Consensus & infrastructure

Data Availability

The guarantee that data a system depends on can actually be retrieved by anyone who needs it, when they need it. A data-availability layer stores data and proves it is retrievable, so applications do not have to trust a single host.

Also known as: DA, DA layer

Data availability is the property that data a system relies on can actually be retrieved when it is needed. It became a headline concern through rollups: a Layer 2 posts compressed transaction data somewhere and asserts a new state, but that assertion is only trustworthy if anyone can fetch the data to check the maths. If the data cannot be retrieved, the state cannot be verified, and users cannot exit. So a data-availability layer does two jobs: it stores the data, and it makes the retrievability itself verifiable.

The technical challenge is proving availability without forcing everyone to download everything. Erasure coding plus sampling is the standard answer: encode the data into fragments so that a random check of a few fragments gives high statistical confidence the whole is recoverable. A network can then guarantee availability while each participant stores only a slice.

For DeAI the relevance is direct. Model weights, training datasets, provenance records, and agent memory all need to be retrievable by parties who do not trust each other and do not want a single company as the gatekeeper. Networks like Walrus market themselves as decentralised storage and data-availability layers precisely so that an AI application can point at a piece of data, prove it is available and unaltered, and let any counterparty verify it independently. Availability is what turns stored bytes into a claim other systems can build on.

Related terms