fix: slow start config add min_weight_percent field to avoid too big edf deadline

api/envoy/config/cluster/v3/cluster.proto

@@ -363,12 +363,17 @@ message Cluster {
     // By tuning the parameter, is possible to achieve polynomial or exponential shape of ramp-up curve.
     //
     // During slow start window, effective weight of an endpoint would be scaled with time factor and aggression:
-    // `new_weight = weight * time_factor ^ (1 / aggression)`,
+    // `new_weight = weight * max(min_weight_percent, time_factor ^ (1 / aggression))`,
     // where `time_factor=(time_since_start_seconds / slow_start_time_seconds)`.
     //
     // As time progresses, more and more traffic would be sent to endpoint, which is in slow start window.
     // Once host exits slow start, time_factor and aggression no longer affect its weight.
     core.v3.RuntimeDouble aggression = 2;
+
+    // Configures the minimum percentage of origin weight that avoids too small new weight,
+    // which may cause endpoints in slow start mode receive no traffic in slow start window.
+    // If not specified, the default is 10%.
+    type.v3.Percent min_weight_percent = 3;
   }
 
   // Specific configuration for the RoundRobin load balancing policy.

docs/root/intro/arch_overview/upstream/load_balancing/slow_start.rst

@@ -15,7 +15,7 @@ During slow start window, load balancing weight of a particular endpoint will be
 
 .. math::
 
-  NewWeight = {Weight*TimeFactor}^\frac{1}{Aggression}
+  NewWeight = {Weight}*{max(MinWeightPercent,{TimeFactor}^\frac{1}{Aggression})}
 
 where,
 
@@ -25,6 +25,8 @@ where,
 
 As time progresses, more and more traffic would be sent to endpoint within slow start window.
 
+:ref:`MinWeightPercent parameter<envoy_v3_api_field_config.cluster.v3.Cluster.SlowStartConfig.min_weight_percent>` specifies the minimum percent of origin weight to make sure the EDF scheduler has a reasonable deadline, default is 10%.
+
 :ref:`Aggression parameter<envoy_v3_api_field_config.cluster.v3.Cluster.SlowStartConfig.aggression>` non-linearly affects endpoint weight and represents the speed of ramp-up.
 By tuning aggression parameter, one could achieve polynomial or exponential speed for traffic increase.
 Below simulation demonstrates how various values for aggression affect traffic ramp-up:

docs/root/version_history/current.rst

@@ -36,6 +36,7 @@ Bug Fixes
 * jwt_authn: fixed the crash when a CONNECT request is sent to JWT filter configured with regex match on the Host header.
 * tcp_proxy: fix a crash that occurs when configured for :ref:`upstream tunneling <envoy_v3_api_field_extensions.filters.network.tcp_proxy.v3.TcpProxy.tunneling_config>` and the downstream connection disconnects while the the upstream connection or http/2 stream is still being established.
 * tls: fix a bug while matching a certificate SAN with an exact value in ``match_typed_subject_alt_names`` of a listener where wildcard ``*`` character is not the only character of the dns label. Example, ``baz*.example.net`` and ``*baz.example.net`` and ``b*z.example.net`` will match ``baz1.example.net`` and ``foobaz.example.net`` and ``buzz.example.net``, respectively.
+* upstream: cluster slow start config add ``min_weight_percent`` field to avoid too big EDF deadline which cause slow start endpoints receiving no traffic, default 10%. This fix is releted to `issue#19526 <https://github.com/envoyproxy/envoy/issues/19526>`_.
 * upstream: fix stack overflow when a cluster with large number of idle connections is removed.
 * xray: fix the AWS X-Ray tracer extension to not sample the trace if ``sampled=`` keyword is not present in the header ``x-amzn-trace-id``.
 

source/common/upstream/load_balancer_impl.cc

@@ -789,7 +789,12 @@ EdfLoadBalancerBase::EdfLoadBalancerBase(
           slow_start_config.has_value() && slow_start_config.value().has_aggression()
               ? absl::optional<Runtime::Double>({slow_start_config.value().aggression(), runtime})
               : absl::nullopt),
-      time_source_(time_source), latest_host_added_time_(time_source_.monotonicTime()) {
+      time_source_(time_source), latest_host_added_time_(time_source_.monotonicTime()),
+      slow_start_min_weight_percent_(slow_start_config.has_value()
+                                         ? PROTOBUF_PERCENT_TO_DOUBLE_OR_DEFAULT(
+                                               slow_start_config.value(), min_weight_percent, 10) /
+                                               100.0
+                                         : 0.1) {
   // We fully recompute the schedulers for a given host set here on membership change, which is
   // consistent with what other LB implementations do (e.g. thread aware).
   // The downside of a full recompute is that time complexity is O(n * log n),
@@ -987,7 +992,8 @@ double EdfLoadBalancerBase::applySlowStartFactor(double host_weight, const Host&
     auto time_factor = static_cast<double>(std::max(std::chrono::milliseconds(1).count(),
                                                     host_create_duration.count())) /
                        slow_start_window_.count();
-    return host_weight * applyAggressionFactor(time_factor);
+    return host_weight *
+           std::max(applyAggressionFactor(time_factor), slow_start_min_weight_percent_);
   } else {
     return host_weight;
   }

source/common/upstream/load_balancer_impl.h

@@ -467,6 +467,7 @@ class EdfLoadBalancerBase : public ZoneAwareLoadBalancerBase,
   const absl::optional<Runtime::Double> aggression_runtime_;
   TimeSource& time_source_;
   MonotonicTime latest_host_added_time_;
+  const double slow_start_min_weight_percent_;
 };
 
 /**