Data Availability Sampling (DAS)

Data availability sampling (DAS) is a technique that lets a blockchain node check, with very high confidence, that all of the data in a block has been published to the network, without downloading the whole block. The node requests a number of small, randomly chosen pieces of the data; if all of them are returned, it is overwhelmingly likely that the full data is available. DAS is a key part of plans to scale blockchains, because it allows block sizes to grow while ordinary computers can still verify the network.

The technique addresses the "data availability problem". A block producer could publish a block header while withholding some of the underlying transactions, preventing others from checking the block or reconstructing the chain's state. This matters most for rollups and other layer 2 networks that post their data to a base chain. DAS is used by Celestia, which launched in 2023 as a network specialised in data availability, and Ethereum introduced a form of it, PeerDAS, in its Fusaka upgrade, activated in December 2025.

Imagine a librarian who claims to have put a thousand-page book on the shelf. Checking every page would take hours. Instead, you ask for 30 random pages. If the librarian had secretly removed a large chunk of the book, at least one of your random requests would almost certainly hit a missing page, and you would catch them out.

The trick that makes this work is that the book is first copied out in a special way, with extra pages added, so that any missing section can be rebuilt as long as at least half of all pages exist. That forces anyone who wants to hide something to remove a lot of pages, which makes random spot checks very effective.

DAS combines two techniques:

• Erasure coding. The block data is extended with redundant pieces using an erasure code, so that the original can be rebuilt from any 50% of the extended data, in a one-dimensional design. To hide even a small part of the original, a producer must therefore withhold at least half of the extended data.
• Commitments. The producer publishes a short cryptographic commitment to the extended data in the block header, such as a polynomial commitment or a Merkle tree root, much as a cryptographic hash function fingerprints data, so each sampled piece can be checked against it.

Each sampling node then requests random pieces with proofs. If enough nodes sample, together they also hold enough pieces to reconstruct the whole block if needed.

If a producer withholds a fraction $f$ of the extended data, the chance that one node's $k$ independent random samples all succeed anyway is:

• $f$ is the share of extended data withheld, at least 0.5 for data to be unrecoverable.
• $k$ is the number of random samples the node requests.

With $f = 0.5$ and $k = 30$, the probability is 0.5 raised to the 30th power, about 0.000000001, or roughly one in a billion. Each extra sample halves the chance of being fooled, so a light node can gain strong assurance by downloading only a tiny fraction of the block.

The Ethereum blockchain's Dencun upgrade in 2024 introduced "blobs", a cheaper type of data for rollups, but initially required nodes to download every blob. PeerDAS, part of the wider danksharding roadmap, lets nodes download and custody only part of each blob's extended data and sample the rest from peers, allowing the number of blobs per block to rise over time. Celestia uses a two-dimensional erasure coding scheme and fraud proofs for incorrectly encoded data, with light nodes sampling directly. Other data availability networks use related techniques. These approaches underpin the "modular" model, in which execution happens on rollups while separate layers provide consensus and data availability, as described under consensus layer.

DAS guarantees that data was published, not that it is valid; checking validity still depends on fraud or validity proofs from optimistic rollups or zero-knowledge rollups. Sampling assumes that enough honest nodes take part and that the network does not let an attacker answer some nodes' requests selectively. The cryptographic commitments used by some designs depend on a trusted setup ceremony, and the peer-to-peer networking needed to serve samples quickly is complex. Published data is also typically kept for only a limited time, so long-term storage depends on other parties. The broader scaling challenge is described under blockchain trilemma and sharding.