arnaucube 01408fba86 | 3 years ago | |
---|---|---|
.github/workflows | 3 years ago | |
testvectors/circom | 3 years ago | |
.gitignore | 3 years ago | |
.golangci.yml | 3 years ago | |
LICENSE | 3 years ago | |
README.md | 3 years ago | |
addbatch_test.go | 3 years ago | |
circomproofs.go | 3 years ago | |
circomproofs_test.go | 3 years ago | |
dbg.go | 3 years ago | |
go.mod | 3 years ago | |
go.sum | 3 years ago | |
hash.go | 3 years ago | |
hash_test.go | 3 years ago | |
helpers_test.go | 3 years ago | |
tree.go | 3 years ago | |
tree_test.go | 3 years ago | |
utils.go | 3 years ago | |
vt.go | 3 years ago | |
vt_test.go | 3 years ago |
arbo: tree in Esperanto.
MerkleTree implementation in Go. Compatible with the circomlib implementation of the MerkleTree, following the specification from https://docs.iden3.io/publications/pdfs/Merkle-Tree.pdf and https://eprint.iacr.org/2018/955.
Allows to define which hash function to use. So for example, when working with zkSnarks the Poseidon hash function can be used, but when not, it can be used the Blake2b hash function, which has much faster computation time.
The method tree.AddBatch
is designed for the cases where there is a big amount of key-values to be added in the tree. It has the following characteristics:
tree.Add
method, most of the intermediate nodes will be recalculated each time that a new leaf is addedAs result, the method tree.AddBatch
goes way faster thant looping over tree.Add
, and can compute the tree with parallelization, so as more available CPUs, faster will do the computation.
// create new database
database, err := db.NewBadgerDB(c.TempDir())
// create new Tree with maxLevels=100 and Blake2b hash function
tree, err := arbo.NewTree(database, 100, arbo.HashFunctionBlake2b)
key := []byte("hello")
value := []byte("world")
// Add a key & value into the merkle tree
err = tree.Add(key, value)
// There are cases where multiple key-values (leafs) are going to be added to a
// Tree, for these cases is more efficient to use:
invalids, err := tree.AddBatch(keys, values)
// generate the merkle proof of a leaf by it's key
value, siblings, err := tree.GenProof(key)
// verify the proof
verified, err := arbo.CheckProof(tree.hashFunction, key, value, tree.Root(), siblings)
if !verified {
fmt.Println("proof could not be verified")
}
// get the value of a leaf assigned to a key
gettedKey, gettedValue, err := tree.Get(key)
// update the value of a leaf assigned to a key
err = tree.Update(key, value)
// dump the tree (the leafs)
dump, err := tree.Dump(nil) // instead of nil, a root to start from can be used
// import the dump into a tree
err = tree.ImportDump(dump)
// print graphviz diagram of the tree
err = tree.PrintGraphviz(nil) // instead of nil, a root to start from can be used
Arbo is designed to be compatible with circom merkle tree's snark-friendly merkletree. The only change needed is the hash function used for the Tree, for example using the Poseidon hash function:
tree, err := arbo.NewTree(database, 32, arbo.HashFunctionPoseidon)
Be aware of the characteristics of this kind of hashes, such as using values inside the finite field used by the hash, and also the computation time.
The interface of arbo uses byte arrays, and for the case of these kind of hashes
(that usually work directly with finite field elements), arbo expects those
values to be represented by little-endian byte arrays. There is a helper method
to convert a *big.Int
to []byte
using little-endian:
bLen := tree.HashFunction().Len()
kBigInt := big.NewInt(100)
// convert *big.Int to byte array
kBytes := arbo.BigIntToBytes(bLen, kBigInt)
// convert byte array to *big.Int
kBigInt2 := arbo.BytesToBigInt(kBytes)