Slide 1

Slide 1 text

SWIM Scalable Weakly Consistent Infection Style Process Group Membership Protocol

Slide 2

Slide 2 text

Paul Hinze phinze

Slide 3

Slide 3 text

Paul Hinze phinze death stare

Slide 4

Slide 4 text

Armon Dadgar armon creator of Serf and Consul

Slide 5

Slide 5 text

ma

Slide 6

Slide 6 text

No content

Slide 7

Slide 7 text

No content

Slide 8

Slide 8 text

Process Group Membership Protocol Who is alive

Slide 9

Slide 9 text

No content

Slide 10

Slide 10 text

Process Group Membership Protocol

Slide 11

Slide 11 text

SWIM Scalable Weakly Consistent Infection Style Process Group Membership Protocol

Slide 12

Slide 12 text

Scalable The SWIM effort is motivated by the unscalability of traditional heartbeating protocols.

Slide 13

Slide 13 text

Heartbeating A B C A A B B Failure Detection + Membership

Slide 14

Slide 14 text

Heartbeating

Slide 15

Slide 15 text

Evaluating Protocols Completeness Speed Accuracy Overhead

Slide 16

Slide 16 text

Evaluating Protocols Completeness Speed Accuracy Overhead heartbeating Yes Limit * Interval High Nodes2 !

Slide 17

Slide 17 text

Key Insight Failure Detection State Updates. Solve separately from

Slide 18

Slide 18 text

Failure Detection ping! ack! A B C D {B,C,D}

Slide 19

Slide 19 text

Failure Detection ping! ack! A B C D {B,C,D}

Slide 20

Slide 20 text

Indirect Ping ping(C)! ack! B C D ping(C)! ping ack fail {B,C,D}

Slide 21

Slide 21 text

Indirect Ping ping(C)! B C D ping(C)! fail ...C is dead! fail

Slide 22

Slide 22 text

Key Insight Failure Detection State Updates. Solve separately from

Slide 23

Slide 23 text

Key Insight State Updates Failure Detection Piggyback onto messages.

Slide 24

Slide 24 text

State Updates ping! (B is dead) ack! (D just joined) A B C D {B,C} {D}

Slide 25

Slide 25 text

Infection Style A B C D {B,C} {A,B,D} {A,B} weakly consistent

Slide 26

Slide 26 text

Evaluating Protocols Completeness Speed Accuracy Overhead SWIM Yes, eventually 1 * Interval High-ish O(N)

Slide 27

Slide 27 text

Improvements Time Bounded Completeness Increased Accuracy

Slide 28

Slide 28 text

Completeness A B {B, C, D, ..., N} N

Slide 29

Slide 29 text

Completeness A B {B, C, D, ..., N} N 1. Shuffle List 2. Iterate

Slide 30

Slide 30 text

Completeness Fixed Time

Slide 31

Slide 31 text

Evaluating Protocols Completeness Speed Accuracy Overhead SWIM Yes, fixed time 1 * Interval High-ish O(N)

Slide 32

Slide 32 text

Accuracy B C D ...C is MAYBE dead! A

Slide 33

Slide 33 text

Accuracy B C D I heard C might be dead. A

Slide 34

Slide 34 text

Accuracy B C D I'm not dead! A

Slide 35

Slide 35 text

Accuracy

Slide 36

Slide 36 text

Evaluating Protocols Completeness Speed Accuracy Overhead SWIM Yes, fixed time 1 * Interval High O(N)

Slide 37

Slide 37 text

Evaluating Protocols Completeness Speed Accuracy Overhead SWIM Yes, fixed time 1 * Interval High O(N)

Slide 38

Slide 38 text

No content

Slide 39

Slide 39 text

Limitations Update Latency Problem Solution Separate Gossip Timer

Slide 40

Slide 40 text

Limitations Cannot Handle Network Partitions Problem Solution Track and Retry Recently Dead Nodes

Slide 41

Slide 41 text

Limitations No Concept of Graceful Leave (vs Failure) Problem Solution Broadcast and Tracking of "Intents"

Slide 42

Slide 42 text

Limitations New Nodes Take Too Long Materialize Initial State Problem Solution Anti-entropy TCP State Syncs

Slide 43

Slide 43 text

Limitations No built-in facility for user data Problem Solution Implement user payloads (ordering via lamport clocks)

Slide 44

Slide 44 text

Limitations No peer metadata, only IP Addresses Problem Solution Inject versioned peer metadata into state messages

Slide 45

Slide 45 text

Limitations No encryption Problem Solution Implement AES-GCM with key rotation

Slide 46

Slide 46 text

56 nodes 2K nodes Performance

Slide 47

Slide 47 text

Implementations Memberlist
 https://github.com/hashicorp/memberlist Serf lib
 https://github.com/hashicorp/serf Serf
 http://www.serfdom.io Consul
 http://www.consul.io

Slide 48

Slide 48 text

Implementations events, queries, and scripts serf CLI serf lib memberlist

Slide 49

Slide 49 text

Implementations service discovery, K/V, health checks consul serf lib memberlist

Slide 50

Slide 50 text

Implementations service discovery, K/V, health checks

Slide 51

Slide 51 text

Thanks Cloud icons by Julien Deveaux from the Noun Project