Spec — Rust Crate — Rust Docs

Caution: Bao is intended to be a cryptographic hash function, but it hasn't yet been reviewed. The output may change prior to the 1.0 release.

Bao (rhymes with bough 🌳) is a general purpose tree hash for files. Here's the full specification. Here's a talk I gave about it in November 2018. What makes a tree hash different from a regular hash? Depending on how many cores you've got in your machine, the first thing you might notice is that it's ten times faster than SHA-2:

Why is bao hash so fast? It's mostly parallelism, multiple threads working on different subtrees at the same time. The demo above is from the 4-core i5-8250U processor in my laptop, but Bao can scale much higher. In-memory benchmarks on a 48-core AWS m5.24xlarge instance hit 91 GB/s of throughput. Bao is also based on BLAKE2b, which was designed to outperform SHA-1, and it uses the fastest SIMD implementation available.

Apart from parallelism, tree hashes make it possible to verify a file piece-by-piece rather than all-at-once. This is done by storing both the input and the entire hash tree together in an encoded file. Clients can stream these files, or do random seeks into them, with the guarantee that every byte they read matches the root hash.

Use case: A secure messaging app might support attachment files by including the hash of an attachment in a message. With a regular hash, the recipient would need to download an entire attachment to verify it, but that's impractical for e.g. large video files. With a Bao hash, the recipient can stream a video, while still verifying each byte as it comes in. (This scenario was in fact the original motivation for the Bao project.)

# Create an input file that's a megabyte of random data.
> head -c 1000000 /dev/urandom > f

# Convert it into a Bao encoded file.
> bao encode f f.bao

# Compare the size of the two files. The encoding overhead is small.
> stat -c "%n %s" f f.bao | column -t
f       1000000
f.bao   1015624

# Note that the `bao hash` of the input file is the same as the
# `bao hash --encoded` of the encoded file, but the latter is faster.
> bao hash f
[some hash...]
> bao hash --encoded f.bao
[the same hash...]
> hash=`bao hash --encoded f.bao`

# Stream decoded bytes from the encoded file, using the hash above.
> cmp f <(bao decode $hash f.bao)

# Observe that using the wrong hash to decode results in an error. This
# is also what will happen if we use the right hash but corrupt some
# bytes in the encoded file.
> bad_hash=`echo $hash | sed s/a/b/`
> cmp f <(bao decode $bad_hash f.bao)
Error: Custom { kind: InvalidData, error: StringError("hash mismatch") }
cmp: EOF on /proc/self/fd/11 which is empty

Encoded files support random seeking, but seeking might not be available or efficent over a network. (Note that one seek in the content usually requires several seeks in the encoding, as the decoder traverses the hash tree level-by-level.) In these situations, rather than e.g. hacking a seek interface into your HTTP client, you can instead request an encoded slice. A slice contains some part of the original file and the parts of the hash tree needed to verify it. Creating a slice requires seeking over the full encoding, but the recipient can then stream the slice without needing to seek. Decoding a slice uses the same root hash as regular decoding, so it doesn't require any preparation in advance from the sender or the recipient.

Use case: A BitTorrent-like application could fetch different slices of a file from different peers, without needing to define the slices ahead of time. Or a distributed file storage application could request random slices of an archived file from its storage providers, to prove that they're honestly storing the file, without needing to prepare or store challenges for the future.

# Using the encoded file from above, extract a 100 KB slice from
# somewhere in the middle. We'll use start=500000 (500 KB) and
# count=100000 (100 KB).
> bao slice 500000 100000 f.bao f.slice

# Look at the size of the slice. It contains the 100 KB of content plus
# some overhead. Again, the overhead is small.
> stat -c "%n %s" f.slice
f.slice 104584

# Using the same parameters we used to create the slice, plus the same
# hash we got above from the full encoding, decode the slice.
> bao decode-slice $hash 500000 100000 f.slice > f.slice.out

# Confirm that the decoded output matches the corresponding section from
# the input file. (Note that `tail` numbers bytes starting with 1.)
> tail --bytes=+500001 f | head -c 100000 > expected.out
> cmp f.slice.out expected.out

# Now try decoding the slice with the wrong hash. Again, this will fail,
# as it would if we corrupted some bytes in the slice.
> bao decode-slice $bad_hash 500000 100000 f.slice
Error: Custom { kind: InvalidData, error: StringError("hash mismatch") }

By default, all of the operations above work with a "combined" encoded file, that is, one that contains both the content bytes and the tree hash bytes interleaved. However, sometimes you want to keep them separate, for example to avoid duplicating a very large input file. In these cases, you can use the "outboard" encoding format, via the --outboard flag:

# Re-encode the input file from above in the outboard mode.
> bao encode f --outboard f.obao

# Compare the size of all these files. The size of the outboard file is
# equal to the overhead of the original combined file.
> stat -c "%n %s" f f.bao f.obao | column -t
f       1000000
f.bao   1015624
f.obao  15624

# Decode the whole file in outboard mode. Note that both the original
# input file and the outboard encoding are passed in as arguments.
> cmp f <(bao decode $hash f --outboard f.obao)

The bao command line utility is published on crates.io as the bao_bin crate. To install it, add ~/.cargo/bin to your PATH and then run:

cargo install bao_bin

To build the binary directly from this repo:

git clone https://github.com/oconnor663/bao
cd bao/bao_bin
cargo build --release
./target/release/bao --help

tests/bao.py is a fully functional second implementation in Python, designed to be as short and readable as possible. It's a good starting point for understanding the algorithms involved, before diving into the Rust code.

The bao library crate includes no_std support if you set default-features = false in your Cargo.toml. Most of the standalone functions that don't obviously depend on std are available. For example, bao::encode::encode is available with a single threaded implementation, but bao::encode::encode_to_vec isn't available. Of the streaming implementations, only hash::Writer is available, because the encoding and decoding implementations rely more on the std::io::{Read, Write, Seek} interfaces. If there are any callers that want to do streaming encoding or decoding under no_std, please let me know, and we can figure out which libcore-compatible traits it makes sense to implement.