Rollups publish their data as blobs, and nodes check it's available by sampling.
The idea: Instead of downloading whole blobs, each node downloads a few random pieces of the erasure-coded data and checks them against the KZG commitments. If enough pieces are available, the whole blob can be rebuilt, so many small random checks add up to confidence that the data was really published.
Why it matters: Makes layer 2 transactions cheap.
Checking data by sampling
To stop the data from being rebuilt, a block producer has to withhold more than half of the extended pieces (gray). Then every random sample has a better than even chance of landing on a gap. A node whose 8 samples all come back has been fooled with probability under 0.4%, and with 30 samples under one in a billion. Simplified: PeerDAS actually samples whole columns of a two-dimensional layout.
Rollups publish their transaction data as blobs, introduced with EIP-4844 in March 2024[1, 2]. Blobs are cheaper than calldata and are pruned after about 18 days[3].
Each blob is committed with KZG. Since Fusaka, PeerDAS lets nodes check availability by sampling pieces of erasure-coded blobs instead of downloading everything[4], and blob parameter forks raised capacity further[5]. Security rests on KZG and on enough nodes sampling honestly.