LibraryConsensus2018Design paperCorpus record
Snowflake to Avalanche: A Novel Metastable Consensus Protocol Family
Avalanche. Team Rocket; later with Maofan Yin, Kevin Sekniqi, Robbert van Renesse and Emin Gün Sirer.
A family of consensus protocols that sample the network repeatedly and tip toward one outcome. The 2018 note was pseudonymous. A later write-up adds named co-authors. Avalanche the network is an implementation, not the sample itself.
Avalanche asks each node to sample a few others, adopt the colour it sees, and repeat. A decision that starts with a small majority is argued to lock in quickly, without a leader and without every node talking to every other.
The five-minute read
Metastability is the metaphor and the mechanism
A network that is slightly more red than blue, and in which each node repeatedly takes the colour of a random sample, tends to fall all the way to red. The paper builds consensus on that tendency.
Sampling replaces the quorum you fully know
A node does not collect a vote from a named committee. It asks a small random set, several times. Confidence goes up when the answers agree.
Snowflake, Snowball, Avalanche
The paper is a family. Snowflake decides a single bit. Snowball adds confidence counters. Avalanche applies the idea to a growing set of transactions, represented as a graph of conflicts.
Leaders are absent on purpose
There is no elected block proposer in the core subsampling loop. That is the scaling claim. It is also why the paper has to argue safety from the sampling math rather than from a quorum intersection proof alone.
The graph still needs an application
Avalanche consensus on conflicting transactions is not, by itself, a smart-contract chain. The later platform that carries the name added subnets, virtual machines and a staking layer. Keep the paper and the platform apart.
One action, walked through
- A node hears a new transaction and notes which accepted transactions it conflicts with.
- It samples a small number of random peers and asks what they currently prefer.
- If the sample strongly prefers one side, the node adopts that preference and increases its confidence.
- It repeats. Confidence decays or resets when the sample disagrees, under the paper's counters.
- Once confidence stays above a threshold for long enough, the node accepts the transaction and rejects its conflicts.
The argument, unpacked
Why a slight edge is enough
The interesting claim is not that unanimous networks agree. It is that a modest initial advantage, refreshed by random samples, runs away to a decision. The parameters — sample size, quorum size inside the sample, and the confidence threshold — are the product. Change them and you have a different probability of a stuck network.
Sybil resistance is not inside the sampler
Sampling assumes you are asking honest or at least stake-weighted peers rather than an attacker who invented a million identities. The paper's family needs a membership story. Proof of work, proof of stake, or a permissioned set can be that story. The sampler does not provide it.
Conflict sets are the programming model
Two transactions conflict or they do not. That fits payments. It fits less obviously a contract call that touches shared state in subtle ways. A chain that wants rich state has to define conflict before it can reuse this loop.
What has to be true
- Peer samples are drawn from a set the attacker cannot cheaply flood.
- The network is large enough, and samples are independent enough, for the metastability argument to apply.
- Parameters match a region where the paper's probabilities hold. Untuned thresholds are not 'the Avalanche paper'.
- Conflicts are explicit. Hidden shared-state conflicts will be accepted as if they were independent.
What happened after the paper
The Avalanche platform used the family as its consensus story and then built a staking network, subnets and a virtual machine around it. Those are subsequent designs. A study of the 2018 paper should be able to explain a confidence counter without mentioning a token price.
What to check before you use the idea
- What is the sample size, the quorum inside it, and the finality threshold?
- How does a new node know whom it is allowed to sample?
- How is a conflict defined for the state the product actually keeps?
- What happens when the network splits and both sides pass their confidence threshold?
Terms
- Metastability
- A state that is locally stable but, given a small push and repeated sampling, runs to one of two extremes.
- Subsampling
- Asking a few random peers instead of collecting a vote from the full set.
- Confidence
- A counter that rises when samples agree and is punished when they do not. Acceptance waits on this counter.
- Conflict set
- The transactions that cannot all be accepted. The protocol is choosing a side of this set.
The problem the paper names
Quorum protocols want every correct node to hear every vote. That does not get cheaper as the network grows. The paper asks whether a node that only asks a small random sample, over and over, can still end up with the same decision as everyone else.
What the design proposes
- Snowflake and Snowball are binary decision protocols built from repeated subsampling.
- Confidence increases when consecutive samples agree. A node can flip early and then lock.
- Avalanche applies the idea to a chain of conflicts, so competing transactions are decided by the same metastable dynamic.
How the mechanism is specified
- Safety is probabilistic and depends on the sample size, the network assumption, and the fraction of correct participants.
- There is no standing leader in the core family. Leadership, if a network adds it, is a later choice.
- The protocol does not by itself decide fees, virtual machines or subnet policy.
What this page does not treat as proven
- Do not quote the paper as a measured throughput of the Avalanche network.
- Metastability is not finality in the classical BFT sense. The paper's own confidence notion has to be read, not translated into a slogan.
- Parameter choice can break the intuition. Sampling is not automatically 'more decentralised'.
Why a venture studio still reads it
Interesting for high-churn decision systems where a full mesh is the bottleneck. Useless as a slide that says 'subsampling, therefore scale' without the sample size and the adversary.
This is Blockchain Lab's reading of a public design paper. It is not the paper, not a copy of it, and not an offer of tokens, equity, custody or a partnership. Later network behaviour can diverge from the text. Nothing here is investment, legal or technical advice.
Research status: Design paper. Last reviewed: 1 October 2026. This is a reading of a public paper, not investment, legal or security advice.