The blog article introducing it is really good. It’s also one of the few remaining actually good websites, loads fast no bloat and still has good “fancy” design.

To summarize:

• Massively parallel and redundant over network (including with geographic separation i.e. high pings)
• Intended for large files (MB or larger)
• Exabyte scale (exabytes of data, trillions of files, millions of clients)
• Understands tiered storage (like using HDDs and SSDs appropriately)
• No odd hardware or software requirements (Written in C++ and Go, works on anything linux works on, doesn’t need a service)
• Communicates via TCP; accessed via their linux kernel module or the ternfs network api
• Immutable (no modifying only deleting) with snapshotting
• Slow directory creation and removal
• No permissions (“other services” should handle those

The immutability and parallelism as well as tiering are similar to btrfs, though all more extreme.

This does sound quite ideal (though overkill) for a dedicated media nas setup.

Edit: a few more points:

• Syncs bewteen locations are async, so there will be loss of recent data if an entire location is lost
• Writes are authorized at a single primary location, so don’t get lower latency for closer datacenters though perhaps better speeds (they plan to change this)
• The client talks to different components (servers) directly using their IPv4 (for private setups this is too expensive if the filesystem is meant to be accessible on the internet)
• The kernel module isn’t posix compliant and needs programs to write their files start to end (which is common but not universal), else programs need to write to temp files first

I’m very interested in this