Having task to utilize old servers as a backup distributed file system like MooseFS I see a collision in whole concept:
DFS is fault-tolerant and keep track of HDD status by itself, so any file system carrying data is better used without RAID. Although, OS should be fault-tolerant to HDD failure and should have no probable problems to boot. Using RAID is the best way to achieve that.
In my setup this leaves me poor options:
- Use two HDD in RAID-1 for OS and other HDDs in JBOD for data, what takes out half of 4xLFF server HDD bays.
- Use small partition in either LVM or MD mode with a big partition for data, what may bring problems with booting, restoring OS after HDD failure and also break best practices of dedicated HDD for data storage.
- Network booted hosts with all disks dedicated to storage.
Network booting bring other options up:
- Hosts can be booted in "root-on-NFS" or in "root-on-iSCSI" mode what make OS files access relatively slow.
- Hosts can be booted from networked squash file into RAM what make OS fast, but unable to save local data.
My questions are:
Is there a DFS architecture which have no local data stored on a storage server (not metadata server)?
What may happen to DFS if local files (not the stored chunks) are relatively slow?
If there is a way to load OS into RAM, but still have it saved on network backend?
