Trivial Description
The last line of org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.gpu.GpuResourceAllocator#recoverAssignedGpus is this:
usedDevices.put(gpuDevice, containerId);
We should have an info (or if not info, at least a debug) level log to indicate that a container is allocated to a GPU device during recovery.
Please also check recovery related code, maybe there can be some room for improvement around logging.