Skip to content
Financial Context Wiki

Data Availability Sampling (DAS)

A verification-first guide to probabilistic data-availability sampling, erasure coding, commitments, sampling models, PeerDAS cells and columns, custody, peer independence and reconstruction limits.

Updated

For educational purposes only; not investment or security advice. DAS conclusions are probabilistic and depend on exact encoding, sampling, networking, client, fork and adversary assumptions.

Direct answer

Data Availability Sampling lets a node request a protocol-defined subset of authenticated, erasure-coded data units and make a probabilistic local judgment about whether the committed object was sufficiently published to be recoverable. The node does not download the full object. A successful sample increases confidence only under the stated encoding, reconstruction threshold, sample selection, peer diversity, timing and adversary model.

A KZG or cell proof binds a returned unit to a commitment; it does not prove that enough other units exist. A local sampling pass does not automatically prove network-wide reconstruction, execution validity, consensus finality, rollup settlement, user exit or permanent archiving. PeerDAS is one concrete Ethereum design with fork-specific columns, cells, custody groups, sampling, networking and fork-choice rules; its current parameters are not universal DAS constants.

How it works

  1. Pin the exact object and specification snapshot: chain, network, fork, client and spec revision; slot, block root and blob commitments; original and extended dimensions; cell, row, column and custody-group definitions; and the availability decision that the node must make.
  2. Verify the encoding and reconstruction rule. Record the erasure-extension algorithm, commitment scheme, number of original and encoded units, minimum units needed to recover the object and how malformed extensions are rejected. Sampling is meaningful only if withholding enough units to block reconstruction creates a detectable region.
  3. Define the sample universe and procedure. Record population size, unavailable or withheld set, sample count, with- or without-replacement selection, randomness source, uniqueness, custody duties, peer selection, timeout, retry and deduplication. Custody assignments and random samples are different evidence.
  4. State the probability model before calculating it. For a toy fixed withheld fraction f and s independent uniform samples with replacement, P_miss = (1 - f)^s. Without replacement, use the finite-population ratio of combinations. Correlated peers, biased randomness, eclipsing, adaptive serving and retry selection invalidate the simple model.
  5. Fetch from diverse peers within the protocol deadline. Verify canonical header and blob inclusion, commitment, index, KZG cell proof and response bytes; record missing, invalid and late responses separately. Repeating one column or asking the same control plane is not new independent evidence.
  6. Apply the named protocol’s local availability and fork-choice rules, then separately record network propagation and reconstruction, execution validity, safe and finalized consensus states, serving retention and archive status. A local pass must not be promoted into claims the protocol does not make.
  7. Rehearse withholding, malformed coding, invalid proofs, eclipse and Sybil attacks, selective or adaptive service, correlated sampling, peer and subnet partitions, timeouts, pruning, parameter and blob-schedule changes, reorganization and client disagreement. Actually reconstruct from collected units and preserve requests, responses, proofs, versions and archive evidence.

Worked examples

  • Independent sampling with replacement. In a toy model, 50% of units are withheld and a node makes 30 independent uniform samples with replacement. Miss probability is 0.5^30 = 0.0000000009313225746; detection probability is 99.9999999069%. This is a teaching model, not a live PeerDAS or another network’s service guarantee.
  • Targeting a miss probability. In the same toy independent model with 25% withheld, require miss probability at most 0.000001. The sample count is ceil(log(0.000001) / log(0.75)) = 49. At 48 samples, miss probability is 0.75^48 = 0.0000010067940558701114, still above the target; at 49, it is 0.75^49 = 0.0000007550955419025835.
  • Sampling without replacement. A finite population has 128 columns, 64 serviceable and 64 withheld. Sampling 8 distinct columns without replacement misses the withheld set with probability C(64,8) / C(128,8) = 0.0030958642767920487; detection probability is 99.6904135723%. This differs from 0.5^8 = 0.00390625 and still assumes uniform selection and static withholding.
  • Fulu parameter snapshot. In the Fulu specification checked on 2026-08-12, NUMBER_OF_COLUMNS = 128, SAMPLES_PER_SLOT = 8, NUMBER_OF_CUSTODY_GROUPS = 128 and CUSTODY_REQUIREMENT = 4. The minimum sample share is 8 / 128 = 6.25%; minimum custody share is 4 / 128 = 3.125%. If a node custodies 12 groups and sampling size is the greater of 8 and 12, it requests 12 / 128 = 9.375%. These are versioned protocol parameters, not permanent or cross-protocol guarantees.

Risks

  • Sampling the wrong chain, network, fork, slot, block or data object.
  • Trusting a stale, noncanonical or reorganized header and commitment.
  • Using malformed or incorrectly extended erasure-coded data.
  • Accepting an invalid KZG, cell or inclusion proof.
  • Treating commitment authenticity as proof of availability.
  • Collecting too few samples for the stated threat model.
  • Applying a with-replacement formula to without-replacement sampling or vice versa.
  • Using biased, predictable or manipulable randomness.
  • Counting duplicate units or retries as independent samples.
  • Sampling correlated peers, subnets, custody groups or control planes.
  • Suffering eclipse or Sybil attacks that isolate the sampler.
  • Allowing selective or adaptive serving after sample choices are known.
  • Misclassifying timeout, congestion or client failure as data withholding.
  • Confusing deterministic custody assignments with random sampling.
  • Assuming a withheld fraction that does not match the reconstruction threshold.
  • Failing reconstruction because of CPU, memory, bandwidth or implementation limits.
  • Encountering gossip, request-response, subnet or cross-seeding failures.
  • Extrapolating a local pass to network-wide availability and recovery.
  • Confusing DAS with execution validity, consensus finality, settlement or exit.
  • Missing retention, archive, fallback, fork, blob schedule or parameter changes.

Common misconceptions

  • One successful sample proves the complete object is available.
  • A KZG or cell proof is itself a data-availability proof.
  • More requests always provide independent evidence.
  • DAS validates execution and makes the chain final.
  • Data available now will remain permanently archived.

Sources

Navigation

Search the wiki...