EIP-8070: Sparse Blobpool

Discussion topic for EIP-8070 [PR]

Description

Introduce custody-aligned sampling in the EL blobpool to vacate bandwidth.

Abstract

This proposal introduces the sparse blobpool, a construction that brings cell-level, custody-aligned sampling in the Execution Layer (EL). For every new type 3 (blob-carrying) transaction, an EL node fetches full blob payloads only with probability p = 0.15, and otherwise it merely samples the blobs, using the same custody assignment as its Consensus Layer (CL) counterpart. For full nodes, this means downloading as little as 1/8 of the data (8 out of 128 cells), so that the average bandwidth consumption compared to the (current) full blobpool is 0.15 + 0.85/8 ~ 0.25, a ~4x reduction. The choice of p = 0.15 balances reducing bandwidth consumption with guaranteeing the full propagation of txs, by ensuring that for each blob tx there exists a large connected backbone of nodes that have the full blob payload. At an individual node level, p = 0.15 translates to 98.6% probability of least 3/50 neighbours holding the full blob payload, only 0.03% chance of total unavailability. The sampling performed with probability 1 - p = 0.85 enables streamlined data availability checks during block validation, as well as enhancing the availability of the data.

1 Like

Thanks for adding the rationale for why normative peer scoring is not a good idea. I remember that in the GossipSub paper a ‘covert flash attack’ was described, in which malicious nodes spend some time being honest to increase their peer score and then in a coordinated way one day all at once stop their message distribution / behave maliciously. I agree that adding sampling noise and actively trying to get random new peer connections (to avoid eclipse attacks) is the safest way forward

I don’t get how extra noise adds any security. Peers can act normally to eclipse a node then not serve more than 50 percent of a blob to block da.

GetCells / Cells: the request id position is left open by the EIP text, and clients have split on it

EIP-8070 writes GetCells (0x14) and Cells (0x15) without a request id:

GetCells: [[hash_0: B_32, hash_1: B_32, ...], cell_mask: B_16]
Cells:    [[hash_0: B_32, hash_1: B_32, ...], cells: [[cell_0_0: B_2048, ...], ...], cell_mask: B_16]

while PooledTransactions, a few lines above, is written with one ([request_id: P, [tx_0, tx_1, ...]]). That leaves two readings, and both are implemented today:

Reading GetCells on the wire Implementations (commit checked)
Request id first, the fields follow directly [request_id, [hash, ...], cell_mask] geth (0d06140), Nethermind (fcea02a), Besu (29abdb4)
EIP-2481 shape: request id, then the EIP’s body as one nested list [request_id, [[hash, ...], cell_mask]] reth main (d5991a4)

The first reading is what devp2p caps/eth.md specifies in its eth/72 section (ethereum/devp2p#263): [request-id: P, [txhash₁: B_32, ...], cells: B_16] and [request-id: P, [txhash₁: B_32, ...], [[cell₁: B_2048, ...], ...], cells: B_16].

The second applies EIP-2481’s rule, which wraps the old command as the second element. That rule was written for the eth/65 messages EIP-2481 lists, and later messages do not all follow it: eth/70 GetReceipts (EIP-7975) puts firstBlockReceiptIndex directly after the request id.

The two layouts do not decode into each other, in either direction, for either message. With geth’s rlp package and geth’s packet types, a nested GetCells fails with rlp: expected input string or byte on Hashes[0]; decoding a flat one into the nested shape fails with expected input list on the hash list. Example: request id 42, one hash 0x…1234, cells 0, 8 and 127:

flat   (53 B): f42ae1a000000000000000000000000000000000000000000000000000000000000012349001010000000000000000000000000080
nested (54 B): f52af3e1a000000000000000000000000000000000000000000000000000000000000012349001010000000000000000000000000080

reth had a change moving to the flat layout, with vectors generated from geth (paradigmxyz/reth#27095); it was closed on 2026-10-07 without being merged.

Could the authors confirm that the flat layout is the intended one, and write request_id into both schemas as devp2p does? For example:

GetCells: [request_id: P, [hash_0: B_32, hash_1: B_32, ...], cell_mask: B_16]
Cells:    [request_id: P, [hash_0: B_32, hash_1: B_32, ...], cells: [[cell_0_0: B_2048, ...], ...], cell_mask: B_16]

Since devp2p already specifies this layout and three clients implement it, this looks like a clarification rather than a wire change.