Since weeks, I have a #robotframework service provisioned with #humanitec on a #k8s cluster - and every 2 days I have to redeploy the service. From logs it looked like playwright was eating memory. Thought about scheduling redeployments, improving load-balancing.

In the end turned out to be simple: one of the tasks did not call CLOSE BROWSER thus the browser instances remained as zombies and waited for the main thread (= the service) to end.