A data availability blockchain with sub-linear full block validation

Over the past few months the idea of using blockchains as data availability “engines” to post state transitions, rather than execute them, has gotten more traction lately. However much of this discussion hasn’t considered what an optimal chain designed exclusively for data availability and nothing else would look like, and what properties it would have. Such a chain could potentially be implemented as an Ethereum sidechain. Last year I proposed one. I’ve done some further work on this and uploaded a draft paper at https://arxiv.org/abs/1905.09274 and a prototype at https://github.com/musalbas/lazyledger-prototype. I’ll summarise it here.

Reducing block verification to data availability verification

We can build a blockchain where the only rule that determines whether a block is valid or not is whether the data behind that block is available or not. Consequently, consensus nodes and block producers don’t have to process or care at all what messages are actually inside the block. You can thus reduce the problem of block verification to data availability verification. We know how to do data availability verification of blocks with a O(\sqrt{\mathsf{blocksize}} + log(\sqrt{\mathsf{blocksize}})) bandwidth cost, using data availability proofs that use erasure coding and sampling. Nodes who want to verify that a block is available need to sample a fixed number of chunks to get their desired data availability probability, plus Merkle proofs for those chunks (from row/column roots) which are each log(\sqrt{\mathsf{blocksize}}) sized, plus 2\sqrt{\mathsf{blocksize}} row and column Merkle roots.

Note that this isn’t quite magic, because you still need to assume a minimum number of nodes making sample requests, so that the network can recover the full block, just like in a peer-to-peer file-sharing network such as BitTorrent. So if you want to increase your block size, you either need to make sure that there are enough nodes in your network, or increase the number of samples each node makes (which would no longer make it sub-linear), or perhaps a mixture of both.

What greatly interests me about this, is that you therefore have a system whose throughput increases by adding nodes that are not participating in the consensus or producing blocks. If you increase the number of nodes making samples, you can increase maximum safe block size. This would be similar to for example if the throughput of Bitcoin increased by adding more non-mining full nodes. Effectively by removing any need to do state transition execution to verify blocks, and making data availability the only prerequisite of block validity, we can have similar scalability properties to peer-to-peer filesharing systems like BitTorrent, e.g. more nodes = more capacity. (Note that you can’t have the system automatically “count” nodes and adjust block size based on that since they could easily be Sybils, however.)

Sovereign client-side applications

Since no transactions are executed by block producers, and invalid transactions may be posted on the chain, all transaction execution is done by the actual end-users of each application (smart contract), similar to how Mastercoin uses Bitcoin as a base layer for posting messages. However, we could create an application model that supports arbitrary applications defined by clients who are using the blockchain as a place to just post messages. What’s kind of neat about this is that because all the application logic is executed client-side, applications can be written in any arbitrary language or environment, since they’re not executed by the chain.

The main important principal here is application state sovereignty: client nodes must be able to execute all of the messages relevant to the applications they use to compute the state for their applications, without needing to execute messages from other applications, unless other specific applications are explicitly declared as dependencies (in which case the dependency application’s messages are considered relevant).

This means that if there are two applications A and B using the chain that are “independent” from each other, users of application A should not have to download or process the data of B, so if B suddenly becomes more popular, the workload for users of application A stays roughly the same, and vice versa.

Philosophically, this can be thought of as a system of ‘virtual’ sidechains that live on the same chain, in the sense that transactions associated with each application only need to be processed by users of those applications, similar to the fact that only users of a specific sidechain need to process transactions of that sidechain. However, because all applications share the same chain, the availability of the data of all their transactions are equally and uniformly guaranteed by the same consensus group, unlike in traditional sidechains where each sidechain may have a different (smaller) consensus group.

An interesting consequence of this is that as application logic is defined and interpreted off-chain, if some users of an application wanted to update an application’s logic or state, they can do so without requiring a hard fork of the chain. If other users don’t follow the update, the two groups of users will have different states for the same applications, which is like an “off-chain” hard fork.

Efficient per-application message retrieval

To help clients/end-users get the complete set of messages/transactions per block for their applications, without having to download the messages of other applications, we can construct our block headers to allow nodes with the complete set of data (storage node) to facilitate this.

Firstly, each application assigns itself a ‘namespace identifier’, such that any message M included in a block can be parsed by some function \mathsf{namespace}(M) = \mathsf{nid} to get its namespace identifier \mathsf{nid}. Then, when computing the Merkle root of all the messages in the block to use in the block header, we can use an ordered Merkle tree where each node is flagged by the range of namespaces that can be found in its reachable leafs. In the example Merkle tree below, \mathsf{namespace}(M) simply returns the first four hex characters in the message. If the Merkle tree is constructed incorrectly (e.g. not ordered correctly or flagged with incorrect namespaces), then the data availability proof should fail (e.g. there would be a fraud proof of incorrectly generated extended data).

image image

Clients can then query storage nodes for messages of specific namespaces, and can easily verify that the storage node has returned the complete set of messages for a block by reading the flags in the nodes of the proofs.

DoS-resistance

Transaction fees and maximum block sizes are possible under this model without effecting application state sovereignty; see paper for more details. This is necessary to disincentivise users from flooding other applications’ namespaces with garbage.

tl;dr/summary

We can have a dedicated data availability blockchain with the following properties:

  • “Fully verifying” a block (in the sense of a full node) has a O(\sqrt{\mathsf{blocksize}} + log(\sqrt{\mathsf{blocksize}})) bandwidth cost, using data availability proofs based on erasure coding and sampling.
  • Similar to a peer-to-peer file-sharing network, the transaction throughput of the blockchain can “securely” increase by adding more nodes to the network that are not participating in block production or consensus.
  • Users can download and post state transitions to the blockchain for their own “sovereign” applications, without downloading the state transitions for other applications.
  • Changing application logic does not require a hard fork of the chain, as application logic is defined and interpreted off-chain.
15 Likes

I’m nearly done with the paper, but have a question.

What happens when a data chain block is validated but then erased far in the future? Does it invalidate the entire chain? How are storage nodes guaranteed to persist old blocks entirely, forever?

1 Like

This same problem is present in normal blockchains as well. The paper doesn’t address it, nor should it.

1 Like

Really? This is literally a paper about data availability. It seems silly to gloss over what happens when it fails to do so.

Current blockchains do indeed have similar issues, but you’re right it should probably be mentioned in the paper.

If a (full) node accepts a block as valid, but it and the network later loses that data, then new nodes can no longer be bootstrapped, since they won’t be able to get the necessary data to validate the chain. Ripple apparently has this issue, as the first 30,000 blocks from its chain are missing. :rofl:

The main purpose of data availability proofs is to prevent block producers from not releasing data, by checking that the public network has the data at a certain point in time. After that, it is assumed that there will always be at least one public copy of the data somewhere.

I would argue that this design actually makes it less likely for the network to lose the data, compared to current blockchain designs. This is because with this design, nodes with low resources can contribute to the storage of the blockchain by storing block samples. With current designs, you would typically have to run a full node and store everything to contribute to storage, at least directly anyway.

3 Likes

It’s nice that this functionality is built-in at the protocol level, though we can do something similar with normal blockchain systems at the client layer (i.e., non-consensus). Just have nodes by default store random erasure code samples of old blocks instead of all blocks, as they do now.

With the added storage burden, valid blocks could contain logarithmically many backlinks to valid block headers. Maybe this would allow consistent bootstrapping at any height if block data is lost (maybe carrying succinct state forward in time).

@musalbas what other ideas/problems are you working through in relation to this?

With the added storage burden, valid blocks could contain logarithmically many backlinks to valid block headers. Maybe this would allow consistent bootstrapping at any height if block data is lost (maybe carrying succinct state forward in time).

How would that prevent data being lost? Even if you have backlinks to headers, that doesn’t guarantee the data behind the headers is available.

@musalbas what other ideas/problems are you working through in relation to this?

At the moment I’m planning to redraft the paper to focus more on the scale-out property I mentioned, where on-chain transaction capacity can be increased by adding nodes. This could be seen as an alternative “sharding” strategy.

But I think there’s two things to think about in relation to the application model:

  1. How can you build light clients for specific applications given that state commitments aren’t verified on-chain by the consensus?
  2. Is it possible to translate all blockchain use case into the sovereign application model, where you can’t force other contracts to take a dependency on your contract? What are the limitations? It’s not obvious for example how you could have a contract that has its own balance with another token contract - how can the contract withdraw tokens to some address, without requiring the token contract to validate the state of the contract withdrawing the tokens? (e.g. in Ethereum, msg.sender can be a contract, but it’s not clear how that would be possible with this design without requiring other contracts to verify the state of the msg.sender contract)

No I only meant for reliably bootstrapping to the current state, if previous state is deleted.

Great post!

  1. How can you build light clients for specific applications given that state commitments aren’t verified on-chain by the consensus?

I guess the most of DAG-based consensus protocols have similar issues about light clients.
Out of curiosity, is there any requirements for a protocol to support light clients other than verified state commitments?

I had always assumed a model for light clients that relied on verified state commitments (or verified transaction commitments). How would it work otherwise?

1 Like