On the other hand, I've run multi-hundred-node HPC stateless HPC clusters long term off an (rpm-based) NFS root with little trouble once I could get the cluster vendors out of the picture. Places like the Tri-Labs appear to -- or have done? -- with larger systems, even if they need something like https://computation.llnl.gov/projects/spindle at scale. What flattened NFS wasn't executables but i/o of data files, BLAST being the worst offender; staging those in array jobs was definitely worthwhile.
I strongly prefer the trades-off of stateless systems in that context, though I might prefer to do it off a parallel filesystem, and clearly your kilometreage varies. [From experience ~10 years ago, I'd say avoid GPFS more than NFS, but doubtless it's improved, and I hold no brief for NFS.]
I've found that even for unix-unix, samba is just more reliable with more reasonable error behaviour.
To me samba has just too much magic, especially when mixing unix, OSX and windows clients.
magic is just technology we don't understand. I use cifs over nfs as well.
With excel locking, os x trashes, etc.? What good configuration guide would you suggest for a mixed Linux, OS X and Windows environment with MS Office?
This magic only solves have the problems
fruit:resource = xattr
fruit:metadata = netatalk
fruit:locking = none
fruit:encoding = private
delete veto files = yes
veto files = /.snapshot/.windows/.mac/.TemporaryItems/.DS_Store
Even before that, in a supercomputing context (2002-), when writing MPI "wrapper" software that would be used on end-user clusters with all kinds of unfortunate network setups, I avoided a lot of trouble by copying the executable to the local node before running it.
This isn't a new issue. And the traditional solutions work.