reverseproxy: Improve hashing LB policies with HRW (#4724)

* reverseproxy: Improve hashing LB policies with HRW

Previously, if a list of upstreams changed, hash-based LB policies
would be greatly affected because the hash relied on the position of
upstreams in the pool. Highest Random Weight or "rendezvous" hashing
is apparently robust to pool changes. It runs in O(n) instead of
O(log n), but n is very small usually.

* Fix bug and update tests

1 files changed, 17 insertions, 12 deletions

diff --git a/modules/caddyhttp/reverseproxy/selectionpolicies.go b/modules/caddyhttp/reverseproxy/selectionpolicies.go
index 001f7f806..125a07f9a 100644
--- a/modules/caddyhttp/reverseproxy/selectionpolicies.go
+++ b/modules/caddyhttp/reverseproxy/selectionpolicies.go
@@ -514,21 +514,26 @@ func leastRequests(upstreams []*Upstream) *Upstream {
 	return best[weakrand.Intn(len(best))]
 }
 
-// hostByHashing returns an available host
-// from pool based on a hashable string s.
+// hostByHashing returns an available host from pool based on a hashable string s.
 func hostByHashing(pool []*Upstream, s string) *Upstream {
-	poolLen := uint32(len(pool))
-	if poolLen == 0 {
-		return nil
-	}
-	index := hash(s) % poolLen
-	for i := uint32(0); i < poolLen; i++ {
-		upstream := pool[(index+i)%poolLen]
-		if upstream.Available() {
-			return upstream
+	// Highest Random Weight (HRW, or "Rendezvous") hashing,
+	// guarantees stability when the list of upstreams changes;
+	// see https://medium.com/i0exception/rendezvous-hashing-8c00e2fb58b0,
+	// https://randorithms.com/2020/12/26/rendezvous-hashing.html,
+	// and https://en.wikipedia.org/wiki/Rendezvous_hashing.
+	var highestHash uint32
+	var upstream *Upstream
+	for _, up := range pool {
+		if !up.Available() {
+			continue
+		}
+		h := hash(s + up.String()) // important to hash key and server together
+		if h > highestHash {
+			highestHash = h
+			upstream = up
 		}
 	}
-	return nil
+	return upstream
 }
 
 // hash calculates a fast hash based on s.