Skip to content
This repository has been archived by the owner on Oct 4, 2022. It is now read-only.

Latest commit

 

History

History
751 lines (465 loc) · 23.7 KB

lecture5.md

File metadata and controls

751 lines (465 loc) · 23.7 KB
Raw

class: middle, center, title-slide

Large-scale Data Systems

Lecture 5: Consensus



Prof. Gilles Louppe
g.louppe@uliege.be

Today

  • Most important abstraction in distributed systems: consensus.
  • Builds upon broadcast and failure detectors.
  • From consensus, we will build:
    • total order broadcast
    • replicated state machines
    • ... and almost all higher level distributed fault-tolerant applications!

class: middle, center, black-slide

.width-80[]

???

Bump on Francois' question: is reality as complicated as the lasagna of abstractions we have been defining?

  • Not necessarily, and we could certainly implement services in a more compact and optimized way.
  • However, what our theoretical framework allows us to do properly is to ensure correctness.
  • In all previous lectures, we have defined self-contained components with a limited number of features, which allowed to properly analyze them and prove their correctness.
  • We also went from simple to more and more complicated components by building in a bottom-up fashion and relying on things for which correctness was already ensured.
  • Directly starting from a complicated set of requirements and implementing everything in one go is possible, but would be much more difficult to achieve, debug and analyze.

Also, this core set of abstractions is here to stay. We will study some applications that make use of those principles, but they likely to be obsolete in 5 to 10 years. In fact, some of them already are. The theory is not.

class: middle

Consensus

class: middle

Consensus is the problem of making processes all agree on one of the values they propose.

.center.width-100[] .caption[How do we reach an agreement?]

class: middle

Motivation

Solving consensus is key to solving many problems in distributed computing:

  • synchronizing replicated state machines;
  • electing a leader;
  • managing group membership;
  • deciding to commit or abort distributed transactions.

Any algorithm that helps multiple processes maintain common state or to decide on a future action, in a model where processes may fail, involves solving a consensus problem.

Consensus

.width-90[]

.exercise[Which is safety, which is liveness?]

???

  • Termination: liveness
  • Validity: safety
  • Integrity: safety
  • Agreement: safety

class: middle

Sample execution

.center.width-100[]

.exercise[Does this satisfy consensus?]

???

Yes

Uniform consensus

.width-90[]

class: middle

Sample execution

.center.width-100[]

.exercise[Does this satisfy uniform consensus?]

???

No

class: center, middle

.width-100[]

class: middle

Are we done then? No!

  • The FLP impossibility result holds for asynchronous systems only.
  • Consensus can be implemented in synchronous and partially synchronous systems. (We will prove it!)
  • The result only states that termination cannot be guaranteed.
    • Can we have other guarantees while maintaining a high probability of termination?

class: middle

Consensus in fail-stop

Hierarchical consensus

Asumptions

  • Assume a perfect failure detector (synchronous system).
  • Assume processes $1, ..., N$ form an ordered hierarchy as given by $\text{rank}(p)$.
    • $\text{rank}(p)$ is a unique number between $1$ and $N$ (e.g., the pid).

class: middle

Algorithm

  • Hierarchical consensus ensures that the correct process with the lowest rank imposes its value on all the other processes.
    • If $p$ is correct and rank $1$, it imposes its values on all other processes by broadcasting its proposal.
    • If $p$ crashes immediately and $q$ is correct and rank 2, then it ensures that $q$'s proposal is decided.
    • The core of the algorithm addresses the case where $p$ is faulty but crashes after sending some of its proposal messages and $q$ is correct.
  • Hierarchical consensus works in rounds, with a rotating leader.
    • At round $i$, process $p$ with rank $i$ is the leader.
    • It decides its proposal and broadcasts it to all processes.
    • All other processes that reach round $i$ wait before taking any actions, until they deliver this message or until they detect the crash of $p$.
      • upon which processes move to the next round.

class: middle

.width-70[]

???

  • What if a process does not propose?

class: middle

Execution without failure

.center.width-100[]

class: middle

Execution with failure (1)

.center.width-100[]

.exercise[- Uniform consensus?

  • How many failures can be tolerated?]

???

  • Not uniform consensus.
  • $N-1$ failures at most.

class: middle

Execution with failure (2)

.center.width-100[]

class: middle

Correctness

  • Termination: Every correct process eventually decides some value.
    • Every correct node makes it to the round it is leader in.
      • If some leader fails, completeness of the FD ensures progress.
      • If leader correct, validity of BEB ensures delivery.
  • Validity: If a process decides $v$, then $v$ was proposed by some process.
    • Always decide own proposal or adopted value.
  • Integrity: No process decides twice.
    • Rounds increase monotonically.
    • A node only decides once in the round it is leader.
  • Agreement: No two correct processes decide differently.
    • Take correct leader with minimum rank $i$.
      • By termination, it will decide $v$.
      • It will BEB $v$:
        • Every correct node gets $v$ and adopts it.
        • No older proposals can override the adoption.
        • All future proposals and decisions will be $v$.

Hierarchical uniform consensus

  • Same as hierarchical consensus, but must ensure uniform agreement.
  • A round consists of two communication steps:
    • The leader BEB broadcasts its proposal
    • The leader collects acknowledgements
  • Upon reception of all acknowledgements, RB broadcast the decision and decide at delivery.
    • This ensures that if a decision is made (at a faulty or correct process), then this decision will be made at all correct processes.
    • Processes proceed to the next round only if the current leader fails.

???

Different from Hierarchical consensus in the sense that we dont necessarily have to wait N rounds.

class: middle

.width-70[]

class: middle

.width-70[]

???

[Q] Why is reliable non-uniform broadcast sufficient to have uniform consensus?

Consider now the uniform agreement property, and assume that two processes decide differently. This can only be possible if two processes rb-broadcast decision messages with two different proposal values. Consider any two processes p and q such that rank (q) > rank (p) and suppose that p and q rb-broadcast two decision values v and v', respectively. Because of the strong accuracy property of the failure detector, process q must have adopted v before reaching round rank (q). Hence, it holds v = v', which contradicts the assumption.

class: middle

Consensus in fail-noisy

class: middle

Would hierarchical consensus work with an eventually perfect failure detector?

  • A false suspicion (i.e., a violation of strong accuracy) might lead to the violation of agreement.
  • Not suspecting a crashed process (i.e., a violation of strong completeness) might lead to the violation of termination.

.center.width-100[]

class: middle

Towards consensus...

We will build a consensus component in fail-noisy by combining three abstractions:

  1. an eventual leader detector
  2. an epoch-change abstraction
  3. an epoch consensus abstraction

Eventual leader detector ($\Omega$)

.center[]

.exercise[This abstraction can be implemented from an eventually perfect failure detector. How?]

???

Elect as leader the correct process with the minimal rank. Eventually the set of correct processes will be accurate at all correct processes, resulting in eventual agreement.

Epoch-Change ($ec$)

  • Let us define an Epoch-Change abstraction, whose purpose it is to signal a change of epoch corresponding to the election of a leader.
  • An indication event StartEpoch contains:
    • an epoch timestamp $ts$
    • a leader process $l$.

???

This component is emitting only.

class: middle

.center[]

Leader-based Epoch-Change

  • Every process $p$ maintains two timestamps:
    • a timestamp $lastts$ of the last epoch that it locally started;
    • a timestamp $ts$ of the last epoch it attempted to start as a leader.
  • When the leader detector makes $p$ trust itself, $p$ adds $N$ to $ts$ and broadcasts a NewEpoch message with $ts$.
  • When $p$ receives a NewEpoch message with parameter $newts > lastts$ from $l$ and $p$ most recently trusted $l$, then $p$ triggers a StartEpoch message.
  • Otherwise, $p$ informs the aspiring leader $l$ with a NACK that a new epoch could not be started.
  • When $l$ receives a NACK and still trusts itself, it increments $ts$ by $N$ and tries again to start a new epoch.

class: middle

.width-60[]

???

[Q] How many failures can be tolerated? $N-1$

class: middle

Sample execution (1)

.center.width-100[]

class: middle

Sample execution (2)

.center.width-100[]

class: middle

.exercise[

  • What if $p_1$ fails only later, some time after the second bebDeliver event?
  • What if instead of crashing, $p_1$ eventually trusts $p_2$?
  • Could $p_1$ and $p_2$ keep bouncing NACKs to each other?]

Epoch consensus ($ep$)

  • Let us define an epoch consensus abstraction, whose purpose is similar to consensus, but with the following simplifications:
    • Epoch consensus represents an attempt to reach consensus.
      • The procedure can be aborted when it does not decide or when the next epoch should already be started by the higher-level algorithm.
    • Every epoch consensus instance is identified by an epoch timestamp $ts$ and a designated leader $l$.
    • Only the leader proposes a value. Epoch consensus is required to decide only when the leader is correct.
  • An instance must terminate when the application locally triggers an Abort event.
  • The state of the component is initialized
    • with a higher timestamp than that of all instances it initialized previously;
    • with the state of the most recently locally aborted epoch consensus instance.

class: middle

.center[]

class: middle

.center[]

Read/Write Epoch consensus

  • Let us initialize the Read/Write Epoch consensus algorithm with the state of the most recently aborted epoch consensus instance.
    • The state contains a proposal $val$ and its associated timestamp $valts$.
    • Passing the state to the next epoch consensus serves the validity and lock-in properties.
  • The algorithm involves two rounds of messages from the leader to all processes.
    • The leader writes its proposal value to all processes, who store the epoch timestamp and the value in their state, and acknowledge this to the leader.
    • When the leader receives enough acknowledgements, it decides this value.
    • However, if the leader of some previous epoch already decided some value $val$, then no other value should be decided (to not violate lock-in).
    • To prevent this, the leader first reads the state of the processes, which return State messages.
    • The leader receives a quorum of State messages and choses the value that comes with the highest timestamp, if one exists.
    • The leader decides and broadcasts its decision to all processes, which then decide too.

class: middle

class: middle

class: middle

Sample execution (1)

.center.width-100[]

class: middle

Sample execution (2)

.center.width-100[]

???

Note how in the read phase p1 could proceed as soon as it obtains a quorum.

Quorum: minimum number of votes that a distributed transaction must obtain in order to perform an operation.

class: middle

Sample execution (3)

.center.width-100[]

.exercise[What is wrong in this execution?]

???

$p1$ should not proceed since 2 is not > 2.

class: middle

Sample execution (4a)

.center.width-100[]

class: middle

Sample execution (4b)

.center.width-100[]

class: middle

Correctness

Assume a majority of correct processes, i.e. $N > 2f$, where $f$ is the number of crash faults.

  • Lock-in: If a correct process has ep-decided $v$ in an epoch consensus with timestamp $ts' \leq ts$, then no correct process ep-decides a value different from $v$.
    • If some process ep-decided $v$ at $ts' < ts$, then it decided after receiving a Decided message with $v$ from leader $l'$ of epoch $ts'$.
    • Before sending this message, $l'$ had broadcast a Write containing $v$ and collected Accept messages.
    • These responding processes set their variables $val$ to $v$ and $valts$ to $ts'$.
    • At the next epoch, the leader sent a Write message with the previous $(ts', v)$ pair and collected Accept messages.
    • This pair has the highest timestamp with a non-null value.
    • This implies that the leader of this epoch can only ep-decides $v$.
    • This argument can be continued until $ts$, establishing lock-in.

class: middle

  • Validity: If a correct process ep-decides $v$, then $v$ was ep-proposed by the leader $l'$ of some epoch consensus with timestamp $ts' \leq ts$ and leader $l'$.
    • If some process ep-decides $v$, it is because this value was delivered from a Decided message.
    • Furthermore, every process stores in $val$ only the value received in a Write message from the leader.
    • In both cases, this value comes from tmpval of the leader.
    • In any epoch, the leader sets tmpval only to the value it ep-proposed or to some value it received in a State message from another process.
    • By backward induction, $v$ was ep-proposed by the leader in some epoch $ts' \leq ts$.
  • Uniform agreement + integrity: No two processes ep-decide differently + Every correct process ep-decides at most once.
    • $l$ sends the same value to all processes in the Decided message.
  • Termination: If the leader $l$ is correct, has ep-proposed a value, and no correct process aborts this epoch consensus, then every correct process eventually ep-decides some value.
    • When $l$ is correct and no process aborts the epoch, then every process eventually receives a Decide message and ep-decides.

???

[Q] What may go wrong if we do not assume a majority of correct processes?

Leader-Driven consensus

  • Let us now combine the epoch-change and the epoch consensus abstractions to form the leader-driven consensus algorithm.
  • We will write the glue to repeatedly run epoch consensus until epoch changes stabilize and all decisions are taken.
  • The algorithm provides uniform consensus in fail-noisy.

class: middle

Leader-driven consensus runs through a sequence of epochs, triggered by StartEpoch events from the epoch-change primitive:

  • The current epoch timestamp is $ets$ and the associated leader is $l$.
  • The StartEpoch events determine the timestamp $newts$ and the leader $newl$ of the next epoch consensus instance to start.
  • To switch from one epoch consensus to the next, the algorithm aborts the running epoch consensus instance, obtains its state and initializes the next epoch consensus instance with it.
  • As soon as a process has obtained a proposal value $v$ for consensus and is the leader of the current epoch, it ep-proposes this value for epoch consensus.
  • When the current epoch ep-decides a value, the process also decides this value for consensus.
  • The process continue to participate in the consensus to help other processes decide.

class: middle

class: middle

class: middle

Sample execution

.center.width-100[]

???

Every process proposes a different value and then starts epoch 6 with leader q. The open circle depicts that process q is the leader and the arrows show that it drives the message exchange in the epoch consensus instance, which is implemented by Algorithm 5.6.

Process q writes its proposal x, but only process r receives it and sends an ACCEPT message before the epoch ends; hence, process r updates its state to (6, x). Epoch 8 with leader s starts subsequently and the processes abort the epoch consensus instance with timestamp 6.

At process r, epoch 8 starts much later, and r neither receives nor sends any message before the epoch is aborted again. Note that the specification of epoch-change would also permit that process r never starts epoch 8 and moves from epoch 6 to 11 directly. Process s is the leader of epoch 8. It finds no highest value different from ⊥ and writes its own proposal z. Subsequently, it sends a DECIDED message and process p ep-decides and uc-decides z in epoch 8.

Note that p, q, and s now have state (8, z). Then process s crashes and epoch 11 with leader r starts. The state of r is still (6, x); but it reads value z from p or q, and therefore writes z. As r is correct, all remaining processes ep-decide z in this epoch consensus instance and, consequently, also q and r uc-decide z.

It could also have been that process p crashed immediately after deciding z; in this case, the remaining processes would also have decided z due to the lock-in and agreement properties of epoch consensus. This illustrates that the algorithm provides uniform consensus.

class: middle

Correctness

  • Validity: If a process decides $v$, then $v$ was proposed by some process.
    • A process uc-decides $v$ only when it has ep-decided $v$ in the current epoch consensus.
    • Every decision can be attributed to a unique epoch and to a unique instance of epoch consensus.
    • Let $ts^*$ be the smallest timestamp of an epoch consensus in which some process decides $v$.
    • According to the validity property of epoch consensus, this means $v$ was ep-proposed by the leader of some epoch whose timestamp is a most $ts^*$.
    • Since a process only ep-proposes $val$ when $val$ has been uc-proposed for consensus, the validity property follows for processes that uc-decide in epoch $ts^*$.
    • The argument extends to $ts > ts^*$ because the lock-in property of epoch consensus forces processes to ep-decide $v$ only, which in turn make them uc-decide.

class: middle

  • Uniform agreement: No two processes decide differently.
    • Every decision attributed to an ep-decision of some epoch consensus instance.
    • If two correct processes decide when they are in the same epoch, then the uniform agreement of epoch consensus ensures the decisions are the same.
    • If they decide in different epoch, the lock-in property establishes uniform agreement.

class: middle

  • Integrity: No process decides twice.
    • The decided flag in the algorithm prevents multiple decisions.
  • Termination: Every correct process eventually decides some value.
    • Because of eventual leadership of the epoch-change primitive, there is some epoch with timestamp $ts$ and leader $l$ such that no further epoch starts and $l$ is correct.
    • From that instant, no further abortions are triggered.
    • The termination property of epoch consensus ensures that every correct process eventually ep-decides, and therefore uc-decides.

.grid[ .kol-3-4[

Paxos

Leader-driven consensus is a modular formulation of the Paxos consensus algorithm by Leslie Lamport. ] .kol-1-4[.circle.width-100[]] ]

.center.width-80[]

class: center, black-slide, middle

<iframe width="640" height="400" src="https://www.youtube.com/embed/s8JqcZtvnsM?cc_load_policy=1&hl=en&version=3" frameborder="0" allowfullscreen></iframe>

???

Formulation is different but equivalent.

  • Leader election, epoch change =p Promise phase
  • Lock-in = commitment

class: middle

Total order broadcast

Total order broadcast ($tob$)

  • The total-order (reliable) broadcast (also known as atomic broadcast) abstraction ensures that all processes deliver the same messages in a common global order.
  • Total-order broadcast is the key abstraction for maintaining consistency among multiple replicas that implement one logical service.

class: middle, center

.width-100[]

Consensus-based TOB

  • Messages are first disseminated using a reliable broadcast instance.
    • No particular order is imposed on the messages.
    • At any point in time, it may be that no two processes have the same sets of unordered messages.
  • The processes use consensus to decide on one set of messages to be delivered, order the messages in this set, and finally deliver them.

class: middle

.width-80[]

class: middle

Sample execution

.center.width-90[]

Replicated state machines ($rsm$)

  • A state machine consists of variables and commands that transform its state and produce some output.
  • Commands are deterministic programs, such that the outputs are solely determined by the initial state and the sequence of commands.
  • A state machine can be made fault-tolerant by replicating it on different processes.
  • This can now be easily implemented simply by disseminating all commands to execute using a uniform total-order broadcast primitive.

.alert[ This gives a generic recipe to make any deterministic program distributed, consistent and fault-tolerant! ]

class: middle, center

class: middle

.center[]

TOB-based Replicated state machines

Summary

  • Consensus is the problem of making processes all agree on one of the values they propose.
  • The FLP impossibility result states that no consensus protocol can be proven to always terminate in an asynchronous system.
  • In fail-stop, Hierarchical Consensus provides an implementation based on broadcast and failure detection.
  • In fail-noisy, Leader-Driven Consensus achieves consensus by repeatedly running epoch consensus until all decisions are taken.
  • The consensus primitive greatly simplifies the implementation of any fault-tolerant consistent distributed system.
    • Total-order broadcast
    • Replicated state machines

class: end-slide, center count: false

The end.

References

  • Fischer, Michael J., Nancy A. Lynch, and Michael S. Paterson. "Impossibility of distributed consensus with one faulty process." Journal of the ACM (JACM) 32.2 (1985): 374-382.
  • Lamport, Leslie. "The part-time parliament." ACM Transactions on Computer Systems (TOCS) 16.2 (1998): 133-169.
  • Lamport, Leslie. "Paxos made simple." ACM Sigact News 32.4 (2001): 18-25.