SKALE Consensus

The purpose of SKALE chain is to order transactions into blocks and then process them by EVM.

SKALE chain is composed of a fixed set of N network nodes that process user transactions in the following phases

The diagram below illustrates typical transaction flow:

Note that in addition to normal block processing, a node can receive blocks through block catchup mechanism.

Block catchup means that the node does not participate in block consensus. Instead, it simply downloads committed blocks from other nodes, verifying block signatures. block catchup typically happens when a node is powered on after being offline. Block catchup can also be used by third-party nodes that do not participate in core chain, such a archive nodes.

block consensus and block catchup run in parallel. This means that every node in addition in normal block consensus procedure makes periodic random connections to other nodes, to attempt to download ready committed blocks.

The blockchain provides a guarantee that every transaction is included into the chain only once. This means, in particular, that when a node commits a block, the node will remove the transactions included in the block from the pending transaction queue.

Each node runs skaled, SKALE software blockchain agent.

skaled is composed of:

SKALE is provably secure. This means one can prove two qualities of the blockchain

Provable security means that under certain mathematical assumptions, SKALE chain will always be consistent and lively, no matter what the attacker does.

The mathematical assumptions for provable security are specified below.

Since real Internet sometimes drops messages on the way without delivering them, the eventual delivery guarantee is achieved in practice by retransmissions. The sending node will make multiple attempts to transfer message to the receiving node, until the transfer is successful and is confirmed by the receiving node.

Each sending node maintains a separate outgoing message queue for each receiving node. To schedule a message for delivery to a particular node, message is placed into the corresponding outgoing message queue.

Each outgoing message queue is serviced by a separate program thread. The thread reads messages from the queue and attempts to transfer them to the destination node. If the destination node temporarily does not accept messages, the thread will keep initiating transfer attempts until the message is delivered. The destination node can, therefore, temporarily go offline without causing messages to be lost.

Since there is a dedicated message sending thread for each destination node, messages are sent independently. Failure of a particular destination node to accept messages will not affect receipt of messages by other nodes.

In the remainder of this document, anywhere where it is specified that a message is sent from node A to B, we mean reliable independent delivery as described above.

Each node stores consensus state. For each round of consensus, consensus state includes the set of proposed blocks, as well as the state variables of the protocols used by the consensus round.

The state is stored in non-volatile memory and preserved across reboots.

During A, a node will temporarily become unavailable. After a reboot, messages destined to the node will be delivered to the node. Therefore, a reboot does not disrupt operation of asynchronous consensus.

Since consensus protocol state is not lost during a reboot, a node reboot will be interpreted by its peers as a temporarily slowdown of network links connected to the node.

A is an event, where a node loses all of parts of the consensus state. For instance, a node can lose received block proposals or values of protocol variables.

A hard crash can happen in case of a software bug or a hardware failure. It also can happen if a node stays offline for a very long time. In this case, the outgoing message queues of nodes sending messages to this node will overflow, and the nodes will start dropping older messages. This will lead to a loss of a protocol state.

This specification specifies one hour as a default lifetime of a message which has been placed into an outgoing queue. Messages older than one hour may be dropped from the message queues. A reboot, which took less than an hour is, therefore, guaranteed to be a a normal reboot.

Hard crashes are permitted by the consensus protocol, as long as not too many nodes crash at the same time. Since a crashed node does not conform to the consensus protocol, it counts as a Byzantine node for the consensus round, in which the state was lost. Therefore, only a limited number of concurrent hard crashes can exist at a given moment in time. The sum of crashed nodes and byzantine nodes can not be more than t in the equation (1). Then the crash is qualified as a limited hard crash.

During a limited hard crash, other nodes continue block generation and consensus. The blockchain continues to grow. When a crashed node is back online, it will sync its blockchain with other nodes using a catchup procedure described in this document, and start participating in consensus.

A widespread crash is a crash where the sum of crashed nodes and Byzantine nodes is more than $t$.

During a widespread crash a large proportion of nodes or all nodes may lose the state for a particular round and consensus progress may stall. The blockchain, therefore, may lose its liveliness.

Security of the blockchain will be preserved, since adding a new block to blockchain requires a supermajority threshold signature of nodes, as described later in this document.

The simplest example of a widespread crash is when more than 1/3 of nodes are powered off. In this case, consensus will stall. When the nodes are back online, consensus will start working again.

In real life, a widespread crash can happen due to to a software bug affecting a large proportion of nodes. As an example, after a software update all nodes in an schain may experience the same bug.

In a case of a catastrophic failure a separate failure resolution protocol is used to restart consensus.

First, nodes will detect a catastrophic failure by detecting absence of new block commits for a long time.

Second, nodes will execute a failure recovery protocol that utilizes Ethereum main chain for coordination. Each node will stop consensus operation. The nodes will then sync their blockchains replicas, and agree on time to restart consensus.

Finally, after a period of mandatory silence, nodes will start consensus at an agreed time point in the future.

Each node stores a sequence of blocks. Blocks are constructed from transactions submitted by users.

The following properties are guaranteed:

An honest node is a node that behaves according to the rules described in this document. A Byzantine node can behave in arbitrary way, including doing nothing at all.

The goal of a Byzantine node is to either violate the liveliness property of the protocol by preventing the blockchain from committing new blocks or violate the consistency property of the protocol by making two different nodes commit two different blocks having the same block ID.

It is assumed that out of N total nodes, $t$ nodes are Byzantine, where less the following condition is satisfied.

or

The above condition is well known in the consensus theory. There is a proof that shows that secure asynchronous consensus is impossible for larger values of $t$.

It is easy to show that if a security proof works for a certain number of Byzantine nodes, it will work for a fewer Byzantine nodes. Indeed, an honest node can always be viewed as a Byzantine node that decided to behave honestly. Therefore, in proofs, we always assume that the system has the maximum allowed number of Byzantine nodes

In this case the number of honest nodes is

Note, that it is beneficial to select N in such a way that is divisible by 3. Otherwise an increase in N does not lead to an increase in the maximum allowed number of Byzantine nodes.

As an example, for we get . For we get too, so an increase in N does not improve Byzantine tolerance.

In this specification, we assume that the N is always selected in such a way that is divisible by 3.

In this case, expressions simplify as follows

Consensus uses voting rounds. It is, therefore, important to proof some basic mathematical properties of voting.

Typically, a node will vote by signing a value and transmitting it to other nodes. To count votes, a receiving node will count received signatures for a particular value v.

The number of Byzantine nodes is less than a simple majority of honest nodes.

This directly follows from the fact that , and, therefore, a simple majority of honest nodes is

We define supermajority as a vote of at least nodes.

A vote of all honest nodes is a supermajority.

Proof: this comes from the fact that .

If a particular message was signed by a supermajority vote, at least a simple majority of honest nodes signed this message

Even if all Byzantine nodes participate in a supermajority vote, the number of honest votes it needs to receive is

which is exactly the simple majority of honest nodes s.

If honest nodes are required to never sign conflicting messages, two conflicting messages can not be signed by a supermajority vote.

Proof: lets A and B be two conflicting messages. Since a particular honest node will sign either A or B, both A and B can not get simple majority of honest nodes. Since a supermajority vote requires participation of a simple majority of honest nodes, both A and B can not reach a supermajority, even if Byzantine nodes vote for both.

A supermajority vote, is, therefore, an important conflict avoidance mechanism. If a message is signed by a supermajority vote, it is guaranteed that no conflicting messages exist. As an example, if a block is signed by a supermajority vote, it is guaranteed that no other block with the same block ID exists.

Our protocol uses threshold signatures for supermajority voting.

Each node is supposed to be in possession of BLS private key share PKSI. Initial generation of key shares is performed using joint-Feldman Distributed Key Generation (DKG) algorithm that is described in this document. DKG algorithm is executed when an schain is created.

Nodes are able to collectively issue supermajority threshold signatures on messages, where the threshold value is equal to the supermajority vote . For instance for N = 16, the threshold value is 11.

BLS threshold signatures are implemented as described in the paper of by Boldyreva. BLS threshold signatures require a choice of elliptic curve and group pairing. We use elliptic curve (altBN256) and group pairing (optimal-Ate) implemented in Ethereum Constantinople release.

To verify the signature, one uses BLS public key PK. This key is computed during the initial DKG algorithm execution. The key is stored in SKALE manager contract on Ethereum mainnet and is available to anyone.

Each user transaction T is assumed to be an Ethereum-compatible transaction, represented as a sequence of bytes.

Each block is a byte string, which includes a header followed by a body.

Block header is a JSON object that includes the following:

Note: All integers in this spec are unsigned 64-bit integers unless specified otherwise.

BLOCK BODY is a concatenated transactions array of all transactions in the block.

Block hash is calculated by taking 256-bit Keccack hash of block header concatenated with block body, while omitting CURRENT BLOCK HASH, CURRENT BLOCK SIG, and CURRENT BLOCK TSIG from the header. The reason why these fields are omitted is because they are not known at the time block is hashed and signed.

Note: Throughout this spec we use SHA-3 as a secure hash algorithm.

A node or a third party can verify the block by verifying a threshold signature on it and also verifying the previous block hash stored in the block. Since the threshold signature is a supermajority threshold signature and since any honest node will only sign a single block at a particular block ID, no two blocks with the same block ID can get a threshold signature. This provides security against forks.

A block starts as a block proposal. A block proposal has the same structure as a block, but has the threshold signature element unset.

Node concurrently make proposals for a given block ID. A node can only make one block proposal for a given block ID.

Once a block proposal is selected to become a block by consensus, it is signed by a supermajority of nodes. A signed proposal is then committed to the end of the chain on each node.

Each node will keep a pending transactions queue. The first node that receives a transaction will attempt to propagate it to all other nodes in the queue. A user client software may also directly submit the transaction to all nodes.

When a node commits a block to its blockchain, if will remove the matching transactions from the transaction queue.

Each transaction requires payment of a gas fee, compatible with ETH gas fee. The gas fee can be paid in native currency of the SKALE chain (sFUEL) or in Proof of Work. The gas price is adjusted after each committed block. It is decreased if the block has been underloaded, meaning that the number of transactions in the block is less than 70 percent of the maximum number of transactions per block, and is increased if the block has been overloaded.

Typically pending queues of all nodes will have similar sets of messages, with small differences due to network propagation times.

When node A needs to send to node B a block proposal P, A does need the send the actual transactions that compose P. A only needs to send transaction hashes, and then B will reconstruct the proposal from hashes by matching hashes to messages in its pending queue.

In particular, for each transaction hash in the block proposal, the receiving node will match the hash to a transaction in its pending queue. Then, for transactions not found in the pending queue, the receiving node will send a request to the sending node. The sending node will then send the bodies of these transactions to the receiving node. After that the receiving node will then reconstruct the block proposal.

For a particular node, the blockchain consists of a range of committed blocks Bi starting from B0 end ending with BTIPID, where TIPID is the ID of the largest known committed block. Block ids are sequential positive integers. Blocks are stored in non-volatile storage.

New blocks a created by running consensus rounds. Each round corresponds to a particular BLOCKID.

At the beginning of a consensus round, each node makes a block proposal.

When a consensus round completes for a particular block, one of block proposals wins and is signed using a supermajority signature, becoming a committed block.

Due to a randomized nature of consensus, the is a small probability that consensus will agree on an empty block instead of agreeing on any of the proposed blocks. In this case, an empty block is pre-committed to a blockchain.

There are two ways, in which blockchain on a particular node grows and TIPID is incremented:

Normal consensus operation: during normal consensus, a node constantly participates in consensus rounds, making block proposals and then committing the block after the consensus round commits.

Catchup: a separate catchup agent is continuously running on a node. The catchup engine is continuously making random sync connections to other nodes. During a sync both nodes sync their blockchains and block proposal databases.

If during catchup, node A discovers that node B has a larger value of TIPID, A will download the missing blocks range from B, and commit it to its chain after verifying supermajority threshold signatures on the received blocks.

Note that both normal and catchup operation append blocks to the blockchain. The catchup procedure intended to catchup after hard crashes.

When the node comes online from a hard crash, it will immediately start participating in the consensus for new blocks by accepting block proposals and voting according to consensus mechanism, but without issuing its own block proposals. Since a block proposal requires hash of the previous block, a node will only issue its own block proposal for a particular block id once it a catch up procedure moves the TIPID to a given block id.

Liveliness property is guaranteed under hard crashes if the following is true: normal consensus guarantees liveliness properly, catch-up algorithm guarantees eventual catchup, and if the number of nodes in a hard crashed state at a given time plus the number of Byzantine nodes is less or equal N ⅓.

Since the normal consensus algorithm is resilient to having Byzantine nodes, normal consensus will still proceed if we count crashed nodes as Byzantine nodes and guarantee that the total number of Byzantine nodes is less than . When a node that crashed joins the system back, it will immediately start participating in the new consensus rounds. For the consensus rounds that it missed, it will use the catchup procedure to download blocks from other nodes.

A node is required to create a block proposal directly after its TIPID moves to a new value. TIPID will be incremented by $1$ once a previous consensus round completes. TIPID will also move, if the catchup agent appends blocks to the blockchain.

To create a block a node will:

Note that the node does not remove transactions from the pending queue at the time of proposal. The reason for this is that at the proposal time there is no guarantee that the proposal will be accepted.

MAX BLOCK SIZE is the maximum size of the block body in bytes. Currently we use MAX BLOCK SIZE = 8 MB. FUTURE: We may consider self-adjusting block size to target a particular average block commit time, such as 1s.

BEACON TIME is time between empty block creation. If no-one is submitting transactions to the blockchain, empty beacon blocks will be created. Beacon blocks are used to detect normal operation of the blockchain. The current value of BEACON TIME is 3s.

Once a node creates a block proposal it will communicate it to other nodes using the data data availability protocol described below.

The data availability protocol guarantees that if the the protocol completes successfully, the message is transferred to the supermajority of nodes.

The five-step protocol is described below:

Data Availability Receipt Requirement In further consensus steps, any node voting for proposal P is required to include S in the vote. Honest nodes will ignore all votes that do not include the supermajority signature S.

The protocol used above guarantees data availability, meaning that any proposal P that wins consensus will be available to any honest nodes. This is proven in steps below.

Liveliness. If A is honest, than the five-step protocol above will always complete. By completion of the protocol we mean that all honest nodes will receive S. Byzantine nodes will not be able to stall the protocol.

By properties of the send operation discussed in Section 1.2 all sends in Step 1-3 are performed in parallel. In step 4 node A waits to receive signature shares for the supermajority of nodes. This step will always take fine time, even if Byzantine nodes do not reply. This comes from the fact that there is a supermajority of honest nodes. In step 5 S will be added to outgoing message queues of all nodes. Since honest nodes do accept messages, S will ultimately be delivered to all honest nodes as described in Section 1.2.

If a proposal has a supermajority signature, it is was communicated to and stored on the simple majority of honest nodes.

The proof directly follows from Lemma 3, and from the fact that an honest node B only signs the proposal after B has received and stored the proposal.

If a proposal wins consensus and is to be committed to the blockchain, then any honest node X that does not have the proposal can efficiently retrieve it.

First, a proposal will not pass consensus without having a supermajority signature. This comes from the fact that all nodes voting for the proposal will need to include S in the vote.

By the properties of binary Byzantine agreement protocol of Mostéfaoui at al., a proposal can win consensus only if at least one honest node votes for the proposal. A proposal without a signature will never win consensus, since an honest node will never vote for it.

Therefore, if a proposal won consensus, it is guaranteed to have a supermajority signature.

Second by previous lemma, if a proposal has a supermajority signature, any honest node can retrieve it. This completes the proof.

The protocol discussed above is important because it guarantees that if a proposal wins consensus, all honest nodes can get this proposal from other honest nodes and add it to the blockchain.

The consensus described above uses an Asynchronous Binary Byzantine Agreement (ABBA) protocol (ABBA). We currently use ABBA from Mostéfaoui et. all. Any other ABBA protocol P can be used, as long as it has the following properties

Note that, an ABBA protocol typically outputs a random number COMMON COIN as a byproduct of its operation. We use this COMMON COIN as a random number source.

A consensus round R is executed for each BLOCKID and has the following properties:

Liveliness: each consensus round R will always produce a block in a finite time.

The proof follows from the fact that each R runs N parallel versions of ABBA binary consensus, and from the liveliness property of the ABBA consensus

Consistency: each consensus round will produce the same result P on all nodes

This follows from the consistency property of the ABBA consensus and from the fact that the consensus round algorithm is deterministic and does not depend on the node where it is executed.

Data Availability: the winning proposal P is available to any honest node.

This follows from the fact, that ABBA will not return consensus yes vote unless at least one honest node initially votes yes, and from the fact that an honest node will not vote yes unless it has a data availability proof (threshold signature S).

For each node, 33 percent of the storage is assigned to blockchain, 33 percent to EVM and 33 to the rest of the system, such as consensus state.

If blockchain storage is exhausted, the old blocks will be deleted to free storage in increments of 1024 blocks.

If EVM/Solidity storage is exhausted, EVM will start throwing \\"OutOfStorage\\" errors until storage is freed.

If consensus storage is exhausted, the consensus agent will start erasing items such as messages in the message outgoing queues, in the order of item age, from oldest to newest.

The goal is to provide EVM/Solidity compatibility, except the cases documented in this specification. The compatibility is for client software, in particular Metamask, Truffle, Web3js and Web3py.

Once a block is finalized on the chain, it is passed to EVM, and each transaction is sequentially executed by the EVM one after another. We currently use unmodified Ethereum EVM, therefore there should not be compatibility issues. Once Ethereum finalizes EWASM version of EVM, we will be able to plug in in.

EVM has pluggable storage backend database to store EVM/Solidity variables we simplified and sped up the storage by using LevelDB from Google. Each variable in EVM is stored as a key value in LevelDB where the key is the sha3 hash of the virtual memory address and the value is the 256 bit value of the variable. In EVM all variables have 256 bits.

We do not charge users gas for transactions.

We do have a protection against Denial of Service attacks.

Each transaction needs to submit proof of work (PoW) proportional to the amount of gas that the transaction would have used if we would charge for transactions. We are currently using the same PoW algorithm as Ethereum.

This PoW is calculated in the browser or other client that submits a transaction and is passed together with the transaction. If the transaction does not include the required PoW it will be rejected.

We are still researching the formula for k. Ideally k should go down if the chain is underloaded and increase if the chains starts to be overloaded.

The goal is to provide compatible JSON client API for client software such as Web3js, Web3py, Metamask and Truffle.

Existing clients such Web3js connect to a single node, which creates security problem for Solidity read requests that read variables.

Transactions involve a consensus of the entire blockchain, but Solidity read requests interact with a single node. Therefore, an malicious node, such as Infura, can prove a user incorrect information on, e.g. the amount of funds the user has in possession.

Therefore, in the future we will need to add multi-node requests where the first node that receives the request passes it to all others and collects a tsig.