►
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
Deep Dive: Dragonfly - Jin Zhang, Alibaba Cloud & Haibing Zhou, eBay
Dragonfly is an intelligent P2P based container image distribution system which provides a native image distribution solution for cloud native scenarios. In this session, we will cover some advanced features of Dragonfly in details, such as image distribution in a cloud environment, the scheduling strategy in P2P distribution, highly-available deployments and so on. Besides these, we also encourage developers and end-users to extend Dragonfly's usage with other softwares in cloud native ecosystem. Furthermore, the team would love to get feedback from users and contributors about current features and future roadmap.
https://sched.co/MPjN
A
So
hi
everyone
welcome
to
the
Dragon
Friday
obsession.
My
name
is
Jo
and
I'm
from
I'm
from
eBay
infrastructure
engineering
team.
Today
I'm
going
to
talk
about
the
dragonfly,
an
intelligent,
p2p
image,
distribution,
fat,
distribution
solution
before
we
go
deep
in
so
in
this
section
we
have
three
parts
one
we
are
going
to
go
through
the.
Why
we
need
image,
why
we
need
a
specific
system
for
doing
image,
distribution
and
file
distribution.
A
A
So
before
we
go
into
the
detail,
let's
look
at
several
cases
where
image
distribution
of
any
fat
distribution
can
be
a
problem.
Firstly,
let's
say
you
have
a
rotor
machines.
The
first
job
you
need
to
do
is
to
install
the
operating
system
of
the
machines.
No
matter
is
Linux
or
some
different
districts,
and
once
the
operating
system
is
running,
you
need
to
install
additional
software
to
run
your
application.
One
example
right
so
one
example
is
JVM.
You
need
to
run
Java
applications.
A
Another
example
is
you
have
to
run
containers
you
need
to
install
one
of
the
container
run
time.
No
matter
is
Kakaako
or
its
container
deep,
and
then
you
will
have
your
application
so
that,
based
on
what's
that,
you
are
using
if
you're,
if
you're,
writing
java
application,
you
have
to
dispute
your
java
archives
to
all
the
machines.
If
you're
running,
if
you're
using
node
years,
you
have
to
dispute,
you
know
TS
modules,
it
runs
some
adds
up
like
go.
You
made
these
people
proteins
and
that's
not
the
end
of
story.
A
You
have
your
application
running
and
you
still
sometimes
you
need
to
distribute
your
configurations
and
configurations
does
not
have
to
be
small.
For
example,
if
I
learning
a
machine
learning
training
job,
you
have,
you
have
to
dispute
the
models
file
and
we
are
running
a
search
engine.
You
have
to
dispute
the
index
data
to
add
closed
as
to
your
application.
In
that
case,
the
search
engine
and
the
image
disputing
can
be
another
problem
as
well.
A
Today
we
are
all
talking
about
cloud
native,
but
that's
not
always
possible,
or
at
least
not
fast
enough
for
you
to
move
in
from
your
leaky
application
to
crown
native
application.
So
you
may
have
to
do
a
simulation
where
your
package
of
the
software
you're
running
services
into
a
fat
container.
That
basically
means
you
are
on
your
container
at
the
VM,
and
in
that
case
the
image
can
be
very,
very
large.
A
In
the
second
case,
if
you
have
a
remote
data
center,
so
you
have
your
service
running
in
your
primary
data
center,
but
you
also
need
to
dispute
some
of
the
image
like
age
if
you're
running
pop
and
finally
in
kubernetes,
you
sometimes
have
to
deploy
a
demo
sets.
So
demon
said
basically
is
communities
concept.
We
all
you
want
to
run.
A
A
part
of
the
community
is
null
so,
let's
say
if
we
have
five
five
thousand
nodes
and
you
want
to
create
a
new
like
say
the
more
internal
agent,
the
secret
security
detector,
and
that
can
be
very
time-consuming
when
you
want
to
quickly
deploy
it
to
of
the
nose.
So
that
means
definitely
we
need
to
improve
our
current
distribution
infrastructure.
Let's
take
a
look
how
it
is
today.
A
Basically,
you
have
several
rails
at
the
bottom.
Is
the
storage
and
then
you
build
like
a
religious
service
or
maven
repository
on
top
of
the
storage,
and
then
you
create
a
lot
of
parents
are
too
exhausted
as
a
service,
and
then
the
current
talk
to
the
load
balancer.
So
looking
at
this
structure
is
very
traditional,
it's
very
simple,
but
that
doesn't
solve
our
scalability
problem,
as
we
mentioned
in
previous
two
slides.
So
what
you
can
do,
probably
not
in
loader
parents
that
you
can
configure
the
rack
server
returned.
A
So
in
that
case
the
the
data
can
go
from
the
service
to
crash
directly,
and
if
you
can't
control
the
service,
you
can
add
some
retraction.
So
all
the
data
does
not
have
to
go
through
the
service
in
order
to
serving
crowns.
The
data
can
goes
from
the
soil
to
crunch
the
directory,
and
also
you
can
enhance
the
storage
last.
A
A
So
that's
the
church,
that's
how
the
distribution
it
works.
You
have
a
centralized
server.
Every
files,
cost
average
current
request
the
server
and
a
gadget
copy.
But
if
we
can
do
like
this,
why
would
current
get
one
of
the
current
get
is
copy
and
sure
which
between
each
others?
In
that
case
you
don't
have
the
load
and
the
centralized
service
become
much
much
decreased
and
that's
probably
a
solved
problem
in
as
Ariel
so
like
BitTorrent,
it
solves
how
you
can
dispute
media
music
and
recently,
the
ipfs
where
you
run
decentralize
the
web.
A
A
Hey
I
know,
I
want
to
create
tasks
to
you
and
it's
basically,
what
super
node
does
is
create
a
record
in
its
memory
and
gives
you
a
task
idea
in
later
request.
The
FCAT
can
use
this
task
ID
to
pull
the
status,
so
the
state
basically
gives
you
a
list
appears.
We
are
you
care
where
you
can
request
from
different
peers
to
download
the
file,
so
the
craft
gets
the
list
and
send
the
request
to
each
peer
once
downloaded
successfully
the
currents
update
to
a
super
nodes
say:
hey
I
know:
I
have
this
node.
A
A
So,
in
addition,
in
your
case
well,
the
the
class
is
fresh.
No
one
has
the
file
the
super
node,
acting
as
the
first
step
to
download
from
the
sauce
and
serving,
as
the
first
appear,
so
I'm
going
to
talk
about
each
of
the
components
in
more
detail.
First
LS
to
get
super
node,
as
we
just
talked
super
node,
is
nothing
but
a
bunch
of
api's.
A
So
it
takes
to
take
the
peer
registration
takes
task
equation
and
also
gives
you
the
pieces,
information
and
update
it.
Update
the
pieces
from
beyond
each
of
the
know,
but
what's
becoming
more
interesting,
is
the
scheduling
part?
How
does
super
node
select
list
appears
that
is
more
efficiency
to
you.
A
A
The
most
lister
load
to
peer,
which
you
can
request
from
and
also
it
can
do
the
seeding.
If
no
one
has
the
file
we
can
acting
as
the
first
one,
the
supernova
can
acting
as
first
one
to
request
from
sauce
and
then
serving
up
serving
to
different
clients.
So
what
is
difficut?
The
FK
is
pretty
much
like
target
occur.
It's
just
simple
CRI,
but
beyond
that
it
can
also
do
some
funny
things
like
rate
limiting
so
actually
rate.
A
Limiting
is
very
important
when,
when
you
have
some
network
intensive
application
running
on
that
note,
you
will
do
not
want
to
consume
all
of
the
bandwidth,
so
by
using
rate
limiting
we
can
reduce
the
impact
to
the
running
application.
On
that
note
and
to
checksum,
of
course,
you
don't
want
to
download
the
wrong
file
right
and
the
last
one
is
the
seeding.
So
that's
the
basic
appeal
functionality,
but
looking
at
that,
how
can
you
solve
the
image
distribution
problem?
Docker
does
not
know
that
you
forget.
Continuity
does
not
know
TF
get.
A
How
can
you
enhance
the
image
pouring
performance
and
that's
with
that
part?
We
need
a
proxy,
so
the
proxy
needs
to
so
the
taça
demon
can
talk
to
the
proxy
processing.
Talk
to
the
TF
gets
downloads
file.
So
and
that
exactly
what
the
FDA
or
the
FDA
is,
but
before
we
talk
to
this,
let's
see
how
different
container
runtimes
lets
you
configure
the
proxy.
So
for
darker,
you
have
a
option
to
say
you
can
specify
the
rotation
mirror
where
you
can.
A
When
you
download
some
page
images
from
docker
hub,
you
can
specify
a
far
more
close
location
to
you
and
for
community
as
well.
It
can
actually
suppose
private
image
mirror
where
you
can
convict
a
different
mirror
for
different
registries
and
all
if
you
can
control
the
source
code
of
your
research
service,
you
can
add
in
retraction
instead
of
downloaded
from
storage,
you
can
retract
to
a
proxy
and
TF
demo.
Is
that
your
proxy,
which
dragon
fire
has
jack
fry
is
simply
as
a
reversal
is
a
reverse
proxy?
A
B
Hello,
my
name
is
John.
Ting
comes
from
Alibaba
Group,
one
of
the
members
of
dragon
flying
team
next
via
analysis.
How
does
junk
fly
pass
the
padding
in
image
pulling
images
when
we
introduce
young
fly
via
released
many
futures
to
show
how
excellent
young
fly
is,
and
one
of
the
very
important
future
is
knowing
inversely
about
support
for
all
kinds
of
container
technologies.
B
B
B
Each
statement
in
Java
file
will
be
created
as
a
major
image
layer,
and
these
layers
ask
during
the
blob
portion
of
the
registry
hit
by
digestion
and
pulling
a
layer
is
carried
out
by
our
standard
HTTP
request,
so
the
process
of
pulling
a
major
centers
around
getting
this
two
components:
Manifesta
and
the
layers.
So
let's
say
how
to
get
these
two
components.
B
B
There
is
a
very
important
filter
of
evades
the
offence
layers
layers
in
our
manifests.
His
lister
of
loyalties,
filters,
including
digest
when
the
manifest
is
enhance.
The
client
master
verify
the
signature
to
ensure
the
names
and
the
layers
are
valid
once
confirmed.
The
kind
of
we
are
use,
digester
to
download
layers
of
image.
A
pooling
of
imagery
is
carried
out
by
a
standard
HTTP
request
and
he
stammers
downloading
files
from
remote
ntp
server.
B
The
uir
is
as
follows,
and
the
digester
is
listing
the
manifest
this
basic
example
of
pulling
major
layer
from
harbor
registry
deployed
by
myself.
Nasa,
says
example
response
the
rebounce
response.
Coder
is
hundred.
It
means
that
the
kind
he
can
download
the
layout
directory
from
the
rest
free
to
allowed
for
incremental
downloads
grant
requests
should
be
exported.
B
B
B
B
The
sector,
the
second
example-
is
pulling
a
layer
from
official
registry.
It
has
two
parts
which
is
much
different
from
a
previous
one.
The
first
request
to
registry
gets
a
307
response
code.
It
tells
the
kind
you
should
follow.
Those
with
drag
to
request
another
server
for
pulling
the
layers
see.
Another
service
is
Poland
out
by
the
response,
header
location
and
his
HTTP
request.
B
Lcpl
server,
and
the
second
request
is
loading
the
image
layer
fire
from
the
FTPS
server.
From
this
example,
we
know
that
the
blood
portion
of
the
doors
doors
in
many
layers
can
be
another
different
and
bond
from
red
string,
and
this
is
a
third
example
of
pulling
a
layer.
It
is
much
similar
with
the
previous
vamp
bounces.
The
only
one
difference
here
is
that
the
location
server
is,
it
is
a
HTTP
server.
So,
in
summary,
we
learned
the
whole
process
of
pouring
an
image.
B
First,
early
papers
are
manifest
by
HTTP
request,
secondly,
papers,
layer,
fires
by
several
HTTP
requests.
So
let's
say
how
does
drunk
flying
poet
by
default,
trying
to
fly
all
in
downloads,
individual
layer
of
ice,
because
the
layer
files
that
are
usually
larger
and
pulling
them
are
very
time
consuming,
especially
in
norske
actually
pulling
their
files
is
stammers
downloading,
comma
files
from
remote
file
servers?
Of
course,
we
can
also
configure
chunk
fly
to
download
the
manifesto
if
we
want
to
spit
out
in
some
scenarios.
B
B
The
house
to
show
the
houses
showed
here
is
a
component
of
junk
flying
name.
The
TF
demon
reintroduced
the
above
first
first
way
is
three
meters.
When
you
do
actually
use
talk
to
a
mentor
named
Doug
poem
into
name,
you
can
use
you.
Can
you
this
restroom
mirror?
You
can
see
entity
abdomen
or
the
resident
mirror
to
pour
images
in
a
second
away,
HTTP
proxy,
when
the
blood
portion
of
the
registry
is
an
HTTP
server?
Oh,
it's
a
redirected
to
a
TCP
server.
You
can
use
Jeff
Niemann
or
the
HTTP
proxy
to
prune
image.
B
B
Next
of
a
tag,
look
as
the
whole
process
of
pulling
image
by
drunk
fly
from
this
picture.
We
can
say
there
are
three
components
between
dock
D
and
registry.
The
comments
are
soup,
node,
D,
F
get
and
depth.
Demon
is
a
proxy
pitcher
dock
D
and
territory
for
stole
a
dog
D
stands
requested
to
resolutely
in
factor
all
requests
from
dog
D
goes
through
the
node
parks
process.
It
is
TF
demon.
B
Secondly,
the
abdomen
catches.
All
the
requests
filter
out
is
appallingly
our
request
and
use
the
Afghan
together
to
premier
to
pool
else
subtly.
Super
node
gets
a
layer
downloading
tasks
from
the
Afghans
and
to
check
the
local
cache.
If
his
not
exist
or
expired,
it
will
download
the
layer
from
remote
registry
and
then
catch
it
catch
it
on
its
local
disk.
Then,
whenever
a
piece
of
the
layer
is
downloaded.
B
There
we
are,
they
will
be
transformed
in
the
pupae
Network
and
the
super
node.
We
are
interact
with
peers
to
replace
the
image
has
pieces
now
and
which
pieces
the
pure
should
download
after
the
peers
get
pieces
met
info
from
soup
nodes
there
we
are
download
pieces
from
the
other
peers
them.
Pieces
are
transported
among
our
peers
appears
and
all
layers
are
downloaded.
Then
the
image
since
the
whole
image
pooling
is
finished.
This
is
the
whole
process
of
pulling
image
by
drunk
flying
here
is
a
demo.