size + 1)))
}
cost *= rootStats.Size
if clog.V(3) {
clog.Infof("And: %v Root: %v Total Cost: %v Best: %v", it.UID(), root.UID(), cost, bestCost)
}
if cost < bestCost {
best = root
bestCost = cost
}
}
if clog.V(3) {
clog.Infof("And: %v Choosing: %v Best: %v", it.UID(), best.UID(), bestCost)
}
// TODO(barakmich): Optimization of order need not stop here. Picking a smart
// Contains() order based on probability of getting a false Contains() first is
// useful (fail faster).
// Put the best iterator (the one we wish to Next()) at the front...
out = append(out, best)
// ... push everyone else after...
for _, it := range its {
if _, canNext := it.(graph.Nexter); !canNext {
continue
}
if it != best {
out = append(out, it)
}
}
// ...and finally, the difficult children on the end.
return append(out, bad...)
}
