Another comment suggested that whatever cluster management automation they have in place may be configured to suspend services during low load to reduce costs. But if usage drops unexpectedly low then you enter a state that has never been tested. It may be that some services are (incorrectly) suspended across the board and causing failures.