Autoscaler is not scaling down
I have auto scaling on along with restart time tracking, and efficiency tracking, rollback enabled. The scale-up factor is 6.0 and scale-down factor is 0.8; stats window is 15 minutes. When the jobs run they go through a few iterations of scaling up from initial parallelism of 1 for each node to up to 50, but when all data is processed and back-pressure is 0 it’s still staying at that scale. No attempt is made to scale down. I thought it may be trying to scale down, failing and rolling back to the last known good conf, but apparently it isn’t.
What gives? I need unused resources returned back to k8s…