►
Description
Technical View: Comparison of Container Orchestration and Management Systems - Lei Zhang, HyperHQ
The large-scale cluster orchestration management has been evolved to a new age which is represented by open source projects like Kubernetes working with containers like Docker. This presentation will provide a detailed comparison between the design and implementation of Kubernetes, Swarmkit, Mesos, and legacy Swarm+Compose from a technical view. We will also explain the pros & cons of each system, the scenarios where users might want to use one or combination of these systems. What's more, the last part of this presentation will talk about the design of Borg and how it deeply influenced the idea of Kubernetes and other projects. We hope audiences can learn the way we figure out these differences besides the comparison result.
About Lei Zhang
Lei is a Phd candidate and the feature maintainer of Kubernetes project. He once worked for cloud team in VMware and Baidu, and now for open source community which mainly focuses on Kubernetes scheduling and kubelet on the upstream as well as maintaining HyperContainer runtime which is also known as Hypernetes & kubernetes/frakti project. As a active community advocator, Lei is the Microsoft MVP of 2016, the Top Star Speaker of InfoQ Container Conference in 2015, and once published the book "Docker and Kubernetes Under The Hood" which is the best seller of container cloud area in China.
Join us for KubeCon + CloudNativeCon in Barcelona May 20 - 23, Shanghai June 24 - 26, and San Diego November 18 - 21! Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy and all of the other CNCF-hosted projects.
A
Today,
I
would
talk
about
the
content
orchestration.
The
management
so
I'll
give
you
the
completion
from
the
technical
view
to
talk
about
from
the
implementation
instead
of
just
listing
functions
of
all
of
the
projects,
so
I'm
hurry
and
I
work
for
hypothesis
and
I'm
a
fish
maintainer
kubernetes.
So
let's
begin
our
talk.
Let's,
let's
go
about
this
talk.
We
all
our
used
three
typical
projects
here,
which
are
open
sourced
the
kubernetes
from
ACF
and
darker
from
Doc,
Corporation
and
Mase's
from
Apache
foundation.
A
So
composer
and
swarm
will
not
be
included
in
today's
station
because
those
two
projects
are
callable
agencies
so
and
on
the
right-hand.
Ladies,
always
we're
not
also
be
included,
because
if
scope
is
much
larger
than
the
the
orchestration
management,
so
let's
begin
so.
First
of
all,
let's
talk
about
idea
and
architecture
about
this
little
project.
So,
let's
begin
from
kubernetes.
In
my
opinion,
the
kubernetes
is
designed
to
help
you
to
build
right
things
with
containers
by
using
by
following
its
unique
concepts
and
accommodation.
A
So
it's
more
like
us
being
framework
in
container
ecosystem,
and
this
design
is
quite
simple,
who
is
composed
by
master
node
and
worker
nodes
and
both
of
them
nodes
are
composed
by
several
independent
binaries.
So
this
has
some
benefits.
For
example,
we
have
the
ability
of
modular
a
transparent
and
easy
to
manage,
but
also
has
some
disadvantages
like
it
will
be
a
little
bit
more
complicated
to
deploy
the
system.
That
is
this.
A
This
problem
has
been
solved
and
a
little
better
when
1.7
point
4
is
released,
so
you
have
could
be
the
main
to
deploy
the
cluster
and
another
philosophy
of
kubernetes.
Is
that
it
encouraged
user
to
use
plugins
to
do
the
network
kind
of
persistent
volume
stuff
so,
instead
of
building
something
so
it's
different
and
although
the
whole
architecture
of
crew
brain
is
I
think
is
actually
driven
by
control
loops.
A
Let's
see
why,
for
example,
if
you
want
to
create
a
podium
kubernetes,
so
you
just
say:
request
API
server
and
after
that
it
will
create
an
object.
The
report
object
in
the
LED
and
the
next
step.
It
no
need
to
call
a
scheduler
or
Kubla
to
do
something.
We
cause
scheduler
itself
is
actually
event
loop,
it
comes
it
comes.
It
keeps
it
running
to
trake
if
there's
a
new
pole
created.
So
if
you're
trying
to
bite
a
new
pole.
A
Being
the
specific
note
by
using
this
algorithm
schedule
as
schedule
algorithm
so
after
a
poll
to
the
body
to
the
node
in
turn
unto
call
the
crew
bed
again,
it
because
crippling
the
still
adjusted
running
its
own
independent
control
loop
to
check
if
there's
a
Newport
party
with
me.
So
if
yes
I
will
finish
the
information
that
pulled
to
to
this
machine,
I
start
all
the
container
inside
it.
So
that's
how
you
can
start
or
create
and
start
up
old.
A
So
after
your
pod
has
been
created,
the
controller
manager
will
be
responsible
for
making
making
sure
that
everything-
okay,
for
example,
of
the
replication
controller,
it
will
keep
shaking
your
desired
number
of
instances
against
the
real
numbers
in
your
cluster.
So
that
is
a
so-called
a
reconcile
process.
It
also
implemented
by
a
control
loop
right.
So
you
can
see
here
that
all
of
the
mechanism
of
the
controller
is
actually
fully
in
the
basic
controls
theory
which
you,
the
feedback
loop
to
decide
what
to
do
next.
A
So
it
is
the
best
model
for
all
of
the
controllers
in
kubernetes
and,
of
course,
for
the
swamp.
We
talk
straighter,
so,
let's
go
through
the
darker
part,
and
dr.
now
has
the
building
cost
of
spout
for
docker
container,
which
the
powered
by
swamp
cake
and
the
swamp
key
design
is
very
similar
to
kubernetes,
which
also
composed
by
a
master
node
and
worker
nodes.
But
all
of
those
notes,
all
the
components
all
the
nodes
are
built
into
a
single
binary.
That's
that
is
different
from
kubernetes
and
the
owner
hand.
A
It
use
a
build
in
that
store.
So
let's
begin
with
that,
be
this
architecture
when
you
create
a
service
in
darker,
so
in
the
future,
docker
will
not
encourage
you
to
create
a
single
container.
It
even
cover
you
to
create
a
service.
So
if
you
quit
studies,
that
request
will
go
be
handled
by
the
API
player
and
then
create
a
service
object
in
this
build
in
that
store,
and
these
deaths,
though,
is
actually
implemented
by
raft
protocol
and,
together
with
the
member
db2
story,
object
so
up
to
that
slum.
A
Kid
also
maintains
for
event
loops
the
first
event
to
the
actuator,
which
is
very
like
replication
controller
in
urban
area.
So
it
will
create
tasks
from
the
you're
just
creating
new
services,
and
the
task
is
actually
a
container.
For
example,
I
want
to
start
a
container.
That
is
the
task
and
after
your
altar
eggs
of
traded
credit
task,
it
will
keep
running.
The
Rican
style
persist
to
take.
A
If
there
are
right
numbers
of
red
number
of
parts
clear
running
unit
clusters,
though
it's
pretty
much
like
a
replica
controller
and
a
second
event,
looping
swamp,
it
is
allocator
it
responsible
for
allocating
Ostara
resources,
for
example,
network
to
the
tasks
to
container.
So
it
will
call
a
dark
network
api
to,
as
I,
for
example,
word
IP
abode.
The
IP
addresses
for
container
to
all
of
your
resources
at
one
of
your
services
and
tasks,
and
after
your
task
has
been
allocated,
we
network.
Let
me
it
will
be
ready
to
do
a
schedule.
A
The
scheduler,
which
is
also
invent
loop.
It
was
can
all
the
unlike
tasks
and
use
a
node
heap
to
do
this
scheduler
and
it
was
search
to
the
hip
to
fight
a
based
on
that,
which
means
your
requirement
of
your
users
constraints
and
had
light
hits
the
were
clubbed.
So
that's
how
the
algorithm
schedule
algorithm
work,
singing
some
kid,
and
the
constraints
are
very
simple
here,
just
composed
by
three
kites.
For
example,
we
need
the
node
in
ready
Staters.
A
We
need
a
node
node
with
enough
resources
and
we
also
need
to
meet
the
user
requirement
just
like
world
label
and
ylabel
director
can
do
in
kubernetes
and
next,
the
last.
Meanwhile,
an
event
loop
here
is
dispatcher.
So
when
your
task
has
been
a
lie
to
a
node,
it
need
to
know
how
to
run
a
task
on
the
note.
So
it's
a
little
responsibility
over
the
dispatcher
dispatch.
Her
is
actually
connected
to
every
node
by
using
a
long
run
in
GRCC
stream.
A
So
whenever
tasks
are
ready
to
run,
it
will
stand
your
task
information
to
this
stream
to
the
worker
node
to
to
a
darker
working
at
least
one
worker
node
and
the
work
node
also
stand
back,
hardly
to
say
yes
I'm
alive
and
the
dispatcher
will
send
tasks
to
node.
So
if
we
put
all
of
these
things
together,
we'll
have
a
some
kid
architecture,
so
we
all
see
that
in
every
work
node,
there
is
also
executor
to
run
your
tasks
to
document.
A
So
it
will
use
adapter
to
translate
your
information
in
a
task
to
dog
her
demo
MPI.
So
that's
how
basically,
how
storm
kid
works.
So,
let's
go
to
an
Mason
I
like
to
describe
myself
as
a
distributed
system
kernel,
because
it's
just
like
a
treaty
or
cluster
as
a
single
computer
right
and
it
has
originated
I
for
to
run
Big
Data
job
and
the
co
I
do
make.
A
So
they
actually
the
two-level
scheduling,
which
is
enable
the
fly
great
resource
sharing
and
by
mixed
running
multiple
frameworks
in
your
cluster,
because
you
can
see
here
that
different
frameworks
in
class,
they
actually
had
different
resource
usage
pattern.
So
when
you
mix
them
together,
you
would
have
more
possibilities
to
improve
your
result.
Inflation
see
that's
his
original
basic
idea,
mesas
project
designer
Meadows,
a
tape
called
composed
by
a
composed
by
a
master
slave
Andrew
keeper
to
a
coordination,
and
still
it
is
using
two-level
scheduling.
A
That
means
every
framework
in
missus
should
have
his
own
scheduler
and
it's
called
comes
bounding
as
killed
her
to
run
a
task.
That
also
means
it's
it.
That
also
means
it
need
to
compare.
You
cannot
compare
methods
to
cooperate
directly.
You
should
come
here,
for
example,
measures
with
marathon
with
the
kubernetes,
because
all
the
extraction
and
management
function
actually
provided
by
the
are
player,
a
player
framework
which
is
marathon
and
missus.
You
consider
whether
to
do
basically
run
her
part,
and
we
can
talk
about
it
later.
A
So,
basically,
a
very
good
example:
former
Berkeley
talk
about
how
mesas
works,
so
measures
will
keep
shaking
all
the
way
below
senior
caster
and
it
will
offer
all
of
the
available
resources
to
up
layer
up
layers
to
a
player
framework,
so
that
part
is
coded
so
code
to
the
resource
offer.
So
it
was
ten
recent
offered
with
frameworks
one
by
one
and
the
resource
offer
actually
just
looks
like
a
profile
of
your
available
sauce
of
every
node.
A
So
it
was
then
this
to,
for
example,
a
player
framework
and
to
the
a
player
framework
will
use
its
own
scheduler
to
decide
which
node
I
will
choose.
So
if
it,
except
through
this
resource
offered,
it
will
stand
the
task
to
the
compounding
slave
node.
So
the
slave
node
will
be
responsible
for
the
running,
launching
executor,
who
run
a
job
to
run
a
task
yeah,
that's
basically
how
missus
works
and
people
may
be
curious
them.
A
What
would
hit
what
marry
the
daughter
so
actually
before
1.0
release
on
missus,
we
use
a
token
divisor,
which
actually
implemented
by
amazes
dog
executors
here.
It
will
generate
tasks
from
your
dog
English
and
run
the
tasks
in
executors,
actually
translate
it
to
a
dog
API
Kodaka
demon,
but
after
one
pull,
the
arrow
makes
it
released.
Actually
it
has
its
own
message
container
wider,
which
you
can
just
imagine
like
a
replacement
of
a
Nakajima.
So
now
my
sister
not
need
to
rely
on
dark
demon
because
it
can
do
everything
the
doctor
can
do
yeah.
A
So
let's
Richard
should
check
one
here,
so
we
can
see
post,
kubernetes
and
a
swamp.
Kid
are
using
a
you.
Mental
loop
treatment
and
communities
is
using
HDD,
but
dark
zone.
Kid
actually
do
the
building
dead
store
and
both
kubernetes
and
missiles
are
now
can
be
always
multiple
container
runtime
and
multiple
can
any
images
and
pop
talked
about
dogs.
One
kid
will
stick
to
dock
Tina,
so
yeah,
that's
checkpoint,
and
when
talking
about
the
building,
that,
though
I
think
there
are
some
benefits
and
disadvantages,
it
will
be
easy
to
set
up.
A
It's
super
easy
to
set
up
and
it
has
fewer
run
trips
and
it
he
either
to
do
some
performance.
Turning,
for
example,
you
can
build
a
local
index
for
your
building
that
building
memory
dB,
but
there
also
some
ps1
interests
like
I,
think
the
most
important
thing
that
it
will.
It
is
very
hard
to
debark
its
kind
of
system.
If
you
want
to
do
a
redevelop,
you
want
to
change
in
a
datastore
of
swamp
kid.
A
Well,
data,
profit
and
then
item
information
is
that
it
is
not
easy
to
maintain,
because
it
is
the
lack
of
some
management
API
like
backup,
restore
monitoring.
Just
you
can
just
check
the
EDC,
the
administration
guide
to
say
yeah.
There
are
a
lot
of
things
missing,
mean
some
kids
data
stores.
That's
a
professor
in
to
make
a
choice
here.
The
second
thing
is
about
a
control
panel.
The
control
panel
is
actually
help
to
do
oxidation
and
management,
so
it's
defined
when
I
want
to
do
next
through
the
whole
automatic
workflow.
A
So
there
are
a
lot
of
important
things,
but
I
just
like
just
want
to
talk
about
five
of
them.
The
most
important
thing
of
a
control
panel,
I
think,
is
a
workload
management.
It's
just
like
a
hub
manager,
for
example
a
web
server
with
two
applications
and
the
abilities
Italy
deployment
to
do
that
and
in
Turkish
wonky
it
used
service
again.
A
I
need
to
say
that
darker,
we
are
encouraging
you
to
use
service
instead
with
containers
in
the
future
and
the
four
measures,
the
marathon
and
a
to
use
a
vacation
to
do
that,
and
both
kubernetes
and
Messrs
can
support
the
version
control
of
this
deployment,
but
in
talks
on
Kate
is
not
about
it
yet,
but
it
was
to
put
in
the
future,
of
course,
and
the
kubernetes
deployment
is
quite
interesting.
A
sinkhole,
for
example.
A
If
you
create
a
deployment
here,
for
example-
and
you
will
get
the
desired
number
of
the
old
deployment
and
the
carbon
number
of
your
deployment
and
the
up
to
date
means
how
many
instances
we
have
are
running
latest
version
and
the
available
means
how
many
instances
are
running
the
right
version
in
your
cluster.
So
it's
quite
simple
and
you
can
also
use
a
crew
person
he'll
edit
to
change
the
items
of
this
object.
That
means
Cooper
City
arrow
will
open
the
editor
for
you
to
open
the
object
starting
LCD.
A
So
you
can
change
an
item
value
and
if
you
save
your
modification,
it
will
trigger
a
rolly
update
automatically.
So
kubernetes
will
cure
old,
pure
old
instances
that
new
Onkyo
wong
stern.
You
want
to
do
that.
Yes,
that's
only
update
and
equals
au
state
image
and
scale
to
change
the
deployment
of
course,
and
since
it
has
a
working
honcho
of
your
deployment,
so
you
can
use
throughout
command
to
see
the
history
and
you
undo
command
to
reset
to
a
specific
version,
but
in
dog
some
kid.
A
It
is
quite
simple
now,
because
it's
in
philosophy
of
some
case
they
want
to
keep
everything
single.
So
it's
just
about
some
basic
control.
Controller
commands
like
your
own
start
and
but
it
also
supposed
rolling
update.
But
this
update
actually
because
docker
in
the
using
command
Ally
for
everything
right,
so
it
it
has
to
provide
you
some
options.
Some
duplicated
option,
like
you
may
have
environment
add,
but
you
also
have
environmental
remove
because
it
does
not
figure
out.
It
cannot
figure
out
what
you
want
to
do
in
the
update.
A
So
in
it
is
talk,
some
kid
and
in
marathon
it
is
use
application,
which
is
very
start,
updated
things,
and
is
it
only
also
support
only
update,
but
please
note
that
the
is
solely
Mrs
Mary
stone
actually
respect
the
dependency
of
applications.
It
will
only
update
your
applications,
one
by
one
in
the
right
order,
so
yeah,
it's
powerful
and
it
hosted
version
control
and
in
close
to
roll
back
to
a
specific
version
which
name
it
deployment.
A
So
any
modification
you
made
tuna
messes
up
marathon
application
will
generate
a
new
deployment
and
the
next
thing
is
sacred
management,
so
it
describes
how
we
can
store
or
equipted
credential
data
into
LCD
and
then
use
a
container
and
opposed
to
consume
it.
So
kubernetes
implemented
this
bicycle
volume,
so
in
polled
you
can
just
consume
those
sacred
credential
data
by
using,
for
example,
environmental,
libel
or
a
warning.
5:18
are
supporting
stalkers
donkey
yet,
but
it
is
under
discussion,
so
I
think,
maybe
later
and
foremost
a
marathon.
A
The
next
thing
is
scheduling
and
it
is
implementing
in
kubernetes
by
using
the
horizontal
or
poulard
scalar,
and
it
used
CPU
usage
as
a
target
resource
by
default.
But
the
sting
intimation
today
that
the
equipment
is
actually
going
to
do
a
custom
metric.
For
example,
we
can
specify
that
I
want
to
use
resource,
which
name
of
the
QPS
and
the
tablet
value
of
the
temp.
So
you
need
to
tell
kubernetes
how
can
it
read
data
right?
A
So
you
need
to
provide
a
URL
to
eight
and
also
need
to
tell
communities
how
can
it
filled
out
the
data
from
the
response?
But
actually,
if
you
are
using
things
like
a
permit
use,
it
will
become
much
simpler
because
primitives
and
kubernetes
are
all
blown
to
the
same
safe
project.
So
they
will
share
the
same
metric
data
structure.
So
you
need
you
just
need
to
provide
a
URL
and
the
kubernetes
will
know
how
to
read
metric
data
from
that
URL.
A
So
I
think
that's
a
powerful
foundation
right
and
the
awesome
kid
is
not
supported
yet,
but
still
under
discussion
and
the
marathon
visit
implemented
by
a
standalone
autoscaler.
So
it
will
read
the
internal
metric
data
from
ASOS
and
to
decide
which
one
to
use
I
do
track
the
value
through
to
do
the
autoscaler
to
do
all
the
scheduling
so
yeah,
that's
pretty
quickly
simple
and
the
last
part
is
the
state
for
workloads
and
it
is
implemented
in
kubernetes
by
Penn,
State
I
think
you
can
know
it.
A
So
in
this
case,
kubernetes
will
provide
every
Piku
bins
of
stable
membership
and
volumes.
Just
like
my
show
here
is
a
cathedral
to
custard
instances.
So
all
of
these
distances
will
have
their
own
index
unique
index
and
with
their
opposite
of
volume.
So
if
you
kill
instance
or
restart
instance,
here
it
will
start
with
its
own
index.
A
So
if
we
resource
offer
has
accepted
by
state
of
application,
it
will
not
be
send
to
other
frameworks,
even
though
the
instance
is
it
related.
So
that's
reserved
so
it.
So
that's
why
it.
You
can
use
this
part
of
resource
later,
and
this
is
it's
a
same
thing
about
today,
personal
volume.
It
will
reuse
existing
volume,
even
though
you
recreate
or
restarting
instance,
and
it
using
a
reservation
able
to
just
like
index
to
to
say
yeah
this.
A
This
poor
person
volume
belong
to
me
until
you
label
to
do
that,
and
the
third
part
is
about
a
service
discovery
and
a
load
balance.
So
basically,
this
is
very
important
in
your
system
because
you
to
know
how
to
fatty
or
your
container
how
to
file
upload,
but
in
kubernetes
this
is
all
implementing
the
bio
of
IP
tables
based
solutions.
So,
for
example,
let's
begin
from
the
internal
traffic,
your
internal
traffic
will
reach
a
proto
role.
A
A
The
this
group
of
tables
will
choose
will
rather
need
choose
a
bank
and
a
rule
in
your
cluster
that
that
egg-
actually,
this
at
least
Bank
and
a
rule,
will
be
responsible
for
peak
a
real
boat
to
redirect
your
traffic
to
a
real
boat,
the
IP
address
of
a
rowboat.
So
that
is
how
the
kubernetes
service
works.
All
based
load
based
on
IP
tables,
and
you
can
see
the
problem
because
all
of
the
IP
tables
are
created
with
all
the
node
in
the
cluster.
So
how
can
I
reach
my
portal
from
outside
the
world
right?
A
So
Furman
is
actually
provided
you
a
lot
of
solutions
and
basically
the
idea
is
saying
that
is:
you
can
assign
IP
address
which
can
route
to
any
node
in
a
cluster
and
use
that
a
IP
address
at
the
photo.
So
it
can
be
a
floating
IP
right.
It
can
also
be
the
real
IP
address
of
any
node
right,
so
yeah,
this
all
works
and
the
right
hand.
A
If
your
criminis
is
running
on
open
standard
updates
or
GCE
physical
water
I
mean
so
you
can
use
an
external
load
balancer,
so
it
will
create
you
ice
layer,
load
balancer
to
do
the
same
thing
and
there
also
a
building
ingress
here
which
should
implement
implemented
by
injects
and
Asia
epoxy.
So
if
you
use
that
just
just
like
this
example,
your
request
will
be
firstly
posted
by
this
pot
and
then
redirect
to
the
internal
service,
so
the
stings
next
are
always
them.
A
Just
as
you
have
the
alternate
protocol
here
and
all
right
hand,
since
ingress
is
the
l7
solution,
so
you
can
do
things
like
SSL
on
that
part
and
for
this
for
name
service
in
this
it
has
a
built-in
sky
DNS
support
for
to
do
that
and
I
think
there
is
nothing
special
here,
but
in
darker
some
kid
I
never
say
that
it's
solution
is
much
more
complicated.
So
please
follow
me.
Please
follow
me
here
we
can
begin
from
the
internal
traffic.
A
So
after
you
create
a
service
in
your
swamp,
kill
cluster,
it
will
create
a
contrast,
data
box
for
every
service
than
container,
so
the
standard
box
is
actually
a
network
namespace,
so
in
standard
namespace
it
will
have
IP
tables
nine.
Previous
world's
coffee.
Girls,
though,
when
you
stand
there
internal
traffic,
it
will
be
first
translated
by
internal
DNS
from
its
name
to
a
world
high
P.
So
this
word
IP
is
also
maintained
by
IP
tables,
just
like
what
kubinashi
did
and
after
that
it
will
resist.
A
Ip
tables
rules
will
redirect
your
request
to
the
IP
based
model,
which
knows
how
many,
how
many
containers
they
have
for
specific
service.
So
this
is
very
well
balanced
happens
and
it
will
pick
back
in
a
container
for
you
to
do
the
redirection,
so
it
will
use
it.
It
has
to
use
the
user
defied
overlay
network
to
do
that
to
reach
other
node.
So
this
is
internal
traffic,
but
for
the
old
static
traffic
photostatic
stays
the
dog
stomp.
It
actually
has
a
totally
independent
solution,
which
name
is
routing
mesh.
A
So
if
you
send
a
request
from
the
outside
world,
you
should
reach
upon
magic
rule
on
the
node
and
a
port
mapping
rule.
It
will
extend
the
request
to
so
called
ingress
stand
box,
and
these
ten
bucks
exist
on
every
node
in
your
cluster.
So
instead
in
that
sandbox
there
is
a
group
of
tables
to
redirect
your
request
to
the
IPS
model,
and
this
IPS
model
also
knows
how
many
containers
they
have
for
specific
service.
A
So
if
you
do
load
the
balance
part
just
like
you
know
traffic
before,
but
it
has
to
use
a
ingress
overlay
Network
here,
because
it
also
need
to
know
how
to
connect
to
each
container
running
on
any
nodes
in
a
cluster.
So
you
can
see
here
the
dark
zone.
Kid
has
to
maintain
at
least
two
overlay
networks
here,
the
ingress
wrong
and
the
UAD
far
away,
and
they
had
a
student
end
to
cosmic
sandbox
here,
one
for
ingress,
not
one
for
a
either
credited
servicing
container.
A
So
that's
the
load,
balance
partying
in
Carcassonne
kid
and
for
name
service
is
user
embedded
in
server
here.
So
any
measure
Samaritan.
However,
it
is
much
easier
because
even
using
of
actually
parks
about
based
developer,
balanced
load
balancer
here,
so
everything
just
happens
on
a
tree
proxy
site
and
for
the
external
stays
is
just
the
same
question.
How
can
you
expose
this
a
tree
proxy?
To
also
were
so
basically
so
in
most
cases,
a
public
IP
will
be
enough
and
the
phone
name
service
and
marathon
and
using
a
message
building
units
to
do
that.
A
So
it's
quite
simple:
let's
switch
to
check
pond
here
and
the
incorporate
posting
crew
blades
and
swampy.
They
are
all
using
IP
table
best
work,
IP
to
fold
other
portal
and
button
load
balance
apart.
Their
difference.
Cobra
is
a
still
using
a
tablespace
solution,
the
running
mode
I
mean
and
but
in
stalkers,
wanted
in
the
using
IP
us
to
do
not
Venice
and
for
the
external
observer,
ok,
X
nog
says
the
equipment
is
provider
UV
stable
way,
which
actually
allows
you
to
reach
IP
address
of
a
node
of
node
first
and
in
dark.
A
Some
kid,
however,
make
a
union
so
called
really
mesh,
which
we
require
the
extra
rest,
Ingres
overlaying,
the
word
for
you
to
do
and
the
mass
of
the
marathon.
The
unit
focuses
those
things
become
much
simple
and
the
Koopalings
watch
its
HDD
to
update
all
of
evitable
rules
on
the
node
visually
bound
by
could
be
proxy,
but
in
dog
some
kid.
It
is
using
underlayer
gossipy
protocol
to
to
advocate
to
the
information
about
the
network
and
rules,
so
it
donít
watch
its
rot
is
robbed
buildings
and
dead
store.
A
So
that's
difference
and
in
mesos
marathon
is
using
a
standalone.
Updater
use
the
extra
proxy
template
to
do
that.
So,
let's
go
to
the
scheduling
part
and
in
this
part
I
think
I
had
lots
of
things
who
said
because
kubernetes
actually
use
pod
as
a
schedule
in
unit
4,
but
it's
a
unique.
But
why
so?
I
like
review
this
example,
because
a
lot
of
people
want
to
run
a
java
application
right
used
to
you
need
to
package
your
welfare
to
Tom
kite.
A
So
in
kubernetes
you
cute
and
justify
this
waffle,
this
war
container
as
a
unique
container.
So
it
will
be
started
before
all
of
your
user
containers
and
you
can
ask
it
in
a
command
to
copy
or
refer
to
a
shared
volume,
and
after
that,
after
that
container
exit,
your
user
container,
which
the
Tomcat
can
just
load
it
waffle
from
the
shared
volume
and
just
audio
application.
So
you
don't
need
to
package
to
your
waffle
in
your
thumb
case
and
you
don't
need
to
use
the
person
volume
to
Stoli
Stoli
waffle
to
load
in
later
yeah.
A
So
that's
how
pole
works
is
allowed
to
do
something
which
is
called
a
so-called
content
design
pattern
to
decouple
your
containers
right
yeah.
So
that's
a
power
up
hold
and
that's
why
couplets
you
did
it
here,
actual
Asian
lesson
learned
from
pork
and
kubernetes
also
supposed
motifs
motive,
schedulers
that
we
can
specify
which
scheduler
I
want
to
use
for
this
pod
by
YouTube,
adjusted,
adding
annotation
your
llamo
in
coffee
llamo,
and
it
also
supposed
to
curious
liars
accuse
tires
which
actually,
the
co
idea
of
a
bog.
A
The
quality
of
a
bog
is
that
it
allows
you
to
mix
up.
Different
applications
may
depend
priority,
so
you
can
do
resource
preemption
here.
So
that's
one
of
the
criteria
bug
and
it
is
the
same
with
the
kubernetes.
So
it
has
three
kinds
of
priority.
Actually
and
for
example,
here
is
a
basketball
pope
which
actually,
you
can
see,
request
the
stick
to
form
a
memory,
mb
memory,
but
it's
a
limited
in
128
and
be
so
it
actually
over-communicate
64
MB
memory.
In
this
case.
A
That
also
means
this
code
can
be
killed
whenever
this
noted
under
memory
pressure
yeah.
So
that's
the
curious
tires.
Actually,
more
mobile
features
are
introduced
into
schedulers,
for
example
the
currency
cash,
which
is
what
I'm
doing
today
and
the
potable
result
boundary,
which
also
a
cool
idea
of
the
container
in
Google.
So
yeah.
That's
are
all
interesting
features
which
can
make
it
possible
for
you
to
run
your
workloads
in
much
larger
scale.
Cluster,
an
intox
own.
A
Kid
is
scheduled
to
quite
simple,
very
simple,
as
I
described
before,
and
it
is
using
the
very
simple
node
team
to
choose
the
best
node,
which
meets
your
requirements.
The
constraint,
acquirement
and,
of
course,
choose
a
node
to
be.
The
latest
like
with
the
lightest,
were
clubs.
So
it's
quite
simple
and
it
does
not
support
multi
schedule
or
accuse
tires
yet,
but
I
think
in
the
future,
swampy
scheduler
will
become
much
more
powerful
because
it
is
keep
boring
earlier
from
racers
and
incorporates
it
right
and
I.
A
It
just
just
would
select
what
you
did
even
bark
and
the
for
huge
tires
of
course,
nice
about
it,
and
actually
this
about
much
much
much
more
features
in
scary,
part
of
their
abilities
and
some
kid
like
toss,
conviction
and
data
locality,
max
min
fairness,
priority
and
or
forgiving
and
delays
delays
scheduling
so
that
that
feature
those
features
are
all
the
secret.
Why
methods
can
supposed
so
many?
A
So
many
so
many
nodes,
you
know
casters
yeah,
so
that
is
what
kubernetes
is
catching
up
and
another
point
today
that
missus
is
only
project
here
which
can
help
you
to
run
the
Big
Data
work
lives,
both
verbal
and
hey
stand
back.
Some
kid
cannot
do
it
very
well
and
the
so
late
switch
summer
part
and
the
over
case
over
user
case
in
the
hybrid
days
which
she
described
as
a
talker
down
a
right
way.
So,
in
order
philosophy,
we
trying
to
rap
over
our
cloud
container
as
a
single
container
for
girl
group.
A
Just
the
wrong
hyper
wrong
to
start
container
and
you
can
do
link
you
can
use.
You
can
attach
a
five
you
to
a
container
directory,
so
it
looks
very
much
like
poker
right,
but
actually
hybridization
the
power
to
buy
kubernetes
and
people
in
hyper
also
maintains
a
lot
of
Cobras
feature.
So,
let's
see
why
so
you
know
regular
case,
we
think
that
the
extensibility-
it
really
matters
and
you
hybridize,
that
you
are
using
the
so
called
hibernate
his
project
to
do
everything.
A
So
it
is
actually
using
hyper
container
originally
based
on
hypervisor
container
to
as
a
coffin
runtime.
So
it
is
implemented
by
the
content
or
on
camera
interface
of
a
kubernetes,
and
we
have
over
multi-tenant
network
based
on
usual.
So
it's
different
from
container
network
and
vessel.
We
have
also
custom
standard
planning,
you
safe
banking,
to
support
over
precision
water
volume
and
we
use
the
over
customer
customized
a
cherry
proxy
based
service
for
a
multi-tenant
service
in
your
cluster.
A
So
in
all
of
these
cases,
I
I
think
it's
very
hard
for
me
to
imagine.
How
can
I
implement
those
features
in
Kuala,
Lumpur,
Pakistan,
kit
or
measures,
because
it
will
be
much
harder
since
kubernetes
a
lot
of
plugins
and
its
delight?
Is
the
cup
hold
two
for
us
to
do
that?
But
it's
not
encasing
missus
and
dark
zone.
Kids,
so
I
think
the
kübra
is
it
should
extensible
and
configure
and
configurable.
If
you
want
to
do
things
like
this,
so
here's
some
summary
just
my
personal
idea.
A
So
if
you
are
an
individual
or
organization
trying
to
find
something
that
works
and
a
simple
family
and
just
works,
so
I
think
doctors
from
keep
it
really
good.
But
if
you
are
cooperation
which
has
a
Twitter
scope
cursor
to
manage,
for
example,
10,000
nodes
right
or
you
are
at
10-
you
are
10
to
you-
are
prepared
to
run
some
data
workloads
like
a
spark,
Hadoop
or
storm.
A
You
need
two
measures,
of
course,
and
but,
but
if
you
just
like
me,
just
like
hyper,
you
want,
you
just
need
an
infrastructure
layer
and
you
want
to
build
your
own
system
on
top
of
it
in
right
way.
By
saying
in
what
way
I
mean
you
want
your
system
to
be
maintainable,
to
easy
to
do.
Refactoring
and
robust,
so
I
think
that,
in
that
case,
equipment
is
the
best
choice,
because
eighty
desire
to
allow
you
to
run
things
to
build
things
on
top
of
it.
Well,
this
is
my
talk,
so
anyone
have
any
questions.
A
A
A
A
For
now
we
don't
have
to.
We
don't
have
very
complicated
network
policy,
because
we
adjusted
what
we
have
done
ingested
to
implement
alert
to
isolation.
That
is
over
for
our
network.
For
now
and
but
in
future
we
will
have
some,
for
example,
in
the
secret
policy
to
for
the
go
on
portfolio
to
limit
your
batteries
or
to
change
in
the
batteries.
But
for
now
we
don't
have
that
who
do
not
have
yet
yet
yeah.
A
I've
container
is
implemented
by
hypervisor
to
replace
the
isolation
layer
of
container
and
but
it'll
use,
docker
image
so
just
replace
the
group
namespace
by
unit
hypervisor,
but
we
still
you
dog,
docker
image
and
we
have
our
own
kernel
running
image,
which
will
be
very
small
and
very
quick.
So
you
can
start
above
hype
container
just
brings
reading
within
I
think
one.
Second,
it's
very
quick.
So
it's
basically
using
Apple
technology
but
is
not
like
a
full-blown
work
machine.
It's
just
like
a
tiny
work
machine
and
we
use
this
hyper
container
as
a
pod.
A
A
A
Okay,
so
if
there
are
no
more
questions,
I
think
this
is
the
end
of
my
more
question.
No,
this
the
end
of
my
talk.
So
so
you
know
you
know,
follow
me
through
my
email.
If
you
have
anything
and
the
question
about
the
hypervisor
best
container
and
you
can
find
my
github
ID
and
treat
hurt,
so
we
can
connect
into
me
very
easily.
Thank
you.