[pm90]

The original architecture seemed overly complex IMO. Maintaining a fleet of HAProxys and their configs seems daunting... it seems like they're gaining both flexibility (maintaining configs easily) and observability (request/response metrics for thirft requests)

[ckwang]

Per-service proxy deployments are a bit complex for the infrastructure but provide a nice abstraction for the service and service developers themselves. The configuration scheme is indeed daunting, which is what we're hoping Envoy and its xDS APIs + centralized configs can help us solve for developer teams.