Tornado File System Architecture
A radical storage architecture where files are never written to disk — data is chunked into HTTP/3 datagrams and continuously streamed through a ring of servers using QUIC.
- Author
- Radesh Govind
- Created
- July 30, 2026
Motivation
Traditional file systems assume storage means “data at rest on a disk.” But disk introduces latency, wear, and a static attack surface. What if data never stopped moving? A file system where storage is the network itself — data lives in fiber optic cables and RAM buffers, continuously circulating through a ring of servers.
This is not a cache layer or a CDN. This is storage as motion.
Proposal
The Wind
A ring of at least 3 servers that receive QUIC streams and immediately forward them to the next node, keeping data perpetually in motion. There is no “write to disk” step — data exists only in transit and in RAM buffers during forwarding.
Protocol Layer
- QUIC multiplexed streams — one stream per file, enabling fine-grained flow control
- Unreliable datagrams (RFC 9221) — for file chunks that can tolerate loss, reducing latency
- User-space control — fine-grained bandwidth throttling per file/stream
Storage Capacity
Storage capacity is bounded by the Bandwidth-Delay Product:
Capacity = Bandwidth × Round-Trip Latency
For a 100 Gbps ring with 10ms RTT: ~125 MB of in-flight data. Scaling requires more bandwidth, more nodes (increasing RTT), or both.
Reliability
Data in motion means packet loss = data loss. The system uses Forward Error Correction (FEC) to rebuild files from inevitable packet loss without retransmission delays. Each file is encoded with redundant chunks so that any k of n chunks are sufficient to reconstruct the original.
File Operations
- Read — tap into the ring and capture the stream for the requested file
- Write — inject a new stream into the ring, which begins circulating immediately
- Delete — stop forwarding the stream; data dissipates as packets drain from the ring
Alternatives Considered
- RAM disk (tmpfs) — data at rest in memory, single point of failure, no network distribution
- Distributed cache (Redis Cluster) — data at rest in RAM, not in motion, requires explicit replication
- Network-attached storage (NFS/iSCSI) — traditional disk-backed, adds network latency on top of disk latency
Decision
Pending — draft stage.