[griffinmyers]

I hit this issue a year ago, in a slightly different setup. We were trying to expose an internal ECS Service to other services via an NLB. Things worked great, as long as the sending and receiving services didn't end up on the same EC2 instance. Unfortunately this occurred fairly often, since the ECS scheduler wasn't aware of that bizarre constraint (maybe we could have made it aware, but, that seems like a pretty brittle way to fix things).

Our resolution was to just fall back to an ELB. A bummer in that it didn't support dynamic port registration (the service could only be on a single EC2 instance at a time, since it has a static port), but joyful in that we didn't have spurious failures as an artifact of task packing.