* state -> step * vote -> v * New version of the algorithm and the proof * New version of the algorithm and the proofs * Added algorithm description * Add algorithm description * Add introduction * Add conclusion * Add conclusion file * fix warnings (caption was defined twice) - only the latter is used anyways (centers captions) - this makes it possible to autom. building the paper * Update grammar * s/state_p/step_p * Address Ismail's comments * intro: language fixes * definitions: language fixes * consensus: various fixes * proof: some fixes * try to improve reviewability * \eq -> = * textwrap to 79 * various minor fixes * proof: fix itemization * proof: more minor fixes * proof: timeouts are functions * proof: fixes to lemma6 * Intro changes and improve title page * Add Marko and Ming to acks * add readme * Format algorithm correctly Clarify condition semantic and timeouts Improve descriptions * patform -> platform * Ensure that rules are mutually exclusive - various clarifications and small improvements * Release v0.6 * small nits for smoother readabilitypull/7804/head
@ -1,25 +1,7 @@ | |||||
# Tendermint-spec | |||||
# Tendermint Spec | |||||
The repository contains the specification (and the proofs) of the Tendermint | |||||
consensus protocol. | |||||
## How to install Latex on Mac OS | |||||
MacTex is Latex distribution for Mac OS. You can download it [here](http://www.tug.org/mactex/mactex-download.html). | |||||
Popular IDE for Latex-based projects is TexStudio. It can be downloaded | |||||
[here](https://www.texstudio.org/). | |||||
## How to build project | |||||
In order to compile the latex files (and write bibliography), execute | |||||
`$ pdflatex paper` <br/> | |||||
`$ bibtex paper` <br/> | |||||
`$ pdflatex paper` <br/> | |||||
`$ pdflatex paper` <br/> | |||||
The generated file is paper.pdf. You can open it with | |||||
`$ open paper.pdf` | |||||
This repository contains the paper describing the Tendermint consensus | |||||
algorithm, including formal proofs of its safety and liveness. | |||||
For the pdf, see the [latest | |||||
release](https://github.com/tendermint/spec/releases). |
@ -1,249 +0,0 @@ | |||||
\section{Tendermint} | |||||
\label{sec:tendermint} | |||||
\newcommand\Disseminate{\textbf{Disseminate}} | |||||
\newcommand\Proposal{\mathsf{PROPOSAL}} | |||||
\newcommand\ProposalPart{\mathsf{PROPOSAL\mbox{-}PART}} | |||||
\newcommand\PrePrepare{\mathsf{INIT}} | |||||
\newcommand\Prevote{\mathsf{PREVOTE}} | |||||
\newcommand\Precommit{\mathsf{PRECOMMIT}} | |||||
\newcommand\Decision{\mathsf{DECISION}} | |||||
\newcommand\ViewChange{\mathsf{VC}} | |||||
\newcommand\ViewChangeAck{\mathsf{VC\mbox{-}ACK}} | |||||
\newcommand\NewPrePrepare{\mathsf{VC\mbox{-}INIT}} | |||||
\newcommand\coord{\mathsf{proposer}} | |||||
\newcommand\newHeight{newHeight} | |||||
\newcommand\newRound{newRound} | |||||
\newcommand\nil{nil} | |||||
\newcommand\prevote{prevote} | |||||
\newcommand\precommit{precommit} | |||||
\newcommand\commit{commit} | |||||
\newcommand\timeoutPropose{timeoutPropose} | |||||
\newcommand\timeoutPrevote{timeoutPrevote} | |||||
\newcommand\timeoutPrecommit{timeoutPrecommit} | |||||
\newcommand\proofOfLocking{proof\mbox{-}of\mbox{-}locking} | |||||
\begin{algorithm}[htb!] | |||||
\def\baselinestretch{1} | |||||
\scriptsize\raggedright | |||||
\begin{algorithmic}[1] | |||||
\SHORTSPACE | |||||
\INIT{} | |||||
\STATE $height_p := 0$ \COMMENT{current height, or consensus instance we are currently executing} | |||||
\STATE $round_p := 0$ \COMMENT{current round number} | |||||
\STATE $step_p \in \set{\propose=0, \prevote=1, \precommit=2, \commit=3}$, initially $\propose$ \COMMENT{current round step} | |||||
\STATE $lockedValue := nil$ | |||||
\STATE $lockedRound := -1$ | |||||
\STATE $validValue := nil$ | |||||
\STATE $validRound := -1$ | |||||
\ENDINIT | |||||
\SPACE | |||||
\UPON{start} | |||||
\STATE $StartRound(0)$ | |||||
\ENDUPON | |||||
\SPACE | |||||
\FUNCTION{$StartRound(round)$} \label{line:tab:startRound} | |||||
\STATE $round_p \assign round$ | |||||
\STATE $state_p \assign \propose$ | |||||
\IF{$\coord(height_p, round_p) = p$} | |||||
\IF{$lockedValue_p \neq \nil$} | |||||
\STATE $proposal \assign lockedValue$ | |||||
\ELSIF{$validValue_p \neq \nil$} | |||||
\STATE $proposal \assign validValue$ | |||||
\ELSE | |||||
\STATE $proposal \assign getValue()$ | |||||
\ENDIF | |||||
\STATE \PBroadcast\ $\li{\Proposal,height_p, round_p, proposal}$ to all \label{line:tab:send-proposal} | |||||
\ENDIF | |||||
\STATE \textbf{after} $\timeoutPropose$ execute $OnTimeoutPropose(height_p, round_p)$ | |||||
\ENDFUNCTION | |||||
\SPACE | |||||
\UPON{receiving $\li{\Proposal,height_p,round_p, proposal}$ \From\ $\coord(height_p,round_p)$ \With\ $state_p = \propose$} \label{line:tab:recvProposal} | |||||
\IF{$lockedValue_p \neq \nil$} \label{line:tab:hasLockedValue} | |||||
\STATE \PBroadcast \ $\li{\Prevote,height_p,round_p,lockedValue_p}$ to all \label{line:tab:send-locked-value} | |||||
\ELSIF{$valid(proposal))$} \label{line:tab:locked-value-nil} | |||||
\STATE \PBroadcast \ $\li{\Prevote,height_p,round_p,proposal}$ to all \label{line:tab:send-prevote} | |||||
\ELSE | |||||
\STATE \PBroadcast \ $\li{\Prevote,height_p,round_p,\nil}$ to all \label{line:tab:send-prevote-nil} | |||||
\ENDIF | |||||
\STATE $state_p \assign \prevote$ \label{line:tab:setStateToPrevote} | |||||
\ENDUPON | |||||
\SPACE | |||||
\FUNCTION{$OnTimeoutPropose(height,round)$} \label{line:tab:onTimeoutPropose} | |||||
\IF{$height = height_p \wedge round = round_p \wedge state_p = \propose$} | |||||
\STATE \PBroadcast \ $\li{\Prevote,height_p,round_p,\nil}$ to all | |||||
\STATE $state_p \assign \prevote$ | |||||
\ENDIF | |||||
\ENDFUNCTION | |||||
\SPACE | |||||
\UPON{receiving $\li{\Prevote,height_p, round_p,*}$ \From\ at least $2f+1$ processes \With\ | |||||
$state_p \le \prevote$} \label{line:tab:recvAny2/3Prevote} | |||||
\STATE \textbf{after} $\timeoutPrevote$ execute $OnTimeoutPrevote(height_p, round_p)$ \label{line:tab:timeoutPrevote} | |||||
\ENDUPON | |||||
\SPACE | |||||
\UPON{receiving $\li{\Prevote,height_p, round_p,v}$ \From\ at least $2f+1$ processes \With\ | |||||
$state_p \le \prevote$} \label{line:tab:recvPrevote} | |||||
\IF{$v \neq \nil$} | |||||
\STATE $lockedValue_p \assign v$ \label{line:tab:setLockedValue} | |||||
\STATE $lockedRound_p \assign round_p$ \label{line:tab:setLockedRound} | |||||
\STATE \PBroadcast \ $\li{\Precommit,height_p,round_p,lockedValue_p}$ to all \label{line:tab:send-precommit} | |||||
\ELSE | |||||
\STATE \PBroadcast \ $\li{\Precommit,height_p,round_p,\nil}$ to all \label{line:tab:send-precommit-nil} | |||||
\ENDIF | |||||
\STATE $state_p \assign \precommit$ \label{line:tab:setStateToCommit} | |||||
\ENDUPON | |||||
\SPACE | |||||
\UPON{receiving $\li{\Prevote,height_p,round,v}$ \From\ at least $2f+1$ processes \With\ | |||||
$round > lockedRound_p$} \label{line:tab:unlockRule} | |||||
\IF{$v \neq lockedValue_p$} | |||||
\STATE $lockedValue_p \assign \nil$ | |||||
\STATE $lockedRound_p \assign -1$ | |||||
\ENDIF | |||||
\ENDUPON | |||||
\SPACE | |||||
\UPON{receiving $\li{\Prevote,height_p,round,v}$ \From\ at least $2f+1$ processes \With\ | |||||
$v \neq \nil \wedge round > validRound_p$} \label{line:tab:validValueRule} | |||||
\STATE $validValue_p \assign v$ \label{line:tab:setValidValue} | |||||
\STATE $validRound_p \assign round$ \label{line:tab:setValidRound} | |||||
\ENDUPON | |||||
\SPACE | |||||
\FUNCTION{$OnTimeoutPrevote(height,round)$} \label{line:tab:onTimeoutPrevote} | |||||
\IF{$height = height_p \wedge round = round_p \wedge state_p = \prevote$} | |||||
\STATE \PBroadcast \ $\li{\Precommit,height_p,round_p,\nil}$ to all \label{line:tab:precommit-nil-onTimeout} | |||||
\STATE $state_p \assign \precommit$ | |||||
\ENDIF | |||||
\ENDFUNCTION | |||||
\SPACE | |||||
\UPON{receiving $\li{\Precommit,height_p,round,v}$ \From\ at least $2f+1$ processes \With\ $state_p < \commit$} \label{line:tab:onPrecommitRule} | |||||
\IF{$vote \neq \nil$} | |||||
\STATE $decide(height_p, v)$ \label{line:tab:decide} | |||||
\STATE$height_p \assign height_p + 1$ | |||||
\STATE reset $round_p$, $step_p$, $lockedRound_p$, $lockedValue_p$, $validValue_p$ and $validRound$ to init values | |||||
\STATE $StartRound(0)$ | |||||
\ENDIF | |||||
\IF{$round_p = round$} | |||||
\STATE $StartRound(round_p+1)$ | |||||
\ENDIF | |||||
\ENDUPON | |||||
\SPACE | |||||
\UPON{receiving $\li{\Precommit,height_p,round_p,*}$ \From\ at least $2f+1$ processes \With\ $state_p = \precommit$} \label{line:tab:startTimeoutPrecommit} | |||||
\STATE \textbf{after} $\timeoutPrecommit$ execute $OnTimeoutPrecommit(height_p, round_p)$ | |||||
\ENDUPON | |||||
\SPACE | |||||
\FUNCTION{$OnTimeoutPrecommit(height,round)$} \label{line:tab:onTimeoutPrecommit} | |||||
\IF{$height = height_p \wedge round = round_p \wedge state_p = \precommit$} | |||||
\STATE $StartRound(round_p+1)$ | |||||
\ENDIF | |||||
\ENDFUNCTION | |||||
\SPACE | |||||
\UPON{receiving message \textbf{with} $height = height_p$ and $round > round_p$ \From\ at least $f+1$ processes} \label{line:tab:skipRounds} | |||||
\STATE $StartRound(round)$ | |||||
\ENDUPON | |||||
\end{algorithmic} | |||||
\caption{Tendermint consensus algorithm} | |||||
\label{alg:tendermint} | |||||
\end{algorithm} | |||||
In this section we present a new Byzantine consensus algorithm called Tendermint\footnote{https://github.com/tendermint/tendermint} | |||||
(see Algorithm~\ref{alg:tendermint}). | |||||
The algorithm requires $N > 3f$, i.e., faulty processes have together the voting power that is smaller than one third of the total voting power. For simplicity we present the algorithm for the case $N = 3f + 1$. We use the notation "upon receiving message $\li{TAG, h, r, v}$ from at least $X$ processes", to denote that a message with the given field values (* denotes any value) has been received from number of processes with the aggregate voting power equals at least $X$. | |||||
The upon rules of Algorithm~\ref{alg:tendermint} are executed atomically. | |||||
Algorithm shares the basic mechanisms called the locking and the unlocking of a value with the DLS algorithm for authenticated faults (the Algorithm 2 from \cite{DLS88:jacm}), but the algorithm structure and the implementation of these mechanisms is different in Tendermint. | |||||
Tendermint consensus algorithm is optimized for the gossip based communication so the consensus layer creates a minimum number of information that needs to be gossiped. Processes exchange only three message types: $\Proposal$, $\Prevote$ and $\Precommit$, and there is only a single mode of execution, i.e., there is no separation between the normal and the recovery mode, which is the case in PBFT-like protocols (e.g., \cite{CL02:tcs}, \cite{Ver09:spinning} or \cite{Cle09:aardvark}). We believe that this makes the protocol simpler to understand and implement correctly, as one of the Tendermint goals is proposing \emph{understandable} Byzantine consensus protocol following Raft path~\cite{Ongaro14:raft}. | |||||
The algorithm proceeds in a sequence of rounds, where each round has a dedicated coordinator (called proposer). The assignment scheme of rounds to coordinators is known to all processes and is given as a function $\coord(height, round)$ returning the coordinator for round $r$ in the consensus instance $height$. In the algorithm~\ref{alg:tendermint} processes agree (decides) on a sequence of values (solves multiple instances problem) by sequentially executing one consensus instances after the other. | |||||
Tendermint decomposes the consensus problem into four relatively independent mechanisms (or subproblems) which are discussed in the following subsections: | |||||
\begin{itemize} | |||||
\item locking mechanism | |||||
\item unlocking mechanism | |||||
\item finding the right value to propose | |||||
\end{itemize} | |||||
\subsection{Locking mechanism} | |||||
\label{sec:locking} | |||||
A process in Tendermint initially does not have any preference about what the decided value should be ($lockedValue := nil$). The process learns about possible decision value from the coordinator of the current round. As a proposer might be a faulty process, different processes | |||||
might receive different suggestions. Using the locking mechanism, correct processes try to converge to a single value within the single round boundary\footnote{We will explain in the part related to safety how we will build on this mechanism to ensure safety across round boundaries.}. | |||||
If during the locking phase, a sufficient number of correct processes succeed in agreeing on the common value, then they will later decide on that value. | |||||
More formally, the locking mechanism ensures that two correct processes can lock only a single value in some round $r$ (they can however potentially lock different values in distinct rounds). In Algorithm~\ref{alg:tendermint}, a correct process $p$ locks a value $v$ in round $r$ by setting a variable $lockedValue_p$ to $v$ and a variable $lockedRound_p$ to $r$ (see line~\ref{line:tab:setLockedValue} and \ref{line:tab:setLockedRound}). A correct process can have only a single value locked at a particular point of the algorithm execution. | |||||
However, before a value is decided, a process might change his mind and unlock its current value using the unlocking mechanism. | |||||
Locking phase starts by a process suggesting what value should be locked in the current round. This is achieved by sending a $\Prevote$ message to all processes. If a correct process has locked some value $v$ ($lockedValue = v \neq \nil$), he votes for $v$ and ignores the suggested value by the proposer (see lines~\ref{line:tab:hasLockedValue}-\ref{line:tab:send-locked-value}). Otherwise, a process might accept the value suggested by the proposer if the proposal is valid\footnote{Validity of a value is an application specific and defined by the function \emph{valid()}.}, and vote for that value (see lines~\ref{line:tab:locked-value-nil}-\ref{line:tab:setStateToPrevote}). In case the suggested value is invalid or a process has not received any proposal from the current coordinator within $\timeoutPropose$ from the start of the current round, it votes with $\nil$ value (see line~\ref{line:tab:send-prevote-nil} and line \ref{line:tab:onTimeoutPropose}). Note that $\nil$ is special value that can not be locked, and it is used to inform others that the algorithm step has not completed successfully; in case of the locking phase, a process has not received a valid value from the proposer on time. | |||||
A process locks value $v$ if it receives at least $2f+1$ $\Prevote$ messages for the value $v$ in the current round (see lines~\ref{line:tab:recvPrevote}-\ref{line:tab:setLockedRound}). A correct process sends only a single $\Prevote$ message in every round. Algorithm relies on the $state$ variable to prevent correct processes from sending more than one $\Prevote$ message in a single round. As a correct process sends a single $\Prevote$ message in a round, and $f$ is the total voting power of faulty processes, two correct processes cannot receive $2f+1$ $\Prevote$ messages for different values in a single round. So if two correct processes lock some value in a given round $r$, then it must be the same value. We will prove this formally in the section~\ref{sec:proof}. | |||||
The locking phase ends by processes informing others if they locked some value in the current round (or not) by sending $\Precommit$ message. If a process has locked some value $v$ in the current round, then this value is sent in the $\Precommit$ message (see line~\ref{line:tab:send-precommit}); otherwise $\nil$ is sent (line~\ref{line:tab:send-precommit-nil} and \ref{line:tab:precommit-nil-onTimeout}). A correct process sends only single $\Precommit$ message in every round. Similarly as with the $\Prevote$ message, the algorithm relies on the $state$ variable to prevent correct processes from sending more than one $\Precommit$ message in a single round. | |||||
\subsection{Unlocking mechanism} | |||||
\label{sec:unlocking} | |||||
Initially, a correct process does not lock any value. During the algorithm execution, processes might end up locking different values in different rounds. As we discussed in the locking section (section~\ref{sec:locking}), a correct process that has locked some value will vote | |||||
only for that value at the beginning of the locking phase. With correct processes voting for different values during the locking phase, it is not possible to reach an agreement what value should be locked, and later decided. Without unlocking mechanism in place, the algorithm could loop forever in that state, and no decision will be reached. The unlocking mechanism is therefore essential for the algorithm termination as it allows processes to release locks and try to converge to a common value in the following rounds. In Algorithm~\ref{alg:tendermint}, a correct process $p$ unlocks a value by setting $lockedValue_p$ to $\nil$ and $lockedRound_p$ to $-1$. | |||||
A correct process $p$ releases its locked value if he observes that some correct process might\footnote{We say that other process \emph{might} have locked different value as we do not get a direct message from that process claiming that he locked a value with the corresponding proof of locking. We infer this information by observing $\Prevote$ messages.} have locked a different value in the round $r$ higher than the current $lockedRound_p$ ($r > lockedRound_p$). This is the case if a process $p$ receives at least $2f+1$ $\Prevote$ messages with value $v' \neq lockedValue_p$ and $round > lockedRound_p$ (see the rule starting at line~\ref{line:tab:unlockRule}). The process also implicitly releases a lock on its current value by locking a different value at lines~\ref{line:tab:setLockedValue} and \ref{line:tab:setLockedRound}. | |||||
The unlocking mechanism, together with the underlying gossip layer, ensures the following properties: | |||||
\begin{itemize} | |||||
\item (i) if a correct process $p$ locks a value $v$ in a round $r$ during a good period (period of synchrony), then at the end of the round $r$ all correct processes will lock either $v$ or they will not lock any value | |||||
\item (ii) if a correct process $p$ locks a value $v$ in a round $r$, then every correct process $q$ that has locked a different value in some of the previous rounds, i.e., $lockedValue_q \neq v$ and $lockedRound_q < r$, will eventually unlock it's value. | |||||
\end{itemize} | |||||
From the property (ii) follows that after a sufficiently big good period in which no correct process manages to lock a value, all correct processes will lock either some common value $v$ or they will not lock any value (where $v$ is the value locked in the highest round $r$ by some correct process). This follows from the fact that processes keep resending messages at the gossip layer, so during good periods, $\Prevote$ messages that led to the correct process locking the value $v$ in the round $r$ are received by all correct processes. Therefore, they either stay with $v$ as its locked value, or unlock its current value. | |||||
Both properties ensure that during a sufficiently long good period, there will eventually be a round, such that at the beginning of this round, all correct processes will lock the same value or they will not lock any value. As we will explain in the following section (\ref{sec:findRighValue}), this is one of the essential mechanisms for ensuring termination of the Tendermint consensus algorithm. | |||||
\subsection{Finding the right value to propose} | |||||
\label{sec:findRighValue} | |||||
As we discussed in section~\ref{sec:locking}, a correct process in Tendermint votes always for its locked value (if it exists) or for the \emph{valid} value suggested by the coordinator of the current round. In order to reach a decision in some round $r$ it is necessary that all correct processes vote for the same value in that round (we do not depend on faulty processes for termination). Then correct processes can lock that value, and later decide. Tendermint relies on the coordinator to ensure that processes that have not locked any value, vote for the same value. As a correct process always votes for its locked value, then several conditions need to be fulfilled just before the round $r$ starts so we can achieve such preferable scenario: | |||||
\begin{enumerate} | |||||
\item if a correct process $p$ has locked some value $v$ at the beginning of the round $r$, then all other correct processes have either locked $v$ or they have not locked any value. \label{enum:termination:v-is-locked} | |||||
\item the proposer is a correct process, and a value suggested by the proposer is $v$. \label{enum:termination:v-is-proposed} | |||||
\end{enumerate} | |||||
In case those conditions hold, all correct processes will lock $v$ in the round $r$ and later decide. | |||||
The existence of a round where the condition \ref{enum:termination:v-is-locked}) holds, follows from the unlocking mechanism. The first part of the condition \ref{enum:termination:v-is-proposed}) follows from the fact that $\coord$ function eventually provides a correct proposer. The second part of the condition \ref{enum:termination:v-is-proposed}) is more tricky as the proposer needs to suggest exactly the value that is locked by correct processes. If the proposer has locked $v$ at the beginning of the round $r$, then it trivially follows from the condition \ref{enum:termination:v-is-locked}). Otherwise, we need an additional mechanism that ensures that the correct proposer suggests $v$. | |||||
In order to ensure this condition, Tendermint has an additional mechanism for determining what is the good value to suggest such that this condition holds. It is implemented using variables $validValue$ and $validRound$. Using these two variables every process tracks the highest round ($validRound$) in which correct process might have locked some value ($validValue$). So whenever process observes at least $2f+1$ $\Prevote$ messages with $round > validRound_p$ (see the rule at line~\ref{line:tab:validValueRule}) for some value $v \neq \nil$, | |||||
it updates $validValue$ and $validRound$ variables. | |||||
This mechanism, together with the underlying gossip layer, ensures the following properties: | |||||
\begin{itemize} | |||||
\item (i) if a correct process $p$ locks a value $v$ in a round $r$ during a good period (period of synchrony), then at the end of the round $r$ all correct processes will have $validValue$ equal to $v$ and $validRound$ equal to $r$. | |||||
\item (ii) if no correct process $p$ locks some value during a sufficiently long good period, then all correct processes will eventually have $validValue$ equal to the same value $v$. | |||||
\end{itemize} | |||||
If during a good period, no correct process locks a value during a sufficiently long period (so the property i) does not hold), then | |||||
all correct processes will have $validValue$ equal to $v$ due to the following reasoning: During this time period, all correct processes will receive delayed $\li{Prevote}$ messages from previous rounds (because of resending messages at the gossip layer). Now let denote with $r$ and $v$, a round and a value, such that there are at least $2f+1$ $\li{Prevote}$ messages with $r$ and $v$, and $r$ is the highest among such $(*, r)$ pairs. Then all correct processes will receive those messages and set $validValue$ to $v$. | |||||
In both cases, during sufficiently long good period, there will eventually be a round, such that at the beginning of this round, all correct processes will have $validValue$ equal to the same value; and in case some correct process has locked some value $v$, then the $validValue$ will be equal to $v$. | |||||
The unlocking mechanism and the mechanism for determining the right proposal value ensures that during a period of synchrony, there will be a round $r$ such that (i) all correct processes either lock the same value $v$ or they do not lock any value, and (ii) the $validValue$ is equal to $v$ at all correct processes. If a proposer of such round $r$ is correct, then he proposes $v$, all correct processes vote for $v$ during the locking phase; therefore they lock $v$, send a $\Precommit$ message with $v$ and later decide. The formal proof of termination is provided in section~\ref{sec:proof}. | |||||
@ -0,0 +1,25 @@ | |||||
# Tendermint-spec | |||||
The repository contains the specification (and the proofs) of the Tendermint | |||||
consensus protocol. | |||||
## How to install Latex on Mac OS | |||||
MacTex is Latex distribution for Mac OS. You can download it [here](http://www.tug.org/mactex/mactex-download.html). | |||||
Popular IDE for Latex-based projects is TexStudio. It can be downloaded | |||||
[here](https://www.texstudio.org/). | |||||
## How to build project | |||||
In order to compile the latex files (and write bibliography), execute | |||||
`$ pdflatex paper` <br/> | |||||
`$ bibtex paper` <br/> | |||||
`$ pdflatex paper` <br/> | |||||
`$ pdflatex paper` <br/> | |||||
The generated file is paper.pdf. You can open it with | |||||
`$ open paper.pdf` | |||||
@ -0,0 +1,16 @@ | |||||
\section{Conclusion} \label{sec:conclusion} | |||||
We have proposed a new Byzantine-fault tolerant consensus algorithm that is the | |||||
core of the Tendermint BFT SMR platform. The algorithm is designed for the wide | |||||
area network with high number of mutually distrusted nodes that communicate | |||||
over gossip based peer-to-peer network. It has only a single mode of execution | |||||
and the communication pattern is very similar to the "normal" case of the | |||||
state-of-the art PBFT algorithm. The algorithm ensures termination with a novel | |||||
mechanism that takes advantage of the gossip based communication between nodes. | |||||
The proposed algorithm and the proofs are simple and elegant, and we believe | |||||
that this makes it easier to understand and implement correctly. | |||||
\section*{Acknowledgment} | |||||
We would like to thank Anton Kaliaev, Ismail Khoffi and Dahlia Malkhi for comments on an earlier version of the paper. We also want to thank Marko Vukolic, Ming Chuan Lin, Maria Potop-Butucaru, Sara Tucci, Antonella Del Pozzo and Yackolley Amoussou-Guenou for pointing out the liveness issues | |||||
in the previous version of the algorithm. Finally, we want to thank the Tendermint team members and all project contributors for making Tendermint such a great platform. |
@ -0,0 +1,397 @@ | |||||
\section{Tendermint consensus algorithm} \label{sec:tendermint} | |||||
\newcommand\Disseminate{\textbf{Disseminate}} | |||||
\newcommand\Proposal{\mathsf{PROPOSAL}} | |||||
\newcommand\ProposalPart{\mathsf{PROPOSAL\mbox{-}PART}} | |||||
\newcommand\PrePrepare{\mathsf{INIT}} \newcommand\Prevote{\mathsf{PREVOTE}} | |||||
\newcommand\Precommit{\mathsf{PRECOMMIT}} | |||||
\newcommand\Decision{\mathsf{DECISION}} | |||||
\newcommand\ViewChange{\mathsf{VC}} | |||||
\newcommand\ViewChangeAck{\mathsf{VC\mbox{-}ACK}} | |||||
\newcommand\NewPrePrepare{\mathsf{VC\mbox{-}INIT}} | |||||
\newcommand\coord{\mathsf{proposer}} | |||||
\newcommand\newHeight{newHeight} \newcommand\newRound{newRound} | |||||
\newcommand\nil{nil} \newcommand\id{id} \newcommand{\propose}{propose} | |||||
\newcommand\prevote{prevote} \newcommand\prevoteWait{prevoteWait} | |||||
\newcommand\precommit{precommit} \newcommand\precommitWait{precommitWait} | |||||
\newcommand\commit{commit} | |||||
\newcommand\timeoutPropose{timeoutPropose} | |||||
\newcommand\timeoutPrevote{timeoutPrevote} | |||||
\newcommand\timeoutPrecommit{timeoutPrecommit} | |||||
\newcommand\proofOfLocking{proof\mbox{-}of\mbox{-}locking} | |||||
\begin{algorithm}[htb!] \def\baselinestretch{1} \scriptsize\raggedright | |||||
\begin{algorithmic}[1] | |||||
\SHORTSPACE | |||||
\INIT{} | |||||
\STATE $h_p := 0$ | |||||
\COMMENT{current height, or consensus instance we are currently executing} | |||||
\STATE $round_p := 0$ \COMMENT{current round number} | |||||
\STATE $step_p \in \set{\propose, \prevote, \precommit}$ | |||||
\STATE $decision_p[] := nil$ | |||||
\STATE $lockedValue_p := nil$ | |||||
\STATE $lockedRound_p := -1$ | |||||
\STATE $validValue_p := nil$ | |||||
\STATE $validRound_p := -1$ | |||||
\ENDINIT | |||||
\SHORTSPACE | |||||
\STATE \textbf{upon} start \textbf{do} $StartRound(0)$ | |||||
\SHORTSPACE | |||||
\FUNCTION{$StartRound(round)$} \label{line:tab:startRound} | |||||
\STATE $round_p \assign round$ | |||||
\STATE $step_p \assign \propose$ | |||||
\IF{$\coord(h_p, round_p) = p$} | |||||
\IF{$validValue_p \neq \nil$} \label{line:tab:isThereLockedValue} | |||||
\STATE $proposal \assign validValue_p$ \ELSE \STATE $proposal \assign | |||||
getValue()$ | |||||
\label{line:tab:getValidValue} | |||||
\ENDIF | |||||
\STATE \Broadcast\ $\li{\Proposal,h_p, round_p, proposal, validRound_p}$ | |||||
\label{line:tab:send-proposal} | |||||
\ELSE | |||||
\STATE \textbf{schedule} $OnTimeoutPropose(h_p, | |||||
round_p)$ to be executed \textbf{after} $\timeoutPropose(round_p)$ | |||||
\ENDIF | |||||
\ENDFUNCTION | |||||
\SPACE | |||||
\UPON{$\li{\Proposal,h_p,round_p, v, -1}$ \From\ $\coord(h_p,round_p)$ | |||||
\With\ $step_p = \propose$} \label{line:tab:recvProposal} | |||||
\IF{$valid(v) \wedge (lockedRound_p = -1 \vee lockedValue_p = v$)} | |||||
\label{line:tab:accept-proposal-2} | |||||
\STATE \Broadcast \ $\li{\Prevote,h_p,round_p,id(v)}$ | |||||
\label{line:tab:prevote-proposal} | |||||
\ELSE | |||||
\label{line:tab:acceptProposal1} | |||||
\STATE \Broadcast \ $\li{\Prevote,h_p,round_p,\nil}$ | |||||
\label{line:tab:prevote-nil} | |||||
\ENDIF | |||||
\STATE $step_p \assign \prevote$ \label{line:tab:setStateToPrevote1} | |||||
\ENDUPON | |||||
\SPACE | |||||
\UPON{$\li{\Proposal,h_p,round_p, v, vr}$ \From\ $\coord(h_p,round_p)$ | |||||
\textbf{AND} $2f+1$ $\li{\Prevote,h_p, vr,id(v)}$ \With\ $step_p = \propose \wedge (vr \ge 0 \wedge vr < round_p)$} | |||||
\label{line:tab:acceptProposal} | |||||
\IF{$valid(v) \wedge (lockedRound_p \le vr | |||||
\vee lockedValue_p = v)$} \label{line:tab:cond-prevote-higher-proposal} | |||||
\STATE \Broadcast \ $\li{\Prevote,h_p,round_p,id(v)}$ | |||||
\label{line:tab:prevote-higher-proposal} | |||||
\ELSE | |||||
\label{line:tab:acceptProposal2} | |||||
\STATE \Broadcast \ $\li{\Prevote,h_p,round_p,\nil}$ | |||||
\label{line:tab:prevote-nil2} | |||||
\ENDIF | |||||
\STATE $step_p \assign \prevote$ \label{line:tab:setStateToPrevote3} | |||||
\ENDUPON | |||||
\SPACE | |||||
\UPON{$2f+1$ $\li{\Prevote,h_p, round_p,*}$ \With\ $step_p = \prevote$ for the first time} | |||||
\label{line:tab:recvAny2/3Prevote} | |||||
\STATE \textbf{schedule} $OnTimeoutPrevote(h_p, round_p)$ to be executed \textbf{after} $\timeoutPrevote(round_p)$ \label{line:tab:timeoutPrevote} | |||||
\ENDUPON | |||||
\SPACE | |||||
\UPON{$\li{\Proposal,h_p,round_p, v, *}$ \From\ $\coord(h_p,round_p)$ | |||||
\textbf{AND} $2f+1$ $\li{\Prevote,h_p, round_p,id(v)}$ \With\ $valid(v) \wedge step_p \ge \prevote$ for the first time} | |||||
\label{line:tab:recvPrevote} | |||||
\IF{$step_p = \prevote$} | |||||
\STATE $lockedValue_p \assign v$ \label{line:tab:setLockedValue} | |||||
\STATE $lockedRound_p \assign round_p$ \label{line:tab:setLockedRound} | |||||
\STATE \Broadcast \ $\li{\Precommit,h_p,round_p,id(v))}$ | |||||
\label{line:tab:precommit-v} | |||||
\STATE $step_p \assign \precommit$ \label{line:tab:setStateToCommit} | |||||
\ENDIF | |||||
\STATE $validValue_p \assign v$ \label{line:tab:setValidRound} | |||||
\STATE $validRound_p \assign round_p$ \label{line:tab:setValidValue} | |||||
\ENDUPON | |||||
\SHORTSPACE | |||||
\UPON{$2f+1$ $\li{\Prevote,h_p,round_p, \nil}$ | |||||
\With\ $step_p = \prevote$} | |||||
\STATE \Broadcast \ $\li{\Precommit,h_p,round_p, \nil}$ | |||||
\label{line:tab:precommit-v-1} | |||||
\STATE $step_p \assign \precommit$ | |||||
\ENDUPON | |||||
\SPACE | |||||
\UPON{$2f+1$ $\li{\Precommit,h_p,round_p,*}$ for the first time} | |||||
\label{line:tab:startTimeoutPrecommit} | |||||
\STATE \textbf{schedule} $OnTimeoutPrecommit(h_p, round_p)$ to be executed \textbf{after} $\timeoutPrecommit(round_p)$ | |||||
\ENDUPON | |||||
\SPACE | |||||
\UPON{$\li{\Proposal,h_p,r, v, *}$ \From\ $\coord(h_p,r)$ \textbf{AND} | |||||
$2f+1$ $\li{\Precommit,h_p,r,id(v)}$ \With\ $decision_p[h_p] = \nil$} | |||||
\label{line:tab:onDecideRule} | |||||
\IF{$valid(v)$} \label{line:tab:validDecisionValue} | |||||
\STATE $decision_p[h_p] = v$ \label{line:tab:decide} | |||||
\STATE$h_p \assign h_p + 1$ \label{line:tab:increaseHeight} | |||||
\STATE reset $lockedRound_p$, $lockedValue_p$, $validRound_p$ and $validValue_p$ to initial values | |||||
and empty message log | |||||
\STATE $StartRound(0)$ | |||||
\ENDIF | |||||
\ENDUPON | |||||
\SHORTSPACE | |||||
\UPON{$f+1$ $\li{*,h_p,round, *, *}$ \textbf{with} $round > round_p$} | |||||
\label{line:tab:skipRounds} | |||||
\STATE $StartRound(round)$ \label{line:tab:nextRound2} | |||||
\ENDUPON | |||||
\SHORTSPACE | |||||
\FUNCTION{$OnTimeoutPropose(height,round)$} \label{line:tab:onTimeoutPropose} | |||||
\IF{$height = h_p \wedge round = round_p \wedge step_p = \propose$} | |||||
\STATE \Broadcast \ $\li{\Prevote,h_p,round_p, \nil}$ | |||||
\label{line:tab:prevote-nil-on-timeout} | |||||
\STATE $step_p \assign \prevote$ | |||||
\ENDIF | |||||
\ENDFUNCTION | |||||
\SHORTSPACE | |||||
\FUNCTION{$OnTimeoutPrevote(height,round)$} \label{line:tab:onTimeoutPrevote} | |||||
\IF{$height = h_p \wedge round = round_p \wedge step_p = \prevote$} | |||||
\STATE \Broadcast \ $\li{\Precommit,h_p,round_p,\nil}$ | |||||
\label{line:tab:precommit-nil-onTimeout} | |||||
\STATE $step_p \assign \precommit$ | |||||
\ENDIF | |||||
\ENDFUNCTION | |||||
\SHORTSPACE | |||||
\FUNCTION{$OnTimeoutPrecommit(height,round)$} \label{line:tab:onTimeoutPrecommit} | |||||
\IF{$height = h_p \wedge round = round_p$} | |||||
\STATE $StartRound(round_p + 1)$ \label{line:tab:nextRound} | |||||
\ENDIF | |||||
\ENDFUNCTION | |||||
\end{algorithmic} \caption{Tendermint consensus algorithm} | |||||
\label{alg:tendermint} | |||||
\end{algorithm} | |||||
In this section we present the Tendermint Byzantine fault-tolerant consensus | |||||
algorithm. The algorithm is specified by the pseudo-code shown in | |||||
Algorithm~\ref{alg:tendermint}. We present the algorithm as a set of \emph{upon | |||||
rules} that are executed atomically\footnote{In case several rules are active | |||||
at the same time, the first rule to be executed is picked randomly. The | |||||
correctness of the algorithm does not depend on the order in which rules are | |||||
executed.}. We assume that processes exchange protocol messages using a gossip | |||||
protocol and that both sent and received messages are stored in a local message | |||||
log for every process. An upon rule is triggered once the message log contains | |||||
messages such that the corresponding condition evaluates to $\tt{true}$. The | |||||
condition that assumes reception of $X$ messages of a particular type and | |||||
content denotes reception of messages whose senders have aggregate voting power at | |||||
least equal to $X$. For example, the condition $2f+1$ $\li{\Precommit,h_p,r,id(v)}$, | |||||
evaluates to true upon reception of $\Precommit$ messages for height $h_p$, | |||||
a round $r$ and with value equal to $id(v)$ whose senders have aggregate voting | |||||
power at least equal to $2f+1$. Some of the rules ends with "for the first time" constraint | |||||
to denote that it is triggered only the first time a corresponding condition evaluates | |||||
to $\tt{true}$. This is because those rules do not always change the state of algorithm | |||||
variables so without this constraint, the algorithm could keep | |||||
executing those rules forever. The variables with index $p$ are process local state | |||||
variables, while variables without index $p$ are value placeholders. The sign | |||||
$*$ denotes any value. | |||||
We denote with $n$ the total voting power of processes in the system, and we | |||||
assume that the total voting power of faulty processes in the system is bounded | |||||
with a system parameter $f$. The algorithm assumes that $n > 3f$, i.e., it | |||||
requires that the total voting power of faulty processes is smaller than one | |||||
third of the total voting power. For simplicity we present the algorithm for | |||||
the case $n = 3f + 1$. | |||||
The algorithm proceeds in rounds, where each round has a dedicated | |||||
\emph{proposer}. The mapping of rounds to proposers is known to all processes | |||||
and is given as a function $\coord(h, round)$, returning the proposer for | |||||
the round $round$ in the consensus instance $h$. We | |||||
assume that the proposer selection function is weighted round-robin, where | |||||
processes are rotated proportional to their voting power\footnote{A validator | |||||
with more voting power is selected more frequently, proportional to its power. | |||||
More precisely, during a sequence of rounds of size $n$, every process is | |||||
proposer in a number of rounds equal to its voting power.}. | |||||
The internal protocol state transitions are triggered by message reception and | |||||
by expiration of timeouts. There are three timeouts in Algorithm \ref{alg:tendermint}: | |||||
$\timeoutPropose$, $\timeoutPrevote$ and $\timeoutPrecommit$. | |||||
The timeouts prevent the algorithm from blocking and | |||||
waiting forever for some condition to be true, ensure that processes continuously | |||||
transition between rounds, and guarantee that eventually (after GST) communication | |||||
between correct processes is timely and reliable so they can decide. | |||||
The last role is achieved by increasing the timeouts with every new round $r$, | |||||
i.e, $timeoutX(r) = initTimeoutX + r*timeoutDelta$; | |||||
they are reset for every new height (consensus | |||||
instance). | |||||
Processes exchange the following messages in Tendermint: $\Proposal$, | |||||
$\Prevote$ and $\Precommit$. The $\Proposal$ message is used by the proposer of | |||||
the current round to suggest a potential decision value, while $\Prevote$ and | |||||
$\Precommit$ are votes for a proposed value. According to the classification of | |||||
consensus algorithms from \cite{RMS10:dsn}, Tendermint, like PBFT | |||||
\cite{CL02:tcs} and DLS \cite{DLS88:jacm}, belongs to class 3, so it requires | |||||
two voting steps (three communication exchanges in total) to decide a value. | |||||
The Tendermint consensus algorithm is designed for the blockchain context where | |||||
the value to decide is a block of transactions (ie. it is potentially quite | |||||
large, consisting of many transactions). Therefore, in the Algorithm | |||||
\ref{alg:tendermint} (similar as in \cite{CL02:tcs}) we are explicit about | |||||
sending a value (block of transactions) and a small, constant size value id (a | |||||
unique value identifier, normally a hash of the value, i.e., if $\id(v) = | |||||
\id(v')$, then $v=v'$). The $\Proposal$ message is the only one carrying the | |||||
value; $\Prevote$ and $\Precommit$ messages carry the value id. A correct | |||||
process decides on a value $v$ in Tendermint upon receiving the $\Proposal$ for | |||||
$v$ and $2f+1$ voting-power equivalent $\Precommit$ messages for $\id(v)$ in | |||||
some round $r$. In order to send $\Precommit$ message for $v$ in a round $r$, a | |||||
correct process waits to receive the $\Proposal$ and $2f+1$ of the | |||||
corresponding $\Prevote$ messages in the round $r$. Otherwise, | |||||
it sends $\Precommit$ message with a special $\nil$ value. | |||||
This ensures that correct processes can $\Precommit$ only a | |||||
single value (or $\nil$) in a round. As | |||||
proposers may be faulty, the proposed value is treated by correct processes as | |||||
a suggestion (it is not blindly accepted), and a correct process tells others | |||||
if it accepted the $\Proposal$ for value $v$ by sending $\Prevote$ message for | |||||
$\id(v)$; otherwise it sends $\Prevote$ message with the special $\nil$ value. | |||||
Every process maintains the following variables in the Algorithm | |||||
\ref{alg:tendermint}: $step$, $lockedValue$, $lockedRound$, $validValue$ and | |||||
$validRound$. The $step$ denotes the current state of the internal Tendermint | |||||
state machine, i.e., it reflects the stage of the algorithm execution in the | |||||
current round. The $lockedValue$ stores the most recent value (with respect to | |||||
a round number) for which a $\Precommit$ message has been sent. The | |||||
$lockedRound$ is the last round in which the process sent a $\Precommit$ | |||||
message that is not $\nil$. We also say that a correct process locks a value | |||||
$v$ in a round $r$ by setting $lockedValue = v$ and $lockedRound = r$ before | |||||
sending $\Precommit$ message for $\id(v)$. As a correct process can decide a | |||||
value $v$ only if $2f+1$ $\Precommit$ messages for $\id(v)$ are received, this | |||||
implies that a possible decision value is a value that is locked by at least | |||||
$f+1$ voting power equivalent of correct processes. Therefore, any value $v$ | |||||
for which $\Proposal$ and $2f+1$ of the corresponding $\Prevote$ messages are | |||||
received in some round $r$ is a \emph{possible decision} value. The role of the | |||||
$validValue$ variable is to store the most recent possible decision value; the | |||||
$validRound$ is the last round in which $validValue$ is updated. Apart from | |||||
those variables, a process also stores the current consensus instance ($h_p$, | |||||
called \emph{height} in Tendermint), and the current round number ($round_p$) | |||||
and attaches them to every message. Finally, a process also stores an array of | |||||
decisions, $decision_p$ (Tendermint assumes a sequence of consensus instances, | |||||
one for each height). | |||||
Every round starts by a proposer suggesting a value with the $\Proposal$ | |||||
message (see line \ref{line:tab:send-proposal}). In the initial round of each | |||||
height, the proposer is free to chose the value to suggest. In the | |||||
Algorithm~\ref{alg:tendermint}, a correct process obtains a value to propose | |||||
using an external function $getValue()$ that returns a valid value to | |||||
propose. In the following rounds, a correct proposer will suggest a new value | |||||
only if $validValue = \nil$; otherwise $validValue$ is proposed (see | |||||
lines~\ref{line:tab:isThereLockedValue}-\ref{line:tab:getValidValue}). | |||||
In addition to the value proposed, the $\Proposal$ message also | |||||
contains the $validRound$ so other processes are informed about the last round | |||||
in which the proposer observed $validValue$ as a possible decision value. | |||||
Note that if a correct proposer $p$ sends $validValue$ with the $validRound$ in the | |||||
$\Proposal$, this implies that the process $p$ received $\Proposal$ and the | |||||
corresponding $2f+1$ $\Prevote$ messages for $validValue$ in the round | |||||
$validRound$. | |||||
If a correct process sends $\Proposal$ message with $validValue$ ($validRound > -1$) | |||||
at time $t > GST$, by the \emph{Gossip communication} property, the | |||||
corresponding $\Proposal$ and the $\Prevote$ messages will be received by all | |||||
correct processes before time $t+\Delta$. Therefore, all correct processes will | |||||
be able to verify the correctness of the suggested value as it is supported by | |||||
the $\Proposal$ and the corresponding $2f+1$ voting power equivalent $\Prevote$ | |||||
messages. | |||||
A correct process $p$ accepts the proposal for a value $v$ (send $\Prevote$ | |||||
for $id(v)$) if an external \emph{valid} function returns $true$ for the value | |||||
$v$, and if $p$ hasn't locked any value ($lockedRound = -1$) or $p$ has locked | |||||
the value $v$ ($lockedValue = v$); see the line | |||||
\ref{line:tab:accept-proposal-2}. In case the proposed pair is $(v,vr \ge 0)$ and a | |||||
correct process $p$ has locked some value, it will accept | |||||
$v$ if it is a more recent possible decision value\footnote{As | |||||
explained above, the possible decision value in a round $r$ is the one for | |||||
which $\Proposal$ and the corresponding $2f+1$ $\Prevote$ messages are received | |||||
for the round $r$.}, $vr > lockedRound_p$, or if $lockedValue = v$ | |||||
(see line~\ref{line:tab:cond-prevote-higher-proposal}). Otherwise, a correct | |||||
process will reject the proposal by sending $\Prevote$ message with $\nil$ | |||||
value. A correct process will send $\Prevote$ message with $\nil$ value also in | |||||
case $\timeoutPropose$ expired (it is triggered when a correct process starts a | |||||
new round) and a process has not sent $\Prevote$ message in the current round | |||||
yet (see the line \ref{line:tab:onTimeoutPropose}). | |||||
If a correct process receives $\Proposal$ message for some value $v$ and $2f+1$ | |||||
$\Prevote$ messages for $\id(v)$, then it sends $\Precommit$ message with | |||||
$\id(v)$. Otherwise, it sends $\Precommit$ $\nil$. A correct process will send | |||||
$\Precommit$ message with $\nil$ value also in case $\timeoutPrevote$ expired | |||||
(it is started when a correct process sent $\Prevote$ message and received any | |||||
$2f+1$ $\Prevote$ messages) and a process has not sent $\Precommit$ message in | |||||
the current round yet (see the line \ref{line:tab:onTimeoutPrecommit}). A | |||||
correct process decides on some value $v$ if it receives in some round $r$ | |||||
$\Proposal$ message for $v$ and $2f+1$ $\Precommit$ messages with $\id(v)$ (see | |||||
the line \ref{line:tab:decide}). To prevent the algorithm from blocking and | |||||
waiting forever for this condition to be true, the Algorithm | |||||
\ref{alg:tendermint} relies on $\timeoutPrecommit$. It is triggered after a | |||||
process receives any set of $2f+1$ $\Precommit$ messages for the current round. | |||||
If the $\timeoutPrecommit$ expires and a process has not decided yet, the | |||||
process starts the next round (see the line \ref{line:tab:onTimeoutPrecommit}). | |||||
When a correct process $p$ decides, it starts the next consensus instance | |||||
(for the next height). The \emph{Gossip communication} property ensures | |||||
that $\Proposal$ and $2f+1$ $\Prevote$ messages that led $p$ to decide | |||||
are eventually received by all correct processes, so they will also decide. | |||||
\subsection{Termination mechanism} | |||||
Tendermint ensures termination by a novel mechanism that benefits from the | |||||
gossip based nature of communication (see \emph{Gossip communication} | |||||
property). It requires managing two additional variables, $validValue$ and | |||||
$validRound$ that are then used by the proposer during the propose step as | |||||
explained above. The $validValue$ and $validRound$ are updated to $v$ and $r$ | |||||
by a correct process in a round $r$ when the process receives valid $\Proposal$ | |||||
message for the value $v$ and the corresponding $2f+1$ $\Prevote$ messages for | |||||
$id(v)$ in the round $r$ (see the rule at line~\ref{line:tab:recvPrevote}). | |||||
We now give briefly the intuition how managing and proposing $validValue$ | |||||
and $validRound$ ensures termination. Formal treatment is left for | |||||
Section~\ref{sec:proof}. | |||||
The first thing to note is that during good period, because of the | |||||
\emph{Gossip communication} property, if a correct process $p$ locks a value | |||||
$v$ in some round $r$, all correct processes will update $validValue$ to $v$ | |||||
and $validRound$ to $r$ before the end of the round $r$ (we prove this formally | |||||
in the Section~\ref{sec:proof}). The intuition is that messages that led to $p$ | |||||
locking a value $v$ in the round $r$ will be gossiped to all correct processes | |||||
before the end of the round $r$, so it will update $validValue$ and | |||||
$validRound$ (the line~\ref{line:tab:recvPrevote}). Therefore, if a correct | |||||
process locks some value during good period, $validValue$ and $validRound$ are | |||||
updated by all correct processes so that the value proposed in the following | |||||
rounds will be acceptable by all correct processes. Note | |||||
that it could happen that during good period, no correct process locks a value, | |||||
but some correct process $q$ updates $validValue$ and $validRound$ during some | |||||
round. As no correct process locks a value in this case, $validValue_q$ and | |||||
$validRound_q$ will also be acceptable by all correct processes as | |||||
$validRound_q > lockedRound_c$ for every correct process $c$ and as the | |||||
\emph{Gossip communication} property ensures that the corresponding $\Prevote$ | |||||
messages that $q$ received in the round $validRound_q$ are received by all | |||||
correct processes $\Delta$ time later. | |||||
Finally, it could happen that after GST, there is a long sequence of rounds in which | |||||
no correct process neither locks a value nor update $validValue$ and $validRound$. | |||||
In this case, during this sequence of rounds, the proposed value suggested by correct | |||||
processes was not accepted by all correct processes. Note that this sequence of rounds | |||||
is always finite as at the beginning of every | |||||
round there is at least a single correct process $c$ such that $validValue_c$ | |||||
and $validRound_c$ are acceptable by every correct process. This is true as | |||||
there exists a correct process $c$ such that for every other correct process | |||||
$p$, $validRound_c > lockedRound_p$ or $validValue_c = lockedValue_p$. This is | |||||
true as $c$ is the process that has locked a value in the most recent round | |||||
among all correct processes (or no correct process locked any value). Therefore, | |||||
eventually $c$ will be the proper in some round and the proposed value will be accepted | |||||
by all correct processes, terminating therefore this sequence of | |||||
rounds. | |||||
Therefore, updating $validValue$ and $validRound$ variables, and the | |||||
\emph{Gossip communication} property, together ensures that eventually, during | |||||
the good period, there exists a round with a correct proposer whose proposed | |||||
value will be accepted by all correct processes, and all correct processes will | |||||
terminate in that round. Note that this mechanism, contrary to the common | |||||
termination mechanism illustrated in the | |||||
Figure~\ref{ch3:fig:coordinator-change}, does not require exchanging any | |||||
additional information in addition to messages already sent as part of what is | |||||
normally being called "normal" case. | |||||
@ -0,0 +1,117 @@ | |||||
\section{Definitions} \label{sec:definitions} | |||||
\subsection{Model} | |||||
We consider a system of processes that communicate by exchanging messages. | |||||
Processes can be correct or faulty, where a faulty process can behave in an | |||||
arbitrary way, i.e., Byzantine faults. We assume that each process | |||||
has some amount of voting power (voting power of a process can be $0$). | |||||
Processes in our model are not part of a single administrative domain; | |||||
therefore we cannot enforce a direct network connectivity between all | |||||
processes. Instead, we assume that each process is connected to a subset of | |||||
processes called peers, such that there is an indirect communication channel | |||||
between all correct processes. Communication between processes is established | |||||
using a gossip protocol \cite{Dem1987:gossip}. | |||||
Formally, we model the network communication using the \emph{partially | |||||
synchronous system model}~\cite{DLS88:jacm}: in all executions of the system | |||||
there is a bound $\Delta$ and an instant GST (Global Stabilization Time) such | |||||
that all communication among correct processes after GST is reliable and | |||||
$\Delta$-timely, i.e., if a correct process $p$ sends message $m$ at time $t | |||||
\ge GST$ to correct process $q$, then $q$ will receive $m$ before $t + | |||||
\Delta$\footnote{Note that as we do not assume direct communication channels | |||||
among all correct processes, this implies that before the message $m$ | |||||
reaches $q$, it might pass through a number of correct processes that will | |||||
forward the message $m$ using gossip protocol towards $q$.}. Messages among | |||||
correct processes can be delayed, dropped or duplicated before GST. | |||||
Spoofing/impersonation attacks are assumed to be impossible at all times due to | |||||
the use of public-key cryptography. The bound $\Delta$ and GST are system | |||||
parameters whose values are not required to be known for the safety of our | |||||
algorithm. Termination of the algorithm is guaranteed within a bounded duration | |||||
after GST. In practice, the algorithm will work correctly in the slightly | |||||
weaker variant of the model where the system alternates between (long enough) | |||||
good periods (corresponds to the \emph{after} GST period where system is | |||||
reliable and $\Delta$-timely) and bad periods (corresponds to the period | |||||
\emph{before} GST during which the system is asynchronous and messages can be | |||||
lost), but consideration of the GST model simplifies the discussion. | |||||
We assume that process steps (which might include sending and receiving | |||||
messages) take zero time. Processes are equipped with clocks so they can | |||||
measure local timeouts. All protocol messages are signed, i.e., when a correct | |||||
process $q$ receives a signed message $m$ from its peer, the process $q$ can | |||||
verify who was the original sender of the message $m$. | |||||
The details of the Tendermint gossip protocol will be discussed in a separate | |||||
technical report. For the sake of this paper it is sufficient to assume that | |||||
messages are being gossiped between processes and the following property holds | |||||
(in addition to the partial synchrony network assumptions): | |||||
\begin{itemize} \item \emph{Gossip communication:} If a correct process $p$ | |||||
receives some message $m$ at time $t$, all correct processes will receive | |||||
$m$ before $max\{t, GST\} + \Delta$. \end{itemize} | |||||
%Messages that are being gossiped are created by the consensus layer. We can | |||||
%think about consensus protocol as a content creator, which %defines what | |||||
%messages should be disseminated using the gossip protocol. A correct | |||||
%process creates the message for dissemination either i) %explicitly, by | |||||
%invoking \emph{send} function as part of the consensus protocol or ii) | |||||
%implicitly, by receiving a message from some other %process. Note that in | |||||
%the case ii) gossiping of messages is implicit, i.e., it happens without | |||||
%explicit send clause in the consensus algorithm %whenever a correct | |||||
%process receives some messages in the consensus algorithm\footnote{If a | |||||
%message is received by a correct process at %the consensus level then it | |||||
%is considered valid from the protocol point of view, i.e., it has a | |||||
%correct signature, a proper message structure %and a valid height and | |||||
%round number.}. | |||||
%\item Processes keep resending messages (in case of failures or message loss) | |||||
%until all its peers get them. This ensures that every message %sent or | |||||
%received by a correct process is eventually received by all correct | |||||
%processes. | |||||
\subsection{State Machine Replication} | |||||
State machine replication (SMR) is a general approach for replicating services | |||||
modeled as a deterministic state machine~\cite{Lam78:cacm,Sch90:survey}. The | |||||
key idea of this approach is to guarantee that all replicas start in the same | |||||
state and then apply requests from clients in the same order, thereby | |||||
guaranteeing that the replicas' states will not diverge. Following | |||||
Schneider~\cite{Sch90:survey}, we note that the following is key for | |||||
implementing a replicated state machine tolerant to (Byzantine) faults: | |||||
\begin{itemize} \item \emph{Replica Coordination.} All [non-faulty] replicas | |||||
receive and process the same sequence of requests. \end{itemize} | |||||
Moreover, as Schneider also notes, this property can be decomposed into two | |||||
parts, \emph{Agreement} and \emph{Order}: Agreement requires all (non-faulty) | |||||
replicas to receive all requests, and Order requires that the order of received | |||||
requests is the same at all replicas. | |||||
There is an additional requirement that needs to be ensured by Byzantine | |||||
tolerant state machine replication: only requests (called transactions in the | |||||
Tendermint terminology) proposed by clients are executed. In Tendermint, | |||||
transaction verification is the responsibility of the service that is being | |||||
replicated; upon receiving a transaction from the client, the Tendermint | |||||
process will ask the service if the request is valid, and only valid requests | |||||
will be processed. | |||||
\subsection{Consensus} \label{sec:consensus} | |||||
Tendermint solves state machine replication by sequentially executing consensus | |||||
instances to agree on each block of transactions that are | |||||
then executed by the service being replicated. We consider a variant of the | |||||
Byzantine consensus problem called Validity Predicate-based Byzantine consensus | |||||
that is motivated by blockchain systems~\cite{GLR17:red-belly-bc}. The problem | |||||
is defined by an agreement, a termination, and a validity property. | |||||
\begin{itemize} \item \emph{Agreement:} No two correct processes decide on | |||||
different values. \item \emph{Termination:} All correct processes | |||||
eventually decide on a value. \item \emph{Validity:} A decided value | |||||
is valid, i.e., it satisfies the predefined predicate denoted | |||||
\emph{valid()}. \end{itemize} | |||||
This variant of the Byzantine consensus problem has an application-specific | |||||
\emph{valid()} predicate to indicate whether a value is valid. In the context | |||||
of blockchain systems, for example, a value is not valid if it does not | |||||
contain an appropriate hash of the last value (block) added to the blockchain. |
@ -0,0 +1,138 @@ | |||||
\section{Introduction} \label{sec:tendermint} | |||||
Consensus is a fundamental problem in distributed computing. It | |||||
is important because of it's role in State Machine Replication (SMR), a generic | |||||
approach for replicating services that can be modeled as a deterministic state | |||||
machine~\cite{Lam78:cacm, Sch90:survey}. The key idea of this approach is that | |||||
service replicas start in the same initial state, and then execute requests | |||||
(also called transactions) in the same order; thereby guaranteeing that | |||||
replicas stay in sync with each other. The role of consensus in the SMR | |||||
approach is ensuring that all replicas receive transactions in the same order. | |||||
Traditionally, deployments of SMR based systems are in data-center settings | |||||
(local area network), have a small number of replicas (three to seven) and are | |||||
typically part of a single administration domain (e.g., Chubby | |||||
\cite{Bur:osdi06}); therefore they handle benign (crash) failures only, as more | |||||
general forms of failure (in particular, malicious or Byzantine faults) are | |||||
considered to occur with only negligible probability. | |||||
The success of cryptocurrencies and blockchain systems in recent years (e.g., | |||||
\cite{Nak2012:bitcoin, But2014:ethereum}) pose a whole new set of challenges on | |||||
the design and deployment of SMR based systems: reaching agreement over wide | |||||
area network, among large number of nodes (hundreds or thousands) that are not | |||||
part of the same administrative domain, and where a subset of nodes can behave | |||||
maliciously (Byzantine faults). Furthermore, contrary to the previous | |||||
data-center deployments where nodes are fully connected to each other, in | |||||
blockchain systems, a node is only connected to a subset of other nodes, so | |||||
communication is achieved by gossip-based peer-to-peer protocols. | |||||
The new requirements demand designs and algorithms that are not necessarily | |||||
present in the classical academic literature on Byzantine fault tolerant | |||||
consensus (or SMR) systems (e.g., \cite{DLS88:jacm, CL02:tcs}) as the primary | |||||
focus was different setup. | |||||
In this paper we describe a novel Byzantine-fault tolerant consensus algorithm | |||||
that is the core of the BFT SMR platform called Tendermint\footnote{The | |||||
Tendermint platform is available open source at | |||||
https://github.com/tendermint/tendermint.}. The Tendermint platform consists of | |||||
a high-performance BFT SMR implementation written in Go, a flexible interface | |||||
for | |||||
building arbitrary deterministic applications above the consensus, and a suite | |||||
of tools for deployment and management. | |||||
The Tendermint consensus algorithm is inspired by the PBFT SMR | |||||
algorithm~\cite{CL99:osdi} and the DLS algorithm for authenticated faults (the | |||||
Algorithm 2 from \cite{DLS88:jacm}). Similar to DLS algorithm, Tendermint | |||||
proceeds in | |||||
rounds\footnote{Tendermint is not presented in the basic round model of | |||||
\cite{DLS88:jacm}. Furthermore, we use the term round differently than in | |||||
\cite{DLS88:jacm}; in Tendermint a round denotes a sequence of communication | |||||
steps instead of a single communication step in \cite{DLS88:jacm}.}, where each | |||||
round has a dedicated proposer (also called coordinator or | |||||
leader) and a process proceeds to a new round as part of normal | |||||
processing (not only in case the proposer is faulty or suspected as being faulty | |||||
by enough processes as in PBFT). | |||||
The communication pattern of each round is very similar to the "normal" case | |||||
of PBFT. Therefore, in preferable conditions (correct proposer, timely and | |||||
reliable communication between correct processes), Tendermint decides in three | |||||
communication steps (the same as PBFT). | |||||
The major novelty and contribution of the Tendermint consensus algorithm is a | |||||
new termination mechanism. As explained in \cite{MHS09:opodis, RMS10:dsn}, the | |||||
existing BFT consensus (and SMR) algorithms for the partially synchronous | |||||
system model (for example PBFT~\cite{CL99:osdi}, \cite{DLS88:jacm}, | |||||
\cite{MA06:tdsc}) typically relies on the communication pattern illustrated in | |||||
Figure~\ref{ch3:fig:coordinator-change} for termination. The | |||||
Figure~\ref{ch3:fig:coordinator-change} illustrates messages exchanged during | |||||
the proposer change when processes start a new round\footnote{There is no | |||||
consistent terminology in the distributed computing terminology on naming | |||||
sequence of communication steps that corresponds to a logical unit. It is | |||||
sometimes called a round, phase or a view.}. It guarantees that eventually (ie. | |||||
after some Global Stabilization Time, GST), there exists a round with a correct | |||||
proposer that will bring the system into a univalent configuration. | |||||
Intuitively, in a round in which the proposed value is accepted | |||||
by all correct processes, and communication between correct processes is | |||||
timely and reliable, all correct processes decide. | |||||
\begin{figure}[tbh!] \def\rdstretch{5} \def\ystretch{3} \centering | |||||
\begin{rounddiag}{4}{2} \round{1}{~} \rdmessage{1}{1}{$v_1$} | |||||
\rdmessage{2}{1}{$v_2$} \rdmessage{3}{1}{$v_3$} \rdmessage{4}{1}{$v_4$} | |||||
\round{2}{~} \rdmessage{1}{1}{$x, [v_{1..4}]$} | |||||
\rdmessage{1}{2}{$~~~~~~x, [v_{1..4}]$} \rdmessage{1}{3}{$~~~~~~~~x, | |||||
[v_{1..4}]$} \rdmessage{1}{4}{$~~~~~~~x, [v_{1..4}]$} \end{rounddiag} | |||||
\vspace{-5mm} \caption{\boldmath Proposer (coordinator) change: $p_1$ is the | |||||
new proposer.} \label{ch3:fig:coordinator-change} \end{figure} | |||||
To ensure that a proposed value is accepted by all correct | |||||
processes\footnote{The proposed value is not blindly accepted by correct | |||||
processes in BFT algorithms. A correct process always verifies if the proposed | |||||
value is safe to be accepted so that safety properties of consensus are not | |||||
violated.} | |||||
a proposer will 1) build the global state by receiving messages from other | |||||
processes, 2) select the safe value to propose and 3) send the selected value | |||||
together with the signed messages | |||||
received in the first step to support it. The | |||||
value $v_i$ that a correct process sends to the next proposer normally | |||||
corresponds to a value the process considers as acceptable for a decision: | |||||
\begin{itemize} \item in PBFT~\cite{CL99:osdi} and DLS~\cite{DLS88:jacm} it is | |||||
not the value itself but a set of $2f+1$ signed messages with the same | |||||
value id, \item in Fast Byzantine Paxos~\cite{MA06:tdsc} the value | |||||
itself is being sent. \end{itemize} | |||||
In both cases, using this mechanism in our system model (ie. high | |||||
number of nodes over gossip based network) would have high communication | |||||
complexity that increases with the number of processes: in the first case as | |||||
the message sent depends on the total number of processes, and in the second | |||||
case as the value (block of transactions) is sent by each process. The set of | |||||
messages received in the first step are normally piggybacked on the proposal | |||||
message (in the Figure~\ref{ch3:fig:coordinator-change} denoted with | |||||
$[v_{1..4}]$) to justify the choice of the selected value $x$. Note that | |||||
sending this message also does not scale with the number of processes in the | |||||
system. | |||||
We designed a novel termination mechanism for Tendermint that better suits the | |||||
system model we consider. It does not require additional communication (neither | |||||
sending new messages nor piggybacking information on the existing messages) and | |||||
it is fully based on the communication pattern that is very similar to the | |||||
normal case in PBFT \cite{CL99:osdi}. Therefore, there is only a single mode of | |||||
execution in Tendermint, i.e., there is no separation between the normal and | |||||
the recovery mode, which is the case in other PBFT-like protocols (e.g., | |||||
\cite{CL99:osdi}, \cite{Ver09:spinning} or \cite{Cle09:aardvark}). We believe | |||||
this makes Tendermint simpler to understand and implement correctly. | |||||
Note that the orthogonal approach for reducing message complexity in order to | |||||
improve | |||||
scalability and decentralization (number of processes) of BFT consensus | |||||
algorithms is using advanced cryptography (for example Boneh-Lynn-Shacham (BLS) | |||||
signatures \cite{BLS2001:crypto}) as done for example in SBFT | |||||
\cite{Gue2018:sbft}. | |||||
The remainder of the paper is as follows: Section~\ref{sec:definitions} defines | |||||
the system model and gives the problem definitions. Tendermint | |||||
consensus algorithm is presented in Section~\ref{sec:tendermint} and the | |||||
proofs are given in Section~\ref{sec:proof}. We conclude in | |||||
Section~\ref{sec:conclusion}. | |||||
@ -0,0 +1,280 @@ | |||||
\section{Proof of Tendermint consensus algorithm} \label{sec:proof} | |||||
\begin{lemma} \label{lemma:majority-intersection} For all $f\geq 0$, any two | |||||
sets of processes with voting power at least equal to $2f+1$ have at least one | |||||
correct process in common. \end{lemma} | |||||
\begin{proof} As the total voting power is equal to $n=3f+1$, we have $2(2f+1) | |||||
= n+f+1$. This means that the intersection of two sets with the voting | |||||
power equal to $2f+1$ contains at least $f+1$ voting power in common, \ie, | |||||
at least one correct process (as the total voting power of faulty processes | |||||
is $f$). The result follows directly from this. \end{proof} | |||||
\begin{lemma} \label{lemma:locked-decision_value-prevote-v} If $f+1$ correct | |||||
processes lock value $v$ in round $r_0$ ($lockedValue = v$ and $lockedRound = | |||||
r_0$), then in all rounds $r > r_0$, they send $\Prevote$ for $id(v)$ or | |||||
$\nil$. \end{lemma} | |||||
\begin{proof} We prove the result by induction on $r$. | |||||
\emph{Base step $r = r_0 + 1:$} Let's denote with $C$ the set of correct | |||||
processes with voting power equal to $f+1$. By the rules at | |||||
line~\ref{line:tab:recvProposal} and line~\ref{line:tab:acceptProposal}, the | |||||
processes from the set $C$ can't accept $\Proposal$ for any value different | |||||
from $v$ in round $r$, and therefore can't send a $\li{\Prevote,height_p, | |||||
r,id(v')}$ message, if $v' \neq v$. Therefore, the Lemma holds for the base | |||||
step. | |||||
\emph{Induction step from $r_1$ to $r_1+1$:} We assume that no process from the | |||||
set $C$ has sent $\Prevote$ for values different than $id(v)$ or $\nil$ until | |||||
round $r_1 + 1$. We now prove that the Lemma also holds for round $r_1 + 1$. As | |||||
processes from the set $C$ send $\Prevote$ for $id(v)$ or $\nil$ in rounds $r_0 | |||||
\le r \le r_1$, by Lemma~\ref{lemma:majority-intersection} there is no value | |||||
$v' \neq v$ for which it is possible to receive $2f+1$ $\Prevote$ messages in | |||||
those rounds (i). Therefore, we have for all processes from the set $C$, | |||||
$lockedValue = v$ and $lockedRound \ge r_0$. Let's assume by a contradiction | |||||
that a process $q$ from the set $C$ sends $\Prevote$ in round $r_1 + 1$ for | |||||
value $id(v')$, where $v' \neq v$. This is possible only by | |||||
line~\ref{line:tab:prevote-higher-proposal}. Note that this implies that $q$ | |||||
received $2f+1$ $\li{\Prevote,h_q, r,id(v')}$ messages, where $r > r_0$ and $r | |||||
< r_1 +1$ (see line~\ref{line:tab:cond-prevote-higher-proposal}). A | |||||
contradiction with (i) and Lemma~\ref{lemma:majority-intersection}. | |||||
\end{proof} | |||||
\begin{lemma} \label{lemma:agreement} Algorithm~\ref{alg:tendermint} satisfies | |||||
Agreement. \end{lemma} | |||||
\begin{proof} Let round $r_0$ be the first round of height $h$ such that some | |||||
correct process $p$ decides $v$. We now prove that if some correct process | |||||
$q$ decides $v'$ in some round $r \ge r_0$, then $v = v'$. | |||||
In case $r = r_0$, $q$ has received at least $2f+1$ | |||||
$\li{\Precommit,h_p,r_0,id(v')}$ messages at line~\ref{line:tab:onDecideRule}, | |||||
while $p$ has received at least $2f+1$ $\li{\Precommit,h_p,r_0,id(v)}$ | |||||
messages. By Lemma~\ref{lemma:majority-intersection} two sets of messages of | |||||
voting power $2f+1$ intersect in at least one correct process. As a correct | |||||
process sends a single $\Precommit$ message in a round, then $v=v'$. | |||||
We prove the case $r > r_0$ by contradiction. By the | |||||
rule~\ref{line:tab:onDecideRule}, $p$ has received at least $2f+1$ voting-power | |||||
equivalent of $\li{\Precommit,h_p,r_0,id(v)}$ messages, i.e., at least $f+1$ | |||||
voting-power equivalent correct processes have locked value $v$ in round $r_0$ and have | |||||
sent those messages (i). Let denote this set of messages with $C$. On the | |||||
other side, $q$ has received at least $2f+1$ voting power equivalent of | |||||
$\li{\Precommit,h_q, r,id(v')}$ messages. As the voting power of all faulty | |||||
processes is at most $f$, some correct process $c$ has sent one of those | |||||
messages. By the rule at line~\ref{line:tab:recvPrevote}, $c$ has locked value | |||||
$v'$ in round $r$ before sending $\li{\Precommit,h_q, r,id(v')}$. Therefore $c$ | |||||
has received $2f+1$ $\Prevote$ messages for $id(v')$ in round $r > r_0$ (see | |||||
line~\ref{line:tab:recvPrevote}). By Lemma~\ref{lemma:majority-intersection}, a | |||||
process from the set $C$ has sent $\Prevote$ message for $id(v')$ in round $r$. | |||||
A contradiction with (i) and Lemma~\ref{lemma:locked-decision_value-prevote-v}. | |||||
\end{proof} | |||||
\begin{lemma} \label{lemma:agreement} Algorithm~\ref{alg:tendermint} satisfies | |||||
Validity. \end{lemma} | |||||
\begin{proof} Trivially follows from the rule at line | |||||
\ref{line:tab:validDecisionValue} which ensures that only valid values can be | |||||
decided. \end{proof} | |||||
\begin{lemma} \label{lemma:round-synchronisation} If we assume that: | |||||
\begin{enumerate} | |||||
\item a correct process $p$ is the first correct process to | |||||
enter a round $r>0$ at time $t > GST$ (for every correct process | |||||
$c$, $round_c \le r$ at time $t$) | |||||
\item the proposer of round $r$ is | |||||
a correct process $q$ | |||||
\item for every correct process $c$, | |||||
$lockedRound_c \le validRound_q$ at time $t$ | |||||
\item $\timeoutPropose(r) | |||||
> 2\Delta + \timeoutPrecommit(r-1)$, $\timeoutPrevote(r) > 2\Delta$ and | |||||
$\timeoutPrecommit(r) > 2\Delta$, | |||||
\end{enumerate} | |||||
then all correct processes decide in round $r$ before $t + 4\Delta + | |||||
\timeoutPrecommit(r-1)$. | |||||
\end{lemma} | |||||
\begin{proof} As $p$ is the first correct process to enter round $r$, it | |||||
executed the line~\ref{line:tab:nextRound} after $\timeoutPrecommit(r-1)$ | |||||
expired. Therefore, $p$ received $2f+1$ $\Precommit$ messages in the round | |||||
$r-1$ before time $t$. By the \emph{Gossip communication} property, all | |||||
correct processes will receive those messages the latest at time $t + | |||||
\Delta$. Correct processes that are in rounds $< r-1$ at time $t$ will | |||||
enter round $r-1$ (see the rule at line~\ref{line:tab:nextRound2}) and | |||||
trigger $\timeoutPrecommit(r-1)$ (see rule~\ref{line:tab:startTimeoutPrecommit}) | |||||
by time $t+\Delta$. Therefore, all correct processes will start round $r$ | |||||
by time $t+\Delta+\timeoutPrecommit(r-1)$ (i). | |||||
In the worst case, the process $q$ is the last correct process to enter round | |||||
$r$, so $q$ starts round $r$ and sends $\Proposal$ message for some value $v$ | |||||
at time $t + \Delta + \timeoutPrecommit(r-1)$. Therefore, all correct processes | |||||
receive the $\Proposal$ message from $q$ the latest by time $t + 2\Delta + | |||||
\timeoutPrecommit(r-1)$. Therefore, if $\timeoutPropose(r) > 2\Delta + | |||||
\timeoutPrecommit(r-1)$, all correct processes will receive $\Proposal$ message | |||||
before $\timeoutPropose(r)$ expires. | |||||
By (3) and the rules at line~\ref{line:tab:recvProposal} and | |||||
\ref{line:tab:acceptProposal}, all correct processes will accept the | |||||
$\Proposal$ message for value $v$ and will send a $\Prevote$ message for | |||||
$id(v)$ by time $t + 2\Delta + \timeoutPrecommit(r-1)$. Note that by the | |||||
\emph{Gossip communication} property, the $\Prevote$ messages needed to trigger | |||||
the rule at line~\ref{line:tab:acceptProposal} are received before time $t + | |||||
\Delta$. | |||||
By time $t + 3\Delta + \timeoutPrecommit(r-1)$, all correct processes will receive | |||||
$\Proposal$ for $v$ and $2f+1$ corresponding $\Prevote$ messages for $id(v)$. | |||||
By the rule at line~\ref{line:tab:recvPrevote}, all correct processes will send | |||||
a $\Precommit$ message (see line~\ref{line:tab:precommit-v}) for $id(v)$ by | |||||
time $t + 3\Delta + \timeoutPrecommit(r-1)$. Therefore, by time $t + 4\Delta + | |||||
\timeoutPrecommit(r-1)$, all correct processes will have received the $\Proposal$ | |||||
for $v$ and $2f+1$ $\Precommit$ messages for $id(v)$, so they decide at | |||||
line~\ref{line:tab:decide} on $v$. | |||||
This scenario holds if every correct process $q$ sends a $\Precommit$ message | |||||
before $\timeoutPrevote(r)$ expires, and if $\timeoutPrecommit(r)$ does not expire | |||||
before $t + 4\Delta + \timeoutPrecommit(r-1)$. Let's assume that a correct process | |||||
$c_1$ is the first correct process to trigger $\timeoutPrevote(r)$ (see the rule | |||||
at line~\ref{line:tab:recvAny2/3Prevote}) at time $t_1 > t$. This implies that | |||||
before time $t_1$, $c_1$ received a $\Proposal$ ($step_{c_1}$ must be | |||||
$\prevote$ by the rule at line~\ref{line:tab:recvAny2/3Prevote}) and a set of | |||||
$2f+1$ $\Prevote$ messages. By time $t_1 + \Delta$, all correct processes will | |||||
receive those messages. Note that even if some correct process was in the | |||||
smaller round before time $t_1$, at time $t_1 + \Delta$ it will start round $r$ | |||||
after receiving those messages (see the rule at | |||||
line~\ref{line:tab:skipRounds}). Therefore, all correct processes will send | |||||
their $\Prevote$ message for $id(v)$ by time $t_1 + \Delta$, and all correct | |||||
processes will receive those messages the by time $t_1 + 2\Delta$. Therefore, | |||||
as $\timeoutPrevote(r) > 2\Delta$, this ensures that all correct processes receive | |||||
$\Prevote$ messages from all correct processes before their respective local | |||||
$\timeoutPrevote(r)$ expire. | |||||
On the other hand, $\timeoutPrecommit(r)$ is triggered in a correct process $c_2$ | |||||
after it receives any set of $2f+1$ $\Precommit$ messages for the first time. | |||||
Let's denote with $t_2 > t$ the earliest point in time $\timeoutPrecommit(r)$ is | |||||
triggered in some correct process $c_2$. This implies that $c_2$ has received | |||||
at least $f+1$ $\Precommit$ messages for $id(v)$ from correct processes, i.e., | |||||
those processes have received $\Proposal$ for $v$ and $2f+1$ $\Prevote$ | |||||
messages for $id(v)$ before time $t_2$. By the \emph{Gossip communication} | |||||
property, all correct processes will receive those messages by time $t_2 + | |||||
\Delta$, and will send $\Precommit$ messages for $id(v)$. Note that even if | |||||
some correct processes were at time $t_2$ in a round smaller than $r$, by the | |||||
rule at line~\ref{line:tab:skipRounds} they will enter round $r$ by time $t_2 + | |||||
\Delta$. Therefore, by time $t_2 + 2\Delta$, all correct processes will | |||||
receive $\Proposal$ for $v$ and $2f+1$ $\Precommit$ messages for $id(v)$. So if | |||||
$\timeoutPrecommit(r) > 2\Delta$, all correct processes will decide before the | |||||
timeout expires. \end{proof} | |||||
\begin{lemma} \label{lemma:validValue} If a correct process $p$ locks a value | |||||
$v$ at time $t_0 > GST$ in some round $r$ ($lockedValue = v$ and | |||||
$lockedRound = r$) and $\timeoutPrecommit(r) > 2\Delta$, then all correct | |||||
processes set $validValue$ to $v$ and $validRound$ to $r$ before starting | |||||
round $r+1$. \end{lemma} | |||||
\begin{proof} In order to prove this Lemma, we need to prove that if the | |||||
process $p$ locks a value $v$ at time $t_0$, then no correct process will | |||||
leave round $r$ before time $t_0 + \Delta$ (unless it has already set | |||||
$validValue$ to $v$ and $validRound$ to $r$). It is sufficient to prove | |||||
this, since by the \emph{Gossip communication} property the messages that | |||||
$p$ received at time $t_0$ and that triggered rule at | |||||
line~\ref{line:tab:recvPrevote} will be received by time $t_0 + \Delta$ by | |||||
all correct processes, so all correct processes that are still in round $r$ | |||||
will set $validValue$ to $v$ and $validRound$ to $r$ (by the rule at | |||||
line~\ref{line:tab:recvPrevote}). To prove this, we need to compute the | |||||
earliest point in time a correct process could leave round $r$ without | |||||
updating $validValue$ to $v$ and $validRound$ to $r$ (we denote this time | |||||
with $t_1$). The Lemma is correct if $t_0 + \Delta < t_1$. | |||||
If the process $p$ locks a value $v$ at time $t_0$, this implies that $p$ | |||||
received the valid $\Proposal$ message for $v$ and $2f+1$ | |||||
$\li{\Prevote,h,r,id(v)}$ at time $t_0$. At least $f+1$ of those messages are | |||||
sent by correct processes. Let's denote this set of correct processes as $C$. By | |||||
Lemma~\ref{lemma:majority-intersection} any set of $2f+1$ $\Prevote$ messages | |||||
in round $r$ contains at least a single message from the set $C$. | |||||
Let's denote as time $t$ the earliest point in time a correct process, $c_1$, triggered | |||||
$\timeoutPrevote(r)$. This implies that $c_1$ received $2f+1$ $\Prevote$ messages | |||||
(see the rule at line \ref{line:tab:recvAny2/3Prevote}), where at least one of | |||||
those messages was sent by a process $c_2$ from the set $C$. Therefore, process | |||||
$c_2$ had received $\Proposal$ message before time $t$. By the \emph{Gossip | |||||
communication} property, all correct processes will receive $\Proposal$ and | |||||
$2f+1$ $\Prevote$ messages for round $r$ by time $t+\Delta$. The latest point | |||||
in time $p$ will trigger $\timeoutPrevote(r)$ is $t+\Delta$\footnote{Note that | |||||
even if $p$ was in smaller round at time $t$ it will start round $r$ by time | |||||
$t+\Delta$.}. So the latest point in time $p$ can lock the value $v$ in | |||||
round $r$ is $t_0 = t+\Delta+\timeoutPrevote(r)$ (as at this point | |||||
$\timeoutPrevote(r)$ expires, so a process sends $\Precommit$ $\nil$ and updates | |||||
$step$ to $\precommit$, see line \ref{line:tab:onTimeoutPrevote}). | |||||
Note that according to the Algorithm \ref{alg:tendermint}, a correct process | |||||
can not send a $\Precommit$ message before receiving $2f+1$ $\Prevote$ | |||||
messages. Therefore, no correct process can send a $\Precommit$ message in | |||||
round $r$ before time $t$. If a correct process sends a $\Precommit$ message | |||||
for $\nil$, it implies that it has waited for the full duration of | |||||
$\timeoutPrevote(r)$ (see line | |||||
\ref{line:tab:precommit-nil-onTimeout})\footnote{The other case in which a | |||||
correct process $\Precommit$ for $\nil$ is after receiving $2f+1$ $Prevote$ for | |||||
$\nil$ messages, see the line \ref{line:tab:precommit-v-1}. By | |||||
Lemma~\ref{lemma:majority-intersection}, this is not possible in round $r$.}. | |||||
Therefore, no correct process can send $\Precommit$ for $\nil$ before time $t + | |||||
\timeoutPrevote(r)$ (*). | |||||
A correct process $q$ that enters round $r+1$ must wait (i) $\timeoutPrecommit(r)$ | |||||
(see line \ref{line:tab:nextRound}) or (ii) receiving $f+1$ messages from the | |||||
round $r+1$ (see the line \ref{line:tab:skipRounds}). In the former case, $q$ | |||||
receives $2f+1$ $\Precommit$ messages before starting $\timeoutPrecommit(r)$. If | |||||
at least a single $\Precommit$ message from a correct process (at least $f+1$ | |||||
voting power equivalent of those messages is sent by correct processes) is for | |||||
$\nil$, then $q$ cannot start round $r+1$ before time $t_1 = t + | |||||
\timeoutPrevote(r) + \timeoutPrecommit(r)$ (see (*)). Therefore in this case we have: | |||||
$t_0 + \Delta < t_1$, i.e., $t+2\Delta+\timeoutPrevote(r) < t + \timeoutPrevote(r) + | |||||
\timeoutPrecommit(r)$, and this is true whenever $\timeoutPrecommit(r) > 2\Delta$, so | |||||
Lemma holds in this case. | |||||
If in the set of $2f+1$ $\Precommit$ messages $q$ receives, there is at least a | |||||
single $\Precommit$ for $id(v)$ message from a correct process $c$, then $q$ | |||||
can start the round $r+1$ the earliest at time $t_1 = t+\timeoutPrecommit(r)$. In | |||||
this case, by the \emph{Gossip communication} property, all correct processes | |||||
will receive $\Proposal$ and $2f+1$ $\Prevote$ messages (that $c$ received | |||||
before time $t$) the latest at time $t+\Delta$. Therefore, $q$ will set | |||||
$validValue$ to $v$ and $validRound$ to $r$ the latest at time $t+\Delta$. As | |||||
$t+\Delta < t+\timeoutPrecommit(r)$, whenever $\timeoutPrecommit(r) > \Delta$, the | |||||
Lemma holds also in this case. | |||||
In case (ii), $q$ received at least a single message from a correct process $c$ | |||||
from the round $r+1$. The earliest point in time $c$ could have started round | |||||
$r+1$ is $t+\timeoutPrecommit(r)$ in case it received a $\Precommit$ message for | |||||
$v$ from some correct process in the set of $2f+1$ $\Precommit$ messages it | |||||
received. The same reasoning as above holds also in this case, so $q$ set | |||||
$validValue$ to $v$ and $validRound$ to $r$ the latest by time $t+\Delta$. As | |||||
$t+\Delta < t+\timeoutPrecommit(r)$, whenever $\timeoutPrecommit(r) > \Delta$, the | |||||
Lemma holds also in this case. \end{proof} | |||||
\begin{lemma} \label{lemma:agreement} Algorithm~\ref{alg:tendermint} satisfies | |||||
Termination. \end{lemma} | |||||
\begin{proof} Lemma~\ref{lemma:round-synchronisation} defines a scenario in | |||||
which all correct processes decide. We now prove that within a bounded | |||||
duration after GST such a scenario will unfold. Let's assume that at time | |||||
$GST$ the highest round started by a correct process is $r_0$, and that | |||||
there exists a correct process $p$ such that the following holds: for every | |||||
correct process $c$, $lockedRound_c \le validRound_p$. Furthermore, we | |||||
assume that $p$ will be the proposer in some round $r_1 > r$ (this is | |||||
ensured by the $\coord$ function). | |||||
We have two cases to consider. In the first case, for all rounds $r \ge r_0$ | |||||
and $r < r_1$, no correct process locks a value (set $lockedRound$ to $r$). So | |||||
in round $r_1$ we have the scenario from the | |||||
Lemma~\ref{lemma:round-synchronisation}, so all correct processes decides in | |||||
round $r_1$. | |||||
In the second case, a correct process locks a value $v$ in round $r_2$, where | |||||
$r_2 \ge r_0$ and $r_2 < r_1$. Let's assume that $r_2$ is the highest round | |||||
before $r_1$ in which some correct process $q$ locks a value. By Lemma | |||||
\ref{lemma:validValue} at the end of round $r_2$ the following holds for all | |||||
correct processes $c$: $validValue_c = lockedValue_q$ and $validRound_c = r_2$. | |||||
Then in round $r_1$, the conditions for the | |||||
Lemma~\ref{lemma:round-synchronisation} holds, so all correct processes decide. | |||||
\end{proof} | |||||
@ -1,54 +0,0 @@ | |||||
\section{Definitions} | |||||
\label{sec:definitions} | |||||
\subsection{Model} | |||||
We consider a system composed of $n$ server processes $\Pi = \{ 1, \dots, n\}$. | |||||
Server processes can be correct or faulty, where a faulty process can behave in an arbitrary way, | |||||
i.e., we consider Byzantine faults. We assume that each process has some amount of voting power, and we denote with $N$ the total voting power of all processes in the system. | |||||
Furthermore, we assume that the total voting power of faulty processes is bounded with a system parameter $f$. | |||||
Processes communicate by exchanging messages. We assume a model where processes are not part of the single administration domain; therefore we cannot enforce a direct network connectivity between all server processes. | |||||
Instead, we assume that each server process is connected to a small number of processes called peers, | |||||
such that there is an indirect communication channel between all correct server processes. Communication between processes is established using gossip protocol. | |||||
We assume that processes communicate over wide-area network. Formally, we model the network communication using the partially synchronous system model~\cite{DLS88:jacm}, or rather a slightly weaker variant of this model: we assume that the system alternates between good periods (during which the system is synchronous) and bad periods (during which the system is asynchronous and messages can be lost). During good periods, there is a bound $\Delta$ such that all | |||||
communication among correct processes is reliable and $\Delta$-timely, i.e., if a correct process | |||||
$p$ sends message $m$ at time $t$ to a correct process $q$, then $q$ will receive $m$ | |||||
before $t+\Delta$. Note that as we do not assume direct communication channels among all correct processes, this implies that before the message $m$ reaches $q$, it will pass through a number of | |||||
correct servers that will forward the message $m$ using gossip protocol towards $q$. | |||||
Messages among correct processes can be | |||||
delayed, dropped or duplicated during bad periods. Spoofing/impersonation attacks are assumed to be impossible also during bad periods. | |||||
The bound $\Delta$ is a system parameter whose value is not required to be known for the safety of the algorithm presented. However, the timing bounds derived for our algorithm, and thus the guaranteed latency requires knowledge of $\Delta$. | |||||
We assume that process steps (which might include sending and receiving messages) take zero time. | |||||
Processes are equipped with clocks so they can measure local timeouts. | |||||
We assume integrity of direct communication channels between peers, that is if a process $p$ received a message $m$ from process $q$, then $q$ sent message $m$ to $p$ before. All protocol messages are signed, i.e., when a correct process $q$ receives a signed message $m$ from its peer, the process $q$ can verify who was the original message sender. | |||||
The details of the Tendermint gossip protocol will be discussed in a separate technical report. For the sake of this report it is sufficient to assume the following properties provided by the gossip protocol (in addition to the properties about partial synchrony of the network stated above): | |||||
\begin{itemize} | |||||
\item Messages that are being gossiped come from the consensus layer. We can think about consensus protocol as a content creator, which defines what messages should be disseminated using the gossip protocol. A correct process creates the message for dissemination either i) explicitly, by invoking \emph{send} function as part of the consensus protocol or ii) implicitly, by receiving a message from some other process. Note that in the case ii) gossiping of messages is implicit, i.e., it happens without explicit send clause in the consensus algorithm whenever a correct process receives some messages in the consensus algorithm\footnote{If a message is received by a correct process at the consensus level then it is considered valid from the protocol point of view, i.e., it has a correct signature, a proper message structure and a valid height and round number.}. | |||||
\item Processes keep resending messages (in case of failures or message loss) until all its peers get them. This ensures that every message sent or received by a correct process is eventually received by all correct processes. | |||||
\end{itemize} | |||||
\subsection{Consensus} | |||||
\label{sec:consensus} | |||||
\newcommand{\propose}{\mathsf{propose}} | |||||
\newcommand{\decide}{\mathsf{decide}} | |||||
We consider a variant of the Byzantine consensus problem called Validity Predicate-based Byzantine consensus that is motivated by blockchain systems~\cite{GLR17:red-belly-bc}. The problem is defined by an agreement, a termination, and a validity | |||||
property. | |||||
\begin{itemize} | |||||
\item \emph{Agreement:} No two correct processes decide on different values. | |||||
\item \emph{Termination:} All correct processes eventually decide on a value. | |||||
\item \emph{Validity:} A decided value is valid, i.e., it satisfies the predefined predicate denoted \emph{valid()}. | |||||
\end{itemize} | |||||
This variant of the Byzantine consensus problem has an application-specific valid() predicate to indicate whether a value is valid. In the context of blockchain systems, for example, a value is not valid if it does not contain an appropriate hash of the last value (block) added to the blockchain. | |||||
@ -1,2 +0,0 @@ | |||||
\section{Formal proof of Tendermint consensus algorithm} | |||||
\label{sec:proof} |