How libp2p works:

An oversimplified version of how gossip works is:

1. A “message” is submitted to libp2p to be distributed to all nodes
2. The message is identified by its hash, and the hash is distributed widely
3. Each node keeps track of all the hashes it has (IHaves) and all the hashes it has seen but doesn’t have the full message yet (IWant)
4. Each missing message is requested individually and added to IHaves

The current problem:

From the description above, you can see that the scaling properties of libp2p are dependent on the *number of messages *****(Rather that the total bytes). libp2p works well for other blockchain projects because they have a relatively smaller number of large messages

• Modern L1 chains with 0.1s block times only need to gossip 10 blocks / s
• Ethereum, which gossips individual Txns, has to gossip only (~300 txns/block = 25 txns/s)

In contrast, Farcaster, which gossips each individual message is already gossiping 1,000 messages/s, with peaks exceeding 10,000 messages/s (with each message being quite small, ~220 bytes)

This means the libp2p libraries are keeping track of 100s of thousands of hashes, and requesting each one individually, running into scaling limitations

Potential Solutions

To scale gossip, we need to either (1) Reduce the number of messages sent on the gossip network or (2) Use an alternate channel to distribute individual messages

1. Use “Bundles”

We’ll invent a new idea of a “bundle”, which is the set of all messages submitted only via source=rpc in the same second to a node. The node, at the end of the second, will collect all successfully merged messages into a “bundle” and gossip out the whole bundle.