I wonder if it’d make sense to instrument the integration test stuff with a subprocess memory monitor and then assert mem utilization doesn’t exceed a certain hardcoded threshhold that we just continually drive down?