Glossary · Consensus

Consensus

Published 28 Apr 2025 · Updated 3 Nov 2026 · 16 min read

Consensus is the set of rules and messages by which the nodes of a blockchain or distributed ledger agree which transactions are valid and in what order. It keeps one shared history even when some nodes fail or lie, within the protocol's fault assumptions. Agreement requires enough participants under those rules, not unanimous approval.

  • Consensus needs a protocol-defined quorum, not agreement from every node.
  • Proof of work and proof of stake set voting rights, while agreement rules select one history.
  • In November 2026, finality is probabilistic on Bitcoin, two epochs on Ethereum and one block on CometBFT.

What consensus means in a ledger

Copies of a ledger sit on computers that receive transactions in different orders. Some computers may be offline or dishonest. A distributed ledger needs rules for selecting records and their sequence despite those differences. Consensus supplies those rules. The definitions below show how ISO 22739:2024, MiCA and NIST IR 8202 describe agreement, transaction validation and a shared state.

SourceDateDefinition, in its own wordsWhat it adds
ISO 22739:2024DateJanuary 2024Definition, in its own words"agreement among distributed ledger technology (DLT) nodes that: a transaction is validated; the distributed ledger contains a consistent set and ordering of records of validated transactions"What it addsValidation and order. "Consensus does not necessarily mean that all DLT nodes agree."
MiCA, Article 3(1)(3)Date31 May 2023Definition, in its own words"the rules and procedures by which an agreement is reached, among DLT network nodes, that a transaction is validated"What it addsThe EU's legal term, also used in the DLT Pilot Regime
NIST IR 8202DateOctober 2018Definition, in its own words"A process to achieve agreement within a distributed system on the valid state."What it addsWhen versions conflict, most networks treat the longer chain as correct
ethereum.orgDateUpdated 22 April 2026Definition, in its own words"the complete stack of ideas, protocols and incentives that enable a distributed set of nodes to agree on the state of a blockchain"What it addsProof of work and proof of stake are Sybil resistance, not the whole protocol
Bitcoin white paperDate31 October 2008Definition, in its own words"The majority decision is represented by the longest chain, which has the greatest proof-of-work effort invested in it."What it adds"Proof-of-work is essentially one-CPU-one-vote."
Lamport, Shostak and PeaseDateJuly 1982Definition, in its own words"All loyal generals decide upon the same plan of action."What it addsAgreement despite traitors: 3m + 1 generals cope with m traitors when messages are unsigned

Each full node checks signatures and balances itself. A consensus mechanism chooses one order and history among valid but conflicting alternatives. Two transactions can each pass those checks separately while spending the same funds. Ordering them prevents a double spend because both cannot stand in the accepted history. ISO 22739:2024 explicitly says consensus need not include every DLT node. Nodes can see different newest blocks for a while and still agree on the history the rules accept.

In the 1982 Byzantine Generals problem, commanders must agree on one plan despite traitors sending misleading messages. Lamport, Shostak and Pease showed that unsigned messages need at least 3m + 1 generals to tolerate m traitors. With signed messages, their problem is solvable with any number of generals and traitors. Ledger nodes take the commanders' place. A validator sends proposals or votes, while the consensus protocol defines which combinations establish an agreed history.

Safety, liveness and fault models

A fault threshold belongs to a specific model, not to blockchains in general. Consensus rules distinguish stopped participants from dishonest ones. The table uses f, the number of faulty nodes tolerated. Safety and liveness depend on timing assumptions as well as participant numbers.

Fault modelNodes needed for f faultsExampleSource
Crash or omission, partially synchronous networkNodes needed for f faults2f + 1ExampleRaft: five servers keep working with two downSourceDwork, Lynch and Stockmeyer, April 1988. Raft, May 2014
Byzantine, unsigned messagesNodes needed for f faults3f + 1ExampleFour generals cope with one traitorSourceLamport, Shostak and Pease, July 1982
Byzantine, signed messages, synchronous networkNodes needed for f faultsAny numberExampleSolvable for any number of traitorsSourceLamport, Shostak and Pease, July 1982
Byzantine, partially synchronous, signed or notNodes needed for f faults3f + 1ExamplePBFT, Tendermint and HotStuff: four validators tolerate oneSourceDwork, Lynch and Stockmeyer, April 1988. PBFT, February 1999
Any fault, fully asynchronous, deterministic protocolNodes needed for f faultsNo protocol always decidesExampleOne crashed process can block a decision foreverSourceFischer, Lynch and Paterson, April 1985

Safety means correct nodes never decide differently. Liveness means they eventually decide. Crash faults stop nodes, while Byzantine fault tolerance covers arbitrary behavior, including contradictory messages. Crash consensus needs 2f + 1 nodes under partial synchrony. Byzantine consensus needs 3f + 1, with 2f + 1 votes per quorum. Two such quorums overlap in at least f + 1 nodes, including an honest one. Crash majorities also overlap. Raft tolerates two crashes among five servers. GoQuorum's documentation stresses that Raft handles crashes only, assuming an honest leader.

The 1985 FLP result rules out guaranteed deterministic consensus under unbounded message delays with even one possible crash. Partial synchrony, formalized in 1988, allows timing bounds that are initially unknown or only eventually hold. PBFT preserves safety without synchrony but needs it for liveness. Timeouts enable retries. In November 2026, CometBFT's default proposal timeout starts at three seconds and grows by half a second each round. When agreement fails, Stellar and the XRP Ledger stop rather than diverge. Ethereum can keep producing blocks without finalizing.

Who votes and how votes agree

An open network cannot treat every identity as an equal vote. A consensus mechanism therefore needs both participation rules and an agreement rule. Work, stake or admission determines influence. Fork choice and voting then determine which proposed history the participants accept.

Ethereum's documentation, updated 22 April 2026, describes proof of work and proof of stake as Sybil resistance mechanisms and block author selectors. Neither alone is a complete consensus protocol. Spending energy or providing collateral creates an economic deterrent to Sybil attacks. Permissioned systems instead restrict admission. This separates the cost or permission needed to participate from the rules that turn eligible participants' messages into one decision.

In November 2026, Bitcoin pairs work with most cumulative work, while Ethereum pairs stake with Gasper's LMD-GHOST fork choice and validator votes. Solana pairs stake with Alpenglow's Votor voting. Hedera combines admitted, stake-weighted council nodes with hashgraph virtual voting. Avalanche's Snowman uses repeated, stake-weighted subsampled voting. Stellar uses federated quorums, while XRP Ledger participants use UNL lists. Cardano's developer portal contrasts Raft's honest-node assumption with Ouroboros's Byzantine model. Praos selects the longest chain. Eligibility alone therefore does not identify the consensus algorithm.

How a protocol reaches one decision

A consensus protocol turns proposals into decisions through messages and waiting rules. Choose Nakamoto, Raft, PBFT, Tendermint or HotStuff, then select 4, 7, 10, 31, 100 or 180 validators. Step through each exchange to see senders, recipients, the waiting quorum, tolerated faults and the running message count.

Figure 1 · Page team model from the protocol papersOne decision, message by messageChange the participant count or fail the leader, then step through the messages needed to reach one decision.
Protocol
Validators
Leader
Steps to one decision4 validators, 3 steps
Step 1 of 3: Pre-prepareMessages
Who sends
Primary to 3 backups
Messages in this step
3
Running total
3 of 24
Waits for
Nothing yet
The primary proposes an order for the request.
Messages for this decision24Between replicas, normal case
One-way network delays3Until the decision, not counting timeouts
Faults tolerated1 ByzantineQuorum: 3 of 4
Same decision, four protocolsMessages per decision at 4 validators, log scaleRaft6HotStuff24PBFT24Tendermint27

Nakamoto uses blocks and confirmations, not voting rounds. There is no leader to replace when block production pauses.

One leader collects majority acknowledgements before committing. This consensus model tolerates crash faults, not dishonest behavior.

Three phases include all-to-all prepare and commit messages. A timeout triggers a view change to replace the primary.

Propose, prevote and precommit form a round. After a timeout, participants advance to another round with the next proposer.

Votes go to the leader, so message counts grow linearly. Basic HotStuff uses additional phases before committing.

Message counts are the page team's model from protocol papers and ignore gossip, batching and signature aggregation.

PBFT's 1999 protocol uses pre-prepare, prepare and commit. Preparing requires 2f backup prepares matching the pre-prepare, and committing requires 2f + 1 commits. Tendermint and its fork CometBFT use propose, prevote and precommit, with more than two thirds of voting power needed to commit. Raft instead commits an entry once its originating leader replicates it to a majority. The 1999 paper's PBFT file system was only 3% slower than an unreplicated one.

All-to-all voting drives the page team's 100-participant model to 19,800 messages for PBFT and 19,899 for Tendermint. Basic HotStuff needs 792, with votes collected by a leader. Chained HotStuff overlaps consecutive blocks' phases, committing through a three-chain pattern. In November 2026, Flow documents Jolteon, and Aptos uses AptosBFT from the same lineage. Monad compares MonadBFT with pipelined HotStuff-family consensus protocols. Sui's Mysticeti instead uses a directed acyclic graph. These model counts are not measured traffic.

Probabilistic, economic and absolute finality

Acceptance does not always rule out chain forks. Probabilistic finality reduces reversal risk, economic finality attaches penalties, and deterministic or absolute finality excludes reversal within stated assumptions. At 05:01 UTC on 3 November 2026, Ethereum's newest block was 79 slots, or 15.8 minutes, ahead of its finalized checkpoint.

Figure 2 · Bitcoin white paper, section 11Confirmations and reversal probabilitySet hash power and blocks built on top to compare modeled catch-up probability, waiting time and deep reorganizations with economic and deterministic finality.
10%
5 blocks
0.091%Chance the attacker ever catches upAbout 1 in 1,094
By depthAttacker share 10%, longer bars mean likelier1 block20.5%2 blocks5.1%3 blocks1.3%6 blocks0.024%12 blocks8.9 × 10⁻⁸24 blocks1.3 × 10⁻¹⁴48 blocksBelow 10⁻¹⁵120 blocksBelow 10⁻¹⁵
Your wait at 10-minute blocks50 min5 blocks on top
Blocks for less than 0.1%5At a 10% share, white paper table
Wait for that depth50 minExpected, at the 10-minute target
Deep reorganizations by majority attackers5 to 7 Jan 2019Ethereum ClassicReorganizations up to 123 blocks deep, 219,500 ETC double spentJan to Feb 2020Bitcoin GoldEight reorganizations by 5 Feb, about 12,858 BTG double spent31 Jul to Aug 2020Ethereum ClassicMore than 3,600 blocks replaced, 807,260 ETC double spent
Where finality is not probabilisticCometBFTFinal after one block once more than two thirds of voting power precommit itEthereumFinal after two epochs, 12.8 minutes. Reverting costs at least a third of staked ETH

At 50%, no confirmation count prevents eventual catch-up, consistent with Ethereum Classic's and Bitcoin Gold's 2020 majority attacks.

At 25% to 45% attacker hash power, reaching below 0.1% probability takes 15 to 340 blocks.

Below 25% attacker hash power, catch-up probability falls quickly as confirmations accumulate in the white paper's model.

Probabilities use the Bitcoin white paper's formula, computed in this page with 10-minute blocks.

The Bitcoin white paper gives a 10% hash-power attacker about a one-in-1,094 catch-up chance after six confirmations. At 30%, falling below 0.1% takes 24 blocks. In November 2026, Ethereum finality takes two epochs, calculated as 12.8 minutes, or about 15 minutes in Ethereum's documentation. Reverting finalized history requires losing at least one third of staked ETH. Casper FFG's accountable safety means conflicting finalized checkpoints require at least one third of validators to violate slashing conditions.

CometBFT finalizes a block once over two thirds of voting power precommits it. Fabric guarantees that validated blocks are final, and Canton documents no forks, reorganizations or rollbacks. Solana's Alpenglow, live since 21 October 2026, finalizes in one round with 80% of stake or in two with 60% certificates, targeting about 150 ms. Confirmed and finalized now describe the same state. Polkadot's GRANDPA finalizes all blocks up to one that more than two thirds of validators attest to. Ethereum research targets a few tens of seconds, with no fork date.

Network consensus in November 2026

The table separates agreement rules, participation rules and finality for twelve networks on 3 November 2026. A block time or slot target is not a promise of settlement. Hedera orders events rather than blocks, and Canton has no single chain that all nodes replicate, so their timing needs different labels.

NetworkAgreement ruleSybil resistanceBlock or slot timeFinality as documentedValidators or producers
BitcoinAgreement ruleMost cumulative work (Nakamoto)Sybil resistanceProof of workBlock or slot time600 s target, 598.4 s over 14 blocks to 974,344Finality as documentedProbabilistic, 6 confirmations in about an hourValidators or producersAny miner
EthereumAgreement ruleGasper: LMD-GHOST and Casper FFGSybil resistanceProof of stake, 32 ETH minimumBlock or slot time12 s slotsFinality as documentedTwo epochs, 12.8 min (about 15 min per ethereum.org)Validators or producers814,678 active validators, 44.3M ETH staked (3 Nov 2026)
SolanaAgreement ruleAlpenglow's Votor since 21 Oct 2026Sybil resistanceProof of stakeBlock or slot time200 ms targetFinality as documented80% in one round or 60% in two, ~150 ms targetValidators or producers659 staked validators (Solana Compass, 3 Nov 2026)
Cosmos HubAgreement ruleCometBFT (Tendermint)Sybil resistanceProof of stakeBlock or slot time5.72 s measuredFinality as documentedOne block, more than 2/3 precommitsValidators or producers180 bonded (2 Nov 2026)
PolkadotAgreement ruleBABE production, GRANDPA finalitySybil resistanceNominated proof of stakeBlock or slot time6 s slotsFinality as documentedMore than 2/3 attest to a chainValidators or producers600 in the active set
Avalanche C-ChainAgreement ruleSnowman (k 20, α 15, β 20)Sybil resistanceProof of stake, 2,000 AVAXBlock or slot timeDynamic since 19 Nov 2025Finality as documentedAbout 1 s per its docsValidators or producers556 (2 Nov 2026)
CardanoAgreement ruleOuroboros Praos, longest chainSybil resistanceProof of stake poolsBlock or slot time1 s slots, about 20 s per blockFinality as documentedPractical in minutes, settled after about 12 hValidators or producers2,856 pools with stake (2 Nov 2026)
BNB Smart ChainAgreement ruleParlia with fast finality votesSybil resistanceStake-elected validator setBlock or slot time0.45 s since 14 Jan 2026Finality as documentedWithin two blocks, about 1.125 sValidators or producers45 active (21 Cabinet, 24 Candidates)
XRP LedgerAgreement ruleXRPL consensus, 80% of a UNLSybil resistanceEach server's UNLBlock or slot timeA ledger every several secondsFinality as documentedValidated ledger, 3 to 6 s settlementValidators or producersSet per server
HederaAgreement ruleHashgraph, virtual votingSybil resistanceCouncil nodes weighted by stakeBlock or slot timeNo blocks at the consensus layerFinality as documented3 to 5 s absolute finalityValidators or producers30 consensus nodes (3 Nov 2026)
AptosAgreement ruleAptosBFT (Jolteon family)Sybil resistanceProof of stakeBlock or slot timeNot stated in the docs readFinality as documentedOrdering after 3 network hopsValidators or producers83 active (3 Nov 2026)
Canton NetworkAgreement ruleBFT ordering by Super Validators, then a mediator verdictSybil resistanceAdmission by Super Validator sponsorsBlock or slot timeNo chain all nodes replicateFinality as documentedMediator verdict, no rollbacksValidators or producers847 validators, 61 Super Validator nodes (3 Nov 2026)

Counts reflect different trackers and read times, not identical categories of participation. Public trackers also differ on Solana's count. Alpenglow went live there on 21 October 2026. Canton's Global Synchronizer uses BFT ordering run by Super Validators in November 2026. Its documentation and release notes describe the ordering protocol differently, and no MainNet change date is published. Finality times in the table are each network's documented figures, not measurements.

How consensus fails in practice

Consensus can fail by accepting conflicting records or by making no progress. A partition exposes that distinction without requiring a dishonest participant. The model separates block production from finality and matches each rule to a dated incident from 2010 to 2025.

Figure 3 · Page team model, simulation of the Ethereum specsSplit a network in twoChoose a consensus family, split ratio and duration to see each side's progress, recovery and the property that gives way.
Family
Split
1 day
Side A, 60%Finality paused
Blocks proposed, expected
4,320
Finality
Under two thirds of stake. The leak restores finality after 13.1 to 13.8 days.
Side B, 40%Finality paused
Blocks proposed, expected
2,880
Finality
Under two thirds of stake. The leak restores finality after 27.0 to 27.1 days.
When the network healsFork choice picks the heavier chain, the other side's blocks are reorganized out, and finality resumes once votes from both sides count again.Safety: keptLiveness: lost
Closest real case11 to 12 March 2013Bitcoin 0.8 and 0.7 splitBlock 225,430 split nodes by client version. The 0.8 chain had about 60% of the hash power. Pools downgraded to 0.7, the other chain won, and at least one large double spend went through.Source: BIP-50 post-mortem, bitcoin.org alert
Closest real case11 and 12 May 2023Ethereum finality paused twiceBlocks kept coming, but finality stopped for 4 epochs and then for 9 to 10 epochs. The second pause started the inactivity leak. About 28 ETH in penalties were applied.Source: Prysm post-mortem, eth2book
Closest real case5 to 6 June 2024Cosmos Hub haltAn upgrade bug halted the chain at block 20,740,970 at 19:21 UTC. It resumed after about 4 hours and 40 minutes, without a fork.Source: Cosmos Hub post-mortem, 6 June 2024
Closest real case2 November 2020Cloudflare etcd partial partitionA misbehaving switch caused repeated Raft elections, writes stopped, and the API and dashboard were affected for 6 hours and 33 minutes.Source: Cloudflare blog, 27 November 2020

Both sides keep mining. When connectivity returns, the chain with less cumulative work loses its blocks through reorganization.

During a long equal split, blocks continue and finality pauses. The inactivity leak restores it on both sides after weeks, requiring social recovery.

A side above two thirds of voting power keeps committing. Otherwise, the chain halts without a fork.

The majority side can elect a leader and commit entries. A minority cannot commit, and an equal split stops both sides.

This page-team model applies protocol rules. Ethereum leak times come from a simulation of the consensus specifications.

DateNetworkFamilyWhat was lostLength or amount
15 Aug 2010NetworkBitcoinFamilyNakamotoWhat was lostSafety: nodes accepted an invalid blockLength or amountThe good chain overtook it at block 74,691, 53 blocks later
11 to 12 Mar 2013NetworkBitcoinFamilyNakamotoWhat was lostSafety: a split between client versionsLength or amount0.8 side held about 60% of hash power, one large double spend
May 2018, Jan to Feb 2020NetworkBitcoin GoldFamilyNakamotoWhat was lostSafety: 51% attacksLength or amount388,000 BTG stolen, then 12,858 BTG double spent
5 to 7 Jan 2019NetworkEthereum ClassicFamilyNakamotoWhat was lostSafety: 51% attackLength or amount15 reorganizations, 219,500 ETC double spent
31 Jul to Aug 2020NetworkEthereum ClassicFamilyNakamotoWhat was lostSafety: 51% attacksLength or amount807,260 ETC double spent, more than 3,600 blocks replaced
2 Nov 2020NetworkCloudflare etcd (not a blockchain)FamilyRaftWhat was lostLiveness of writesLength or amountAPI and dashboard affected for 6 h 33 min
11 Nov 2020NetworkEthereum (proof of work)FamilyNakamoto, client bugWhat was lostSafety for nodes on old Geth versionsLength or amountAbout 30 blocks on a shorter chain
Sep 2021 to Feb 2024NetworkSolanaFamilyTowerBFT with Proof of HistoryWhat was lostLivenessLength or amountSix outages, 4.5 to about 19.7 hours
11 to 12 May 2023NetworkEthereumFamilyGasperWhat was lostLiveness of finalityLength or amount4 epochs, then 9 to 10 epochs, 28 ETH in penalties
5 to 6 Jun 2024NetworkCosmos HubFamilyCometBFTWhat was lostLivenessLength or amountHalted 4 h 40 min, no fork
Nov 2024, Feb 2025NetworkXRP LedgerFamilyUNL consensusWhat was lostLivenessLength or amountAbout 10 minutes, then about 64 minutes

The 2013 Bitcoin split and Ethereum Classic's January 2019 majority attack show different routes to lost safety. BFT outages instead illustrate lost liveness, including Cosmos Hub's June 2024 halt. Ethereum's May 2023 incidents interrupted finality while blocks continued. The second interruption lasted about 64 minutes and activated the inactivity leak. Recovery also involved people, including Bitcoin pools downgrading in March 2013 and validators coordinating Solana restarts during its 2021 to 2024 outages.

Consensus among known participants and Canton

Admission changes who participates, not the need for agreement. A permissioned ledger can use crash-tolerant or Byzantine consensus, depending on its trust assumptions. Operators, membership authorities, notaries or sequencers still carry responsibilities. Knowing the participants does not make their services incapable of failure.

Fabric v3.0.0 introduced SmartBFT alongside Raft ordering in its 17 September 2024 release. Corda's notary provides uniqueness consensus. Corda Enterprise's highly available notary is a cluster run by one party that tolerates crash faults, and its BFT notary is still experimental. Besu QBFT requires signatures from at least two thirds of admitted proof of authority validators, giving immediate finality. More than one third stopping prevents new blocks. GoQuorum's repository was archived on 5 June 2026. These implementations retain different fault assumptions despite restricted membership.

Canton documentation says the sequencer gives all validators messages in the same order and sees only encrypted data. Each validator confirms or rejects solely from its own view. The mediator collects confirmations from participant nodes hosting signatories, while observers need not confirm. Global Synchronizer Super Validators operate sequencer and mediator nodes with two-thirds BFT consensus for ordering and confirmation. Public counts read on 3 November 2026 showed 847 validators and 61 Super Validator nodes.

What consensus does not guarantee

Agreement is one part of a ledger's behavior. It does not certify outside information or remove software and operational dependencies. These limits apply differently across consensus designs, so a protocol's guarantees should be read together with its assumptions and recovery rules.

Correct external inputs

Consensus orders transactions accepted under ledger rules. It cannot establish whether submitted data accurately describes a real event outside the network.

Resistance to majorities

Nakamoto consensus cannot preserve history against an attacker controlling more than half the hash power. Ethereum Classic's 2020 and Bitcoin Gold's 2018 attacks demonstrate that limit.

Progress during faults

Missing voting power can stop BFT progress. Cosmos Hub halted in June 2024, while Solana's 6 February 2024 outage stopped block finalization for about five hours.

Bug-free clients

Consensus cannot repair a shared validation mistake. Bitcoin's August 2010 bug accepted an invalid block, while client differences split Bitcoin in March 2013 and Ethereum in November 2020.

Recovery without people

Protocol rules do not automate every recovery decision. Ethereum documents social recovery after conflicting finalized forks, while Solana's September 2021 restart required validator coordination.

Transaction privacy

Public consensus exposes transactions to validators rather than supplying confidentiality itself. Canton documentation states that the sequencer and other synchronizer infrastructure see only encrypted data.

How consensus research developed

Consensus research is older than blockchains. Pease, Shostak and Lamport's paper of April 1980 came first, and Lamport's Paxos was submitted in January 1990 and published in May 1998. The eight events below trace the line to Solana's change in October 2026.

  1. July 1982

    The Byzantine Generals problem

    Lamport, Shostak and Pease formalize agreement despite traitors. Their distinction between unsigned and signed messages shows why communication assumptions change the faults a protocol can tolerate.

  2. April 1985

    FLP exposes a timing limit

    Fischer, Lynch and Paterson show that fully asynchronous deterministic agreement cannot guarantee termination with one possible crash. The result does not rule out practical consensus systems.

  3. February 1999

    PBFT makes Byzantine replication practical

    Castro and Liskov present Practical Byzantine Fault Tolerance. Its replicated service uses pre-prepare, prepare and commit, with timeout-triggered view changes when the primary fails to make progress.

  4. January 2009

    Bitcoin starts its shared history

    Bitcoin's 3 January 2009 genesis block follows the 31 October 2008 white paper. Nakamoto consensus connects block publication to accumulated computational work.

  5. May 2014

    Raft and Tendermint emerge

    The 20 May 2014 Raft paper explains replicated logs through an elected leader and majority replication. Tendermint also begins in 2014, addressing agreement under a Byzantine fault model.

  6. October 2017

    Casper introduces accountable checkpoint votes

    Buterin and Griffith publish Casper FFG, connecting checkpoint finalization with accountable safety. Conflicting finalized checkpoints imply that at least one third of validators violated a slashing condition.

  7. September 2022

    The Merge changes Ethereum participation

    The 15 September 2022 Merge moves Ethereum from proof of work to proof of stake. Its beacon chain had started on 1 December 2020.

  8. October 2026

    Alpenglow reaches Solana mainnet

    Solana activates Alpenglow at epoch 1064, starting 21 October 2026. Votor replaces TowerBFT voting, and Proof of History no longer paces block production.

Designing and operating a consensus layer

A protocol name leaves important engineering choices unresolved. Our blockchain development services include custom layer-1 protocol development and novel consensus mechanisms. We distinguish the agreement rules from the operating decisions around them. The five areas below help define the system a team intends to build or run, without treating any protocol as universally suitable.

PixelPlex case icon

PixelPlex has built blockchain software since 2013. Echo, our own protocol, went live in 2017. The Echo case describes its proof-of-weighted-randomness consensus as a Byzantine agreement protocol with a small committee of block producers and fast finality. On Canton, we are a validator operator and run our own production infrastructure that the ecosystem depends on daily.

  1. Fault model and membership

    Specify tolerated failures, voting weights and admission rules. These choices establish whose messages count and which assumptions the consensus layer relies on.
  2. Application finality

    Define whether the application accepts confirmation risk or waits for a finalized decision. Keep that requirement separate from block or slot timing.
  3. Timeouts and connectivity

    Specify expected delays and behavior during lost connectivity. Timeout settings should reflect the protocol's timing assumptions, rather than imply guaranteed progress.
  4. Keys and operations

    Define who controls signing keys and operates participants. Our Canton development service includes support for deploying, configuring and managing participant nodes.
  5. Partitions and upgrades

    Test split-network behavior and version transitions against documented rules. Separate checks for conflicting decisions from checks for recovery of progress.

Discuss the assumptions behind your consensus layer

We can discuss your fault model, participation rules and finality requirements before considering an implementation.

Discuss your requirements

Consensus FAQ

What does consensus mean in blockchain?

Consensus in blockchain is the set of rules by which nodes agree which transactions are valid and in what order. Every node checks signatures and balances itself. Consensus selects one history among valid alternatives, preventing a double spend. It needs a quorum, not every node, as ISO 22739:2024 notes.

How do consensus mechanisms differ from work or stake?

Proof of work and proof of stake decide who may vote or produce blocks, so ethereum.org calls them Sybil resistance mechanisms rather than consensus protocols. The agreement rule determines the accepted history. In November 2026, Bitcoin follows the longest chain with most work, Ethereum uses Gasper's fork choice and Casper FFG votes, and CometBFT uses BFT rounds.

How many faulty nodes can BFT tolerate?

A Byzantine fault tolerant network with n = 3f + 1 validators tolerates f faulty validators, fewer than one third. A network of 4 validators tolerates 1 fault, while 100 tolerate 33. Quorums of 2f + 1 votes preserve safety under partial synchrony. Crash-only protocols such as Raft need 2f + 1 nodes for f crashes.

What is finality on Bitcoin and Ethereum?

Finality is the point after which a transaction cannot be reversed under the protocol's rules. Bitcoin is probabilistic, with six confirmations taking about an hour. In November 2026, Ethereum takes two epochs, 12.8 minutes by the specification, or about 15 minutes per ethereum.org. Reverting its finalized history costs at least one third of staked ETH.

What is a 51% attack on consensus?

A 51% attack happens when one party controls more than half of a proof-of-work chain's hash power and rewrites recent history to double spend. Coinbase reported 15 Ethereum Classic reorganizations during 5 to 7 January 2019, including 12 with double spends totaling 219,500 ETC. The 31 July to 1 August 2020 attack double-spent 807,260 ETC, according to Bitquery.

Which consensus mechanisms do permissioned networks use?

Permissioned networks usually use crash fault tolerant or Byzantine fault tolerant ordering among admitted validators. Fabric has offered Raft ordering since v1.4.1 and added SmartBFT in v3.0.0, released 17 September 2024. In November 2026, Besu's QBFT gives immediate finality with at least two-thirds signatures, while Corda's notary provides uniqueness consensus.

How does consensus work on the Canton Network?

On Canton, a sequencer orders encrypted messages, the validators hosting each signatory confirm their own view, and a mediator declares the result. In November 2026, the Global Synchronizer's Super Validators run BFT ordering with fewer than one third Byzantine orderers tolerated. There are no forks, reorganizations, or rollbacks.
Sources (35)

Get updates about blockchain, technologies and our company

We will process the personal data you provide in accordance with our Privacy policy. You can unsubscribe or change your preferences at any time by clicking the link in any email.

Follow us on social networks and don't miss the latest tech news

  • facebook
  • X
  • linkedin
  • instagram
Stay tuned and add value to your feed