You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We have encountered an issue with Envoy Proxy's cluster when RING_HASH load balancing is enabled along with subset configuration. Specifically, the detection of LBEndpoint overload still uses the active requests of the entire cluster as the denominator.
In this scenario, should the overload detection be based only on the nodes within the subset?
Configure a cluster with RING_HASH load balancing with 4 upstream endpoints, A and B are in the same subset.
Send equal amounts of long-delay requests to endpoint A, C, and D.
Send requests to the subset without hashing key.
Expected Behavior: The overload detection should consider only the active requests of the nodes within the subset, so the load balancer always forward requests to B.
Actual Behavior: The overload detection uses the active requests of the entire cluster, so randomly forward requests to both A and B until A reach the overload threshold.
Description:
We have encountered an issue with Envoy Proxy's cluster when RING_HASH load balancing is enabled along with subset configuration. Specifically, the detection of LBEndpoint overload still uses the active requests of the entire cluster as the denominator.
In this scenario, should the overload detection be based only on the nodes within the subset?
https://github.com/envoyproxy/envoy/blob/a0c96b389d2ef44ff207bb17678a5c5eabdbbadb/source/extensions/load_balancing_policies/common/thread_aware_lb_impl.cc#L199C1-L204C94
Steps to Reproduce:
Expected Behavior: The overload detection should consider only the active requests of the nodes within the subset, so the load balancer always forward requests to B.
Actual Behavior: The overload detection uses the active requests of the entire cluster, so randomly forward requests to both A and B until A reach the overload threshold.
The text was updated successfully, but these errors were encountered: