Efficient Synchronization of State-based CRDTs - Vitor Enes

Efficient Synchronization of State-based CRDTs

Vitor Enes Carlos Baquero Jo˜ ao Leit˜ ao Tuesday 3rd October, 2017

(Universidade do Minho) (Universidade Nova de Lisboa)

Outline

1. Data Replication

2. Conflict-free Replicated Data Types - State-based CRDTs - Efficient Synchronization - Results

1

Data Replication

Data Replication

• Improves availability • available even when some replicas are unreachable

• Improves performance • reduced latency : geo-distributed replicas give us low-latency links • increased throughput: multiple replicas serving data

2

Data Replication

• Improves availability • available even when some replicas are unreachable

• Improves performance • reduced latency : geo-distributed replicas give us low-latency links • increased throughput: multiple replicas serving data

• Several consistency models, e.g.: • Strong Consistency: • illusion of a single-copy by synchronizing replicas on each update • very successful within the data center but • wide-area networks increase latency of requests and decrease system’s throughput

• Eventual Consistency: • synchronize replicas in the background • replicas may diverge

2

Conflict-free Replicated Data Types

3

Operation-based A

B

{}

{}

add

x

/ {x} add x  {x}

4

Operation-based A

B

{}

{}

add

x

/ {x}

add

y

add x  {x}

/ {x, y } add

y

{x, y }

4

Operation-based A

B

{}

{}

add

x

/ {x}

add

y

add x  {x}

/ {x, y }

{x, y , z} ; add

y

{x, y }

add add

z

z

/ {x, y , z}

4

Operation-based A

B

{}

add

x

/ {x}

add

y

add x  {x}

{}

/ {x, y }

{x, y , z} ; add

y

{x, y }

add add

z

z

/ {x, y , z}

State-based A

{}

add

x

/ {x} {x}

B

{}

 {x}

4

Operation-based A

B

{}

add

x

/ {x}

add

y

/ {x, y } add

add x  {x}

{}

{x, y , z} ; y

{x, y }

add add

z

z

/ {x, y , z}

State-based A

B

{}

{}

add

x

/ {x}

add

y

/ {x, y }

{x}

{x,y }

 {x}

{x, y }

4

Operation-based A

B

{}

add

x

/ {x}

add

y

/ {x, y } add

add x  {x}

{}

{x, y , z} ; y

{x, y }

add add

z

z

/ {x, y , z}

State-based A

{}

add

x

/ {x}

add

y

{x}

B

{}

 {x}

/ {x, y }

{x, y , z} ; {x,y }

{x, y }

{x,y ,z}

add

z

/ {x, y , z}

4

Operation-based A

B

{}

add

x

/ {x}

add

y

/ {x, y } add

add x  {x}

{}

{x, y , z} ; y

{x, y }

add add

z

z

/ {x, y , z}

State-based A

{}

add

x

/ {x}

add

y

{x}

B

{}

 {x}

/ {x, y }

{x, y , z} ; {x,y }

{x, y }

{x,y ,z}

add

z

/ {x, y , z}

$ $ full state transmission $ $

4

Efficient Synchronization of State-based CRDTs How can we avoid replica A sending the full state to replica B?

5

Efficient Synchronization of State-based CRDTs How can we avoid replica A sending the full state to replica B? Two fundamental problems 1. A knows something about the state of B: • Goal: A to compute a delta (δ) of minimum size to be sent to B

5

Efficient Synchronization of State-based CRDTs How can we avoid replica A sending the full state to replica B? Two fundamental problems 1. A knows something about the state of B: • Goal: A to compute a delta (δ) of minimum size to be sent to B • Techniques: • avoid sending acknowledged entries in δ-buffer (Classic Delta-based CRDTs) • ... • ...

5

Efficient Synchronization of State-based CRDTs How can we avoid replica A sending the full state to replica B? Two fundamental problems 1. A knows something about the state of B: • Goal: A to compute a delta (δ) of minimum size to be sent to B • Techniques: • avoid sending acknowledged entries in δ-buffer (Classic Delta-based CRDTs) • ... • ...

2. A knows nothing: • Goal: design a protocol that minimizes state transmission

5

Efficient Synchronization of State-based CRDTs How can we avoid replica A sending the full state to replica B? Two fundamental problems 1. A knows something about the state of B: • Goal: A to compute a delta (δ) of minimum size to be sent to B • Techniques: • avoid sending acknowledged entries in δ-buffer (Classic Delta-based CRDTs) • ... • ...

2. A knows nothing: • Goal: design a protocol that minimizes state transmission • Techniques: • ... • ... 5

Solving the first problem

A

{}

add a / {a}

B

{}

add b / {b}

6

Solving the first problem

A

{}

add a / {a} {a}

B

{}

add b / {b}

 {a, b}

6

Solving the first problem

A

{}

add a / {a} {a}

B

add b / {} {b}

 {a, b}

{a, b} ? {a,b}

6

Solving the first problem

A

{}

add a / {a} {a}

B

{}

add b / {b}

 {a, b}

{a, b} ? {a,b}

add x / {a, b, x} {a,b,x}

" {a, b, x}

6

Solving the first problem

A

{}

add a / {a} {a}

B

{}

add b / {b}

 {a, b}

{a, b} ?

add x / {a, b, x} {a,b,x}

" {a, b, x}

{a,b}

Avoid back-propagation of δ-groups A

{}

add a / {a} {a}

B

{}

add b / {b}

 {a, b}

{a, b} ? {a,b}

add x / {a, b, x} {a, b,x}

" {a, b, x} 6

Efficient Synchronization of State-based CRDTs How can we avoid replica A sending the full state to replica B? Two fundamental problems 1. A knows something about the state of B: • Goal: A to compute a delta (δ) of minimum size to be sent to B • Techniques: • avoid sending acknowledged entries in δ-buffer (Classic Delta-based CRDTs) • avoid back-propagation of δ-groups (BP) • ...

2. A knows nothing: • Goal: design a protocol that minimizes state transmission • Techniques: • ... • ... 7

Solving the first problem

A

{}

add a / {a}

B

{}

add b / {b}

C

{}

add c / {c}

8

Solving the first problem

A

{}

add a / {a}

B

{}

add b / {b}

C

add c / {} {c}

{a}

&

{a, b} {a}

 {a, c}

8

Solving the first problem

A

{}

add a / {a}

B

{}

add b / {b}

{a}

&

{a, b} {a}

C

{}

add c / {c}

 {a, c}

{a, b, c} 9 {a,c}

8

Solving the first problem

A

{}

add a / {a}

B

{}

add b / {b}

C

add c / {} {c}

{a}

&

{a, b, c} 8 {a, b} {a}

 {a, c}

{a, b, c} 9

{a,b,c}

{a,c}

8

Solving the first problem

A

{}

add a / {a}

B

{}

add b / {b}

C

add c / {} {c}

{a}

&

{a, b, c} 8 {a, b} {a}

 {a, c}

{a, b, c} 9

{a,b,c}

{a,c}

Remove redundant state in δ-groups A

{}

add a / {a}

B

{}

add b / {b}

C

add c / {} {c}

{a}

&

{a, b, c} 8 {a, b} {a}

 {a, c}

{a, b, c} 9

{a,b,c}

{a,c}

8

Efficient Synchronization of State-based CRDTs How can we avoid replica A sending the full state to replica B? Two fundamental problems 1. A knows something about the state of B: • Goal: A to compute a delta (δ) of minimum size to be sent to B • Techniques: • avoid sending acknowledged entries in δ-buffer (Classic Delta-based CRDTs) • avoid back-propagation of δ-groups (BP) • remove redundant state in δ-groups (RR)

2. A knows nothing: • Goal: design a protocol that minimizes state transmission • Techniques: • ... • ... 9

10

Solving the second problem State-driven synchronization A

{a}

B

{a}

add

add

/ {a, x, y }

x,y

z

/ {a, z}

11

Solving the second problem State-driven synchronization A

{a}

add

/ {a, x, y }

x,y

{a, x, y , z} : {a,z}

B

{a}

add

z

/ {a, z}

11

Solving the second problem State-driven synchronization A

{a}

add

/ {a, x, y }

x,y

{a, x, y , z} : {a,z}

B

{a}

add

z

/ {a, z}

{x,y }

$ {a, x, y , z}

11

Solving the second problem State-driven synchronization A

{a}

add

/ {a, x, y }

x,y

{a, x, y , z} : {a,z}

B

{a}

add

z

/ {a, z}

{x,y }

$ {a, x, y , z}

Digest-driven synchronization A

{a}

B

{a}

add

add

/ {a, x, y }

x,y

z

/ {a, z}

11

Solving the second problem State-driven synchronization A

{a}

add

/ {a, x, y }

x,y

{a, x, y , z} : {a,z}

B

{a}

add

z

/ {a, z}

{x,y }

$ {a, x, y , z}

Digest-driven synchronization A

{a}

add

/ {a, x, y }

x,y

{a, x, y } < dB

B

{a}

add

z

/ {a, z}

11

Solving the second problem State-driven synchronization A

{a}

add

/ {a, x, y }

x,y

{a, x, y , z} : {a,z}

B

{a}

add

z

/ {a, z}

{x,y }

$ {a, x, y , z}

Digest-driven synchronization A

{a}

add

/ {a, x, y }

x,y

{a, x, y } < dB

B

{a}

add

z

/ {a, z}

(dA ,{x,y })

$ {a, x, y , z}

11

Solving the second problem State-driven synchronization A

{a}

add

/ {a, x, y }

x,y

{a, x, y , z} : {a,z}

B

{a}

add

z

/ {a, z}

{x,y }

$ {a, x, y , z}

Digest-driven synchronization A

{a}

add

/ {a, x, y }

x,y

{a, x, y } < dB

B

{a}

add

z

/ {a, z}

{a, x, y , z} : (dA ,{x,y })

$ {a, x, y , z}

{z}

11

Efficient Synchronization of State-based CRDTs How can we avoid replica A sending the full state to replica B? Two fundamental problems 1. A knows something about the state of B: • Goal: A to compute a delta (δ) of minimum size to be sent to B • Techniques: • avoid sending acknowledged entries in δ-buffer (Classic Delta-based CRDTs) • avoid back-propagation of δ-groups (BP) • remove redundant state in δ-groups (RR)

2. A knows nothing: • Goal: design a protocol that minimizes state transmission • Techniques: • state-driven synchronization • digest-driven synchronization 12

13

bit.ly/msc-presentation

13

Efficient Synchronization of State-based CRDTs

Vitor Enes Carlos Baquero Jo˜ ao Leit˜ ao Tuesday 3rd October, 2017

(Universidade do Minho) (Universidade Nova de Lisboa)