David Lorge Parnas “A sign that the Software Engineering profession has matured will be that we lose our preoccupation with the first release and focus

David Lorge Parnas

“A sign that the Software Engineering profession has matured will be that we lose our preoccupation with the first release and focus on the long term health of our products.”

Major contributor to information hiding and modularization

Advocate of software development as an engineering discipline

Including good documentation!

Opponent of “star wars” Fellow of ACM, IEEE

Software Lifecycle

Re

q's

An

aly

sis

De

sig

n

Co

nst

ruct

ion

Test

ing

Tra

nsf

er

Ma

inte

na

nce

Textbook view:

Software Lifecycle

Re

q's

An

aly

sis

De

sig

n

Co

nst

ruct

ion

Test

ing

Tra

nsf

er

Ma

inte

na

nce

Textbook view:

More realistic view:

Re

q's

An

aly

sis

De

sig

nC

on

stru

ctio

nTe

stin

gT

ran

sfe

r

EVOLUTION, MAINTENANCE,AND ADDITIONAL RELEASES

Software Lifecycle

Re

q's

An

aly

sis

De

sig

n

Co

nst

ruct

ion

Test

ing

Tra

nsf

er

Ma

inte

na

nce

Textbook view:

More realistic view:

Re

q's

An

aly

sis

De

sig

nC

on

stru

ctio

nTe

stin

gT

ran

sfe

r

EVOLUTION, MAINTENANCE,AND ADDITIONAL RELEASES

“75% of the

effort in a

software project

is spent on

maintenance”

Meir (Manny) Lehman On team Building first

computers in Israel in 1950s Worked at IBM studying

OS/360 Professor at Imperial College

London Received Harlan Mills award Passed away 29.12.10 in

Jerusalem

“Father of software evolution”

Defined “E-type” systems that are ingrained with their environment

Lehman's 8 Laws describe how such systems evolve

Lehman's Laws

1) Continuing change (adaptation)

2) Increasing complexity (unless refactored)

3) Self regulation (of rate of change)

4) Invariant work rate (inertia)

5) Conservation of familiarity (of users and developers)

6) Continuing growth (more features)

7) Declining quality (unless maintained)

8) Feedback system (at multiple levels)

All projects

Terminating Evolutionary

• Project defined by its goals

• Requirements map to acceptance test

• Develop-deliver-maintain trajectory

• Goals are uncovered as project progresses

• Continue indefinitely as long as useful

• Multiple releases along the way

All projects


• Repeated cycles

• Each one is planned with well-defined goals

Staged model

Perpetual development

• Plan as you go• Concurrent

sub-projects• Fuzzy long-

term goals

All projects


Staged model

Perpetual development

Single delivery

Multiple releases

Continuous deployment

Possible updates

Feasibility

Architecture

Installation

threemilestones

after maintenance

Timescales and Terminology

Waterfall / unified process

Evolutionary development

Agile development

Facebook

Continuous deployment

– Once

– Months

– Weeks

– One day

– < Hour

– delivery

} Release

} Deployment

Maintenance Evolution

Requirements Feature requests

Staged Model

• Each version passes conventional phases

• E.g. full cycles of Unified Process

• New versions branch in parallel to maintenance

Rajlich & Bennett, Computer 2000

Evolution and Agile

All software projects


Agile

Perpetual

• Agile can be terminating

• Agile implies a process

Evolutionary/ perpetual may be process-less e.g. Linux Staged

Perpetual Development Lifecycle

time

size

initial

development


time

size

initial

development

release

Production use & maintenance


time

size

initial

development

release

contin

ued

development


feedback& requests


time

size

initial

development

release

release

contin

ued

development



feedback& requests

usersupgrade


time

size

initial

development

release

release

contin

ued

development

contin

ued

development



feedback& requests

feedback& requests

usersupgrade


time

size

initial

development

release

release

release

contin

ued

development

contin

ued

development



Production & maint.

feedback& requests

feedback& requests

usersupgrade

usersupgrade


time

size

initial

development

release

release

release

contin

ued

development

contin

ued

development

contin

ued

development



Production & maint.

feedback& requests

feedback& requests

usersupgrade

usersupgrade

Linux Example

“Linux is evolution,

not intelligent design”

-- Linus Torvalds

Eric Raymond:The Cathedral and the Bazaar

Planning andmanagement Evolution

Raymond’s Main Insights

• Open-source developers are “scratching a personal itch”, so they know the requirements

• Having real users leads to real feedback

• “with enough eyeballs, all bugs are shallow” (Linus’s Law)

• Users are co-developers, may come up with good ideas

• “Release early, release often” leads to the most rapid progress

Linux Old Release Model

Linux Old Release Model

• Distinction between versions

• Production is stable

• Development is current

• Distinction between releases

• Minor production release for bug/security fixes

• Minor development release every few days

• A major release is a major operation taking months

• Interval between major releases is long

• Two years of delaying new features

Linux New Release Model (2.6)

Linux New Release Model (2.6)

• Only production versions are released

• Development done separately in parallel

• New version every 2-3 months

• Merge window of 2 weeks

• ~2 months of stabilization and preparation

• Few minor releases as needed

• New merge window opened at time of release

• Maintain some long-term versions for stability

Software Growth

Lehman and Turski: A system's complexity grows with time It is harder to modify a more complex system Assuming a constant effort per release,

This leads to a diminished growth rate:

21i

ii s

Ess

3 isi

Software Growth

Godfrey and Tu: Linux (and other systems) is growing at a

super-linear rate

More Data

1994 to 2011

78x growth

(29% annual)

Best models:

Quad.-exponential

Piecewise linear

No theory

660 l/d

1385 l/d

3541 l/d

Growing Growth Rate

• Possible explanation: Positive feedback

• Successful project attracts more developers

• More developers produce more code

• Suggests exponential growth

• Contradicts Brooks’s Law (“adding manpower to a late project makes it later”)

• System is modular so don’t need so much communication

• Good information dissemination via mailing lists etc.

• New developers are typically already experts, not novices

Continuous Deployment Variant

New software released in timescales of minutes, not days or weeks

Each developer immediately deploys whatever he works on Based on passing a battery of automatic tests

(unit + integration + regression) Requires strong framework to control releases

and roll them back if needed Makes the whole notion of a “version”

meaningless

Continuous Deployment Variant

Popular mainly in web-based companies and applications Deploy to web site No need to download and install Immediately used by all

Can be done at different granularities Facebook deploys twice a day Amazon deploys every 11 seconds on average

Facebook Case Study

• No detailed plan to achieve a final, well-defined product

• Engineers work directly on a common trunk

• No separate QA team for testing

• New code is deployed twice a day

• Actual use controlled with Gatekeeper

• Facilitates A/B testing

• Engineers self-select what to work on

• No assignment of blame

Facebook Growth

Quadratic rate

(like Linux and other open source projects)

Facebook Workforce• 6-week bootcamp

• Self-select what group to join

• Managers can't assign people to work

• Hackamonth to try out other groups

• Sustainable work rate

Weekly Push

• All code is reviewed

• Passes ~50,000 regression tests

• Used internally by all FB employees

• Deployed incrementally

• Controlled with gatekeeper

Perpetual Development Benefits

Lead time to first working version is short, and a working version is always available No danger of the project coming to nothing

Real users doing real work are effectively brought into the development cycle Helps to test system functionality and find problems Used to prioritize further development according to

what is really needed Ability to use new technology as it becomes

available

Implications for Development

No fixed goal that has to be reached Goal is to continually improve the system and

maintain is usefulness Monitor system usage to identify inadequacies Prioritize according to user needs Don't plan too far ahead (YAGNI protection)

Use contracts that take longevity into account Support for continued evolution Access to code if company becomes insolvent

Implications for Architecture

Can't decide on architecture based on analysis of all the requirement

Need architecture that accommodates change Two tiers: stable core and evolving libraries Open system like e-commerce site based on web

services Use refactoring May need to abandon project eventually

But may still salvage parts for a followup project

Conservation of Familiarity

One of Lehman's laws of software evolution Limits the rate of progress that can be

sustained Need specialized tools to help new team

members to become familiar with the system Newbies are at a disadvantage because they didn't

see how the system developed Need to capture the history of design decisions

Explaining Monumental Failures

Failures caused by "feature creep" Developers made elaborate and beautiful plans But these plans were obsolete by the time they

were completed Do exactly what is most needed at each instant

Failures caused by successful maintenance Delivered system was good for a very long time But when it is to be replaced, an attempt is made to

do too much at once Make improvements continuously all the time

Experimental Evidence

Experiment conducted by the World Wide Consortium for the Grid (W2COG)

The goal: develop a secure service-oriented architecture system

Traditional approach: standard government acquisition process

Alternative: use a "Limited Technology Experiment" based on evolutionary methods

Both start with same government supplied software baseline

18 Months Later...

Traditional: A concept document

with no functional architecture

Cost $1.5M No concrete

deployment plan or timeline

Evolutionary: Delivered open

architecture prototype addressing 80% of requirements

Cost of $100K Plan to complete in 6

months

Denning, Gunderson, & Hayes-Roth, CACM 12/2008

Bottom Line

Expect to see many more projects using evolutionary and agile methods

Especially in environments challenged by rapid technological progress and rapid change

These ideas are actually not new― However, not articulated well till recently

(except for Gilb)― Contradict traditional engineering approach― Nevertheless work well in practice

Documents

David Lorge Parnas “A sign that the Software Engineering profession has matured will be that we lose our preoccupation with the first release and focus