The ExportSnapshot is an MR job, and as a result of that it will run across your NodeManager hosts. To provide its destination as a local filesystem URI, such as your file:///local_linux_fs_dir would only work if that passed path is visible with the same consistent content across all your cluster hosts.
You can do this perhaps by mounting the same NFS across all hosts, and then using a controlled ExportSnapshot parallelism to write to them without overloading them (limit the # of maps to be low-enough).
If that's not desirable, then you can also opt to run the MR job in local mode, which would still be parallel but limitedly so, by passing -Dmapreduce.framework.name=local to ExportSnapshot before any other option.