Walrus wal

Walrus is a decentralized blob-storage protocol built by Mysten Labs and coordinated by the Sui blockchain. It stores immutable byte arrays—called blobs—including text, images, video, source code, and application data. A blob is represented by a Sui Blob object, while its data is held by a separate committee of Walrus storage nodes. The project is open source under the MystenLabs/walrus repository.

Walrus is designed as a data layer rather than a general-purpose smart-contract chain. Sui supplies the control plane for payments, resource allocation, blob metadata, certificates, shard assignments, and epoch orchestration; storage nodes supply the data plane. Users can interact directly through the CLI or Rust/TypeScript SDKs, or through optional publishers, upload relays, and aggregators.

The protocol emphasizes persistent, independently verifiable, and portable data. Blob IDs are content-derived, so the same bytes produce the same ID and changing bytes produces a new ID. Certified blobs remain available for their purchased storage period, and a Sui event/certificate can be used as evidence of availability. Walrus also supports Walrus Sites, decentralized web hosting whose static assets are stored as blobs.

The native utility token is WAL. WAL pays for storage and write fees and can be delegated to storage nodes. SUI remains necessary for Sui transaction gas. The current Walrus product site additionally presents Walrus as a verifiable data platform for AI builders, including a portable Walrus Memory layer for AI agents.

Conventional blockchains fully replicate state across validators, making them inefficient for large binary objects. Traditional centralized object storage is easier to use but introduces provider dependence, censorship or availability risks, and weak independent verification. Decentralized storage systems must provide availability and integrity even when some operators fail or act maliciously, without requiring every node to keep a complete copy.

Simple erasure coding reduces storage overhead but makes node repair expensive: a replacement may need to download and reconstruct the entire blob before regenerating one lost fragment. Frequent failures and committee churn can therefore erase the network and cost advantages of coding. A further challenge is storage accountability: operators need incentives and asynchronous proof/challenge mechanisms so that a node cannot claim rewards while discarding its assigned data.

Walrus addresses these constraints with two-dimensional RedStuff encoding, authenticated slivers, Sui-managed certificates and epochs, and storage challenges. The design targets a Byzantine setting in which up to one-third of shards may be faulty or malicious, while retaining availability and efficient recovery as committees change.

A client first encodes a blob with RedStuff, an adaptation of the Twin-Code framework using RaptorQ-style linear erasure coding. The blob is arranged into a two-dimensional symbol matrix; primary and secondary encodings produce slivers assigned to shards. Vector commitments and authenticated metadata let nodes and readers verify that symbols correspond to the content commitment. Walrus documentation describes approximately 4.5x total storage overhead, plus a fixed metadata term, rather than full replication across all nodes.

The client purchases a Storage resource for a chosen duration and registers a Blob object on Sui. It sends each storage node its assigned sliver pair. Nodes verify the slivers and sign receipts. Once a two-thirds quorum of receipts is aggregated, the resulting availability certificate (the point of availability) is posted on Sui. The client can then discard its local copy; storage nodes are obligated to retain the encoded data for the certified period.

To read, a client or aggregator obtains blob metadata and requests slivers from the committee. It validates returned slivers against authenticated commitments, collects enough to cross the reconstruction threshold, decodes the bytes, and verifies the result. Because the data is immutable and content-addressed, a reader can detect altered or inconsistently encoded content rather than silently accepting it.

RedStuff's second dimension makes repair proportional to the lost sliver rather than the whole blob: a node requests symbols from other nodes, verifies proofs, and reconstructs its primary and secondary slivers. Epoch reconfiguration transfers responsibility to a new committee while reads and writes remain available. WAL-funded storage is held in a storage fund, rewards are distributed based on operator performance and delegated stake, and the protocol burns 3% of each epoch's total rewards; operator commission can also be burned for slashing, while delegated principal is not slashed.