449 lines
11 KiB
Markdown
449 lines
11 KiB
Markdown
---
|
||
marp: true
|
||
paginate: true
|
||
author: Clément Contet, Laurent Fainsin
|
||
math: katex
|
||
---
|
||
|
||
<style>
|
||
section::after {
|
||
/*custom pagination*/
|
||
content: attr(data-marpit-pagination) ' / ' attr(data-marpit-pagination-total);
|
||
}
|
||
</style>
|
||
|
||
# The RAFT Consensus Algorithm
|
||
|
||
<!-- https://github.com/raft/raft.github.io -->
|
||
<!-- https://ongardie.net/static/coreosfest/slides -->
|
||
|
||
|
||
|
||
<footer>
|
||
Reliable, Replicated, Redundant, And Fault-Tolerant.<br>
|
||
Diego Ongaro,
|
||
John Ousterhout,
|
||
Stanford University
|
||
(2014)
|
||
</footer>
|
||
|
||
<!-- Laurent -->
|
||
|
||
---
|
||
|
||
<!-- Qu'est ce que le "consensus" dans un premier temps ? -->
|
||
|
||
<header>
|
||
|
||
# Consensus ?
|
||
|
||
</header>
|
||
|
||
> Consensus algorithms allow a collection of machines to work as a coherent group that can survive the failures of some of its members. – RAFT authors
|
||
|
||
<!-- Les algorithmes de consensus permettent à un ensemble de machines de fonctionner comme un groupe cohérent qui peut survivre aux défaillances de certains de ses membres. -->
|
||
|
||
<br>
|
||
|
||
- Accord sur l'état partagé (image système unique)
|
||
- Réparation (réplication) autonome en cas de défaillance d'un serveur
|
||
- Une minorité de serveurs HS: pas de problème
|
||
- La majorité des serveurs HS: perte de disponibilité, maintien de la cohérence
|
||
- La clé pour construire des systèmes de stockage cohérents
|
||
|
||
<!-- Laurent -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Architecture typique des systèmes de consensus
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!-- Un ou plusieurs clients. Et plusieurs serveurs formant un groupe de consensus (généralement > 3 et impair). Quand client RPC serv, leader réplique cela sur ses suiveurs et une fois cela validé, répond au client. -->
|
||
|
||
<!--
|
||
|
||
- Replicated log -> replicated state machine
|
||
- All servers execute same commands in same order
|
||
- Consensus module ensures proper log replication
|
||
- System makes progress as long as any majority of servers up
|
||
- Failure model: fail-stop (not Byzantine), delayed/lost msgs
|
||
|
||
-->
|
||
|
||
<!-- Laurent -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Motivation ?
|
||
|
||
</header>
|
||
|
||
<!-- Avant RAFT il existait déjà un algorithme bien connu du nom de Paxos qui dominait sur le reste des algos de consensus. -->
|
||
|
||
<!-- Alors pourquoi faire un nouvel algo ? quels sont les motivations des auteurs ? -->
|
||
|
||
Paxos domine le marché depuis ~25 ans (Leslie Lamport, 1989)
|
||
- Difficile à comprendre
|
||
- Difficile à implémenter
|
||
|
||
> The dirty little secret of the NSDI community is that at most five people really, truly understand every part of Paxos ;-). – [NSDI](https://www.usenix.org/conference/nsdi23) reviewer
|
||
|
||
> There are significant gaps between the description of the Paxos algorithm and the needs of a real-world system…the final system will be based on an unproven protocol. – [Chubby](https://research.google/pubs/pub27897/) authors
|
||
|
||
<!-- Clément -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Paxos ?
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!-- Clément -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Les différences de RAFT ?
|
||
|
||
</header>
|
||
|
||
<!-- Les auteurs de RAFT en avait donc ras le bol de voir cette situation, ils ont donc décidés de créer un nouvel algorithme, en se fondant sur la compréhensibilité -->
|
||
|
||
<!-- Clément -->
|
||
|
||
## Prendre des décisions de conception fondées sur la compréhensibilité
|
||
|
||
<!-- Pour se faire ils ont choisis de décomposer le consensus le plus possible en des petits problèmes trivials à résoudre. -->
|
||
|
||
- Décomposition du problème
|
||
|
||
<!-- Il ont aussi essayé de diminuer l'espace des états possibles lors du fonctionnement de l'algorithme pour le rendre le plus simple possible. -->
|
||
|
||
- Minimiser l'espace des états
|
||
- Traiter plusieurs problèmes avec un seul mécanisme
|
||
- Éliminer les cas particuliers
|
||
- Minimiser le non-déterminisme
|
||
- Maximiser la cohérence
|
||
|
||
<!-- Laurent -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Décomposition du problème
|
||
|
||
</header>
|
||
|
||
1. Élection d'un leader (mandat)
|
||
- Sélectionner un serveur qui sera le leader
|
||
- Détecter les pannes, choisir un nouveau leader
|
||
2. Réplication des logs (fonctionnement normal)
|
||
- Le leader accepte les commandes des clients et les ajoute à son journal
|
||
- Le leader réplique son journal aux autres serveurs (écrase les incohérences)
|
||
3. Sécurité
|
||
- Maintenir la cohérence des journaux
|
||
- Seuls les serveurs dont les journaux sont à jour peuvent devenir des leaders
|
||
|
||
<!-- Laurent -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Minimiser l'espace des états
|
||
|
||
</header>
|
||
|
||
<!--
|
||
Il n'y a que 3 états possibles pour les serveurs !
|
||
Et les changements d'états sont assez simples. -->
|
||
|
||
|
||
|
||
<!--
|
||
|
||
Follower: Passive (but expects regular heartbeats to stay in this state, get a random timeout time at startup), if after some time there is no heartbeat, timeout, change to candidate.
|
||
|
||
(RPC: https://en.wikipedia.org/wiki/Remote_procedure_call)
|
||
|
||
Candidate: Issues RequestVote RPCs to get elected as leader. After results, if majority, becomes new leader. if other leader gets elected or is already present, step down.
|
||
|
||
|
||
Leader: Issues AppendEntries RPCs:
|
||
- Replicate its log
|
||
- Heartbeats to maintain leadership
|
||
|
||
-->
|
||
|
||
<!-- Laurent -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Mandats
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!--
|
||
|
||
- At most 1 leader per term
|
||
- Some terms have no leader (failed election, example: 3 servers, all switch to candidate at same time, no majority)
|
||
- Each server maintains current term value (no global view)
|
||
- Exchanged in every RPC
|
||
- Peer has later term? Update term, revert to follower
|
||
- Incoming RPC has obsolete term? Reply with error
|
||
|
||
Terms identify obsolete information
|
||
|
||
-->
|
||
|
||
<!-- Clément -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Élection du leader
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!-- Clément -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Correction des élections
|
||
|
||
</header>
|
||
|
||
- Sécurité: autoriser au maximum un gagnant par mandat
|
||
- Chaque serveur ne donne qu'un seul vote par mandat (persistant sur disque)
|
||
- Majorité requise pour gagner l'élection
|
||
|
||
- Vivacité: un candidat doit finir par gagner
|
||
- Délais d'expiration des élections aléatoire dans $[T, 2T]$ (e.g. 150-300 ms)
|
||
- Le serveur gagne l'élection en dépassant le délai d'attente avant les autres
|
||
- Fonctionne bien si $T_{\text{diffusion}} \ll T \ll \text{MTBF}$
|
||
- Approche aléatoire plus simple que les autres comme le ranking
|
||
|
||
<!-- Clément -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Démo interactive
|
||
|
||
</header>
|
||
|
||
<iframe
|
||
src="https://raft.github.io/raftscope/index.html"
|
||
frameborder="0"
|
||
scrolling="no"
|
||
height=100%
|
||
width=100%
|
||
></iframe>
|
||
|
||
<!--
|
||
|
||
Normal Operation:
|
||
|
||
- Client sends command to leader
|
||
- Leader appends command to its log
|
||
- Leader sends AppendEntries RPCs to all followers
|
||
- Once new entry committed:
|
||
- Leader executes command in its state machine, returns result to client
|
||
- Leader notifies followers of committed entries in subsequent AppendEntries RPCs
|
||
- Followers execute committed commands in their state machines
|
||
- Crashed/slow followers?
|
||
-Leader retries AppendEntries RPCs until they succeed
|
||
- Optimal performance in common case:
|
||
- One successful RPC to any majority of servers
|
||
|
||
-->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Structure des journaux
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!--
|
||
|
||
- Must survive crashes (store on disk)
|
||
- Entry committed if safe to execute in state machines
|
||
- Replicated on majority of servers by leader of its term (that the goal of the leader)
|
||
|
||
-->
|
||
|
||
<!-- Laurent -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Incohérences des journaux
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!--
|
||
|
||
Logs may be inconsistent after leader change
|
||
- No special steps by new leader:
|
||
- Start normal operation
|
||
- Followers’ logs will eventually match leader
|
||
- Leader’s log is “the truth”
|
||
|
||
Crashes can result in log inconsistencies:
|
||
- Raft minimizes special code for repairing inconsistencies:
|
||
- Leader assumes its log is correct (se pose pas de question sur la validité des ses propres logs, source de vérité)
|
||
- Normal operation will repair all inconsistencies
|
||
|
||
-->
|
||
|
||
<!-- Clément -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Propriété de correspondance des journaux
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!--
|
||
|
||
Goal: high level of consistency between logs
|
||
|
||
- If log entries on different servers have same index and term:
|
||
- They store the same command
|
||
- The logs are identical in all preceding entries
|
||
- If a given entry is committed, all preceding entries are also committed
|
||
|
||
-->
|
||
|
||
<!-- Laurent -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Réparation des journaux par correspondance
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!--
|
||
|
||
AppendEntries RPCs include <index, term> of entry preceding new one(s)
|
||
- Follower must contain matching entry; otherwise it rejects request
|
||
- Leader retries with lower log index
|
||
- Implements an induction step, ensures Log Matching Property
|
||
|
||
-->
|
||
|
||
<!-- Clément -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Changement de composition du cluster
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!-- Clément -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# De nombreuses implémentations
|
||
|
||
</header>
|
||
|
||
| Name | Primary Authors | Language | License |
|
||
| :------------- | :------------------------------------------------ | :--------- | :--------- |
|
||
| etcd/raft | Blake Mizerany, Xiang Li and Yicheng Qin (CoreOS) | Go | Apache 2.0 |
|
||
| go-raft | Ben Johnson (Sky) and Xiang Li (CMU, CoreOS) | Go | MIT |
|
||
| hashicorp/raft | Armon Dadgar (hashicorp) | Go | MPL-2.0 |
|
||
| copycat | Jordan Halterman | Java | Apache2 |
|
||
| LogCabin | Diego Ongaro (Stanford, Scale Computing) | C++ | ISC |
|
||
| akka-raft | Konrad Malawski | Scala | Apache2 |
|
||
| kanaka/raft.js | Joel Martin | Javascript | MPL-2.0 |
|
||
|
||
<!-- Laurent -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Quel degré de hasard est nécessaire pour éviter les votes non concluants ?
|
||
|
||
</header>
|
||
|
||
|
||
|
||
<!-- Clément -->
|
||
|
||
---
|
||
|
||
<header>
|
||
|
||
# Étude: Raft est-il plus simple que Paxos ?
|
||
|
||
</header>
|
||
|
||
<!--
|
||
|
||
- 43 students in 2 graduate OS classes (Berkeley and Stanford)
|
||
- Group 1: Raft video, Raft quiz, then Paxos video, Paxos quiz
|
||
- Group 2: Paxos video, Paxos quiz, then Raft video, Raft quiz
|
||
- Instructional videos:
|
||
- Same instructor (Ousterhout)
|
||
- Covered same functionality: consensus, replicated log, cluster reconfiguration
|
||
- Fleshed out missing pieces for Paxos
|
||
- Videos available on YouTube
|
||
- Quizzes:
|
||
- Questions in 3 general categories
|
||
- Same weightings for both tests
|
||
- Experiment favored Paxos slightly:
|
||
- 15 students had prior experience with Paxos
|
||
|
||
-->
|
||
|
||
|
||
|
||
|
||
<!-- Laurent -->
|