Yes, BLAKE3 is a Merkle tree, and what Bao is doing is conceptually similar to BitTorrent. (BEP 30 is even cited in the BLAKE3 spec.) Bao's goal is to be a general-purpose wire format and implementation library, for any application that needs verified streaming. Applications that use BLAKE3 for generic file hashing today can also start doing verified streaming at any point in the future, without needing any other trusted metadata.
root
/ \
parent1 parent2
/ \ / \
chunk1 chunk2 chunk3 chunk4
Say we build a slice containing only chunk2. The contents of that slice will be: [4096][root][parent1][chunk2]
The parent node parent1 is [hash of chunk1][hash of chunk2], and the root node is [hash of parent1][hash of parent2]. The 8-byte total length on the front (4096) is the same as in the full encoding; it defines the structure of the tree, so the decoder knows how many parent nodes to expect.For more details: https://github.com/oconnor663/bao/blob/master/docs/spec.md
This is my first time looking at the Dat protocol. Thanks for pointing that out.
I just looked over the doc at https://datprotocol.github.io/how-dat-works, particularly the section on the tree hash. It looks pretty reasonable, and definitely similar to BLAKE3.
I think one practical downside of the Dat approach is that root finalization is always done as a separate hashing step. So for example, if I hash the short string "hello world", I have to compute one chunk hash and then one root hash, for a total of two calls to the compression function. Compare that to BLAKE3, which will only do a single compression there, by setting both the chunk flags and the root flag for that one block. This difference is probably totally negligible for Dat's intended use case, where network round trip time is going to completely dwarf hashing time for small inputs. But for a general-purpose hash function you don't want to cut your small input performance in half.
A bigger philosophical difference is that Dat doesn't specify a particular chunk size. Again this makes sense for their use case, since you might want to experiment with fancy things like Rabin fingerprinting. Looks like they discuss that here: https://stackoverflow.com/a/51604110/823869. But for a cryptographic hash, it's important that every file has exactly one hash, and a fixed chunk size is required. Another upside of a fixed chunk size is that you really cut down on metadata; the 8-byte length at the front of an encoding or a slice determines the structure of the entire tree.
- its hash - the hash of the brother so you can derive the hash of the parent - the hash of the uncle so you can derive the grand parent - etc... Up until the last brother before the root, so you can verify the root hash
In bao it seems to be the opposite: the included hashes are the ones from the direct ancestors ? So how can bao trust the whole chain ? Or maybe there's something in BLAKE3 that explains how "Chaining Values" fix it ?