quark - quick STARK!
A decentralized horizontally-scaled state machine that can transfer 1,000,000 unique tokens on Uniswap in a single atomic transaction. How?
- UXTO-like storage. Transactions only lock the parts of state they modify, there is no global state lock like in EVM.
- ...that's distributed. Instead of every node storing the entire state, we partition the data in the same way Google's Bigtable/Spanner scales to trillions of rows.
- ...that's decoupled from execution. Self-contained transactions that prove the storage leaves they updated, in
O(log N)time. Built using Cairo STARK's.
✨✨✨ See the proof-of-concept for more info. ✨✨✨
The blockchain state machine is decoupled into sequencing, execution, and storage layers.
Storage nodes store the persistent state of the blockchain. Storage is modeled as a sorted multidimensional map, where the key is (contract, storage_key), the dimensions are (value, time), and the data is partitioned according to row range.
Execution nodes use the storage network for their memory:
- Reads. Every transaction declares the storage leaves it reads from ahead-of-time, and these are fetched in parallel from the storage network at the beginning of tx execution.
- Writes. The execution node generates a STARK proof of a transaction's state transition. This is used to prove the storage writes are authorised to the storage nodes.
Transactions are scheduled in parallel. In order to maintain a consistent view of the machine's memory, the sequencer provides the total ordering of transactions to both the execution and storage layers.
The memory model is best described as multiversion concurrency control, as there are no global locks. Like Bitcoin and Solana, transactions can be scheduled in parallel if they don't modify the same state. Unlike Bitcoin and Solana, the cost of verifying transactions is vastly cheaper - atop Cairo proofs, tx verification scales sublinearly in number of VM steps - O(log^2 n).
Storage is partitioned according to row range, which makes it cheap to fetch large numbers of storage slots from a single contract, as this only touches at most a few nodes.
More information in this system model.
(1)
Sequencer determines global ordering of txs. Users post txs to the sequencer, which enqueues them.
[tendermint 400tps = 2.5ms)
(2)
Executor generates a trace of the state transition. If we can declaratively declare data dependencies, this can be parallelised. Executor writes to the distributed storage journal (a bigtable db).
Assuming good data locality, parallel trace generation.
[100ms to fetch state in parallel from any number of storage nodes]
[10ms to execute transaction in VM]
[1.6s to generate proof]
[100ms to flush writes to storage journal]
[16ms to verify proofs on distributed storage nodes]
Napkin estimate - 2s for a tx, with no hard ceiling on storage or execution steps.
Throughput estimates:
O(log^2 n) proof verification time
Storing state for a popular ERC20 like USDC -
(u256 address, u256 balance)
row size = 64 bytes
1B rows = 64 GB of state
target partition size = 200MB
num partitions = 320
maximum rows per partition = 200MB/64bytes = 3,125,000 rows