Consensus
Last updated
Last updated
Consensus refers to the procedure Mel uses to produce and decide on a canonical blockchain history.
In broad strokes:
Mel uses a fixed-term proof of stake to decide consensus participants, known as stakers. Stakers must stake SYM, a special PoS token, for integer multiples of 200,000-block (~70-day) epochs, and changes in the effective list of stakers only happen on epoch boundaries.
Each block is decided by a Byzantine fault-tolerant (BFT) consensus algorithm that eventually produces a consensus proof — signatures from stakers owning at least 2/3 of the staked SYM. This produces immediate finality, meaning the block can never be reverted, and Mel can never have "block reorgs". Interestingly, the exact consensus algorithm used is not part of the core protocol rules and anything can be used as long as it produces consensus proofs of the right format; the current implementation, however, uses Streamlette, an extremely simple consensus derived from Streamlet.
The consensus game — the cryptoeconomic mechanism incentivizing consensus correctness — generally trades off other properties (like capital efficiency) for robust long-term economic security. The "carrot" of a collusion-tolerant fee economy allows stakers to extract value primarily through "benign collusion" on fees rather than block rewards or misbehavior. Strict slashing, including a catastrophic form known as "nuking", act as unusually powerful "sticks" against staker misbehavior.
To select who gets to participate in consensus, Mel uses a proof-of-stake system. In short, anyone can lock up a sum of SYM, and while it is locked up, a designated node (the staker, which may or may not be the same person who locked up the SYM) receives voting power in consensus.
How does this work in practice? blockchain history is divided into 200,000-block epochs, conventionally numbered from 0. For example, block 12,345 is in epoch 0, while block 1,968,968 is in epoch 9. Anyone can stake (lock up) a particular SYM for a fixed number epochs, during which this SYM will give a designated staker voting power. This is done by sending a special transaction (of type TxKind::Stake, see TodoYellowPaper) with a SYM-denominated output, with the following metadata:
the beneficiary staker public key that uniquely identifies the staker who receives voting power
the starting epoch of this stake, or the first epoch in which this stake contributes to the staker's voting power. This must be an epoch after the current epoch.
the post-end epoch of this stake, or the epoch after the last epoch that the stake contributes to the staker's voting power.
The SYM — known as an individual stake — is then locked up until the end of the the post-end epoch; after the start of the starting epoch and before the start of the post-end epoch, the staker has voting power.
To illustrate this, let's look at example. Here, Stacy is a staker node, and Alice is someone who staked 100 SYM at block height 900,000 for epochs 5..10, designating Stacy as the beneficiary:
This somewhat weirdly constrained staking system has two important consequences:
Although stakes can be locked at any time, the vote weights only change at epoch boundaries. This turns out to be crucial for mitigating "weak subjectivity" and enabling scalable and trustless light clients.
Each stake is "inactive" for 1 epoch before unlocking — epoch 10 in our illustrated example. This is important for economic security, as we will discuss shortly in the section on incentives.
Given the list of consensus participants produced through proof-of-stake, a consensus protocol still needs to run to actually produce any blocks. The two important properties of Mel's consensus protocol, Synkletos, are its immediate finality and detached consensus proofs.
Most blockchains use some form of "longest-chain" consensus, such as Nakamoto consensus. There, anyone can grow the blockchain by building more blocks on existing blocks and broadcasting them, forming a continually growing "block tree". We then simply define the canonical blockchain as the "longest" or "heaviest" branch of this tree by some metric of long or heavy, and design incentives so that people build on what they see as the canonical chain — this then leads to the canonical branch being overwhelmingly longer than the others.
Although elegant and highly resilient to temporary network faults, longest-chain consensus has a pretty big problem: lack of finality. You can never be sure that what you think is the canonical chain actually is the longest chain, since there always could be a longer chain that you haven't discovered yet, due to network lag, a malicious ISP, or other reasons. There's thus no objective, immutable blockchain history.
In practice, this problem mostly affects recent blocks, which often get "reorganized" when other recent branches turn out to be longer, and most of the time waiting for a while (say for an hour) before trusting a block's content makes the chances of a subsequent reorganization past its height exceedingly unlikely. But the problem is that there's no way to define how long "a while" is — especially when dealing with possibly malicious networks rather than simply lag.
This an especially bad problem for any apps that need to autonomously verify on-chain data from off-chain or other blockchains — crucial to Mel's off-chain composable vision! All of these apps need to either handle the complex state rollback needed to deal with block reorganizations, or introduce large artificial latencies to wait for blocks to be old enough, which may or may not be long enough depending on factors (network latency and security) outside of the app's control.
Mel therefore does not use a longest-chain algorithm. Instead, we use an immediately final BFT consensus. This means that once a block appears on the blockchain, it is permanently canonical and can never be reverted.
This solves all the finality problems related to off-chain composability, especially when combined with the next point — off-chain verifiable proofs that a particular block is canonical.
Every Mel block committed to history has a consensus proof — signatures from stakers whose aggregate vote share for the epoch that the block belongs to exceeds 2/3 of the total vote share.
The interesting thing is that these proofs are detached from the precise mechanism in which the BFT consensus works. The consensus algorithm does not need to produce the consensus proof: it may simply guarantee that all honest participants arrive at the same belief for each block. Each staker can then just "ask around" after the consensus produces a result, and given that the BFT consensus is correct, it's guaranteed that they will eventually gather a consensus proof and be able to broadcast the block to the wider Mel network.
In short, there's a three-step process for creating a block in Mel:
the stakers, among themselves, run a BFT to decide the block
an "asking around" gossip produces a consensus proof
the block, with a consensus proof attached, is broadcast throughout the Mel network and becomes part of history.
This means that nobody other than the stakers need to care about the precise way the BFT works. In particular, changing the consensus algorithm doesn't need a governance upgrade to the blockchain validation rules, which is important given Mel's governance-free ethos and the steady innovation seen in BFT consensus algorithms. For instance, through the history of the Mel betanet we started with a rough implementation of HotStuff, moved to the extremely simple and elegant Streamlet consensus, then to a different instantiation, "Streamlette", optimized for deciding one block in an immediate-finality setting.
Moreover, the simplicity of the consensus proof makes it really easy to verify by off-chain apps other than blockchain nodes. A light client that has the vote weights for the current epochs (which can be known near-trustlessly) can easily verify a claim that a certain block is canonical, by simply checking the signatures in the consensus proof and adding the corresponding weights. This makes fully consensus-verifying light clients much easier to write than those for PoS algorithms with complex finalization rules (say, Ethereum), or even longest-chain proof-of-work blockchains.
While staking decides who gets to create blocks and the consensus algorithm how they create blocks, the consensus game is the incentives behind why anyone would create blocks.
The overall theme of Mel's consensus game is a overwhelming focus on maximizing economic security, especially long-term economic security. This because with governance delegitimized, we cannot really rely on "the community will fork away the bad guys" as a normal-case defense against consensus attacks (though, as we'll see, it's still used as a defense-in-depth for extreme cases). Upholding consensus security needs to be the profitable thing to do for stakers in a wide variety of circumstances.
Towards this end, we have a "carrot" and a "stick".
Unlike most blockchains, Mel studiously avoids non-coordination assumptions in its consensus incentives. This means that it's totally fine for rational stakers to collude and attempt to extract monopoly profits, or worse. Decentralization becomes more a matter of fault tolerance and ensuring that the average staker is economically rational, rather than trying to make collusion impossible, making the consensus oligopoly situations that often occur with both PoS and PoW much less scary.
The way Mel does this is a little surprising: we give stakers all the tools they need to collude in a fee cartel. In a system distantly inspired by EIP-1559, stakers vote on a uniform minimum fee level that all stakers must charge, in a system designed to simulate a "despotic" blockchain controlled entirely by a rational profit-maximizing monopoly — who would turns out to actually behave in a trustworthy manner. This way, incentives to collude to extract more fees disappear, the equilibrium fee is the same no matter the level of collusion, and a whole lot of game-theoretical gnarliness surrounding pathological strategies in not-quite-collusion conditions (e.g. just a few dominant stakers) disappears.
Furthermore, our fee model turns to have the additional benefit of funding staker income almost entirely through fees rather than inflation, making staker income much more closely aligned with the interests of actual users of the blockchain rather than "SYM go up" or even "block reward go up".
Like in other PoS systems, misbehaving stakers are punished by slashing their stake. Anybody observing a staker misbehaving in a provable way (say, by signing two conflicting blocks for the same height) can submit a transaction that slashes their stake. This is also the reason for the mandatory "inactive epoch" before unstaking: to give a chance for slashing evidence to be submitted before the staker leaves.
Two important features separate Mel slashing from most other systems.
First, slashing totally destroys the stake, redistributing it among other stakers. Unlike blockchains like Ethereum that attempt to differentially penalize slashing based on how likely it is that it's part of coordinated attack, any slashing in Mel splits up the entire stake to the other stakers. Not only does this simplify the incentives and increase security, it gives an incentive for stakers to police each other, as they all stand to profit if a staker is slashed!
Second, there is nuking, an exceptional action similar but not quite like slashing, where more than 1/3 of the voting power has behaved maliciously. In this case, it is not possible to slash them on-chain anymore, since the malicious stakers can simply censor anybody who attempts to expose them.
Instead, all nodes and clients are programmed to shut down if given proof of nuking. This makes attacking the network successfully without quickly shutting down the network extremely difficult even with the cooperation of all consensus participants, making the set of adversaries who would benefit from breaking consensus safety much smaller.
If even despite this threat nuking happens, a manual hard-fork recovery would indeed be necessary, but coordinating this would be much easier because people would be forced to recover the nuked network; there is simply no option to go along with the attack. Or in Vitalik's words, successful consensus attacks "default to chaos" rather than "default to victory".
Mel's consensus is a combination of a fixed-term proof-of-stake, a Byzantine fault-tolerant consensus algorithm, and a collusion-tolerant fee economy with strict slashing and "nuking".
Proof-of-stake decides who gets to participate in consensus, and the consensus algorithm decides how blocks are created and provides immediate finality. The consensus game incentivizes stakers to create blocks reliably, with the "carrot" of a collusion-tolerant fee economy and the "stick" of strict slashing and nuking.
The combination of these three elements provides strong economic security against long-term consensus attacks, while also making light clients easy to write and blockchain upgrades easy to implement.
