operational transformation in real-time group editors: issues, algorithms, and achievements

copyright 1998 Chengzheng Sun 2

Outline

Basic concepts, issues, and challenges

Alternative approaches and algorithms– GROVE and dOPT algorithm

– Jupiter algorithm

– adOPTed algorithm

– REDUCE and GOT algorithm

– GOTO algorithm

Major achievements and future directions


What is a real-time CE system ?

Real-time CE (Cooperative Editing) systems allow multiple geographically dispersed users to view and edit a shared document at the same time.

Major technological components:

– Distributed computing Computer networks/Internet

– Interactive computing: Human-Computer interaction

– Collaborative computing Human-Human interaction


One significant challenge

Consistency maintenance under the following constraints:

High responsiveness:– the response to local operations

is as short as a single-user editor

High concurrency:– multiple users are allowed to edit

any parts of the document at any time.

High communication latency:– the Internet environment


Operation dependence relationship

Causal ordering relation “” among operations based on Lamport’s “happen before” relation.

Ob is dependent on Oa iff Oa

Ob

Oa and Ob are independent, i.e., Oa || Ob, iff neither OaOb, nor Ob Oa

O1 O3, O2 O3|| O2

O 1O 1

O 2O 2

O 3O 3


Three inconsistency problems

Divergence Final document contents are different

at all sites.

Causality-violation Execution order is different from the cause-effect order. E.g., O1 O3

Intention-violation The actual effect is different from the intended effect. E.g., || O2

A non-serializable problem

O1O1

O 2O 2

O3O3


Traditional approaches

Turn-taking: – Avoid problems by prohibiting concurrency.

Locking: – Unable to solve any of the three problems.

Serialization: – Unsuitable for resolving intention violation.

Causal ordering: – Unsuitable for resolving divergence, or intention

violation.


Operational transformation: an innovation

Basic idea: an operation is transformed before its execution against previously executed independent operations.

OT AlgorithmO

Executed independent operations

O’

Major properties:•Causally ordered execution and good responsiveness •Convergence and intention preservation


GROVE an dOPT algorithm Ellis et al. (MCC, SIGMOD’89).

Log: a linear data structure for saving executed operations.

dOPT algorithm: For 0 i < N

If O || Log[i] Then T(O, Log[i])

O1O1

O2O2

O3O3T(O2,O1)

T(O1,O2)

T(O1,O2)

Causally ordered execution based on state vector timestamping.

Transformation property: For Oa|| Ob Let O’aOa Ob), O’bOb Oa) Then Oao O’b Ob o O’a


The dOPT puzzle: O2 || (O3 O1)

O2

O1

O3

Site 3 Site 1 Site 2

O1’O1

[ O2, O3’ ]Log

dOPT

T

At site 2

At site 3 and 1

O2’O2

[ O3, O1]Log

dOPT

T T

O2 is context-equivalent to O 3

O1 is context-deferent from O 2


The Jupiter approachNichols et al (Xerox PARC, UIST95)

Server

Star-like communication topology

(0, 0)

(2, 0)

(0, 1)(1, 0)

(2, 2)

(2, 1)

(0, 2)(1, 1)

(1, 2)

serverclient

Initial state

2-dimentional state space


The dOPT puzzle in another form

O1

O3 O2

client server

O1

O3O2 || (O3 O1)


Jupiter solution to the puzzle

(0, 0)

(2, 0)

(0, 1)(1, 0)

serverclient

Initial state

2-dimentional state space

O1

O3 O2

client server

O3 O2

(2, 1)O2’’

O1’

O1 O2’

(1, 1)

O3’


The adOPTed approach Ressel et al (U. of Stuttgart, CSCW96)

Transformation Property 2:

T(T(O, Ob), Oa’) = T(T(O, Oa), Ob’)

Transformation Property 1: Oa Ob’ Ob o Oa’

O a

Ob’=T(Ob,Oa)

OO b

Oa’=T(Oa,Ob)

T(O,Ob)

T(Ob,O)

T(O,Oa)

T(Oa,O)

N-dimentional Interaction Model

S0

S1

S2

Ob’=T(Ob,Oa)

Oa’=T(Oa, Ob)S3

Oa

Ob


adOPTed solution to the puzzle

O2

O1

O3

O1

O 2

O3

O2’’=T(O2’, O1)

O1’=T(O1, O2’)

O3’=T(O3, O2)

O2’=T(O2, O3)

Path taken by site 3 and 1

Path taken by site 2

Site 3 Site 1 Site 2


The REDUCE approach Sun et al (Griffith Uni. TOCHI 5(1), 98)

•Inclusion Transformation (IT)

•Exclusion Transformation (ET):

•Reversibility of IT and ET:

•GOT using 1-dimentional HB:

Oa Oa’

Ob

+

Oa Oa’

Ob

_

Oa_Oa’Oa +

Ob

GOT(IT/ET)

O

HBO’


REDUCE solution to the puzzle

O2

O2 || (O3 O1)

O1

[ O2, O3’ ]

O1’

+ +

HB

GOT

O1

O3

O3

_

_


An optimized algorithm: GOTOSun and Ellis (CSCW98)

O1

[ O2, O3’ ]

O1’

HB

GOTO

O1

O3 O2 +

[ O3, O2’ ]

_

transpose

+


Major achievements (1)

Identification of three major inconsistency problems:– divergence,

– causality violation, and

– intention violation (non-serializable).

Definition of three consistency properties: – convergence,

– causality preservation, and

– intention preservation (most challenging to achieve).



Three types of generic OT algorithms :

– Based on a 1-dimensional history

buffer: dOPT, GOT, GOTO

– Based on a 2-dimensional state

space graph: Jupiter

– Based on an N-dimensional

interaction graph: adOPTed



Two types of application-dependent OT functions:– IT: used by all OT algorithms

– ET: used by OT algorithms based on 1-dimensional data structures

Two transformation properties: – TP1: required by all OT systems

– TP2: required by OT systems with arbitrary communication topology.


Future directions (1)

Formalization, verification, and optimization of OT algorithms.



Experimental evaluation of OT-based systems from end-users’ perspective.



Generalization and application of the OT technology to other groupware and distributed systems, e.g:– collaborative hypertext/graphics/image/multimedia editors

– collaborative CAD, CASE, and database applications

– Internet-based multi-player games

operational transformation in real-time group editors: issues, algorithms, and achievements

Technology

o2 o3 o2 o

o o b o b o o

o2 o1 o2

o1 o3 o3

o bisdependenton o

puzzle o2 o1 o

oa o o b oa

puzzleo2 o2 o3 o1 o1