Still the same issue and solution. Removing the ECSServiceAvarageMemoryUtilization metric immediately 'enables' the scale-in activity, if ECSServiceAvarageCPUUtilization is below threshold. In our case the memory utilization (like 58%) where constantly operating close to the ECSServiceAvarageMemoryUtilization threshold of 60%. This might be a reason, too. I would expect, that there is a hysteresis within the scaling algorithm. Try to iterate with the memory utilization threshold.