Each of these is a real failure, and each is a sequence of tabs and keys in siltide. Nothing here is theory: the steps are what to press and what to conclude from each screen.
A device is held but doing nothing.
The most expensive state a fleet can be in: the scheduler counts the device as taken, the bill counts it as running, and nothing is using it.
A card is slower than its neighbors.
Identical devices in one box should do identical work. When one does not, the reason is usually written down somewhere siltide already reads.
An Xid, an ECC storm or a device that simply stopped answering. The question is always the same: what did it do before it went.
Memory is full and utilization is zero.
A device with its memory taken and nothing computing is usually a process that has allocated and then gone to sleep, or one that died without releasing.