Paxos Made Moderately Complex Made Moderately Simple State machine - - PowerPoint PPT Presentation

▶
paxos made moderately complex made moderately simple
SMART_READER_LITE
LIVE PREVIEW

Paxos Made Moderately Complex Made Moderately Simple State machine - - PowerPoint PPT Presentation

Feb 19, 2024 414 likes •1.12k views

Paxos Made Moderately Complex Made Moderately Simple State machine replication Reminder: want to agree on order of ops Can think of operations as a log Op1 Op2 Op3 Op4 Op5 Op6 S1 S2 Paxos? S3 Put k1 v1 Put k2 v2 Op1 Op2 Op3

slide-1
SLIDE 1

“Paxos Made Moderately Complex” Made Moderately Simple

slide-2
SLIDE 2

State machine replication

Reminder: want to agree on order of ops Can think of operations as a log

Op1 Op2 Op3 Op4 Op5 Op6

slide-3
SLIDE 3

Op1 Op2 Op3 Op4 Op5 Op6

S1 S3 S2

Put k1 v1 Put k2 v2

Paxos?

slide-4
SLIDE 4

Paxos

Paxos = Phase 1

  • Send prepare messages
  • Pick value to accept

Phase 2

  • Send accept messages
slide-5
SLIDE 5

Can we do better?

Phase 1: “leader election”

  • Deciding whose value we will use

Phase 2: “commit”

  • Leader makes sure it’s still leader, commits value

What if we split these phases?

  • Lets us do operations with one round-trip
slide-6
SLIDE 6

Roles in PMMC

Replicas (like learners)

  • Keep log of operations, state machine, configs

Leaders (like proposers)

  • Get elected, drive the consensus protocol

Acceptors (simpler than in Paxos Made Simple!)

  • “Vote” on leaders
slide-7
SLIDE 7

A note about ballots in PMMC

(leader, seqnum) pairs Isomorphic to the system we discussed earlier

1 2 3

0, 4, 8, 12, 16, … 1, 5, 9, 13, 17, … 2, 6, 10, 14, 18, … 3, 7, 11, 15, 19, …

slide-8
SLIDE 8

A note about ballots in PMMC

(leader, seqnum) pairs Isomorphic to the system we discussed earlier

1 2 3

0.0, 1.0, 2.0, 3.0, 4.0, … 0.1, 1.1, 2.1, 3.1, 4.1, … 0.2, 1.2, 2.2, 3.2, 4.2, … 0.3, 1.3, 2.3, 3.3, 4.3, …

slide-9
SLIDE 9

Paxos Made Moderately Complex Made Simple

slide-10
SLIDE 10

Paxos Made Moderately Complex Made Simple

slide-11
SLIDE 11

Acceptors

Acceptor ballot_num: 0 accepted:[]

slide-12
SLIDE 12

Acceptors

Acceptor ballot_num: _ accepted:[] p1a(0.1)

slide-13
SLIDE 13

Acceptors

Acceptor ballot_num: 0.1 accepted:[] p1a(0.1)

slide-14
SLIDE 14

Acceptors

Acceptor ballot_num: 0.1 accepted:[] p1a(0.1) p1b([])

slide-15
SLIDE 15

Acceptors

Acceptor ballot_num: 0.1 accepted:[]

slide-16
SLIDE 16

Acceptors

Acceptor ballot_num: 0.1 accepted:[] p1a(0.0)

slide-17
SLIDE 17

Acceptors

Acceptor ballot_num: 0.1 accepted:[] p1a(0.0) Nope!

slide-18
SLIDE 18

Acceptors

Acceptor ballot_num: 0.1 accepted:[]

slide-19
SLIDE 19

Acceptors

Acceptor ballot_num: 0.1 accepted:[] p2a(<0.1, 0, A>)

slide-20
SLIDE 20

Acceptors

Acceptor ballot_num: 0.1 accepted:[<0.1, 0, A>] p2a(<0.1, 0, A>)

slide-21
SLIDE 21

Acceptors

Acceptor ballot_num: 0.1 accepted:[<0.1, 0, A>] p2a(<0.1, 0, A>) OK!

slide-22
SLIDE 22

Acceptors

Acceptor ballot_num: 0.1 accepted:[<0.1, 0, A>]

slide-23
SLIDE 23

Acceptors

Acceptor ballot_num: 0.1 accepted:[<0.1, 0, A>] p2a(<0.0, 0, B>)

slide-24
SLIDE 24

Acceptors

Acceptor ballot_num: 0.1 accepted:[<0.1, 0, A>] p2a(<0.0, 0, B>) Nope!

slide-25
SLIDE 25

Acceptors

Acceptor ballot_num: 0.1 accepted:[<0.1, 0, A>]

slide-26
SLIDE 26

Acceptors

  • Ballot numbers increase
  • Only accept values from current ballot
  • Never remove ballots
  • If a value v is chosen by a majority on ballot b, then

any value accepted by any acceptor in the same slot

  • n ballot b’ > b has the same value
slide-27
SLIDE 27

Paxos Made Moderately Complex Made Simple

slide-28
SLIDE 28

Paxos Made Moderately Complex Made Simple

slide-29
SLIDE 29

Leader: Getting Elected

Leader active: false ballot_num: 0.0 proposals: []

slide-30
SLIDE 30

Leader: Getting Elected

Leader active: false ballot_num: 0.0 proposals: [] Acceptor Acceptor Acceptor p1a(0.0)

slide-31
SLIDE 31

Leader: Getting Elected

Leader active: false ballot_num: 0.0 proposals: [] Acceptor Acceptor Acceptor Nope! Nope!

slide-32
SLIDE 32

Leader: Getting Elected

Leader active: false ballot_num: 1.0 proposals: [] Acceptor Acceptor Acceptor

slide-33
SLIDE 33

Leader: Getting Elected

Leader active: false ballot_num: 1.0 proposals: [] Acceptor Acceptor Acceptor Or…

slide-34
SLIDE 34

Leader: Getting Elected

Leader active: false ballot_num: 0.0 proposals: [] Acceptor Acceptor Acceptor OK([])! OK([])!

slide-35
SLIDE 35

Leader: Getting Elected

Leader active: true ballot_num: 0.0 proposals: [] Acceptor Acceptor Acceptor

slide-36
SLIDE 36

When to run for office

When should a leader try to get elected?

  • At the beginning of time
  • When the current leader seems to have failed

Paper describes an algorithm, based on pinging the leader and timing out If you get preempted, don’t immediately try for election again!

slide-37
SLIDE 37

Paxos Made Moderately Complex Made Simple

slide-38
SLIDE 38

Paxos Made Moderately Complex Made Simple

slide-39
SLIDE 39

Leader: Handling proposals

Leader active: true ballot_num: 0.0 proposals: [] Acceptor Acceptor Acceptor Replica Op1 should be A (A = “Put k1 v1”)

slide-40
SLIDE 40

Leader: Handling proposals

Leader active: true ballot_num: 0.0 proposals: [<1, A>] Acceptor Acceptor Acceptor Replica

slide-41
SLIDE 41

Leader: Handling proposals

Leader active: true ballot_num: 0.0 proposals: [<1, A>] Acceptor Acceptor Acceptor p2a(<0.0, 1, A>) Replica

slide-42
SLIDE 42

Leader: Handling proposals

Leader active: true ballot_num: 0.0 proposals: [<1, A>] Acceptor Acceptor Acceptor Replica Nope! Nope!

slide-43
SLIDE 43

Leader: Handling proposals

Leader active: false ballot_num: 0.0 proposals: [<1, A>] Acceptor Acceptor Acceptor Replica

slide-44
SLIDE 44

Leader: Handling proposals

Leader active: false ballot_num: 0.0 proposals: [<1, A>] Acceptor Acceptor Acceptor Replica Or…

slide-45
SLIDE 45

Leader: Handling proposals

Leader active: true ballot_num: 0.0 proposals: [<1, A>] Acceptor Acceptor Acceptor Replica OK! OK!

slide-46
SLIDE 46

Leader: Handling proposals

Leader active: true ballot_num: 0.0 proposals: [<1, A>] Acceptor Acceptor Acceptor Replica Replica Replica Op1 is A

slide-47
SLIDE 47

Paxos Made Moderately Complex Made Simple

slide-48
SLIDE 48

Election revisited

Acceptor ballot_num: 2.1 accepted:[<2.1, 1, A>] Leader active: false ballot_num: 3.0 proposals: [<1, B>]

slide-49
SLIDE 49

Election revisited

Acceptor ballot_num: 2.1 accepted:[<2.1, 1, A>] Leader active: false ballot_num: 3.0 proposals: [<1, B>] p1a(3.0)

slide-50
SLIDE 50

Election revisited

Acceptor ballot_num: 3.0 accepted:[<2.1, 1, A>] Leader active: false ballot_num: 3.0 proposals: [<1, B>]

slide-51
SLIDE 51

Election revisited

Acceptor ballot_num: 3.0 accepted:[<2.1, 1, A>] Leader active: false ballot_num: 3.0 proposals: [<1, B>] OK([<2.1, 1, A>])

slide-52
SLIDE 52

Election revisited

Acceptor ballot_num: 3.0 accepted:[<2.1, 1, A>] Leader active: true ballot_num: 3.0 proposals: [<1, A>]

slide-53
SLIDE 53

Leaders

  • Only propose one value per ballot and slot
  • If a value v is chosen by a majority on ballot b, then

any value proposed by any leader in the same slot on ballot b’ > b has the same value

slide-54
SLIDE 54

Paxos Made Moderately Complex Made Simple

slide-55
SLIDE 55

Paxos Made Moderately Complex Made Simple

slide-56
SLIDE 56

Replicas

Op1 Op2 Op3 Op4 Op5 Op6 Put k1 v1 Put k2 v2

Replica

slide-57
SLIDE 57

Replicas

Op1 Op2 Op3 Op4 Op5 Op6 Put k1 v1 Put k2 v2 App k1 v1 App k2 v2

slot_out slot_in Replica

slide-58
SLIDE 58

Replicas

Op1 Op2 Op3 Op4 Op5 Op6 Put k1 v1 Put k2 v2 App k1 v1 App k2 v2

slot_out slot_in Replica Leader decision(3, “App k1 v1”)

slide-59
SLIDE 59

Replicas

Op1 Op2 Op3 Op4 Op5 Op6 Put k1 v1 Put k2 v2 App k1 v1 App k2 v2

slot_out slot_in Replica Leader

slide-60
SLIDE 60

Replicas

Op1 Op2 Op3 Op4 Op5 Op6 Put k1 v1 Put k2 v2 App k1 v1 App k2 v2

slot_out slot_in Replica Leader decision(4, “Put k3 v3”)

slide-61
SLIDE 61

Replicas

Op1 Op2 Op3 Op4 Op5 Op6 Put k1 v1 Put k2 v2 App k1 v1 Put k3 v3

slot_out slot_in Replica Leader

App k2 v2

propose(5, “App k2 v2”)

slide-62
SLIDE 62

Paxos Made Moderately Complex Made Simple

slide-63
SLIDE 63

Reconfiguration

All replicas must agree on who the leaders and acceptors are How do we do this?

slide-64
SLIDE 64

Reconfiguration

All replicas must agree on who the leaders and acceptors are How do we do this?

  • Use the log!
  • Commit a special reconfiguration command
  • New config applies after WINDOW slots
slide-65
SLIDE 65

Reconfiguration

What if we need to reconfigure now and client requests aren’t coming in?

slide-66
SLIDE 66

Reconfiguration

What if we need to reconfigure now and client requests aren’t coming in?

  • Commit no-ops until WINDOW is cleared
slide-67
SLIDE 67

Other complications

State simplifications

  • Can track much less information, esp. on replicas

Garbage collection

  • Unbounded memory growth is bad
  • Lab 3: track finished slots across all instances,

garbage collect when everyone has learned result Read-only commands

  • Can’t just read from replica (why?)
  • But, don’t need their own slot
slide-68
SLIDE 68

Questions

What should be in stable storage?

slide-69
SLIDE 69

Question

What are the costs to using Paxos? Is it practical enough?