►
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
Intro: Dragonfly - Allen Sun & Zuozheng Hu, Alibaba Cloud
As cloud native becomes more and more popular in industry, how to distribute images efficiently and safely is a new challenge for enterprises. Dragonfly is an open source intelligent P2P based image and file distribution system. Its goal is to tackle all distribution problems in cloud native scenarios. Dragonfly focuses on: • Simple: well-defined user-facing API (HTTP), non-invasive to all container engines • Efficient: CDN support, P2P based file distribution to save enterprise bandwidth • Intelligent: host level speed limit, intelligent flow control due to host dection • Secure: block transmission encrytion, HTTPS connection support we will focus on the introduction and live demo of Dragonfly, review and provide solutions for enterprises, including mass distribution, secure transmission, bandwidth cost. Real use cases will be discussed in the session.
https://sched.co/MPha
A
Thank
you,
everybody
for
your
attention
to
come
to
a
session
of
dragon
fly,
introduction,
I'm
Alison
from
Alibaba
cloud
and
my
college
soldier,
who
still
will
give
some
introduction
about
dragonfly
before
my
session.
I
want
to
ask
a
several
questions.
How
many
nodes
do
you
manage
in
one
single
kubernetes
cluster
yeah?
Could
you
give
me
a
number
13,
okay,
perfect?
A
A
A
You
know
because
some
Legion
see
applications.
It
is
very
hard
who
can
to
be
continued
right,
so
the
container
image
may
be
very,
very
large,
maybe
to
get
byte
or
more
than
to
invite
some.
The
agency
enterprises
I
think
still
makes
the
same
issue
so
how
to
distribute
the
image
efficiently
at
your
cluster
at
large
gears.
It
is
a
very
real
question
you
should
answer
at
that
time.
We
try
to
provide
some
solutions
to
adopting
the
p2p
scenarios,
the
energy
distribution
scenarios.
A
It
is
dragonfly.
Okay,
so
here
is
the
very
brief
agenda
of
my
session.
First
I
will
give
some
introduction
about
the
history.
Why
dragonfly?
What
is
dragonfly
and
a
second?
It
is
about
such
dangers?
Actually,
dragonfly
is
already
a
standard
box
project
in
C
and
C
F,
and
we
are
planning
to
try
to
make
it
enter
into
incubation
label
after
maybe
three
months,
and
my
college
will
take
some
brief
introduction
about
adoption
and
some
load
map.
Also,
first,
let
me
introduce
the
history.
A
What
is
dragonfly
dragonfly
is
a
p2p
based
highly
reliable
image
distribution
system
and
it
will
solve
the
bottlenecks.
Importing
images
indices
are
spheres.
Here
is
the
milestone
it
is
born
in
2015.
At
that
time
we
try
to
distribute
files.
You
know
we
should
distribute
files
across
the
region
or
across
the
cluster
and
to
make
it
we
provide
two
fundamental
features
of
dragonfly,
the
first.
Of
course
it
is
peer-to-peer
and
the
second
is
a
CDN,
because
we
should
try
to
make
that
request.
A
A
Dragonfly
has
already
become
a
very,
very
fundamental
tool
of
Industry
/
infrastructure,
Alibaba
Group,
and
it
covers
all
the
scenery,
almost
all
the
sceneries
of
the
image
and
the
felt
distribution,
and
at
that
time
we
open
sourced
the
dragonfly
open
source,
the
dragonfly
and
ad.
We
try
to
make
it
try
to
discuss
with
the
ecosystem,
try
to
discuss
with
the
customer
whether
we
can
try
to
make
it
more
standard
to
provide
the
services
for
the
wider
ecosystem
and
in
the
November
of
2017.
A
A
How
about
history
preheats
the
images
from
that
harbor
to
the
dragonfly?
This
is
what
we
want,
because
we
we
try
to
make
try
to
make
it
more
easier
to
to
save
more
time
when
distributing
images,
and
we
supported
memory
file
systems
at
other
large
gears
in
our
cluster.
We
have
lots
of
nodes.
This
physical
machine
is
equipment
is
very,
very
high,
all
with
SSDs
SDS
disk,
but
we
found
that
we
take
advantages
of
the
high
quality
equipment,
but
under
under
cloud
environment,
lots
of
VM
machines.
A
They
are
take
advantages
of
network
disk
to
to
replace
that
a
local
disk
I/o.
So
we
found
some
some
some
shortcomings
of
dragonfly.
We
try
to
support
memory
file
system
to
to
use
memory
to
to
to
reduce
the
disk,
I/o
yeah
and
in
the
in
the
november
of
2018,
dragonfly,
Jhansi
and
CF
as
a
standard
box
project
and
the
integrator
with
harbor.
A
Actually,
as
these
points,
you
know
in
the
industry,
in
the
ecosystem
data
streams,
there
are
several
p2p
distribution
to
us.
Maybe
you
heard
of
the
kraken
projects
from
uber
right.
Here's,
the
very
funny
story
about
that.
Actually,
yesterday
my
college
and
I
and
the
dragon
team
have
dinner
very
happy
to
discuss
about
it's.
A
Actually,
they're
almost
to
say
almost
as
the
same
kraken
plays
that
more
rows
at
the
registry
say
they
build
a
very
simple
registry
to
build
the
image
registry
and
the
distribution
part,
but
dragonfly
only
focused
on
the
distribution
parts
yeah
and
for
the
p2p
policy.
I
think
they
have
a
tracker
to
make
seed
generate
seeds.
They
have
tracker
to
record
all
the
node
status
and
we
have
a
super
node
to
make
it
very
similar
policies.
A
Actually
we
in
some
enterprises
there
are
closed
sauce.
The
project
which
is
aimed
at
32.
The
image
distribution
like
like
a
project
in
eBay,
say
the
project
to
ebay
is
very
interesting
and
very
fancy.
I've
seen
because
when
scheduling
Co
p2p
talk
to
our
system
is
totally
decentralized,
they
try
to
make
all
the
scheduling
system
scheduling
data
in
the
client
to
spread
use
gossip
to
spread
the
information,
but
in
Jagan
fly
we
have
a
super
node
to
check
everything
so
that
dragonflies
supernova
may
suffer
some
HJ
program.
A
But
currently
we
have
tried
to
Soviet
in
the
upstream,
and
now
we
are
doing
the
refactoring
with
golem
as
a
very
beginning
of
this
project.
We
we
use
three
languages
to
build
these
torques
one
pison
soup,
another
with
Jabbar
and
go
along
with
that
with
the
proxy.
But
now
we
are
trying
to
make
it
unique.
We
said
golem
and
we
support
the
HTTP
registry
and
we
have
more
than
50
adopters,
actually
some
adopt
some
adopters.
They
haven't
talked
to
us
to
tears
here.
A
B
Hello,
Aloha
I'm
as
a
founder
of
a
dragonfly
leading,
is
the
desire
and
the
power
of
development.
Now
let
me
introduce
such
a
conference.
Current
situation.
Dragonfly
has
been
open
source
for
18
months
and
has
joined
the
sensor
for
standard
box
project
since
dragonfly
is
an
image
distribution
system.
She
has
no
doubt
that
distribution
efficiency
is
a
faster
element.
So,
let's
start
with
performance
comparison
on
the
left
is
a
testing
condition.
B
She
had
two
virtual
machines
lonicera
and
the
two
hundred
of
virtual
machines
as
cleanser,
and
that
image
size
test
jitter
is
a
750
mega
pad
on
the
light.
It
is
a
performance
in
comparison,
Tate,
x-axis,
it's
a
number
of
pollutant
loads
and
the
y-axis.
It's
a
very
town,
all
the
time
the
orange
curve
represents.
Well
it
a
well
each
time
I
used
in
dragon
fire
$2
image.
It
can
be
clearly
seen
that,
with
the
expansion
of
download
scholar,
Donna's
hammer
is
steer,
were
stable,
basically
maintained
at
about
18
seconds.
B
B
We
have
good
reason
to
blue
dragon
flag
and
so
as
a
problem
apart
efficiency
and
the
stability
of
image
distribution
at
large
scale.
At
the
same
time,
the
use
of
dragonfly
will
not
bring
any
intrusive
modification
to
the
to
the
canal.
Engine
image
lattice
tree
kubernetes
and
so
of
course,
dragonfly
also
supports
integration
with
the
related
system
in
order
to
generate
more
chemical
reaction
and
the
further
enhancing
is
a
distribution
ability.
B
B
At
present,
more
than
50
companies
have
become
a
topic
of
dragonfly
which
are
distributed
in
various
industries,
including
Alibaba
clan
and
the
financial,
china,
mobile
highway,
lazada
and
so
on.
Dragonfly,
currently
how
modern
set
400
stars
on
github,
as
well
as
sixth
maintenance
and
sell
it
to
contributors
in
addition
to
github,
we
also
have
a
variety
of
for
community
discussion
groups
such
as
Qatar
in
talk
and
we
chart
dragonfly
is
about
to
compete
its
you
factor
with
kaoliang,
and
is
it
a
tiny,
many
new
features?
B
We
warmly
welcome
or
interest
people
to
join
us
to
make
dragonfly
stronger
and
stronger
okay.
We
have
just
introduced
current
situation
of
a
dragon
fire
development
in
the
community.
Next,
let's
focus
on
the
main
principles
and
core
architecture
of
which
I
can
fly.
I
believe.
Will
you
please
be
more
interested
in
in
his
aspect:
okay,
first,
real
odd,
briefly,
alighted
the
process
of
image
distribution
with
dragonflies
pupae
technology,
dragonfly
Grandin
named
TF,
get
is
installed
on
each
node,
which
is
configured
as
a
tunnel
proxy
for
the
canal.
Engine
win.
B
B
When
this
women
node,
let's
see
what
the
legislation
we
passed,
a
flower
mo
planter,
you
were
aware
added
a
clan
to
a
p2p
Network,
which
is
composed
of
Marty
plans
at
onload
in
the
same
image
layer.
At
the
same
time,
the
super
node
were
synchronized.
The
image
of
flow
malaria
at
most
once
in
the
process
of
a
second
ization
I
will
block
downloaded
the
word
translator
a
seed
in
the
local
storage,
which
can
be
immediately
spread
between
multiple
clans
bin
section,
the
finishes
all
the
pieces
of
a
layer.
B
B
Uci
as
in
nodes
that
aware
need
to
download
the
same
image
afterwards,
because
the
pubis
wha
has
been
captured
by
a
CDN
to
see
the
seed
in
patches
can
be
directly
used
for
blocking
propagation
without
recognizing
the
image
of
la
maleta
tree
and
mcinnes
seer
again,
Abu
Ezra,
the
tunneler,
the
late,
where
be
faster
than
the
previous
one,
okay
I
believe
everyone
already
knows
more
about
dragonfly.
Finally,
I'd
like
to
introduce
a
core
texture
of
a
dragonfly
which
is
a
super
nose
architecture.
B
Okay,
by
introducing
super
node
architecture,
you
were
honest
and
how
dragon
Frazee
it's
best
to
improve
tunnel
or
delayed
and
save
a
network
trafficker.
A
super
node
is
mainly
composed
of
spiral
modules,
including
p2p
schedule,
cd-I
manager,
transmission,
controller
therapy,
kita
and
storage
management.
B
Okay,
it's
a
p2p
scheduler
is
responsible
for
assigning
internal
tasker's
to
each
client
to
tasks
and
information
mainly
includes
the
block
number
and
which
planet
corresponding
blocks
were
Donald
Flom.
A
total
of
six
sixty
scheduling
strategy
I
will
introduce
the
fastest
three
strategies:
first,
the
sparseness,
the
nose
preference
for
choosing
blocks
there
are
listed
at
distributed
in
PvP
Network.
B
Okay,
that's
a
fastest
three
strategies
have
been
introduced
and,
as
others
will
not
be
introduced
one
by
one.
Next,
let's
focus
on
CDI
management,
which
has
three
functions
first.
The
second
idea
is
responsible
for
downloading
image
from
a
registry
and
the
generating
block
seeds.
It
also
controls
so
that
the
same
image
we're
all
in
the
turn
of
the
ones.
B
Second,
the
catch
is
responsible
for
verifying
validate
and
the
consistence
of
catch.
In
addition,
catcher
is
also
used
to
pull
a
process
catch.
Tc.
Third,
compressor
is
responsible
for
choosing
choosing
the
most
compressible
blocks
for
compression.
By
this
way,
we
can
obtain
really
high
compression
level
you
which
can
not
only
increase
the
speed
but
also
save
a
network
traffic
or,
to
a
larger
extent.
B
B
Celeste
is
Storage
Management,
which
is
the
memory
responsible
for
astrologer
for
Valley
estate,
including
blocked
I'll,
say
that
method,
soft
link
and
said
statistical
for
distribution
and
so
on.
In
addition,
super
node
can
also
tell
me
what
to
use
memory
file
system
according
according
to
hard
work.
B
If
it's
a
memory
file
system
is
enabled
when
downloading
large
files
data
can
be
laid,
and
the
light
and
directives
lose
a
memory
without
a
disk
IO
she's
not
owning
speed,
opposite
download
efficiency
for
for
large
fast,
but
also
avoided.
The
problem
large
fast
occupy
aioli
sauce
for
a
long
time
through
the
coordination
of
power
modules,
the
super
node
it's
absolute,
absolutely
was
of
his
name
and
initials
super
efficiency
and
stability
for
downloading
image.
B
A
Let's
introduction,
let's
introduce
more
about
seduction.
I
will
show
you
two
cases
about
the
dragonflies
adoption.
The
first
case
is
about
Alibaba
cloud
actually
series
product
which
is
named
the
ACK
Alibaba
cloud,
the
kubernetes
services.
They
provide
the
services
to
the
end-user
to
enjoy
the
services
which
is
based
on
affinities,
the
some
users
they
are
using
Cuban
artists
at
very
large
gears
at
the
middle
skiers
and
say
which
to
distribute
image
very
speed
very
rapidly
and
ACK
to
the
integration
of
giant
line
here
is
that
texture.
A
A
The
first
part
when
you
try
to
try
to
download
images
stock
engine,
we
are
trying
to
try
to
send
a
request,
but
before
sending
the
request
here
is
there
is
a
config
watching
that
series,
and
this
config
watching
is
watching.
That
westerns.
Here
is
the
image
to
the
soup.
Another
cache,
if
your
is
same
as
the
talkity,
so
changing
will
trigger
a
poor
requester
from
lock
to
the
TF
demon,
the
edema
you
can.
You
can
regard
it
as
a
proxy
to
drops
the
docker
image
pooling
request
and
it
can.
A
Okay,
so
DF
demon,
we
are
try
to
offense
the
manifest
affirm
the
registry,
but
the
layers,
but
the
blobs
is
there
in
the
registry.
We,
this
is
the
most
size
of
the
image
we
try
to
distribute
it
and
it
will
trigger
a
process
which
is
named
DF
get
DF
get.
Is
that
is
a
peer
peer
agent.
It
can
download
the
files
from
other
peers
and
it
can
also
provide
downloading
services
for
other
peers.
A
So
it
is
a
p2p
Network
agent,
okay,
after
downloading
the
blobs,
everything
is
okay,
downloading
from
the
CDN
images
kashi
from
the
registry
by
the
super
node,
so
download
from
the
CDN,
and
it
also
provides
the
downloading
services
for
other,
as
appears
so
with
this
architecture,
a
cluster
may
be
more
than
15.
Nodes
can
rapidly
distributed
to
the
image
around
all
the
cluster
it's
very
efficiently.
This
is
this
is
the
practice
in
Alibaba
cloud
and
the
second.
This
is
a
practice
of
China
Mobile
Chan
mobile.
A
A
A
So
when
a
image
is
beaut,
bezier
continuous
integration
system,
they
try
to
push
this
image
to
the
production
environment.
Here
is
a
client.
A
client
is
the
is
a
component
to
send
out
the
control
commands
with
the
command,
and
it
will
push
the
images
to
the
harbor.
Harbor
is
the
registry
and
harbor.
We
are
trying
to
use
the
dragonflies
preheat
API
to
try
to
make
the
image
to
push
to
super
nodes
cache.
A
Then
it
is
ready
for
autumn
for
for
super
in
order
to
have
the
image
after
this
is
done,
then
it
will
try
to
try
to
start
distributing
it.
We're
seeing
the
commands
to
the
thumb,
some
fear
services
and
the
services
will
trigger
trigger
every
image.
Pulling
requests
on
the
single
notes.
Talk
engine
the
dacha
engine
will
try
to
take
advantages
of
the
TF
client,
which
is
DF,
daemon
and
DF
care
to
try
to
construct
a
pure
network
to
distribute
the
image.
A
A
So
it
shows
a
little
bit
the
scalability
of
dragonfly-
and
here
is
the
parts
of
the
roadmap
and
the
dragonflies
go
to
map,
and
we
will
try
to
support
us
security
parts
actually
before
the
before
April
dragonfly
support.
Private
container
image
registry
is
a
little
bit
weak,
but
in
the
new
version
we
make
it.
We
have
made
it
and
a
different
encryption
algorithm
in
data
transmission.
Some
customers,
some
and
the
use
of
dragonflies
they
mention
is
that.
A
Can
you
make
the
data
transmission
encryption
and
as
it
is,
it
is
feature
it
is
an
enhancement
and
the
CN
CF
ecosystem
deploys
the
client
using
demonstrating
kubernetes
and
we've
already
supported
to
deploy
dragonfly
as
ahem
used
using
him
chat
so
I'm
so
Keith
hub
on
the
repo
you
can
see
every
teachers
about
that
I
think
the
Dragonfly
teams
did
a
very
well
is
a
talks,
okay,
so
the
efficient
so
efficiently
efficiency.
We
we
were
planning
to
do
that
dynamically
downloading
rate,
limiting
and
intelligence
scheduling
algorithm,
because
you
know
on
every
node
image.
A
Pooling
is
combined
with
the
container
round
time.
So
if
you
have
applications
running
on
the
node
and
at
the
same
time
you
receive
the
Earth
image
polling
request.
Oh
maybe
some
unexpectable
things
will
happen
right
some
conflict,
but
for
the
distribution
we
try
to
make
it
beers
a
low
and
lower
and
lower
risk.
So
we
try
to
try
to
try
to
take
the
intelligence
scheduling
algorithms
on
the
agent
into
consideration
for
the
openness
we
are
doing
a
refactoring
seen
as
in
the
late
of
May.
A
It
is
one
week
later
we
will
released
a
new
version
of
Gollum
it's.
It
is
0.4
dot,
zero
version.
We
we
wish
that
everyone
who
try
to
test
it
and
highly
user
customers
models,
scalability
and
yeah
and
I've
seen
sceneries.
We
were
try
to
make
dragonfly
to
to
be
fit
ball
on
the
cloud
environment
as
we
know
so,
aubergine
and
the
eBay
team
and
a
dragonfly
team.
First,
we
all
design
tools
in
internal
environment,
internal
environment
and
lots
of
the
lots
of
that
is
on
the
bare
mental
of
the
coma
Sheen.
A
But
the
cloud
environment
is
a
little
bit
different
totally
on
the
infrastructure,
storage
and
the
network.
So
we
will
try
to
make
it
more
feasible
on
that
yeah
and
disease
is
a
brief
introduction
about
the
road
map
actually
dragonfly
sessions
as
these
cubic
on.
We
have
three
sessions.
Yesterday
we
talked
about
in
details
the
integration
of
dragonfly
and
harbor,
and
this
session
is
a
very
brief
dragonfly
intercession.
Tomorrow
we
will
have
a
dragonfly
dip
type.
A
A
A
A
C
A
A
Harbor
team
say
they
have
registries
registry,
have
their
storage
storage
can
do
the
replication
right
and
we
doing
the
distribution
if
the
register,
if
you
have
already
done
the
replication
of
the
registry,
then
dragonfly.
We
are
not
to
this
part,
but
if
you,
if
you
are
not
taking
the
advantages
of
the
registry
in
same
dragonfly,
can
take
over
this
part
actually
in
Alibaba
internal
we
implemented
at
this
part
different
regions,
different
zones.
A
We
can
use
the
replication
feature
of
dragonfly
to
do
the
transmission,
okay,
but
I'm
afraid
this
feature
is
not
in
the
open
source
version.
Why?
Because
a
few
customer
few
and
the
user
requested
this
feature.
So
if
you
want
this
feature
you
can
you
can
put
forward
as
usual
on
the
level
we
can
try
to
make
a
different
decision
with
dragonfly
team
and
harbor
team
to
to
provide
a
force
tech
solution
for
you,
okay,.
C
A
Far
as
I
know,
we
provide
container
registry,
the
container
registry
projecting
Alibaba
cloud.
It
is
named.
Acr
and
Asya
have
I,
seen,
have
lots
of
zones
first
China,
mainland
and
the
single
pole,
but
for
the
Europe
side,
I'm,
not
sure
if
there
is
already
have
one
so
back
to
the
China
mainland,
that
Singapore
say
have
the
replication.
They
have
the
replication
for
the
image
image
in
the
registry,
but
for
the
new
room,
I
am
Not
sure
the
features
in
implementing
on
this
side.
Okay,.
A
A
Okay,
this
is
a
very
good
question.
It
is
about
sand
wrong
time
wrong
time.
Options
for
dragonfly
actually
Japan
fly
is
a
very
general
general
p2p
distribution
tools,
currents
late
talker
is
the
most
popular
popular
one
and
in
Alibaba
we
have
a
nazzer
container
round
time,
which
is
named
the
ponch
container.
We
can
use
dragonflies
as
we're
and
maybe
in
the
future,
the
most
neutral
container
round.
Time
is
container
D
or
cryo,
and
we
support
it
also.
A
D
D
A
A
Yes,
some
most
largest
difference
is
that
internal
in
Alibaba
we
have
a
a
player
beyond
the
super
node
series,
a
series
super
node
controller
to
managers
and
crosses
on
scenes
from
some
scenes,
but
we
only
but
up
the
super
node
and
the
DF
team
and
the
DF
gate
see
are
totally
the
same.
Why
we?
Why
we
keep
it
inside?
A
A
A
Actually,
for
your
case,
I
think
you
can
use
dragonfly
when
you
build
the
way
you
build
the
images
in
the
Builder
machines
or
something
like
that.
You
can
push
the
image
to
the
registry
and
with
another
cluster,
with
the
original
way
you
you
will
try
to
stand
the
requester
from
the
running
cluster
to
pour
images
from
registry,
but
but
with
dragonfly
you
can
deploy
dragonfly
in
this
running
cluster.
You
can
use
the
preheat
tree
preheat
API,
to
try
to
try
to
make
the
super
in
order
to
cast
the
image
from
the
registry.
A
The
NZ
cluster
can
enjoy
the
peer-to-peer
distribution
efficiencies,
so
so
I
think
it
can
work.
Definitely
it
can
work
so
for
the
for
the
network
o
network
issues.
If
you
have
different
clusters
with
one
one
kubernetes
with
the
Federation
policies,
I
think
you
the
better
to
deploy
to
deploy
each
super
node
in
each
cluster.
It's
a
good
way.
Actually
Alibaba.
We
do
this.