This PR reintroduces and further decouples blocks and blobs in EIP-4844,
so as to improve network and processing performance.
Block and blob processing, for the purpose of gossip validation, are
independent: they can both be propagated and gossip-validated
in parallel - the decoupled design allows 4 important optimizations
(or, if you are so inclined, removes 4 unnecessary pessimizations):
* Blocks and blobs travel on independent meshes allowing for better
parallelization and utilization of high-bandwidth peers
* Re-broadcasting after validation can start earlier allowing more
efficient use of upload bandwidth - blocks for example can be
rebroadcast to peers while blobs are still being downloaded
* bandwidth-reduction techniques such as per-peer deduplication are more
efficient because of the smaller message size
* gossip verification happens independently for blocks and blobs,
allowing better sharing / use of CPU and I/O resources in clients
With growing block sizes and additional blob data to stream, the network
streaming time becomes a dominant factor in propagation times - on a
100mbit line, streaming 1mb to 8 peers takes ~1s - this process is
repeated for each hop in both incoming and outgoing directions.
This design in particular sends each blob on a separate subnet, thus
maximising the potential for parallelisation and providing a natural
path for growing the number of blobs per block should the network be
judged to be able to handle it.
Changes compared to the current design include:
* `BlobsSidecar` is split into individual `BlobSidecar` containers -
each container is signed individually by the proposer
* the signature is used during gossip validation but later dropped.
* KZG commitment verification is moved out of the gossip pipeline and
instead done before fork choice addition, when both block and sidecars
have arrived
* clients may verify individual blob commitments earlier
* more generally and similar to block verification, gossip propagation
is performed solely based on trivial consistency checks and proposer
signature verification
* by-root blob requests are done per-blob, so as to retain the ability
to fill in blobs one-by-one assuming clients generally receive blobs
from gossip
* by-range blob requests are done per-block, so as to simplify
historical sync
* range and root requests are limited to `128` entries for both blocks
and blobs - practically, the current higher limit of `1024` for blocks
does not get used and keeping the limits consistent simplifies
implementation - with the merge, block sizes have grown significantly
and clients generally fetch smaller chunks.
Jacek Sieka
Danny Ryan
Hsiao-Wei Wang