容错与 Raft（一）

6.5840 2026 Lecture 6: RSM and Raft (1)

this lecture
  today: replicated state machine, and Raft elections (Lab 3A)
  next: Raft persistence, client behavior, snapshots (Lab more 3B, 3C, 3D)

Goal: high availability
  even if a machine fails, deliver service
    i.e., no down time even if a machine fails
  approach: replication

What kinds of failures can replication deal with?
  Replication is good for "fail-stop" failure of a single replica
    fan stops working, CPU overheats and shuts itself down
    someone trips over replica's power cord or network cable
    software notices it is out of disk space and stops
  Replication may not help with bugs or operator error
    Often not fail-stop
    May be correlated (i.e. some input causes all replicas to crash)
  How about earthquake or city-wide power failure?
    Only if replicas are physically separated

How many replicas?
  Often you want few because many is costly
  But enough to survive failures during repair
    Many systems run with 3-5 replicas, as we will see
  This paper runs with only 2 replicas
    resilient to 1 failure at the time

*** topic: replicated state machine

A popular approach to replication
  Clients send operations to primary,
      primary sequences and sends to backups
  All replicas execute all operations
  If same start state,
    same operations,
    same order,
    deterministic,
    then same end state.
  Example: primary backup in GFS

What if primary fails?
  * Coordinator picks new primary in GFS
  * What if coordinator fails?

Can we have the replicas elect a new primary
  how about two servers, S1 and S2
  if both are up, S1 is in charge, forwards decisions to S2
  if S2 sees S1 is down, S2 takes over as coordinator
  what could go wrong?
  network partition! split brain!

the problem: computers cannot distinguish "server crashed" from "network broken"
  the symptom is the same: no response to a query over the network
  this difficulty seemed insurmountable for a long time
  seemed to require outside agent (a human) to decide when to switch servers
  we'd prefer an automated scheme!

Two partition-tolerant replication schemes were invented around 1990,
  Paxos and View-Stamped Replication
  called "consensus" or "agreement" protocols
  in the last 15 years this technology has seen a lot of real-world use
  the Raft paper is a good introduction to modern techniques

*** topic: state-machine replication with raft

state machine replication with Raft -- Lab 2 + 4 as example:
  [diagram: clients, 3 replicas, k/v layer + state, raft layer + logs]
  Raft is a library included in each replica

time diagram of one client command
  [C, L, F1, F2]
  client sends Put/Get "command" to k/v layer in leader
  k/v layer calls Start() to invoke Raft
    leader's Raft layer adds command to log
    leader sends AppendEntries RPCs to followers
    followers add command to log
  leader waits for replies from a bare majority (including itself)
  entry is "committed" if a majority put it in their logs
    committed means won't be forgotten even if failures
    majority -> will be seen by the next leader's vote requests
    leader "piggybacks" commit info in next AppendEntries
  leader and follower hand commands to k/v layer once entry is committed
    ApplyMsg and applyCh in lab
  leader sends response to client

why the logs?
  the service keeps the state machine state, e.g. key/value DB
    the log is an alternate representation of the same information!
    why both?
  the log orders the commands
    to help replicas agree on a single execution order
    to help the leader ensure followers have identical logs
  the log stores tentative commands until committed
  the log stores commands in case leader must re-send to followers
  the log stores commands persistently for replay after reboot

are the servers' logs exact replicas of each other?
  no: some replicas may lag
  no: we'll see that they can temporarily have different entries
  the good news:
    they'll eventually converge to be identical
    the commit mechanism ensures servers only execute stable entries

Implementation challenges:
  Failures
     network partitions, lost messages, server crashes
  Concurrency
     within a server and between servers
  Result: many possible executions and many corner cases
    many details to work -- figure 2
  Today: electing a new leader, which must handle these challenges

*** topic: leader election (Lab 3A)

why a leader?
  ensures all replicas execute the same commands, in the same order
  (some designs, e.g. Paxos, don't have a leader)

Raft numbers the sequence of leaders
  new leader -> new term
  a term has at most one leader; might have no leader
  the numbering helps servers follow latest leader, not superseded leader

when does a Raft peer start a leader election?
  when it doesn't hear from current leader for an "election timeout"
  increments local currentTerm, tries to collect votes
  note: this can lead to un-needed elections; that's slow but safe
  note: old leader may still be alive and think it is the leader

how to ensure at most one leader in a term?
  (Figure 2 RequestVote RPC and Rules for Servers)
  leader must get "yes" votes from a majority of servers
  each server can cast only one vote per term
    if candidate, votes for itself
    if not a candidate, votes for first that asks (within Figure 2 rules)
  at most one server can get majority of votes for a given term
    -> at most one leader even if network partition
    -> election can succeed even if some servers have failed
  note: again, majority is out of all servers (not just the live servers)

how does a server learn about a newly elected leader?
  the leader sends out AppendEntries heart-beats
    with the new higher term number
  only the leader sends AppendEntries
    only one leader per term
    so if you see AppendEntries with term T, you know who the leader for T is
  the heart-beats suppress any new election
    leader must send heart-beats more often than the election timeout

an election may not succeed for two reasons:
  * less than a majority of servers are reachable
  * simultaneous candidates split the vote, none gets majority

what happens if an election doesn't succeed?
  no heartbeats -> another timeout -> a new election for a new term
  higher term takes precedence, candidates for older terms quit

without special care, elections will often fail due to split vote
  all election timers likely to go off at around the same time
  every candidate votes for itself
  so no-one will vote for anyone else!
  so everyone will get exactly one vote, no-one will have a majority

how does Raft avoid split votes?
  each server adds some randomness to its election timeout period
  [diagram of times at which servers' timeouts expire]
  randomness breaks symmetry among the servers
    one will choose lowest random delay
  hopefully enough time to elect before next timeout expires
  others will see new leader's AppendEntries heartbeats and
    not become candidates
  randomized delays are a common pattern in network protocols

how to choose the election timeout?
  * at least a few heartbeat intervals (in case network drops a heartbeat)
    to avoid needless elections (which can cause an unnecessary pause)
  * short enough to react quickly to failure, avoid long pauses
  * short enough to allow a few re-tries before tester gets upset
    tester requires election to complete in 5 seconds or less
  * random part long enough to let one candidate succeed before next starts

what if old leader isn't aware a new leader is elected?
  perhaps old leader didn't see election messages
  perhaps old leader is in a minority network partition
  new leader means a majority of servers have incremented currentTerm
  either old leader will see new term in a AppendEntries reply and step down
  or old leader won't be able to get a majority of replies
    so old leader won't commit or execute any new log entries
  thus no split brain
  but a minority may accept old server's AppendEntries
    so logs may diverge at end of old term

example log divergence
  a leader crashes before sending AppendEntries to all
    S1: 3
    S2: 3 3
    S3: 3 3
  (the 3s are the term number in the log entry)
  worse: logs might have different commands in same entry!
    after a series of leader crashes, e.g.
        10 11 12 13  <- log entry #
    S1:  3
    S2:  3  3  4
    S3:  3  3  5

  how could this happen?
    S2 is leader in term 3
      appends 10 to S1, S2, and S3
      appends 11 to S2 and S3 (S1 crashed)
    S2 crashes, reboots quickly, and leader in term 4
      appends 12 to its log, and crashes.
    S3 becomes leader in term 5 (with help of S1)
      appends a different entry for 12 to its log

next lecture we will see how Raft handles log divergence
  including an additional restriction on leader election
  you can pass 3A without this restriction

In Search of an Understandable Consensus Algorithm
(Extended Version)
Diego Ongaro and John Ousterhout
Stanford University
Abstract
Raft is a consensus algorithm for managing a replicated
log. It produces a result equivalent to (multi-)Paxos, and
it is as efficient as Paxos, but its structure is different
from Paxos; this makes Raft more understandable than
Paxos and also provides a better foundation for build-
ing practical systems. In order to enhance understandabil-
ity, Raft separates the key elements of consensus, such as
leader election, log replication, and safety, and it enforc es
a stronger degree of coherency to reduce the number of
states that must be considered. Results from a user study
demonstrate that Raft is easier for students to learn than
Paxos. Raft also includes a new mechanism for changing
the cluster membership, which uses overlapping majori-
ties to guarantee safety.
1 Introduction
Consensus algorithms allow a collection of machines
to work as a coherent group that can survive the fail-
ures of some of its members. Because of this, they play a
key role in building reliable large-scale software systems .
Paxos [15, 16] has dominated the discussion of consen-
sus algorithms over the last decade: most implementations
of consensus are based on Paxos or influenced by it, and
Paxos has become the primary vehicle used to teach stu-
dents about consensus.
Unfortunately, Paxos is quite difficult to understand, in
spite of numerous attempts to make it more approachable.
Furthermore, its architecture requires complex changes
to support practical systems. As a result, both system
builders and students struggle with Paxos.
After struggling with Paxos ourselves, we set out to
find a new consensus algorithm that could provide a bet-
ter foundation for system building and education. Our ap-
proach was unusual in that our primary goal was under-
standability: could we define a consensus algorithm for
practical systems and describe it in a way that is signifi-
cantly easier to learn than Paxos? Furthermore, we wanted
the algorithm to facilitate the development of intuitions
that are essential for system builders. It was important not
just for the algorithm to work, but for it to be obvious why
it works.
The result of this work is a consensus algorithm called
Raft. In designing Raft we applied specific techniques to
improve understandability, including decomposition (Raft
separates leader election, log replication, and safety) an d
This tech report is an extended version of [32]; additional m aterial is
noted with a gray bar in the margin. Published May 20, 2014.
state space reduction (relative to Paxos, Raft reduces the
degree of nondeterminism and the ways servers can be in-
consistent with each other). A user study with 43 students
at two universities shows that Raft is significantly easier
to understand than Paxos: after learning both algorithms,
33 of these students were able to answer questions about
Raft better than questions about Paxos.
Raft is similar in many ways to existing consensus al-
gorithms (most notably, Oki and Liskov’s Viewstamped
Replication [29, 22]), but it has several novel features:
• Strong leader: Raft uses a stronger form of leader-
ship than other consensus algorithms. For example,
log entries only flow from the leader to other servers.
This simplifies the management of the replicated log
and makes Raft easier to understand.
• Leader election: Raft uses randomized timers to
elect leaders. This adds only a small amount of
mechanism to the heartbeats already required for any
consensus algorithm, while resolving conflicts sim-
ply and rapidly.
• Membership changes: Raft’s mechanism for
changing the set of servers in the cluster uses a new
joint consensus approach where the majorities of
two different configurations overlap during transi-
tions. This allows the cluster to continue operating
normally during configuration changes.
We believe that Raft is superior to Paxos and other con-
sensus algorithms, both for educational purposes and as a
foundation for implementation. It is simpler and more un-
derstandable than other algorithms; it is described com-
pletely enough to meet the needs of a practical system;
it has several open-source implementations and is used
by several companies; its safety properties have been for-
mally specified and proven; and its efficiency is compara-
ble to other algorithms.
The remainder of the paper introduces the replicated
state machine problem (Section 2), discusses the strengths
and weaknesses of Paxos (Section 3), describes our gen-
eral approach to understandability (Section 4), presents
the Raft consensus algorithm (Sections 5–8), evaluates
Raft (Section 9), and discusses related work (Section 10).
2 Replicated state machines
Consensus algorithms typically arise in the context of
replicated state machines [37]. In this approach, state ma-
chines on a collection of servers compute identical copies
of the same state and can continue operating even if some
of the servers are down. Replicated state machines are
1

Figure 1: Replicated state machine architecture. The con-
sensus algorithm manages a replicated log containing state
machine commands from clients. The state machines process
identical sequences of commands from the logs, so they pro-
duce the same outputs.
used to solve a variety of fault tolerance problems in dis-
tributed systems. For example, large-scale systems that
have a single cluster leader, such as GFS [8], HDFS [38],
and RAMCloud [33], typically use a separate replicated
state machine to manage leader election and store config-
uration information that must survive leader crashes. Ex-
amples of replicated state machines include Chubby [2]
and ZooKeeper [11].
Replicated state machines are typically implemented
using a replicated log, as shown in Figure 1. Each server
stores a log containing a series of commands, which its
state machine executes in order. Each log contains the
same commands in the same order, so each state ma-
chine processes the same sequence of commands. Since
the state machines are deterministic, each computes the
same state and the same sequence of outputs.
Keeping the replicated log consistent is the job of the
consensus algorithm. The consensus module on a server
receives commands from clients and adds them to its log.
It communicates with the consensus modules on other
servers to ensure that every log eventually contains the
same requests in the same order, even if some servers fail.
Once commands are properly replicated, each server’s
state machine processes them in log order, and the out-
puts are returned to clients. As a result, the servers appear
to form a single, highly reliable state machine.
Consensus algorithms for practical systems typically
have the following properties:
• They ensure safety (never returning an incorrect re-
sult) under all non-Byzantine conditions, including
network delays, partitions, and packet loss, duplica-
tion, and reordering.
• They are fully functional ( available) as long as any
majority of the servers are operational and can com-
municate with each other and with clients. Thus, a
typical cluster of five servers can tolerate the failure
of any two servers. Servers are assumed to fail by
stopping; they may later recover from state on stable
storage and rejoin the cluster.
• They do not depend on timing to ensure the consis-
tency of the logs: faulty clocks and extreme message
delays can, at worst, cause availability problems.
• In the common case, a command can complete as
soon as a majority of the cluster has responded to a
single round of remote procedure calls; a minority of
slow servers need not impact overall system perfor-
mance.
3 What’s wrong with Paxos?
Over the last ten years, Leslie Lamport’s Paxos proto-
col [15] has become almost synonymous with consensus:
it is the protocol most commonly taught in courses, and
most implementations of consensus use it as a starting
point. Paxos first defines a protocol capable of reaching
agreement on a single decision, such as a single replicated
log entry. We refer to this subset as single-decree Paxos.
Paxos then combines multiple instances of this protocol to
facilitate a series of decisions such as a log ( multi-Paxos).
Paxos ensures both safety and liveness, and it supports
changes in cluster membership. Its correctness has been
proven, and it is efficient in the normal case.
Unfortunately, Paxos has two significant drawbacks.
The first drawback is that Paxos is exceptionally diffi-
cult to understand. The full explanation [15] is notori-
ously opaque; few people succeed in understanding it, and
only with great effort. As a result, there have been several
attempts to explain Paxos in simpler terms [16, 20, 21].
These explanations focus on the single-decree subset, yet
they are still challenging. In an informal survey of atten-
dees at NSDI 2012, we found few people who were com-
fortable with Paxos, even among seasoned researchers.
We struggled with Paxos ourselves; we were not able to
understand the complete protocol until after reading sev-
eral simplified explanations and designing our own alter-
native protocol, a process that took almost a year.
We hypothesize that Paxos’ opaqueness derives from
its choice of the single-decree subset as its foundation.
Single-decree Paxos is dense and subtle: it is divided into
two stages that do not have simple intuitive explanations
and cannot be understood independently. Because of this,
it is difficult to develop intuitions about why the single-
decree protocol works. The composition rules for multi-
Paxos add significant additional complexity and subtlety.
We believe that the overall problem of reaching consensus
on multiple decisions (i.e., a log instead of a single entry)
can be decomposed in other ways that are more direct and
obvious.
The second problem with Paxos is that it does not pro-
vide a good foundation for building practical implemen-
tations. One reason is that there is no widely agreed-
upon algorithm for multi-Paxos. Lamport’s descriptions
are mostly about single-decree Paxos; he sketched possi-
ble approaches to multi-Paxos, but many details are miss-
ing. There have been several attempts to flesh out and op-
timize Paxos, such as [26], [39], and [13], but these differ
2

from each other and from Lamport’s sketches. Systems
such as Chubby [4] have implemented Paxos-like algo-
rithms, but in most cases their details have not been pub-
lished.
Furthermore, the Paxos architecture is a poor one for
building practical systems; this is another consequence of
the single-decree decomposition. For example, there is lit-
tle benefit to choosing a collection of log entries indepen-
dently and then melding them into a sequential log; this
just adds complexity. It is simpler and more efficient to
design a system around a log, where new entries are ap-
pended sequentially in a constrained order. Another prob-
lem is that Paxos uses a symmetric peer-to-peer approach
at its core (though it eventually suggests a weak form of
leadership as a performance optimization). This makes
sense in a simplified world where only one decision will
be made, but few practical systems use this approach. If a
series of decisions must be made, it is simpler and faster
to first elect a leader, then have the leader coordinate the
decisions.
As a result, practical systems bear little resemblance
to Paxos. Each implementation begins with Paxos, dis-
covers the difficulties in implementing it, and then de-
velops a significantly different architecture. This is time -
consuming and error-prone, and the difficulties of under-
standing Paxos exacerbate the problem. Paxos’ formula-
tion may be a good one for proving theorems about its cor-
rectness, but real implementations are so different from
Paxos that the proofs have little value. The following com-
ment from the Chubby implementers is typical:
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 un-
proven protocol [4].
Because of these problems, we concluded that Paxos
does not provide a good foundation either for system
building or for education. Given the importance of con-
sensus in large-scale software systems, we decided to see
if we could design an alternative consensus algorithm
with better properties than Paxos. Raft is the result of that
experiment.
4 Designing for understandability
We had several goals in designing Raft: it must provide
a complete and practical foundation for system building,
so that it significantly reduces the amount of design work
required of developers; it must be safe under all conditions
and available under typical operating conditions; and it
must be efficient for common operations. But our most
important goal—and most difficult challenge—was un-
derstandability. It must be possible for a large audience to
understand the algorithm comfortably. In addition, it must
be possible to develop intuitions about the algorithm, so
that system builders can make the extensions that are in-
evitable in real-world implementations.
There were numerous points in the design of Raft
where we had to choose among alternative approaches.
In these situations we evaluated the alternatives based on
understandability: how hard is it to explain each alterna-
tive (for example, how complex is its state space, and does
it have subtle implications?), and how easy will it be for a
reader to completely understand the approach and its im-
plications?
We recognize that there is a high degree of subjectiv-
ity in such analysis; nonetheless, we used two techniques
that are generally applicable. The first technique is the
well-known approach of problem decomposition: wher-
ever possible, we divided problems into separate pieces
that could be solved, explained, and understood relatively
independently. For example, in Raft we separated leader
election, log replication, safety, and membership changes.
Our second approach was to simplify the state space
by reducing the number of states to consider, making the
system more coherent and eliminating nondeterminism
where possible. Specifically, logs are not allowed to have
holes, and Raft limits the ways in which logs can become
inconsistent with each other. Although in most cases we
tried to eliminate nondeterminism, there are some situ-
ations where nondeterminism actually improves under-
standability. In particular, randomized approaches intro -
duce nondeterminism, but they tend to reduce the state
space by handling all possible choices in a similar fashion
(“choose any; it doesn’t matter”). We used randomization
to simplify the Raft leader election algorithm.
5 The Raft consensus algorithm
Raft is an algorithm for managing a replicated log of
the form described in Section 2. Figure 2 summarizes the
algorithm in condensed form for reference, and Figure 3
lists key properties of the algorithm; the elements of these
figures are discussed piecewise over the rest of this sec-
tion.
Raft implements consensus by first electing a distin-
guished leader, then giving the leader complete responsi-
bility for managing the replicated log. The leader accepts
log entries from clients, replicates them on other servers,
and tells servers when it is safe to apply log entries to
their state machines. Having a leader simplifies the man-
agement of the replicated log. For example, the leader can
decide where to place new entries in the log without con-
sulting other servers, and data flows in a simple fashion
from the leader to other servers. A leader can fail or be-
come disconnected from the other servers, in which case
a new leader is elected.
Given the leader approach, Raft decomposes the con-
sensus problem into three relatively independent subprob-
lems, which are discussed in the subsections that follow:
• Leader election: a new leader must be chosen when
an existing leader fails (Section 5.2).
• Log replication: the leader must accept log entries
3

Invoked by candidates to gather votes (§5.2).
Arguments:
term candidate’s term
candidateId candidate requesting vote
lastLogIndex index of candidate’s last log entry (§5.4)
lastLogTerm term of candidate’s last log entry (§5.4)
Results:
term currentTerm, for candidate to update itself
voteGranted true means candidate received vote
Receiver implementation:
1. Reply false if term < currentTerm (§5.1)
2. If votedFor is null or candidateId, and candidate’s log is at
least as up-to-date as receiver’s log, grant vote (§5.2, §5.4)
RequestVote RPC
Invoked by leader to replicate log entries (§5.3); also used as
heartbeat (§5.2).
Arguments:
term leader’s term
leaderId so follower can redirect clients
prevLogIndex index of log entry immediately preceding
new ones
prevLogTerm term of prevLogIndex entry
entries[] log entries to store (empty for heartbeat;
may send more than one for efficiency)
leaderCommit leader’s commitIndex
Results:
term currentTerm, for leader to update itself
success true if follower contained entry matching
prevLogIndex and prevLogTerm
Receiver implementation:
1. Reply false if term < currentTerm (§5.1)
2. Reply false if log doesn’t contain an entry at prevLogIndex
whose term matches prevLogTerm (§5.3)
3. If an existing entry conflicts with a new one (same index
but different terms), delete the existing entry and all that
follow it (§5.3)
4. Append any new entries not already in the log
5. If leaderCommit > commitIndex, set commitIndex =
min(leaderCommit, index of last new entry)
AppendEntries RPC
Persistent state on all servers:
(Updated on stable storage before responding to RPCs)
currentTerm latest term server has seen (initialized to 0
on first boot, increases monotonically)
votedFor candidateId that received vote in current
term (or null if none)
log[] log entries; each entry contains command
for state machine, and term when entry
was received by leader (first index is 1)
Volatile state on all servers:
commitIndex index of highest log entry known to be
committed (initialized to 0, increases
monotonically)
lastApplied index of highest log entry applied to state
machine (initialized to 0, increases
monotonically)
Volatile state on leaders:
(Reinitialized after election)
nextIndex[] for each server, index of the next log entry
to send to that server (initialized to leader
last log index + 1)
matchIndex[] for each server, index of highest log entry
known to be replicated on server
(initialized to 0, increases monotonically)
State
All Servers:
• If commitIndex > lastApplied: increment lastApplied, apply
log[lastApplied] to state machine (§5.3)
• If RPC request or response contains term T > currentTerm:
set currentTerm = T, convert to follower (§5.1)
Followers (§5.2):
• Respond to RPCs from candidates and leaders
• If election timeout elapses without receiving AppendEntries
RPC from current leader or granting vote to candidate:
convert to candidate
Candidates (§5.2):
• On conversion to candidate, start election:
• Increment currentTerm
• V ote for self
• Reset election timer
• Send RequestV ote RPCs to all other servers
• If votes received from majority of servers: become leader
• If AppendEntries RPC received from new leader: convert to
follower
• If election timeout elapses: start new election
Leaders:
• Upon election: send initial empty AppendEntries RPCs
(heartbeat) to each server; repeat during idle periods to
prevent election timeouts (§5.2)
• If command received from client: append entry to local log,
respond after entry applied to state machine (§5.3)
• If last log index ≥ nextIndex for a follower: send
AppendEntries RPC with log entries starting at nextIndex
• If successful: update nextIndex and matchIndex for
follower (§5.3)
• If AppendEntries fails because of log inconsistency:
decrement nextIndex and retry (§5.3)
• If there exists an N such that N > commitIndex, a majority
of matchIndex[i] ≥ N, and log[N].term == currentTerm:
set commitIndex = N (§5.3, §5.4).
Rules for Servers
Figure 2: A condensed summary of the Raft consensus algorithm (excluding membership changes and log compaction). The server
behavior in the upper-left box is described as a set of rules t hat trigger independently and repeatedly. Section numbers such as §5.2
indicate where particular features are discussed. A formal specification [31] describes the algorithm more precisely.
4

Election Safety: at most one leader can be elected in a
given term. §5.2
Leader Append-Only: a leader never overwrites or deletes
entries in its log; it only appends new entries. §5.3
Log Matching: if two logs contain an entry with the same
index and term, then the logs are identical in all entries
up through the given index. §5.3
Leader Completeness: if a log entry is committed in a
given term, then that entry will be present in the logs
of the leaders for all higher-numbered terms. §5.4
State Machine Safety: if a server has applied a log entry
at a given index to its state machine, no other server
will ever apply a different log entry for the same index.
§5.4.3
Figure 3: Raft guarantees that each of these properties is true
at all times. The section numbers indicate where each prop-
erty is discussed.
from clients and replicate them across the cluster,
forcing the other logs to agree with its own (Sec-
tion 5.3).
• Safety: the key safety property for Raft is the State
Machine Safety Property in Figure 3: if any server
has applied a particular log entry to its state machine,
then no other server may apply a different command
for the same log index. Section 5.4 describes how
Raft ensures this property; the solution involves an
additional restriction on the election mechanism de-
scribed in Section 5.2.
After presenting the consensus algorithm, this section dis-
cusses the issue of availability and the role of timing in the
system.
5.1 Raft basics
A Raft cluster contains several servers; five is a typical
number, which allows the system to tolerate two failures.
At any given time each server is in one of three states:
leader, follower, or candidate. In normal operation there
is exactly one leader and all of the other servers are fol-
lowers. Followers are passive: they issue no requests on
their own but simply respond to requests from leaders
and candidates. The leader handles all client requests (if
a client contacts a follower, the follower redirects it to th e
leader). The third state, candidate, is used to elect a new
leader as described in Section 5.2. Figure 4 shows the
states and their transitions; the transitions are discusse d
below.
Raft divides time into terms of arbitrary length, as
shown in Figure 5. Terms are numbered with consecutive
integers. Each term begins with an election, in which one
or more candidates attempt to become leader as described
in Section 5.2. If a candidate wins the election, then it
serves as leader for the rest of the term. In some situations
an election will result in a split vote. In this case the term
will end with no leader; a new term (with a new election)
Figure 4: Server states. Followers only respond to requests
from other servers. If a follower receives no communication ,
it becomes a candidate and initiates an election. A candidat e
that receives votes from a majority of the full cluster becomes
the new leader. Leaders typically operate until they fail.
Figure 5: Time is divided into terms, and each term begins
with an election. After a successful election, a single lead er
manages the cluster until the end of the term. Some elections
fail, in which case the term ends without choosing a leader.
The transitions between terms may be observed at different
times on different servers.
will begin shortly. Raft ensures that there is at most one
leader in a given term.
Different servers may observe the transitions between
terms at different times, and in some situations a server
may not observe an election or even entire terms. Terms
act as a logical clock [14] in Raft, and they allow servers
to detect obsolete information such as stale leaders. Each
server stores a current term number, which increases
monotonically over time. Current terms are exchanged
whenever servers communicate; if one server’s current
term is smaller than the other’s, then it updates its current
term to the larger value. If a candidate or leader discovers
that its term is out of date, it immediately reverts to fol-
lower state. If a server receives a request with a stale term
number, it rejects the request.
Raft servers communicate using remote procedure calls
(RPCs), and the basic consensus algorithm requires only
two types of RPCs. RequestV ote RPCs are initiated by
candidates during elections (Section 5.2), and Append-
Entries RPCs are initiated by leaders to replicate log en-
tries and to provide a form of heartbeat (Section 5.3). Sec-
tion 7 adds a third RPC for transferring snapshots between
servers. Servers retry RPCs if they do not receive a re-
sponse in a timely manner, and they issue RPCs in parallel
for best performance.
5.2 Leader election
Raft uses a heartbeat mechanism to trigger leader elec-
tion. When servers start up, they begin as followers. A
server remains in follower state as long as it receives valid
5

RPCs from a leader or candidate. Leaders send periodic
heartbeats (AppendEntries RPCs that carry no log entries)
to all followers in order to maintain their authority. If a
follower receives no communication over a period of time
called the election timeout, then it assumes there is no vi-
able leader and begins an election to choose a new leader.
To begin an election, a follower increments its current
term and transitions to candidate state. It then votes for
itself and issues RequestV ote RPCs in parallel to each of
the other servers in the cluster. A candidate continues in
this state until one of three things happens: (a) it wins the
election, (b) another server establishes itself as leader, or
(c) a period of time goes by with no winner. These out-
comes are discussed separately in the paragraphs below.
A candidate wins an election if it receives votes from
a majority of the servers in the full cluster for the same
term. Each server will vote for at most one candidate in a
given term, on a first-come-first-served basis (note: Sec-
tion 5.4 adds an additional restriction on votes). The ma-
jority rule ensures that at most one candidate can win the
election for a particular term (the Election Safety Prop-
erty in Figure 3). Once a candidate wins an election, it
becomes leader. It then sends heartbeat messages to all of
the other servers to establish its authority and prevent new
elections.
While waiting for votes, a candidate may receive an
AppendEntries RPC from another server claiming to be
leader. If the leader’s term (included in its RPC) is at least
as large as the candidate’s current term, then the candidate
recognizes the leader as legitimate and returns to follower
state. If the term in the RPC is smaller than the candidate’s
current term, then the candidate rejects the RPC and con-
tinues in candidate state.
The third possible outcome is that a candidate neither
wins nor loses the election: if many followers become
candidates at the same time, votes could be split so that
no candidate obtains a majority. When this happens, each
candidate will time out and start a new election by incre-
menting its term and initiating another round of Request-
V ote RPCs. However, without extra measures split votes
could repeat indefinitely.
Raft uses randomized election timeouts to ensure that
split votes are rare and that they are resolved quickly. To
prevent split votes in the first place, election timeouts are
chosen randomly from a fixed interval (e.g., 150–300ms).
This spreads out the servers so that in most cases only a
single server will time out; it wins the election and sends
heartbeats before any other servers time out. The same
mechanism is used to handle split votes. Each candidate
restarts its randomized election timeout at the start of an
election, and it waits for that timeout to elapse before
starting the next election; this reduces the likelihood of
another split vote in the new election. Section 9.3 shows
that this approach elects a leader rapidly.
Figure 6: Logs are composed of entries, which are numbered
sequentially. Each entry contains the term in which it was
created (the number in each box) and a command for the state
machine. An entry is considered committed if it is safe for that
entry to be applied to state machines.
Elections are an example of how understandability
guided our choice between design alternatives. Initially
we planned to use a ranking system: each candidate was
assigned a unique rank, which was used to select between
competing candidates. If a candidate discovered another
candidate with higher rank, it would return to follower
state so that the higher ranking candidate could more eas-
ily win the next election. We found that this approach
created subtle issues around availability (a lower-ranked
server might need to time out and become a candidate
again if a higher-ranked server fails, but if it does so too
soon, it can reset progress towards electing a leader). We
made adjustments to the algorithm several times, but after
each adjustment new corner cases appeared. Eventually
we concluded that the randomized retry approach is more
obvious and understandable.
5.3 Log replication
Once a leader has been elected, it begins servicing
client requests. Each client request contains a command to
be executed by the replicated state machines. The leader
appends the command to its log as a new entry, then is-
sues AppendEntries RPCs in parallel to each of the other
servers to replicate the entry. When the entry has been
safely replicated (as described below), the leader applies
the entry to its state machine and returns the result of that
execution to the client. If followers crash or run slowly,
or if network packets are lost, the leader retries Append-
Entries RPCs indefinitely (even after it has responded to
the client) until all followers eventually store all log en-
tries.
Logs are organized as shown in Figure 6. Each log en-
try stores a state machine command along with the term
number when the entry was received by the leader. The
term numbers in log entries are used to detect inconsis-
tencies between logs and to ensure some of the properties
in Figure 3. Each log entry also has an integer index iden-
6

tifying its position in the log.
The leader decides when it is safe to apply a log en-
try to the state machines; such an entry is called commit-
ted. Raft guarantees that committed entries are durable
and will eventually be executed by all of the available
state machines. A log entry is committed once the leader
that created the entry has replicated it on a majority of
the servers (e.g., entry 7 in Figure 6). This also commits
all preceding entries in the leader’s log, including entrie s
created by previous leaders. Section 5.4 discusses some
subtleties when applying this rule after leader changes,
and it also shows that this definition of commitment is
safe. The leader keeps track of the highest index it knows
to be committed, and it includes that index in future
AppendEntries RPCs (including heartbeats) so that the
other servers eventually find out. Once a follower learns
that a log entry is committed, it applies the entry to its
local state machine (in log order).
We designed the Raft log mechanism to maintain a high
level of coherency between the logs on different servers.
Not only does this simplify the system’s behavior and
make it more predictable, but it is an important component
of ensuring safety. Raft maintains the following proper-
ties, which together constitute the Log Matching Property
in Figure 3:
• If two entries in different logs have the same index
and term, then they store the same command.
• If two entries in different logs have the same index
and term, then the logs are identical in all preceding
entries.
The first property follows from the fact that a leader
creates at most one entry with a given log index in a given
term, and log entries never change their position in the
log. The second property is guaranteed by a simple con-
sistency check performed by AppendEntries. When send-
ing an AppendEntries RPC, the leader includes the index
and term of the entry in its log that immediately precedes
the new entries. If the follower does not find an entry in
its log with the same index and term, then it refuses the
new entries. The consistency check acts as an induction
step: the initial empty state of the logs satisfies the Log
Matching Property, and the consistency check preserves
the Log Matching Property whenever logs are extended.
As a result, whenever AppendEntries returns successfully,
the leader knows that the follower’s log is identical to its
own log up through the new entries.
During normal operation, the logs of the leader and
followers stay consistent, so the AppendEntries consis-
tency check never fails. However, leader crashes can leave
the logs inconsistent (the old leader may not have fully
replicated all of the entries in its log). These inconsisten -
cies can compound over a series of leader and follower
crashes. Figure 7 illustrates the ways in which followers’
logs may differ from that of a new leader. A follower may
Figure 7: When the leader at the top comes to power, it is
possible that any of scenarios (a–f) could occur in follower
logs. Each box represents one log entry; the number in the
box is its term. A follower may be missing entries (a–b), may
have extra uncommitted entries (c–d), or both (e–f). For ex-
ample, scenario (f) could occur if that server was the leader
for term 2, added several entries to its log, then crashed before
committing any of them; it restarted quickly, became leader
for term 3, and added a few more entries to its log; before any
of the entries in either term 2 or term 3 were committed, the
server crashed again and remained down for several terms.
be missing entries that are present on the leader, it may
have extra entries that are not present on the leader, or
both. Missing and extraneous entries in a log may span
multiple terms.
In Raft, the leader handles inconsistencies by forcing
the followers’ logs to duplicate its own. This means that
conflicting entries in follower logs will be overwritten
with entries from the leader’s log. Section 5.4 will show
that this is safe when coupled with one more restriction.
To bring a follower’s log into consistency with its own,
the leader must find the latest log entry where the two
logs agree, delete any entries in the follower’s log after
that point, and send the follower all of the leader’s entries
after that point. All of these actions happen in response
to the consistency check performed by AppendEntries
RPCs. The leader maintains a nextIndex for each follower,
which is the index of the next log entry the leader will
send to that follower. When a leader first comes to power,
it initializes all nextIndex values to the index just after t he
last one in its log (11 in Figure 7). If a follower’s log is
inconsistent with the leader’s, the AppendEntries consis-
tency check will fail in the next AppendEntries RPC. Af-
ter a rejection, the leader decrements nextIndex and retries
the AppendEntries RPC. Eventually nextIndex will reach
a point where the leader and follower logs match. When
this happens, AppendEntries will succeed, which removes
any conflicting entries in the follower’s log and appends
entries from the leader’s log (if any). Once AppendEntries
succeeds, the follower’s log is consistent with the leader’ s,
and it will remain that way for the rest of the term.
If desired, the protocol can be optimized to reduce the
number of rejected AppendEntries RPCs. For example,
when rejecting an AppendEntries request, the follower
7

can include the term of the conflicting entry and the first
index it stores for that term. With this information, the
leader can decrement nextIndex to bypass all of the con-
flicting entries in that term; one AppendEntries RPC will
be required for each term with conflicting entries, rather
than one RPC per entry. In practice, we doubt this opti-
mization is necessary, since failures happen infrequently
and it is unlikely that there will be many inconsistent en-
tries.
With this mechanism, a leader does not need to take any
special actions to restore log consistency when it comes to
power. It just begins normal operation, and the logs auto-
matically converge in response to failures of the Append-
Entries consistency check. A leader never overwrites or
deletes entries in its own log (the Leader Append-Only
Property in Figure 3).
This log replication mechanism exhibits the desirable
consensus properties described in Section 2: Raft can ac-
cept, replicate, and apply new log entries as long as a ma-
jority of the servers are up; in the normal case a new entry
can be replicated with a single round of RPCs to a ma-
jority of the cluster; and a single slow follower will not
impact performance.
5.4 Safety
The previous sections described how Raft elects lead-
ers and replicates log entries. However, the mechanisms
described so far are not quite sufficient to ensure that each
state machine executes exactly the same commands in the
same order. For example, a follower might be unavailable
while the leader commits several log entries, then it could
be elected leader and overwrite these entries with new
ones; as a result, different state machines might execute
different command sequences.
This section completes the Raft algorithm by adding a
restriction on which servers may be elected leader. The
restriction ensures that the leader for any given term con-
tains all of the entries committed in previous terms (the
Leader Completeness Property from Figure 3). Given the
election restriction, we then make the rules for commit-
ment more precise. Finally, we present a proof sketch for
the Leader Completeness Property and show how it leads
to correct behavior of the replicated state machine.
5.4.1 Election restriction
In any leader-based consensus algorithm, the leader
must eventually store all of the committed log entries. In
some consensus algorithms, such as Viewstamped Repli-
cation [22], a leader can be elected even if it doesn’t
initially contain all of the committed entries. These al-
gorithms contain additional mechanisms to identify the
missing entries and transmit them to the new leader, ei-
ther during the election process or shortly afterwards. Un-
fortunately, this results in considerable additional mech a-
nism and complexity. Raft uses a simpler approach where
it guarantees that all the committed entries from previous
Figure 8: A time sequence showing why a leader cannot de-
termine commitment using log entries from older terms. In
(a) S1 is leader and partially replicates the log entry at ind ex
2. In (b) S1 crashes; S5 is elected leader for term 3 with votes
from S3, S4, and itself, and accepts a different entry at log
index 2. In (c) S5 crashes; S1 restarts, is elected leader, an d
continues replication. At this point, the log entry from ter m 2
has been replicated on a majority of the servers, but it is not
committed. If S1 crashes as in (d), S5 could be elected leader
(with votes from S2, S3, and S4) and overwrite the entry with
its own entry from term 3. However, if S1 replicates an en-
try from its current term on a majority of the servers before
crashing, as in (e), then this entry is committed (S5 cannot
win an election). At this point all preceding entries in the l og
are committed as well.
terms are present on each new leader from the moment of
its election, without the need to transfer those entries to
the leader. This means that log entries only flow in one di-
rection, from leaders to followers, and leaders never over-
write existing entries in their logs.
Raft uses the voting process to prevent a candidate from
winning an election unless its log contains all committed
entries. A candidate must contact a majority of the cluster
in order to be elected, which means that every committed
entry must be present in at least one of those servers. If the
candidate’s log is at least as up-to-date as any other log
in that majority (where “up-to-date” is defined precisely
below), then it will hold all the committed entries. The
RequestV ote RPC implements this restriction: the RPC
includes information about the candidate’s log, and the
voter denies its vote if its own log is more up-to-date than
that of the candidate.
Raft determines which of two logs is more up-to-date
by comparing the index and term of the last entries in the
logs. If the logs have last entries with different terms, the n
the log with the later term is more up-to-date. If the logs
end with the same term, then whichever log is longer is
more up-to-date.
5.4.2 Committing entries from previous terms
As described in Section 5.3, a leader knows that an en-
try from its current term is committed once that entry is
stored on a majority of the servers. If a leader crashes be-
fore committing an entry, future leaders will attempt to
finish replicating the entry. However, a leader cannot im-
mediately conclude that an entry from a previous term is
committed once it is stored on a majority of servers. Fig-
8

Figure 9: If S1 (leader for term T) commits a new log entry
from its term, and S5 is elected leader for a later term U, then
there must be at least one server (S3) that accepted the log
entry and also voted for S5.
ure 8 illustrates a situation where an old log entry is stored
on a majority of servers, yet can still be overwritten by a
future leader.
To eliminate problems like the one in Figure 8, Raft
never commits log entries from previous terms by count-
ing replicas. Only log entries from the leader’s current
term are committed by counting replicas; once an entry
from the current term has been committed in this way,
then all prior entries are committed indirectly because
of the Log Matching Property. There are some situations
where a leader could safely conclude that an older log en-
try is committed (for example, if that entry is stored on ev-
ery server), but Raft takes a more conservative approach
for simplicity.
Raft incurs this extra complexity in the commitment
rules because log entries retain their original term num-
bers when a leader replicates entries from previous
terms. In other consensus algorithms, if a new leader re-
replicates entries from prior “terms,” it must do so with
its new “term number.” Raft’s approach makes it easier
to reason about log entries, since they maintain the same
term number over time and across logs. In addition, new
leaders in Raft send fewer log entries from previous terms
than in other algorithms (other algorithms must send re-
dundant log entries to renumber them before they can be
committed).
5.4.3 Safety argument
Given the complete Raft algorithm, we can now ar-
gue more precisely that the Leader Completeness Prop-
erty holds (this argument is based on the safety proof; see
Section 9.2). We assume that the Leader Completeness
Property does not hold, then we prove a contradiction.
Suppose the leader for term T (leader T) commits a log
entry from its term, but that log entry is not stored by the
leader of some future term. Consider the smallest term U
> T whose leader (leader U) does not store the entry.
1. The committed entry must have been absent from
leaderU’s log at the time of its election (leaders never
delete or overwrite entries).
2. leader T replicated the entry on a majority of the clus-
ter, and leader U received votes from a majority of
the cluster. Thus, at least one server (“the voter”)
both accepted the entry from leader T and voted for
leaderU, as shown in Figure 9. The voter is key to
reaching a contradiction.
3. The voter must have accepted the committed entry
from leader T before voting for leader U; otherwise it
would have rejected the AppendEntries request from
leaderT (its current term would have been higher than
T).
4. The voter still stored the entry when it voted for
leaderU, since every intervening leader contained the
entry (by assumption), leaders never remove entries,
and followers only remove entries if they conflict
with the leader.
5. The voter granted its vote to leader U, so leader U’s
log must have been as up-to-date as the voter’s. This
leads to one of two contradictions.
6. First, if the voter and leader U shared the same last
log term, then leader U’s log must have been at least
as long as the voter’s, so its log contained every entry
in the voter’s log. This is a contradiction, since the
voter contained the committed entry and leaderU was
assumed not to.
7. Otherwise, leader U’s last log term must have been
larger than the voter’s. Moreover, it was larger than
T, since the voter’s last log term was at least T (it con-
tains the committed entry from term T). The earlier
leader that created leader U’s last log entry must have
contained the committed entry in its log (by assump-
tion). Then, by the Log Matching Property, leaderU’s
log must also contain the committed entry, which is
a contradiction.
8. This completes the contradiction. Thus, the leaders
of all terms greater than T must contain all entries
from term T that are committed in term T.
9. The Log Matching Property guarantees that future
leaders will also contain entries that are committed
indirectly, such as index 2 in Figure 8(d).
Given the Leader Completeness Property, we can prove
the State Machine Safety Property from Figure 3, which
states that if a server has applied a log entry at a given
index to its state machine, no other server will ever apply a
different log entry for the same index. At the time a server
applies a log entry to its state machine, its log must be
identical to the leader’s log up through that entry and the
entry must be committed. Now consider the lowest term
in which any server applies a given log index; the Log
Completeness Property guarantees that the leaders for all
higher terms will store that same log entry, so servers that
apply the index in later terms will apply the same value.
Thus, the State Machine Safety Property holds.
Finally, Raft requires servers to apply entries in log in-
dex order. Combined with the State Machine Safety Prop-
erty, this means that all servers will apply exactly the same
set of log entries to their state machines, in the same order.
9

5.5 Follower and candidate crashes
Until this point we have focused on leader failures. Fol-
lower and candidate crashes are much simpler to han-
dle than leader crashes, and they are both handled in the
same way. If a follower or candidate crashes, then fu-
ture RequestV ote and AppendEntries RPCs sent to it will
fail. Raft handles these failures by retrying indefinitely;
if the crashed server restarts, then the RPC will complete
successfully. If a server crashes after completing an RPC
but before responding, then it will receive the same RPC
again after it restarts. Raft RPCs are idempotent, so this
causes no harm. For example, if a follower receives an
AppendEntries request that includes log entries already
present in its log, it ignores those entries in the new re-
quest.
5.6 Timing and availability
One of our requirements for Raft is that safety must
not depend on timing: the system must not produce incor-
rect results just because some event happens more quickly
or slowly than expected. However, availability (the ability
of the system to respond to clients in a timely manner)
must inevitably depend on timing. For example, if mes-
sage exchanges take longer than the typical time between
server crashes, candidates will not stay up long enough to
win an election; without a steady leader, Raft cannot make
progress.
Leader election is the aspect of Raft where timing is
most critical. Raft will be able to elect and maintain a
steady leader as long as the system satisfies the follow-
ing timing requirement:
broadcastTime ≪ electionTimeout ≪ MTBF
In this inequality broadcastTime is the average time it
takes a server to send RPCs in parallel to every server
in the cluster and receive their responses; electionTime-
out is the election timeout described in Section 5.2; and
MTBF is the average time between failures for a single
server. The broadcast time should be an order of mag-
nitude less than the election timeout so that leaders can
reliably send the heartbeat messages required to keep fol-
lowers from starting elections; given the randomized ap-
proach used for election timeouts, this inequality also
makes split votes unlikely. The election timeout should be
a few orders of magnitude less than MTBF so that the sys-
tem makes steady progress. When the leader crashes, the
system will be unavailable for roughly the election time-
out; we would like this to represent only a small fraction
of overall time.
The broadcast time and MTBF are properties of the un-
derlying system, while the election timeout is something
we must choose. Raft’s RPCs typically require the recip-
ient to persist information to stable storage, so the broad-
cast time may range from 0.5ms to 20ms, depending on
storage technology. As a result, the election timeout is
likely to be somewhere between 10ms and 500ms. Typical
Figure 10: Switching directly from one configuration to an-
other is unsafe because different servers will switch at dif -
ferent times. In this example, the cluster grows from three
servers to five. Unfortunately, there is a point in time where
two different leaders can be elected for the same term, one
with a majority of the old configuration ( Cold) and another
with a majority of the new configuration ( Cnew).
server MTBFs are several months or more, which easily
satisfies the timing requirement.
6 Cluster membership changes
Up until now we have assumed that the cluster config-
uration (the set of servers participating in the consensus
algorithm) is fixed. In practice, it will occasionally be nec-
essary to change the configuration, for example to replace
servers when they fail or to change the degree of replica-
tion. Although this can be done by taking the entire cluster
off-line, updating configuration files, and then restarting
the cluster, this would leave the cluster unavailable dur-
ing the changeover. In addition, if there are any manual
steps, they risk operator error. In order to avoid these is-
sues, we decided to automate configuration changes and
incorporate them into the Raft consensus algorithm.
For the configuration change mechanism to be safe,
there must be no point during the transition where it
is possible for two leaders to be elected for the same
term. Unfortunately, any approach where servers switch
directly from the old configuration to the new configura-
tion is unsafe. It isn’t possible to atomically switch all of
the servers at once, so the cluster can potentially split int o
two independent majorities during the transition (see Fig-
ure 10).
In order to ensure safety, configuration changes must
use a two-phase approach. There are a variety of ways
to implement the two phases. For example, some systems
(e.g., [22]) use the first phase to disable the old configura-
tion so it cannot process client requests; then the second
phase enables the new configuration. In Raft the cluster
first switches to a transitional configuration we call joint
consensus; once the joint consensus has been committed,
the system then transitions to the new configuration. The
joint consensus combines both the old and new configu-
rations:
• Log entries are replicated to all servers in both con-
figurations.
10

Figure 11: Timeline for a configuration change. Dashed lines
show configuration entries that have been created but not
committed, and solid lines show the latest committed configu-
ration entry. The leader first creates the Cold,new configuration
entry in its log and commits it to Cold,new (a majority of Cold
and a majority of Cnew). Then it creates the Cnew entry and
commits it to a majority of Cnew. There is no point in time in
which Cold and Cnew can both make decisions independently.
• Any server from either configuration may serve as
leader.
• Agreement (for elections and entry commitment) re-
quires separate majorities from both the old and new
configurations.
The joint consensus allows individual servers to transition
between configurations at different times without com-
promising safety. Furthermore, joint consensus allows the
cluster to continue servicing client requests throughout
the configuration change.
Cluster configurations are stored and communicated
using special entries in the replicated log; Figure 11 illus -
trates the configuration change process. When the leader
receives a request to change the configuration from Cold
to Cnew, it stores the configuration for joint consensus
(Cold,new in the figure) as a log entry and replicates that
entry using the mechanisms described previously. Once a
given server adds the new configuration entry to its log,
it uses that configuration for all future decisions (a server
always uses the latest configuration in its log, regardless
of whether the entry is committed). This means that the
leader will use the rules of Cold,new to determine when the
log entry for Cold,new is committed. If the leader crashes,
a new leader may be chosen under either Cold or Cold,new,
depending on whether the winning candidate has received
Cold,new. In any case, Cnew cannot make unilateral deci-
sions during this period.
Once Cold,new has been committed, neither Cold nor Cnew
can make decisions without approval of the other, and the
Leader Completeness Property ensures that only servers
with the Cold,new log entry can be elected as leader. It is
now safe for the leader to create a log entry describing
Cnew and replicate it to the cluster. Again, this configura-
tion will take effect on each server as soon as it is seen.
When the new configuration has been committed under
the rules of Cnew, the old configuration is irrelevant and
servers not in the new configuration can be shut down. As
shown in Figure 11, there is no time when Cold and Cnew
can both make unilateral decisions; this guarantees safety.
There are three more issues to address for reconfigura-
tion. The first issue is that new servers may not initially
store any log entries. If they are added to the cluster in
this state, it could take quite a while for them to catch
up, during which time it might not be possible to com-
mit new log entries. In order to avoid availability gaps,
Raft introduces an additional phase before the configu-
ration change, in which the new servers join the cluster
as non-voting members (the leader replicates log entries
to them, but they are not considered for majorities). Once
the new servers have caught up with the rest of the cluster,
the reconfiguration can proceed as described above.
The second issue is that the cluster leader may not be
part of the new configuration. In this case, the leader steps
down (returns to follower state) once it has committed the
Cnew log entry. This means that there will be a period of
time (while it is committing Cnew) when the leader is man-
aging a cluster that does not include itself; it replicates l og
entries but does not count itself in majorities. The leader
transition occurs when Cnew is committed because this is
the first point when the new configuration can operate in-
dependently (it will always be possible to choose a leader
from Cnew). Before this point, it may be the case that only
a server from Cold can be elected leader.
The third issue is that removed servers (those not in
Cnew) can disrupt the cluster. These servers will not re-
ceive heartbeats, so they will time out and start new elec-
tions. They will then send RequestV ote RPCs with new
term numbers, and this will cause the current leader to
revert to follower state. A new leader will eventually be
elected, but the removed servers will time out again and
the process will repeat, resulting in poor availability.
To prevent this problem, servers disregard RequestV ote
RPCs when they believe a current leader exists. Specif-
ically, if a server receives a RequestV ote RPC within
the minimum election timeout of hearing from a cur-
rent leader, it does not update its term or grant its vote.
This does not affect normal elections, where each server
waits at least a minimum election timeout before starting
an election. However, it helps avoid disruptions from re-
moved servers: if a leader is able to get heartbeats to its
cluster, then it will not be deposed by larger term num-
bers.
7 Log compaction
Raft’s log grows during normal operation to incorpo-
rate more client requests, but in a practical system, it can-
not grow without bound. As the log grows longer, it oc-
cupies more space and takes more time to replay. This
will eventually cause availability problems without some
mechanism to discard obsolete information that has accu-
mulated in the log.
Snapshotting is the simplest approach to compaction.
In snapshotting, the entire current system state is written
to a snapshot on stable storage, then the entire log up to
11

Figure 12: A server replaces the committed entries in its log
(indexes 1 through 5) with a new snapshot, which stores just
the current state (variables x and y in this example). The snap-
shot’s last included index and term serve to position the sna p-
shot in the log preceding entry 6.
that point is discarded. Snapshotting is used in Chubby
and ZooKeeper, and the remainder of this section de-
scribes snapshotting in Raft.
Incremental approaches to compaction, such as log
cleaning [36] and log-structured merge trees [30, 5], are
also possible. These operate on a fraction of the data at
once, so they spread the load of compaction more evenly
over time. They first select a region of data that has ac-
cumulated many deleted and overwritten objects, then
they rewrite the live objects from that region more com-
pactly and free the region. This requires significant addi-
tional mechanism and complexity compared to snapshot-
ting, which simplifies the problem by always operating
on the entire data set. While log cleaning would require
modifications to Raft, state machines can implement LSM
trees using the same interface as snapshotting.
Figure 12 shows the basic idea of snapshotting in Raft.
Each server takes snapshots independently, covering just
the committed entries in its log. Most of the work con-
sists of the state machine writing its current state to the
snapshot. Raft also includes a small amount of metadata
in the snapshot: the last included index is the index of the
last entry in the log that the snapshot replaces (the last en-
try the state machine had applied), and the last included
term is the term of this entry. These are preserved to sup-
port the AppendEntries consistency check for the first log
entry following the snapshot, since that entry needs a pre-
vious log index and term. To enable cluster membership
changes (Section 6), the snapshot also includes the latest
configuration in the log as of last included index. Once a
server completes writing a snapshot, it may delete all log
entries up through the last included index, as well as any
prior snapshot.
Although servers normally take snapshots indepen-
dently, the leader must occasionally send snapshots to
followers that lag behind. This happens when the leader
has already discarded the next log entry that it needs to
send to a follower. Fortunately, this situation is unlikely
in normal operation: a follower that has kept up with the
Invoked by leader to send chunks of a snapshot to a follower.
Leaders always send chunks in order.
Arguments:
term leader’s term
leaderId so follower can redirect clients
lastIncludedIndex the snapshot replaces all entries up through
and including this index
lastIncludedTerm term of lastIncludedIndex
offset byte offset where chunk is positioned in the
snapshot file
data[] raw bytes of the snapshot chunk, starting at
offset
done true if this is the last chunk
Results:
term currentTerm, for leader to update itself
Receiver implementation:
1. Reply immediately if term < currentTerm
2. Create new snapshot file if first chunk (offset is 0)
3. Write data into snapshot file at given offset
4. Reply and wait for more data chunks if done is false
5. Save snapshot file, discard any existing or partial snapshot
with a smaller index
6. If existing log entry has same index and term as snapshot’s
last included entry, retain log entries following it and reply
7. Discard the entire log
8. Reset state machine using snapshot contents (and load
snapshot’s cluster configuration)
InstallSnapshot RPC
Figure 13: A summary of the InstallSnapshot RPC. Snap-
shots are split into chunks for transmission; this gives the fol-
lower a sign of life with each chunk, so it can reset its election
timer.
leader would already have this entry. However, an excep-
tionally slow follower or a new server joining the cluster
(Section 6) would not. The way to bring such a follower
up-to-date is for the leader to send it a snapshot over the
network.
The leader uses a new RPC called InstallSnapshot to
send snapshots to followers that are too far behind; see
Figure 13. When a follower receives a snapshot with this
RPC, it must decide what to do with its existing log en-
tries. Usually the snapshot will contain new information
not already in the recipient’s log. In this case, the followe r
discards its entire log; it is all superseded by the snapshot
and may possibly have uncommitted entries that conflict
with the snapshot. If instead the follower receives a snap-
shot that describes a prefix of its log (due to retransmis-
sion or by mistake), then log entries covered by the snap-
shot are deleted but entries following the snapshot are still
valid and must be retained.
This snapshotting approach departs from Raft’s strong
leader principle, since followers can take snapshots with-
out the knowledge of the leader. However, we think this
departure is justified. While having a leader helps avoid
conflicting decisions in reaching consensus, consensus
has already been reached when snapshotting, so no de-
cisions conflict. Data still only flows from leaders to fol-
12

lowers, just followers can now reorganize their data.
We considered an alternative leader-based approach in
which only the leader would create a snapshot, then it
would send this snapshot to each of its followers. How-
ever, this has two disadvantages. First, sending the snap-
shot to each follower would waste network bandwidth and
slow the snapshotting process. Each follower already has
the information needed to produce its own snapshots, and
it is typically much cheaper for a server to produce a snap-
shot from its local state than it is to send and receive one
over the network. Second, the leader’s implementation
would be more complex. For example, the leader would
need to send snapshots to followers in parallel with repli-
cating new log entries to them, so as not to block new
client requests.
There are two more issues that impact snapshotting per-
formance. First, servers must decide when to snapshot. If
a server snapshots too often, it wastes disk bandwidth and
energy; if it snapshots too infrequently, it risks exhaust-
ing its storage capacity, and it increases the time required
to replay the log during restarts. One simple strategy is
to take a snapshot when the log reaches a fixed size in
bytes. If this size is set to be significantly larger than the
expected size of a snapshot, then the disk bandwidth over-
head for snapshotting will be small.
The second performance issue is that writing a snap-
shot can take a significant amount of time, and we do
not want this to delay normal operations. The solution is
to use copy-on-write techniques so that new updates can
be accepted without impacting the snapshot being writ-
ten. For example, state machines built with functional data
structures naturally support this. Alternatively, the operat-
ing system’s copy-on-write support (e.g., fork on Linux)
can be used to create an in-memory snapshot of the entire
state machine (our implementation uses this approach).
8 Client interaction
This section describes how clients interact with Raft,
including how clients find the cluster leader and how Raft
supports linearizable semantics [10]. These issues apply
to all consensus-based systems, and Raft’s solutions are
similar to other systems.
Clients of Raft send all of their requests to the leader.
When a client first starts up, it connects to a randomly-
chosen server. If the client’s first choice is not the leader,
that server will reject the client’s request and supply in-
formation about the most recent leader it has heard from
(AppendEntries requests include the network address of
the leader). If the leader crashes, client requests will tim e
out; clients then try again with randomly-chosen servers.
Our goal for Raft is to implement linearizable seman-
tics (each operation appears to execute instantaneously,
exactly once, at some point between its invocation and
its response). However, as described so far Raft can exe-
cute a command multiple times: for example, if the leader
crashes after committing the log entry but before respond-
ing to the client, the client will retry the command with a
new leader, causing it to be executed a second time. The
solution is for clients to assign unique serial numbers to
every command. Then, the state machine tracks the latest
serial number processed for each client, along with the as-
sociated response. If it receives a command whose serial
number has already been executed, it responds immedi-
ately without re-executing the request.
Read-only operations can be handled without writing
anything into the log. However, with no additional mea-
sures, this would run the risk of returning stale data, since
the leader responding to the request might have been su-
perseded by a newer leader of which it is unaware. Lin-
earizable reads must not return stale data, and Raft needs
two extra precautions to guarantee this without using the
log. First, a leader must have the latest information on
which entries are committed. The Leader Completeness
Property guarantees that a leader has all committed en-
tries, but at the start of its term, it may not know which
those are. To find out, it needs to commit an entry from
its term. Raft handles this by having each leader com-
mit a blank no-op entry into the log at the start of its
term. Second, a leader must check whether it has been de-
posed before processing a read-only request (its informa-
tion may be stale if a more recent leader has been elected).
Raft handles this by having the leader exchange heart-
beat messages with a majority of the cluster before re-
sponding to read-only requests. Alternatively, the leader
could rely on the heartbeat mechanism to provide a form
of lease [9], but this would rely on timing for safety (it
assumes bounded clock skew).
9 Implementation and evaluation
We have implemented Raft as part of a replicated
state machine that stores configuration information for
RAMCloud [33] and assists in failover of the RAMCloud
coordinator. The Raft implementation contains roughly
2000 lines of C++ code, not including tests, comments, or
blank lines. The source code is freely available [23]. There
are also about 25 independent third-party open source im-
plementations [34] of Raft in various stages of develop-
ment, based on drafts of this paper. Also, various compa-
nies are deploying Raft-based systems [34].
The remainder of this section evaluates Raft using three
criteria: understandability, correctness, and performan ce.
9.1 Understandability
To measure Raft’s understandability relative to Paxos,
we conducted an experimental study using upper-level un-
dergraduate and graduate students in an Advanced Oper-
ating Systems course at Stanford University and a Dis-
tributed Computing course at U.C. Berkeley. We recorded
a video lecture of Raft and another of Paxos, and created
corresponding quizzes. The Raft lecture covered the con-
tent of this paper except for log compaction; the Paxos
13

 0
 10
 20
 30
 40
 50
 60
 0  10  20  30  40  50  60
Raft grade
Paxos grade
Raft then Paxos
Paxos then Raft
Figure 14: A scatter plot comparing 43 participants’ perfor-
mance on the Raft and Paxos quizzes. Points above the diag-
onal (33) represent participants who scored higher for Raft .
lecture covered enough material to create an equivalent
replicated state machine, including single-decree Paxos,
multi-decree Paxos, reconfiguration, and a few optimiza-
tions needed in practice (such as leader election). The
quizzes tested basic understanding of the algorithms and
also required students to reason about corner cases. Each
student watched one video, took the corresponding quiz,
watched the second video, and took the second quiz.
About half of the participants did the Paxos portion first
and the other half did the Raft portion first in order to
account for both individual differences in performance
and experience gained from the first portion of the study.
We compared participants’ scores on each quiz to deter-
mine whether participants showed a better understanding
of Raft.
We tried to make the comparison between Paxos and
Raft as fair as possible. The experiment favored Paxos in
two ways: 15 of the 43 participants reported having some
prior experience with Paxos, and the Paxos video is 14%
longer than the Raft video. As summarized in Table 1, we
have taken steps to mitigate potential sources of bias. All
of our materials are available for review [28, 31].
On average, participants scored 4.9 points higher on the
Raft quiz than on the Paxos quiz (out of a possible 60
points, the mean Raft score was 25.7 and the mean Paxos
score was 20.8); Figure 14 shows their individual scores.
A paired t-test states that, with 95% confidence, the true
distribution of Raft scores has a mean at least 2.5 points
larger than the true distribution of Paxos scores.
We also created a linear regression model that predicts
a new student’s quiz scores based on three factors: which
quiz they took, their degree of prior Paxos experience, and
 0
 5
 10
 15
 20
implement explain
number of participants
Paxos much easier
Paxos somewhat easier
Roughly equal
Raft somewhat easier
Raft much easier
Figure 15: Using a 5-point scale, participants were asked
(left) which algorithm they felt would be easier to implemen t
in a functioning, correct, and efficient system, and (right)
which would be easier to explain to a CS graduate student.
the order in which they learned the algorithms. The model
predicts that the choice of quiz produces a 12.5-point dif-
ference in favor of Raft. This is significantly higher than
the observed difference of 4.9 points, because many of the
actual students had prior Paxos experience, which helped
Paxos considerably, whereas it helped Raft slightly less.
Curiously, the model also predicts scores 6.3 points lower
on Raft for people that have already taken the Paxos quiz;
although we don’t know why, this does appear to be sta-
tistically significant.
We also surveyed participants after their quizzes to see
which algorithm they felt would be easier to implement
or explain; these results are shown in Figure 15. An over-
whelming majority of participants reported Raft would be
easier to implement and explain (33 of 41 for each ques-
tion). However, these self-reported feelings may be less
reliable than participants’ quiz scores, and participants
may have been biased by knowledge of our hypothesis
that Raft is easier to understand.
A detailed discussion of the Raft user study is available
at [31].
9.2 Correctness
We have developed a formal specification and a proof
of safety for the consensus mechanism described in Sec-
tion 5. The formal specification [31] makes the informa-
tion summarized in Figure 2 completely precise using the
TLA+ specification language [17]. It is about 400 lines
long and serves as the subject of the proof. It is also use-
ful on its own for anyone implementing Raft. We have
mechanically proven the Log Completeness Property us-
ing the TLA proof system [7]. However, this proof relies
on invariants that have not been mechanically checked
(for example, we have not proven the type safety of the
specification). Furthermore, we have written an informal
proof [31] of the State Machine Safety property which
is complete (it relies on the specification alone) and rela-
Concern Steps taken to mitigate bias Materials for review [28, 31]
Equal lecture quality Same lecturer for both. Paxos lecture based on and improved from exist-
ing materials used in several universities. Paxos lecture i s 14% longer.
videos
Equal quiz difficulty Questions grouped in difficulty and pai red across exams. quizzes
Fair grading Used rubric. Graded in random order, alternati ng between quizzes. rubric
Table 1: Concerns of possible bias against Paxos in the study, steps t aken to counter each, and additional materials available.
14

0%
20%
40%
60%
80%
100%
 100  1000  10000  100000
cumulative percent
150-150ms
150-151ms
150-155ms
150-175ms
150-200ms
150-300ms
0%
20%
40%
60%
80%
100%
 0  100  200  300  400  500  600
cumulative percent
time without leader (ms)
12-24ms
25-50ms
50-100ms
100-200ms
150-300ms
Figure 16: The time to detect and replace a crashed leader.
The top graph varies the amount of randomness in election
timeouts, and the bottom graph scales the minimum election
timeout. Each line represents 1000 trials (except for 100 tr i-
als for “150–150ms”) and corresponds to a particular choice
of election timeouts; for example, “150–155ms” means that
election timeouts were chosen randomly and uniformly be-
tween 150ms and 155ms. The measurements were taken on a
cluster of five servers with a broadcast time of roughly 15ms.
Results for a cluster of nine servers are similar.
tively precise (it is about 3500 words long).
9.3 Performance
Raft’s performance is similar to other consensus algo-
rithms such as Paxos. The most important case for per-
formance is when an established leader is replicating new
log entries. Raft achieves this using the minimal number
of messages (a single round-trip from the leader to half the
cluster). It is also possible to further improve Raft’s per-
formance. For example, it easily supports batching and
pipelining requests for higher throughput and lower la-
tency. V arious optimizations have been proposed in the
literature for other algorithms; many of these could be ap-
plied to Raft, but we leave this to future work.
We used our Raft implementation to measure the per-
formance of Raft’s leader election algorithm and answer
two questions. First, does the election process converge
quickly? Second, what is the minimum downtime that can
be achieved after leader crashes?
To measure leader election, we repeatedly crashed the
leader of a cluster of five servers and timed how long it
took to detect the crash and elect a new leader (see Fig-
ure 16). To generate a worst-case scenario, the servers in
each trial had different log lengths, so some candidates
were not eligible to become leader. Furthermore, to en-
courage split votes, our test script triggered a synchro-
nized broadcast of heartbeat RPCs from the leader before
terminating its process (this approximates the behavior
of the leader replicating a new log entry prior to crash-
ing). The leader was crashed uniformly randomly within
its heartbeat interval, which was half of the minimum
election timeout for all tests. Thus, the smallest possible
downtime was about half of the minimum election time-
out.
The top graph in Figure 16 shows that a small amount
of randomization in the election timeout is enough to
avoid split votes in elections. In the absence of random-
ness, leader election consistently took longer than 10 sec-
onds in our tests due to many split votes. Adding just 5ms
of randomness helps significantly, resulting in a median
downtime of 287ms. Using more randomness improves
worst-case behavior: with 50ms of randomness the worst-
case completion time (over 1000 trials) was 513ms.
The bottom graph in Figure 16 shows that downtime
can be reduced by reducing the election timeout. With
an election timeout of 12–24ms, it takes only 35ms on
average to elect a leader (the longest trial took 152ms).
However, lowering the timeouts beyond this point violates
Raft’s timing requirement: leaders have difficulty broad-
casting heartbeats before other servers start new elections.
This can cause unnecessary leader changes and lower
overall system availability. We recommend using a con-
servative election timeout such as 150–300ms; such time-
outs are unlikely to cause unnecessary leader changes and
will still provide good availability.
10 Related work
There have been numerous publications related to con-
sensus algorithms, many of which fall into one of the fol-
lowing categories:
• Lamport’s original description of Paxos [15], and at-
tempts to explain it more clearly [16, 20, 21].
• Elaborations of Paxos, which fill in missing details
and modify the algorithm to provide a better founda-
tion for implementation [26, 39, 13].
• Systems that implement consensus algorithms, such
as Chubby [2, 4], ZooKeeper [11, 12], and Span-
ner [6]. The algorithms for Chubby and Spanner
have not been published in detail, though both claim
to be based on Paxos. ZooKeeper’s algorithm has
been published in more detail, but it is quite different
from Paxos.
• Performance optimizations that can be applied to
Paxos [18, 19, 3, 25, 1, 27].
• Oki and Liskov’s Viewstamped Replication (VR), an
alternative approach to consensus developed around
the same time as Paxos. The original description [29]
was intertwined with a protocol for distributed trans-
actions, but the core consensus protocol has been
separated in a recent update [22]. VR uses a leader-
based approach with many similarities to Raft.
The greatest difference between Raft and Paxos is
Raft’s strong leadership: Raft uses leader election as an
essential part of the consensus protocol, and it concen-
15

trates as much functionality as possible in the leader. This
approach results in a simpler algorithm that is easier to
understand. For example, in Paxos, leader election is or-
thogonal to the basic consensus protocol: it serves only as
a performance optimization and is not required for achiev-
ing consensus. However, this results in additional mecha-
nism: Paxos includes both a two-phase protocol for basic
consensus and a separate mechanism for leader election.
In contrast, Raft incorporates leader election directly in to
the consensus algorithm and uses it as the first of the two
phases of consensus. This results in less mechanism than
in Paxos.
Like Raft, VR and ZooKeeper are leader-based and
therefore share many of Raft’s advantages over Paxos.
However, Raft has less mechanism that VR or ZooKeeper
because it minimizes the functionality in non-leaders. For
example, log entries in Raft flow in only one direction:
outward from the leader in AppendEntries RPCs. In VR
log entries flow in both directions (leaders can receive
log entries during the election process); this results in
additional mechanism and complexity. The published de-
scription of ZooKeeper also transfers log entries both to
and from the leader, but the implementation is apparently
more like Raft [35].
Raft has fewer message types than any other algo-
rithm for consensus-based log replication that we are
aware of. For example, we counted the message types VR
and ZooKeeper use for basic consensus and membership
changes (excluding log compaction and client interaction,
as these are nearly independent of the algorithms). VR
and ZooKeeper each define 10 different message types,
while Raft has only 4 message types (two RPC requests
and their responses). Raft’s messages are a bit more dense
than the other algorithms’, but they are simpler collec-
tively. In addition, VR and ZooKeeper are described in
terms of transmitting entire logs during leader changes;
additional message types will be required to optimize
these mechanisms so that they are practical.
Raft’s strong leadership approach simplifies the algo-
rithm, but it precludes some performance optimizations.
For example, Egalitarian Paxos (EPaxos) can achieve
higher performance under some conditions with a lead-
erless approach [27]. EPaxos exploits commutativity in
state machine commands. Any server can commit a com-
mand with just one round of communication as long as
other commands that are proposed concurrently commute
with it. However, if commands that are proposed con-
currently do not commute with each other, EPaxos re-
quires an additional round of communication. Because
any server may commit commands, EPaxos balances load
well between servers and is able to achieve lower latency
than Raft in W AN settings. However, it adds significant
complexity to Paxos.
Several different approaches for cluster member-
ship changes have been proposed or implemented in
other work, including Lamport’s original proposal [15],
VR [22], and SMART [24]. We chose the joint consensus
approach for Raft because it leverages the rest of the con-
sensus protocol, so that very little additional mechanism
is required for membership changes. Lamport’s
α-based
approach was not an option for Raft because it assumes
consensus can be reached without a leader. In comparison
to VR and SMART, Raft’s reconfiguration algorithm has
the advantage that membership changes can occur with-
out limiting the processing of normal requests; in con-
trast, VR stops all normal processing during configura-
tion changes, and SMART imposes an
α-like limit on the
number of outstanding requests. Raft’s approach also adds
less mechanism than either VR or SMART.
11 Conclusion
Algorithms are often designed with correctness, effi-
ciency, and/or conciseness as the primary goals. Although
these are all worthy goals, we believe that understandabil-
ity is just as important. None of the other goals can be
achieved until developers render the algorithm into a prac-
tical implementation, which will inevitably deviate from
and expand upon the published form. Unless developers
have a deep understanding of the algorithm and can cre-
ate intuitions about it, it will be difficult for them to retai n
its desirable properties in their implementation.
In this paper we addressed the issue of distributed con-
sensus, where a widely accepted but impenetrable algo-
rithm, Paxos, has challenged students and developers for
many years. We developed a new algorithm, Raft, which
we have shown to be more understandable than Paxos.
We also believe that Raft provides a better foundation
for system building. Using understandability as the pri-
mary design goal changed the way we approached the de-
sign of Raft; as the design progressed we found ourselves
reusing a few techniques repeatedly, such as decomposing
the problem and simplifying the state space. These tech-
niques not only improved the understandability of Raft
but also made it easier to convince ourselves of its cor-
rectness.
12 Acknowledgments
The user study would not have been possible with-
out the support of Ali Ghodsi, David Mazi` eres, and the
students of CS 294-91 at Berkeley and CS 240 at Stan-
ford. Scott Klemmer helped us design the user study,
and Nelson Ray advised us on statistical analysis. The
Paxos slides for the user study borrowed heavily from
a slide deck originally created by Lorenzo Alvisi. Spe-
cial thanks go to David Mazi` eres and Ezra Hoch for
finding subtle bugs in Raft. Many people provided help-
ful feedback on the paper and user study materials,
including Ed Bugnion, Michael Chan, Hugues Evrard,
16

Daniel Giffin, Arjun Gopalan, Jon Howell, Vimalkumar
Jeyakumar, Ankita Kejriwal, Aleksandar Kracun, Amit
Levy, Joel Martin, Satoshi Matsushita, Oleg Pesok, David
Ramos, Robbert van Renesse, Mendel Rosenblum, Nico-
las Schiper, Deian Stefan, Andrew Stone, Ryan Stutsman,
David Terei, Stephen Y ang, Matei Zaharia, 24 anony-
mous conference reviewers (with duplicates), and espe-
cially our shepherd Eddie Kohler. Werner V ogels tweeted
a link to an earlier draft, which gave Raft significant ex-
posure. This work was supported by the Gigascale Sys-
tems Research Center and the Multiscale Systems Cen-
ter, two of six research centers funded under the Fo-
cus Center Research Program, a Semiconductor Research
Corporation program, by STARnet, a Semiconductor Re-
search Corporation program sponsored by MARCO and
DARP A, by the National Science Foundation under Grant
No. 0963859, and by grants from Facebook, Google, Mel-
lanox, NEC, NetApp, SAP , and Samsung. Diego Ongaro
is supported by The Junglee Corporation Stanford Gradu-
ate Fellowship.
References
[1] B OLOSKY , W. J., B RADSHAW , D., H AAGENS , R. B.,
KUSTERS , N. P., AND LI, P. Paxos replicated state
machines as the basis of a high-performance data store.
In Proc. NSDI’11, USENIX Conference on Networked
Systems Design and Implementation (2011), USENIX,
pp. 141–154.
[2] B URROWS , M. The Chubby lock service for loosely-
coupled distributed systems. In Proc. OSDI’06, Sympo-
sium on Operating Systems Design and Implementation
(2006), USENIX, pp. 335–350.
[3] C AMARGOS , L. J., S CHMIDT , R. M., AND PEDONE , F.
Multicoordinated Paxos. In Proc. PODC’07, ACM Sym-
posium on Principles of Distributed Computing (2007),
ACM, pp. 316–317.
[4] C HANDRA , T. D., G RIESEMER , R., AND REDSTONE , J.
Paxos made live: an engineering perspective. In Proc.
PODC’07, ACM Symposium on Principles of Distributed
Computing (2007), ACM, pp. 398–407.
[5] C HANG , F., D EAN , J., G HEMAWAT, S., H SIEH , W. C.,
WALLACH , D. A., B URROWS , M., C HANDRA , T.,
FIKES , A., AND GRUBER , R. E. Bigtable: a distributed
storage system for structured data. In Proc. OSDI’06,
USENIX Symposium on Operating Systems Design and
Implementation (2006), USENIX, pp. 205–218.
[6] C ORBETT , J. C., D EAN , J., E PSTEIN , M., F IKES , A.,
FROST , C., F URMAN , J. J., G HEMAWAT, S., G UBAREV ,
A., H EISER , C., H OCHSCHILD , P., H SIEH , W., K AN -
THAK , S., K OGAN , E., L I, H., L LOYD , A., M ELNIK ,
S., M WAURA , D., N AGLE , D., Q UINLAN , S., R AO, R.,
ROLIG , L., S AITO , Y., S ZYMANIAK , M., T AYLOR , C.,
WANG , R., AND WOODFORD , D. Spanner: Google’s
globally-distributed database. In Proc. OSDI’12, USENIX
Conference on Operating Systems Design and Implemen-
tation (2012), USENIX, pp. 251–264.
[7] C OUSINEAU , D., D OLIGEZ , D., L AMPORT , L., M ERZ ,
S., R ICKETTS , D., AND VANZETTO , H. TLA + proofs.
In Proc. FM’12, Symposium on F ormal Methods (2012),
D. Giannakopoulou and D. M´ ery, Eds., vol. 7436 of Lec-
ture Notes in Computer Science , Springer, pp. 147–154.
[8] G HEMAWAT, S., G OBIOFF , H., AND LEUNG , S.-T. The
Google file system. In Proc. SOSP’03, ACM Symposium
on Operating Systems Principles (2003), ACM, pp. 29–43.
[9] G RAY, C., AND CHERITON , D. Leases: An efficient fault-
tolerant mechanism for distributed file cache consistency.
In Proceedings of the 12th ACM Ssymposium on Operating
Systems Principles (1989), pp. 202–210.
[10] H ERLIHY , M. P., AND WING , J. M. Linearizability: a
correctness condition for concurrent objects. ACM Trans-
actions on Programming Languages and Systems 12 (July
1990), 463–492.
[11] H UNT, P., K ONAR , M., J UNQUEIRA , F. P., AND REED ,
B. ZooKeeper: wait-free coordination for internet-scale
systems. In Proc ATC’10, USENIX Annual Technical Con-
ference (2010), USENIX, pp. 145–158.
[12] J UNQUEIRA , F. P., R EED , B. C., AND SERAFINI , M.
Zab: High-performance broadcast for primary-backup sys-
tems. In Proc. DSN’11, IEEE/IFIP Int’l Conf. on Depend-
able Systems & Networks (2011), IEEE Computer Society,
pp. 245–256.
[13] K IRSCH , J., AND AMIR , Y. Paxos for system builders.
Tech. Rep. CNDS-2008-2, Johns Hopkins University,
2008.
[14] L AMPORT , L. Time, clocks, and the ordering of events in
a distributed system. Commununications of the ACM 21, 7
(July 1978), 558–565.
[15] L AMPORT , L. The part-time parliament. ACM Transac-
tions on Computer Systems 16 , 2 (May 1998), 133–169.
[16] L AMPORT , L. Paxos made simple. ACM SIGACT News
32, 4 (Dec. 2001), 18–25.
[17] L AMPORT , L. Specifying Systems, The TLA+ Language
and Tools for Hardware and Software Engineers. Addison-
Wesley, 2002.
[18] L AMPORT , L. Generalized consensus and Paxos. Tech.
Rep. MSR-TR-2005-33, Microsoft Research, 2005.
[19] L AMPORT , L. Fast paxos. Distributed Computing 19 , 2
(2006), 79–103.
[20] L AMPSON , B. W. How to build a highly available system
using consensus. In Distributed Algorithms, O. Baboaglu
and K. Marzullo, Eds. Springer-V erlag, 1996, pp. 1–17.
[21] L AMPSON , B. W. The ABCD’s of Paxos. In Proc.
PODC’01, ACM Symposium on Principles of Distributed
Computing (2001), ACM, pp. 13–13.
[22] L ISKOV, B., AND COWLING , J. Viewstamped replica-
tion revisited. Tech. Rep. MIT-CSAIL-TR-2012-021, MIT,
July 2012.
[23] LogCabin source code. http://github.com/
logcabin/logcabin.
17

[24] L ORCH , J. R., A DYA, A., B OLOSKY , W. J., C HAIKEN ,
R., D OUCEUR , J. R., AND HOWELL , J. The SMART
way to migrate replicated stateful services. In Proc. Eu-
roSys’06, ACM SIGOPS/EuroSys European Conference on
Computer Systems (2006), ACM, pp. 103–115.
[25] M AO, Y., J UNQUEIRA , F. P., AND MARZULLO , K.
Mencius: building efficient replicated state machines for
W ANs. In Proc. OSDI’08, USENIX Conference on
Operating Systems Design and Implementation (2008),
USENIX, pp. 369–384.
[26] M AZI `ERES , D. Paxos made practical. http:
//www.scs.stanford.edu/˜dm/home/
papers/paxos.pdf, Jan. 2007.
[27] M ORARU , I., A NDERSEN , D. G., AND KAMINSKY , M.
There is more consensus in egalitarian parliaments. In
Proc. SOSP’13, ACM Symposium on Operating System
Principles (2013), ACM.
[28] Raft user study. http://ramcloud.stanford.
edu/˜ongaro/userstudy/.
[29] O KI , B. M., AND LISKOV, B. H. Viewstamped
replication: A new primary copy method to support
highly-available distributed systems. In Proc. PODC’88,
ACM Symposium on Principles of Distributed Computing
(1988), ACM, pp. 8–17.
[30] O’N EIL , P., C HENG , E., G AWLICK , D., AND ON EIL , E.
The log-structured merge-tree (LSM-tree). Acta Informat-
ica 33, 4 (1996), 351–385.
[31] O NGARO , D. Consensus: Bridging Theory and Practice .
PhD thesis, Stanford University, 2014 (work in progress).
http://ramcloud.stanford.edu/˜ongaro/
thesis.pdf.
[32] O NGARO , D., AND OUSTERHOUT , J. In search of an
understandable consensus algorithm. In Proc ATC’14,
USENIX Annual Technical Conference (2014), USENIX.
[33] O USTERHOUT , J., A GRAWAL , P., E RICKSON , D.,
KOZYRAKIS , C., L EVERICH , J., M AZI `ERES , D., M I-
TRA , S., N ARAYANAN , A., O NGARO , D., P ARULKAR ,
G., R OSENBLUM , M., R UMBLE , S. M., S TRATMANN ,
E., AND STUTSMAN , R. The case for RAMCloud. Com-
munications of the ACM 54 (July 2011), 121–130.
[34] Raft consensus algorithm website.
http://raftconsensus.github.io.
[35] R EED , B. Personal communications, May 17, 2013.
[36] R OSENBLUM , M., AND OUSTERHOUT , J. K. The design
and implementation of a log-structured file system. ACM
Trans. Comput. Syst. 10 (February 1992), 26–52.
[37] S CHNEIDER , F. B. Implementing fault-tolerant services
using the state machine approach: a tutorial. ACM Com-
puting Surveys 22, 4 (Dec. 1990), 299–319.
[38] S HVACHKO , K., K UANG , H., R ADIA , S., AND
CHANSLER , R. The Hadoop distributed file system.
In Proc. MSST’10, Symposium on Mass Storage Sys-
tems and Technologies (2010), IEEE Computer Society,
pp. 1–10.
[39] VAN RENESSE , R. Paxos made moderately complex.
Tech. rep., Cornell University, 2012.
18

Raft FAQ

Q: What do people use Raft for?

A: The most frequent use for Raft (and Paxos) is to build
fault-tolerant "configuration services" whose job is to keep track of
how responsibilities are currently assigned to servers in a large
deployment. This job is particularly sensitive for deployments with
replication; Raft-based configuration services are often used to
select primaries in a way that avoids split brain. The VMware FT
test-and-set server is a simple example of a configuration service.
Chubby, ZooKeeper, and etcd are more powerful fault-tolerant
configuration services based on Raft or Paxos; they are widely used.

Some databases, such as Spanner, CockroachDB, and Lab 3, use Raft or
Paxos to replicate the data. (In contrast, GFS, VMware FT, and Chain
Replication use simpler primary-backup for the data.) Some databases
use Raft or Paxos in two different ways: for the configuration service
that assigns responsibilities to servers (for every shard, who is
currently primary and who are backups), and separately to handle the
data within each shard.

Q: Does Raft sacrifice anything for simplicity?

A: Raft gives up some performance in return for clarity; for example:

* Every operation must be written to disk for persistence; performance
  probably requires batching many operations into each disk write.

* There can only usefully be a single AppendEntries in flight from the
  leader to each follower: followers reject out-of-order
  AppendEntries, and the sender's nextIndex[] mechanism requires
  one-at-a-time. A provision for pipelining many AppendEntries would
  be better.

* The snapshotting design is only practical for relatively small
  states, since it writes the entire state to disk. If the state is
  big (e.g. if it's a big database), you'd want a way to write just
  parts of the state that have changed recently.

* Similarly, bringing recovering replicas up to date by sending them a
  complete snapshot will be slow, needlessly so if the replica already
  has a snapshot that's only somewhat out of date.

* Servers may not be able to take much advantage of multi-core because
  operations must be executed one at a time (in log order).

These could be fixed by modifying Raft, but the result might have less
value as a tutorial.

Q: Is Raft used in real-world software, or do companies generally roll
their own flavor of Paxos (or use a different consensus protocol)?

A: There are several real-world users of Raft: Docker
(https://docs.docker.com/engine/swarm/raft/), etcd (https://etcd.io),
and MongoDB. Other systems said to be using Raft include CockroachDB,
RethinkDB, and TiKV. Maybe you can find more starting at
http://raft.github.io/

On the other hand, many real-world state-machine replication systems
(Google's Chubby, ZooKeeper's ZAB) are derived from the older
Multi-Paxos and Viewstamped Replication protocols.

Q: What is Paxos? In what sense is Raft simpler?

A: There is a protocol called Paxos that allows a set of servers to
agree on a single value. While Paxos requires some thought to
understand, it is far simpler than Raft. Here's an easy-to-read paper
about Paxos:

  http://css.csail.mit.edu/6.824/2014/papers/paxos-simple.pdf

However, Paxos solves a smaller problem than Raft. To build a
real-world replicated service, the replicas need to agree on an
indefinite sequence of values (the client commands), and they need
ways to efficiently recover when servers crash and restart or miss
messages. People have built such systems with Paxos as the starting
point; look up Google's Chubby and Paxos Made Live papers, and
ZooKeeper/ZAB. There is also a protocol called Viewstamped
Replication; it's a good design, and similar to Raft, but the paper
about it is hard to understand.

These real-world protocols are complex, and (before Raft) there was
not a good introductory paper describing how they work. The Raft
paper, in contrast, is relatively easy to read and fairly detailed.
That's a big contribution.

Whether the Raft protocol is inherently easier to understand than
something else is not clear. The issue is clouded by a lack of good
descriptions of other real-world protocols. In addition, Raft
sacrifices performance for clarity in a number of ways; that's fine
for a tutorial but not always desirable in a real-world protocol.

Q: How long had Paxos existed before the authors created Raft?

A: Paxos was invented in the late 1980s. Raft was developed around
2012.

Raft closely resembles a protocol called Viewstamped Replication,
originally published in 1988. There were replicated fault-tolerant file
servers built on top of Viewstamped Replication in the early 1990s,
though not in production use.

A bunch of real-world systems are derived from Paxos: Chubby, Spanner,
Megastore, and Zookeeper/ZAB. Starting in the early 2000s big web
sites and cloud providers needed fault-tolerant services, and Paxos
was dusted off at that time and put into production.

Q: How does Raft's performance compare to Paxos in real-world applications?

A: The fastest Paxos-derived protocols are probably faster than
Raft as described in the paper; have a look at ZAB/ZooKeeper and Paxos
Made Live. On the other hand, etcd3 (using Raft) claims to have
achieved better performance than zookeeper and many Paxos-based
implementations (https://www.youtube.com/watch?v=hQigKX0MxPw).

There are situations where Raft's leader is not so great. If the
datacenters containing replicas and clients are distant from each
other, people sometimes use agreement protocols derived from original
Paxos. The reason is that Paxos has no leader; any replica can start
an agreement; so clients can talk to the replica in their local
datacenter rather than having to talk to a leader in a distant
datacenter. ePaxos is an example.

Q: Why are we learning/implementing Raft instead of Paxos?

A: We're using Raft in 6.824 because there is a paper that clearly
describes how to build a complete replicated service using Raft. I
know of no satisfactory paper that describes how to build a complete
replicated server system based on Paxos.

Q: Are there systems like Raft that can survive and continue to
operate when only a minority of the cluster is active?

A: Not with Raft's properties. But you can do it with different
assumptions, or different client-visible semantics. The basic problem
is split-brain -- the possibility of multiple diverging copies of the
state, caused by multiple subsets of the replicas mutating the state
without being aware of each other. There are two solution approaches
that I know of.

If somehow clients and servers can learn exactly which servers are
live and which are dead (as opposed to live but unreachable due to
network failure), then one can build a system that can function as
long as one is alive, picking (say) the lowest-numbered server known
to be alive. However, it's usually impractical for one computer to
decide if another computer is dead, as opposed to the network losing
the messages between them. One way to do it is to have a human decide:
the human can inspect the servers and decide which are alive and dead.

The other approach is to allow split-brain operation, and to have a
way for servers to reconcile the resulting diverging state after
partitions are healed. This can be made to work for some kinds of
services, but has complex client-visible semantics (usually called
"eventual consistency"). Have a look at the COPS, FuzzyLog, and
Bitcoin papers which are assigned later in the course.

Q: In Raft, the service which is being replicated is not available to
the clients during an election process. In practice how much of a
problem does this cause?

A: The client-visible pause seems likely to be on the order of a tenth of a
second. The authors expect failures (and thus elections) to be rare,
since they only happen if machines or the network fails. Many servers
and networks stay up continuously for months or even years at a time, so
this doesn't seem like a huge problem for many applications.

Q: Are there other consensus systems that don't have leader-election
pauses?

A: There are versions of Paxos-based replication that do not have a leader
or elections, and thus don't suffer from pauses during elections.
Instead, any server can effectively act as leader at any time. The cost
of not having a leader is that more messages are required for each
agreement.

Q: How are Raft and VMware FT related?

A: Raft has no single point of failure, while VMware FT does have a
single point of failure in the form of the test-and-set server. In
that sense Raft is fundamentally more fault-tolerant than VMware FT.
One could fix this by implementing FT's test-and-set server as a
replicated service using Raft or Paxos.

VMware-FT can replicate any virtual machine guest, and thus any
server-style software, even software that has no idea that it is being
replicated. Raft is used as a library integrated into the application
software, which is usually designed specifically to work well with
replication.

Q: Why can't a malicious person take over a Raft server, or forge
incorrect Raft messages?

A: Raft doesn't include defenses against attacks like this. It assumes
that all participants are following the protocol, and that only the
correct set of servers is participating.

A real deployment would have to keep out malicious attackers. The most
straightforward option is to place the servers behind a firewall to
filter out packets from random people on the Internet, and to ensure
that all computers and people inside the firewall are trustworthy.

There may be situations where Raft has to operate on the same network as
potential attackers. In that case a good plan would be to authenticate
the Raft packets with some cryptographic scheme. For example, give each
legitimate Raft server a public/private key pair, have it sign all the
packets it sends, give each server a list of the public keys of
legitimate Raft servers, and have the servers ignore packets that aren't
signed by a key on that list.

Q: The paper mentions that Raft works under all non-Byzantine
conditions. What are Byzantine conditions and why could they make Raft
fail?

A: "Non-Byzantine conditions" means that the servers are fail-stop:
they either follow the Raft protocol correctly, or they halt. For
example, most power failures are non-Byzantine because they cause
computers to simply stop executing instructions; if a power failure
occurs, Raft may stop operating, but it won't send incorrect results
to clients.

Byzantine failure refers to situations in which some computers execute
incorrectly, because of bugs or because someone malicious is
controlling the computers. If a failure like this occurs, Raft may
send incorrect results to clients.

Most of 6.824 is about tolerating non-Byzantine faults. Correct
operation despite Byzantine faults is more difficult; we'll touch on
this topic at the end of the term.

Q: Is the assumption that a Raft cluster is provisioned in the same
physical location or can you deploy peers in
geographically-distributed data centers?

A: The typical deployment is a single data center.  We will see later
some systems that run Paxos across data centers (e.g., Spanner), which
is better done with a leaderless design so that a client can talk to a
local peer (instead of the potentially faraway leader).

Google's Chubby paper reports that their Chubby deployments are
typically a single data center, except for the root Chubby (which
spans geographically-separated data centers).  (Chubby isn't based on
Raft but uses Google's replicated-state machine library, which is
based on Paxos.)

Q: Are there variations of the Raft concensus algorithm that don't
require strict ordering of the operations?  (that is, they don't have
to follow the Leader Completeness Property.)

A: Yes, if you know whether operations commute.  Google "generalized
paxos" or "exploiting commutativity for practical fast replication".

Q: In Figure 1, what does the interface between client and
server look like?

A: Typically an RPC interface to the server. For a key/value storage
server such as you'll build in Lab 3, it's Put(key,value) and
Get(value) RPCs. The RPCs are handled by a key/value module in the
server, which calls Raft.Start() to ask Raft to put a client RPC in
the log, and reads the applyCh to learn of newly committed log
entries.

Q: What if a client sends a request to a leader, but the leader
crashes before sending the client request to all followers, and the
new leader doesn't have the request in its log? Won't that cause the
client request to be lost?

A: Yes, the request may be lost. If a log entry isn't committed, Raft
may not preserve it across a leader change.

That's OK because the client could not have received a reply to its
request if Raft didn't commit the request. The client will know (by
seeing a timeout or leader change) that its request may have been
lost, and will re-send it.

The fact that clients can re-send requests means that the system has
to be on its guard against duplicate requests; you'll deal with this
in Lab 3.

Q: If there's a network partition, can Raft end up with two leaders
and split brain?

A: No. There can be at most one active leader.

A new leader can only be elected if it can contact a majority of servers
(including itself) with RequestVote RPCs. So if there's a partition, and
one of the partitions contains a majority of the servers, that one
partition can elect a new leader. Other partitions must have only a
minority, so they cannot elect a leader. If there is no majority
partition, there will be no leader (until someone repairs the network).

Q: Suppose a new leader is elected while the network is partitioned,
but the old leader is in a different partition. How will the old
leader know to stop committing new entries?

A: The old leader will either not be able to get a majority of
successful responses to its AppendEntries RPCs (if it's in a minority
partition), or if it can talk to a majority, that majority must
overlap with the new leader's majority, and the servers in the overlap
will tell the old leader that there's a higher term. That will cause
the old leader to switch to follower.

Q: When some servers have failed, does "majority" refer to a majority
of the live servers, or a majority of all servers (even the dead
ones)?

A: Always a majority of all servers. So if there are 5 Raft peers in
total, but two have failed, a candidate must still get 3 votes
(including itself) in order to be elected leader.

There are many reasons for this. It could be that the two "failed"
servers are actually up and running in a different partition. From
their point of view, there are three failed servers. If they were
allowed to elect a leader using just two votes (from just the two
live-looking servers), we would get split brain. Another reason is
that we need the majorities of any two leader to overlap at at least
one server, to guarantee that a new leader sees the previous term
number and any log entries committed in previous terms; this requires
a majority out of all servers, dead and alive.

Q: What if the election timeout is too short? Will that cause Raft to
malfunction?

A: A bad choice of election timeout does not affect safety, it only
affects liveness.

If the election timeout is too small, then followers may repeatedly
time out before the leader has a chance to send out any AppendEntries.
In that case Raft may spend all its time electing new leaders, and no
time processing client requests. If the election timeout is too large,
then there will be a needlessly large pause after a leader failure
before a new leader is elected.

Q: Why randomize election timeouts?

A: To reduce the chance that multiple peers simultaneously become
candidates and divide the votes among themselves so that no-one gets a
majority.

Q: Can a candidate declare itself the leader as soon as it receives
votes from a majority, and not bother waiting for further RequestVote
replies?

A: Yes -- a majority is sufficient. It would be a mistake to wait
longer, because some peers might have failed and thus not ever reply.

Q: What network does Raft assume?

A: The network is unreliable: it may lose requests and replies and
delay them.  Raft's RPC library doesn't provide reliability; it is
best effort (e.g,. it sends a request but the network may drop
it). The lab's RPC library provides similar semantics: it may lose
requests, lose replies, delay messages, and entirely disconnect
particular hosts.

Q: What is the purpose of the votedFor check in the requestVote RPC?

A: Two candidates may start an election at the same time for the same
new term. A follower should vote only for one of them.

Q: Can a leader ever stop being a leader except by crashing?

A: Yes. If a leader's CPU is slow, or its network connection breaks,
or loses too many packets, or delivers packets too slowly, the other
servers won't see its AppendEntries RPCs, and will start an election.

Q: When are followers' log entries sent to their state machines?

A: Only after the leader says that an entry is committed, using the
leaderCommit field of the AppendEntries RPC. At that point the
follower can execute (or apply) the log entry, which for us means send
it on the applyCh.

Q: Should the leader wait for replies to AppendEntries RPCs?

A: The leader should send the AppendEntries RPCs concurrently, without
waiting. As replies come back, the leader should count them, and mark
the log entry as committed only when it has replies from a majority of
servers (including itself).

One way to do this in Go is for the leader to send each AppendEntries
RPC in a separate goroutine, so that the leader sends the RPCs
concurrently. Something like this:

  for each server {
    go func() {
      send the AppendEntries RPC and wait for the reply
      if reply.success == true {
        increment count
        if count == nservers/2 + 1 {
          this entry is committed
        }
      }
    } ()
  }

Q: What happens if a half (or more) of the servers die?

A: The service can't make any progress; it will keep trying to elect a
leader over and over. If/when enough servers come back to life with
persistent Raft state intact, they will be able to elect a leader and
continue.

Q: Why is the Raft log 1-indexed?

A: You should view it as zero-indexed, but starting out with an entry
(at index=0) that has term 0. That allows the very first AppendEntries
RPC to contain 0 as PrevLogIndex, and be a valid index into the log.

Q: When the network partitions, won't client requests in minority
partitions be lost?

A: Only the partition with a majority of servers can commit and
execute client operations. The servers in the minority partition(s)
won't be able to commit client operations, so they won't reply to
client requests. Clients will keep re-sending the requests until they
can contact a majority Raft partition, so these clients' requests
won't be lost forever.

Q: Is the argument in 5.4.3 a complete proof?

A: 5.4.3 is not a complete proof. Here are some places to look:

http://ramcloud.stanford.edu/~ongaro/thesis.pdf
http://verdi.uwplse.org/raft-proof.pdf

Q: Are there any limitations to what applications can be built on top of Raft?

A: I think that in order to fit cleanly into a replicated state
machine framework like Raft, the replicated service has to be
self-contained -- it can have private state, and accept commands from
clients that update the state, but it can't contact outside entities
without special precautions. If the replicated application interacts
with the outside world, the outside world has to be able to deal
correctly with repeated requests (due to replication and replay of
logs after reboot), and it has to never contradict itself (i.e. it has
to be careful to send exactly the same answer to all replicas, and to
all re-executions of log entries). That in turn seems to require that
any outside entity that a Raft-based application contacts must itself
be fault-tolerant, i.e. probably has to use Raft or something like it.
That's fairly limiting.

As an example, imagine a replicated online ordering system sending
credit card charging requests to some external credit card processor
service. That external processor will see repeated requests (one or more
from each replica). Will it respond exactly the same way to each
request? Will it charge the credit card more than once? If it does do
the right thing, will it still do the right thing if it crashes and
reboots at an awkward time?

Q: What is the copy-on-write optimization in section 7?

A: The basic idea is for the server to fork() when the service wants
to make a snapshot, giving the child a complete copy of the in-memory
state. If fork() really copied all the memory, and the state was
large, this would be slow. But most operating systems don't copy all
the memory in fork(); instead they mark the memory pages as
"copy-on-write", and make them read-only in both parent and
child. Then the operating system will see a page fault the server
tries to write a page, and the operating system will only copy the
page at that point.  The net effect is usually that the child sees a
copy of its parent process' memory at the time of the fork(), but with
relatively little copying.  (This optimization is not necessary for
your labs.)

Q: Why is it called Raft?

A: https://groups.google.com/g/raft-dev/c/95rZqptGpmU

Q: Is it important for Raft/Paxos to work correctly?

A: Raft/Paxos are often the foundation of a distributed system. For
example, many of Google's services (GFS, Spanner, BigTable, etc.) rely
on Chubby, a configuration service based on Paxos.  As another
example, many compananies use Kubernetes to manage their containers;
Kubernetes in turn stores the configuration information about the
containers in Etcd, a key/value service based on Raft.  Occasionally
issues with systems based on Paxos/Raft materialize and lead to major
outages; see, for example,
https://decentralizedthoughts.github.io/2020-12-12-raft-liveness-full-omission/