xds: WRR rr_fallback should trigger with one endpoint weight

From gRFC A58:
> When less than two subchannels have load info, all subchannels will
> get the same weight and the policy will behave the same as round_robin

Author: ejona86

xds/src/main/java/io/grpc/xds/WeightedRoundRobinLoadBalancer.java

@@ -598,13 +598,15 @@ static final class StaticStrideScheduler {
       if (numWeightedChannels > 0) {
         unscaledMeanWeight = sumWeight / numWeightedChannels;
         unscaledMaxWeight = Math.min(unscaledMaxWeight, (float) (K_MAX_RATIO * unscaledMeanWeight));
-        usesRoundRobin = false;
       } else {
-        // Fall back to round robin if all values are non-positives
-        usesRoundRobin = true;
+        // Fall back to round robin if all values are non-positives. Note that
+        // numWeightedChannels == 1 also behaves like RR because the weights are all the same, but
+        // the weights aren't 1, so it doesn't go through this path.
         unscaledMeanWeight = 1;
         unscaledMaxWeight = 1;
       }
+      // We need at least two weights for WRR to be distinguishable from round_robin.
+      usesRoundRobin = numWeightedChannels < 2;
 
       // Scales weights s.t. max(weights) == K_MAX_WEIGHT, meanWeight is scaled accordingly.
       // Note that, since we cap the weights to stay within K_MAX_RATIO, meanWeight might not

xds/src/test/java/io/grpc/xds/WeightedRoundRobinLoadBalancerTest.java

@@ -1190,13 +1190,17 @@ public void metrics() {
     verifyLongCounterRecord("grpc.lb.wrr.endpoint_weight_not_yet_usable", 1, 2);
     verifyLongCounterRecord("grpc.lb.wrr.endpoint_weight_not_yet_usable", 1, 3);
 
-    // Send each child LB state an ORCA update with some valid utilization/qps data so that weights
-    // can be calculated.
+    // Send one child LB state an ORCA update with some valid utilization/qps data so that weights
+    // can be calculated, but it's still essentially round_robin
     Iterator<ChildLbState> childLbStates = wrr.getChildLbStates().iterator();
     ((WeightedChildLbState)childLbStates.next()).new OrcaReportListener(
         weightedConfig.errorUtilizationPenalty).onLoadReport(
         InternalCallMetricRecorder.createMetricReport(0.1, 0, 0.1, 1, 0, new HashMap<>(),
             new HashMap<>(), new HashMap<>()));
+
+    fakeClock.forwardTime(1, TimeUnit.SECONDS);
+
+    // Now send a second child LB state an ORCA update, so there's real weights
     ((WeightedChildLbState)childLbStates.next()).new OrcaReportListener(
         weightedConfig.errorUtilizationPenalty).onLoadReport(
         InternalCallMetricRecorder.createMetricReport(0.1, 0, 0.1, 1, 0, new HashMap<>(),
@@ -1210,9 +1214,15 @@ public void metrics() {
     // weights were updated
     reset(mockMetricRecorder);
 
-    // We go forward in time past the default 10s blackout period before weights can be considered
-    // for wrr. The eights would get updated as the default update interval is 1s.
-    fakeClock.forwardTime(11, TimeUnit.SECONDS);
+    // We go forward in time past the default 10s blackout period for the first child. The weights
+    // would get updated as the default update interval is 1s.
+    fakeClock.forwardTime(9, TimeUnit.SECONDS);
+
+    verifyLongCounterRecord("grpc.lb.wrr.rr_fallback", 1, 1);
+
+    // And after another second the other children have weights
+    reset(mockMetricRecorder);
+    fakeClock.forwardTime(1, TimeUnit.SECONDS);
 
     // Since we have weights on all the child LB states, the weight update should not result in
     // further rr_fallback metric entries.