►
From YouTube: Deep Dive into Cloud Provider Azure - Pengfei Ni, Microsoft & Brendan Burns, Microsoft
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 into Cloud Provider Azure - Pengfei Ni, Microsoft & Brendan Burns, Microsoft
In this session, we'll dig into the implementation of the Azure cloud provider, and current work to enhance the operations of Kubernetes. Work for Kubernetes 1.17 and designs for the next versions will be discussed. There will be demos of the newest capabilities. You will also learn how the maintainers set up their development environments so you can contribute easily too.
https://sched.co/UagX
A
All
right,
I
think
it
time
for
cheese
data.
So
thanks
everyone
for
coming
to
the
tip
obsession
for
collapsible
area.
My
name
is
Pam
Feeny
I'm
working
in
Microsoft,
eager
and
I'm
many
improving
the
communities
in
current
area,
so
we're
very
glad
to
invite
Brendan
to
come
to
this
session
and
able
to
talk
about
communities
on
public
cloud.
So
let's
welcome
Brandon.
B
B
All
right
cool
all
right.
Here's
the
this
is
the
agenda
for
this
thanks,
so
much
for
coming
and
trekking
out
here
on
and
finding
it
on
short
notice
and
thanks
to
the
audio
folks
for
getting
it
set
up
on
short
notice.
Appreciate
that
I'm
going
to
talk
about
the
past
present
and
future
of
kubernetes
on
public
clouds
and
then
Peng
pays
gonna
come
back
and
give
an
update
and
demo
of
our
work.
B
Is
that
still
that's
so
coming
yeah
all
right
on
moving
the
cloud
provider
out
of
tree
and
with
a
case
engine
to
talk
about
how
you
can
get
involved
and
then
take
any
Q&A
that
that
you
might
have
all
right?
But
first
I
want
to
give
a
brief,
is
no
not
a
brief
history
of
time,
but
a
brief
history
of
cloud
providers.
B
And
the
truth
is
that
kubernetes
itself
was
very
much
built
to
be
run
inside
of
the
cloud
it
was.
It
wasn't
built
to
deploy
software
that
might
happen
to
be
running
on
the
cloud.
It
was
really
built
to
run
inside
of
a
cloud
and
in
fact,
in
early
days,
the
the
goals
in
some
way
were
to
make
the
cloud
easier
to
consume
for
developers.
I.
Think.
If
you
look
at
the
earlier
cloud,
API
s
that
existed
out
there,
they
were
very
infrastructure.
Oriented
they
said
things
like
virtual
machine
virtual
disk
virtual
network.
B
You
know
the
most
interesting
place
where
we
first
ran
into
this
was
in
the
networking
environment,
so
kubernetes
early
on
made
this
interesting
decision,
which
was
that
every
single
pod
in
the
cluster
would
have
its
own
IP
address
right
and
in
many
places
as
we
design
the
system,
we
thought
hey.
You
know
what
we
need
to
meet
people
where
they
are
and
in
a
few
places
we
made
very
opinionated
decisions.
This
was
a
very
opinionated
decision
and
it
basically
was
horrible
for
every
on-premise
network
provider
right.
B
Think
what's
also
interesting
is
that
kubernetes
itself
was
built
by
cloud
providers
right,
I.
Think
if
you
look
at
all
of
the
other
examples
of
software
in
the
orchestration
space,
it
was
mostly
built
by
people
operating
software,
not
necessarily
by
people
providing
the
cloud
and
I
think
that
this
changes
the
perspective.
B
Also,
interestingly
enough,
we
have
sort
of
flipped
around
to
a
point
where
we're
actually
mature
enough
that
we're
actually
taking
back
from
the
cloud
and
so
in
the
work
that
we
are
doing
in
cake
in
the
gatekeeper
project
to
bring
policy
to
kubernetes
we're
actually
looking
back
at
things
that
are
being
done
in,
say,
a
j''r
policy
and
looking
back
at
lessons
learned
around
our
back
around.
How
do
we
do
network
firewalling
and
applying
things
that
have
actually
already
been
built
in
the
cloud
and
API?
B
Is
that
our
customers
said
hey,
you
know
what
we
miss
when
we
as
we
go
into
kubernetes,
we
miss
having
policy,
we
miss
having
you
know,
network
network
firewalls,
bringing
those
we
miss
having
our
back
I
mean
kubernetes
was
built
without
our
back
when
it
first
started.
These
are
fundamental
characteristics
of
what
it's
needed
to
be
in
the
cloud.
B
So
kubernetes
assumes
hey
I'm
gonna
create
a
container
I'm
gonna
mount
a
disk
to
it.
I'm
gonna
delete
the
container
disk
is
gonna
move
no
cloud.
Api
was
built
for
this
right.
Every
single
cloud,
VM
API,
was
built
around
the
idea
that
you,
you
know,
created
a
virtual
machine.
You
attached
a
couple
of
disks
and
then
you
let
it
run
forever
right
and
ever
and
ever
and
ever
and
you
weren't
constantly
ripping
disks
on
and
off
of
machines.
This
is
both
in
the
hypervisor
stack
as
well
as
honestly
in
the
Linux
kernel
itself.
B
So
I,
don't
remember.
If
people
I
don't
know
how
many
people
were
out
using
Linux
when
USB
you
know
was
just
coming
in
plug
and
play
was
just
coming
in,
but
every
single
time
I
plugged
a
disk
or
a
keyboard
or
a
mouse
into
a
Linux
machine
and
then
removed
that
plug.
You
gave
it
a
little
silent
prayer
to
the
kernel
that
it
wouldn't
hang
on
you
right,
because
the
the
kernel
hadn't
been
built
for
it
right.
B
The
kernel
hadn't
been
built
for
the
idea
that
somebody'd
be
ripping
mice
in
on
and
off
a
computer
that
somebody
be
pushing
disks
on
and
off
a
computer,
and
the
same
thing
was
true
in
the
cloud
right
and
so
in
every
cloud
provider
we've
seen
challenges
with
disk
attach
and
detach,
and
you
know
honestly,
of
course,
with
the
azure
cloud
provider.
We're
still
optimizing
that
path
right.
B
Balancer
in
20
minutes
is
fine
if
you
create
it
once
a
year
and
you
never
change
that
load
balancer
ever
again,
but
in
a
world
where
services
come
and
go
where
you're
creating
load
balancers
to
do
continuous
integration
tests,
which
is
what
we
were
doing
at
the
time.
It's
just
not.
Okay
right.
We
there's
this
expectation
around
a
very,
very
dynamic
API.
So
if,
if
cloud
was
a
little
bit,
dynamic,
kubernetes
was
really
dynamic
and
I.
B
Think
what
we've
seen
over
the
last
five
years
is
that
cloud
providers
themselves
have
had
to
step
in
as
no
exception
cloud
providers
have
had
to
step
up
and
improve
the
way
they
deliver
api's.
But
the
good
news
for
the
cloud
providers
is
actually
that
everyone
else
is
going
there
too.
It
wasn't
just
not
just
kubernetes
that
is
driving
this
demand,
although
kubernetes
drives
a
very
significant
fraction
of
all
API
calls
that
end
up
on
Asscher
and
probably
on
other
clouds
too.
B
It's
also
actually
the
same
thing
that
other
people
are
doing
when
they're
doing
continuous
delivery
systems
when
they're
doing
all
sorts
of
other
kinds
of
online
activities.
The
control
plane
for
the
cloud
is
actually
no
longer
really
the
control
plane,
it's
really
becoming
as
essential
as
the
data
plane.
But
as
we
push
forward
into
cloud
providers,
there
is
actually
even
more
problems.
So
the
next
thing
we
ran
into
was
that
not
all
clouds
look
the
same
like
you
squinted
the
clouds
and
you're
like
well
yeah.
B
If
what
you
were
expecting
was
an
IP
address,
suppose
you
were
programming
DNS
to
do
you
know.
Suddenly
you
have
to
use
a
cname
record
instead
of
an
a
record
in
order
to
program.
Dns
and
again,
kubernetes
has
been
sold
on
this
notion
that,
like
it's
portability
right,
it
gives
you
portability
and
abstraction
above
the
clouds,
but
the
clouds
have
this
way
of
leaking
up
and
leaking
through
into
the
cloud
provider
itself.
Another
example
is
things
like
ipv6
right,
Cal
just
gave
a
great
talk
about
ipv6
and
dual
stack,
support
and
kubernetes.
B
Not
every
cloud
provider
supports
ipv6
Asher
does,
but
not
every
cloud
provider
does
and
I.
Imagine
as
they
do
it's
going
to
look
a
little
different
here
and
there
and
so
I
think
these
are
the
places
where
you
know.
We
continue
to
see
challenges,
trying
to
figure
out
what.
Where
do
you
draw
the
abstraction
line?
And
how
do
you
say
you
know
what
it's?
Okay,
that
there
are
these
little
differences
and
yet
still
give
people
a
single,
unified
interface?
B
They
can
program
against
disks
or
another
great
example
right,
since
some
places
disks
can
be
attached
right.
Only
our
disks
have
to
be
attached
exclusively
to
one
machine
other
places
the
disk
can
be
attached
to
read-only
to
a
bunch
of
machines
that
has
changes
the
way
people
build
and
lay
out
their
kubernetes
clusters.
B
So
there's
all
of
these
challenges,
as
you
start
looking
at,
how
do
I
build
a
helm
that
can
run
in
every
single
cloud
if
there
are
these
different
characteristics,
depending
on
the
cloud
right,
so
these
are
all
places
where
this
notion
that
you
could
have
a
cloud
provider
abstraction
and
just
have
everything.
Magically
work
have
sort
of
broken
down
historically
over
time,
then
another
the
one
of
the
biggest
problems
that
that
we've
run
into
lately,
both
in
the
azure
cloud
provider
and
honestly
elsewhere,
is
throttling
right.
B
Every
single
we
have
run
users
out
of
API
quota
on
every
single
cloud
provider
out
there
right
and-
and
you
know
over
time,
I
expect
this
is
this
is
a
recurring
theme
until
the
clouds
actually
become
to
come
to
a
place
where
they
can
provide
effectively
infinite
capacity.
Right
again,
what
you're,
seeing
in
the
cloud
is
that
the
cloud
was
built
for
people
using
CL
eyes
to
occasionally
create
resources.
B
There
is
this
problem
in
kubernetes
itself,
which
is
that
the
cloud
provider
code
is
deeply
attached
to
the
kubernetes
binaries
that
are
built.
This
means
that
improvements
to
the
cloud
provider
like
we've
built
for
the
Azure
cloud
provider
are
tied
to
kubernetes
releases.
It
means
that
we
have
to
cherry-pick
fixes
into
an
old
version
of
kubernetes
in
order
to
make
those
fixes
available
to
users
on
the
agile
kubernetes
service.
It
means
that
we
can't,
you
know,
do
weekly
releases.
We
do
a
weekly
release
of
aks.
B
We
do
a
weekly
release
to
the
azure
kubernetes
service,
but
we
can't
do
a
weekly
release
of
our
cloud
provider
code
right.
So
we've
been
limited
in
our
ability
to
be
agile
by
the
fact
that
this
design
decision
was
made
to
deep
too
tightly
couple,
the
kubernetes
codebase
with
a
cloud
provider
codebase.
It
was
the
right
decision
at
the
time
right.
We
favored
install
simplicity,
we
favored
the
idea
that
you
could
install
one
binary
and
it
would
all
just
work.
We
could
distribute
one
binary
and
it
would
work.
B
I
think
that
the
most
interesting
thing,
the
consequences
of
what
we've
found
is
that
in
the
last
five
years,
we've
gone
from
the
clouds
saying.
Why
are
you
treating
our
api's
in
this
way
to
people
actually
building
features
within
clouds,
specifically
around
capabilities
of
kubernetes?
So
you
know
just
as
a
concrete
example,
we
have
as
your
networking
building
new
features
for
us
to
light
up
private
clusters
for
as
recruit
for
the
azure
kubernetes
service
right.
So
we're
getting
to
a
place
where
supporting
kubernetes
in
Azure
is
a
first-order
requirement
for
our
API
teams.
B
We
have
disks,
we
have
people
working
on
storage
and
disks
saying.
What
do
we
need
to
do
to
make
disks
that
are
natural
for
cloud
native
micro
services?
What
do
we
need
to
do
to
improve,
attach
and
detach
times?
And
things
like
that?
So
we've
gone
through
this
amazing
transition,
I
think
not
just
in
Azure,
but
really
in
all
of
the
clouds,
where
there's
a
realization
that
supporting
these
patterns
is
not
something
that
we
do
sort
of
grudgingly,
but
actually
something
that
we
do
in
tight
conjunction
in
partnership
with
the
teams.
B
We've
seen
a
really
great
partnership
develop
between
the
virtual
machines
scale
sets
team
and
the
agile
kubernetes
service
team,
delivering
together
solutions
for
kubernetes
that
work
well
for
for
our
customers
and
then,
finally,
we're
actually
seeing
the
cloud
do
innovation.
So
recently
we
announced
this
service
called
teleport.
B
So
that's
I
think
a
great
example
of
how
we're
not
just
in
in
our
API
is
but
actually
innovating
in
partnership
with
Azure
storage
in
our
infrastructure,
to
actually
change
the
way
that
people
experience
kubernetes
and
that's
something
that
you
know
would
have
not
existed
three
or
four
years
ago.
I.
Think
the
other
interesting
piece
of
this
is
that
the
cloud
is
also
starting
to
build
on
top
of
kubernetes.
B
So
when
you
look
at
things
like
the
spring
service
that
was
recently
announced
in
Azure,
there's
a
spring
boot
platform
where
you
can
bring
a
Java
app
have
a
spring
experience.
It's
built
on
top
of
aks,
it's
assuming
kubernetes
as
a
building
block
on
which
we're
gonna
deliver
higher
level
services.
Similarly,
with
kubernetes
at
the
edge
sequel,
server
is
saying:
hey
if
you
want
to
run
a
big
data
version
of
sequel
server.
B
A
A
So
Brandon
just
mission,
then
to
all
other
cloud
providers
now
working
on
the
project
called
Auto
chiclet
provider.
So
what
is
auto
chief
was
the
difference
between
the
entry
and
the
out
she
color
provider.
Here's
a
simple
diagram
show
than
the
above
one
is
currently
unit
rate
cloud
provider
and
is
a
blowing
it
out
of
shape.
A
So
when
you're
drunk
clutter
on
the
a
group,
you
would
configure
the
cloud
provider
to
agra
for
cube
later
and
the
controller
manager
with
auto
trade
cloud
provider,
we're
moving
those
components
into
different
pine
arrays
and
the
rania
a
different
part
on
in
cluster.
So
so
now,
with
auto
jacob
later
and
the
COO
controller
manager
would
run
with
cloud
providers
that
you
external
and
we
would
run
a
separator
control
and
component
called
controller
manager,
so
which
would
be
in
charge
of
no
do
lifecycle
the
service
controller
and
also
the
root
controller
for
position
volume.
A
Plugins
are
this
kind
of
foul.
We
would
move
to
CSI
drivers,
so
two
new
components
would
be
introduced
to
to
build
the
safe
drivers.
Folder
for
credential
providers
see
cloud
provider
is
working
on
capture.
Extractor
is
also
from
the
Cuckoo's
Nest,
so
it
is
not
ready
yet
also.
The
last
common
component
is
about
the
cloud
and
node
controller,
so
this
is
what
we
are
working
on.
Key
yield,
the
Intimidator
service.
She
graduate
initialized,
the
node,
so
we
have
created
a
cap
and
have
been
discovered
in
the
cloud
provider.
A
We
have
also
implement
the
field
stay
in
club
for
our
area,
repo.
We
have
the
integrated
under
which
occurs
in
joined.
So
then
I
would
demo
template
her
for
safe
drivers,
we're
also
working
on
our
disk
and
file
and
blob
fields.
Their
drivers.
They
have
been
rich
beta
version
on
Linux
knows,
and
they
shall
be
incurred
here.
The
Internet
has
engine
for
Windows
node.
There
are
some
issues
and
we
are
working
on
for
external
credential
provider.
There
is
a
cap
already
proposed.
A
A
Okay,
because
the
provision
collect
provision
would
require
a
lot
of
time,
so
I
have
already
provision
collector,
but
how
I
would
show
you
how
how
we
showed
clutter
controller
manager,
so
it
simply
enable
the
yield
Cloud
Controller
manager
in
the
occurs
in
your
manufacturer,
and
we
also
can
take
the
profiles
for
master
and
in
node.
So
it
is
simple
enough
clutter
than
with
one
master
Aiden
post
each
ending
pose
would
contain
one
node.
A
So
this
is
the
notes
we
have
cards
for
debt
collector,
okay,
we
have
two
nodes
and
the
wine
map
node
running
in
lid
clutter,
and
we
also
have
deployed
a
multiple
components
on
this
clutter.
So
we
could
check
the
post
on
date
clutter.
So
you
would
say
that
we
have
already
deployed
the
mat
component:
the
API
server,
the
controller
manager,
the
scheduler,
as
we
also
enabled
out
chiclet
provider.
We
also
deployed
the
Cloud
Controller
manager
and
also
the
cloud
node
manager
Kim
aside,
how
inch
node
for
proceeding
volumes.
A
We
also
deployed
the
safe
drivers
about
a
disc
and
a
nail
file.
So
this
is
what
we
need
for
our
tree
cloud
provider,
okay,
so
so
to
attempt
treat
how
the
node
provisioned
and
how
noda
is
initialized.
Let
me
first
delete
a
component
called
cloud
in
the
manager,
so
I
have
a
man
factor
for
the
collateral
manager.
It
generally
give
me
started
running
on
each
node.
It
will
be
tolerant
to
the
no
executive
and
no
scheduler
time,
so
we
could
delete
it.
Frost.
A
Okay,
the
dim
insider
is
now
removed,
so
now
let's
get
the
clutter.
So
this
is
the
IRA
resources
after
I
deployed
the
clutter.
So
a
lot
of
resource
I,
let
you
here
so
the
virtual
machines
formatter,
the
Bush
emergence
cassette
for
two
different
node
poles,
the
papaya
P
addresses
as
well
as
other
results
such
as
security
groups,
the
Routh
table,
the
version
ID
host
and
so
on,
so
G
create
a
new
node.
It
is
simple
enough
to
kill
the
bush
a
minute
ago,
sized
with
new
young
teens.
A
Okay,
then
we
switch
to
the
intently
to
we
would
say,
sign
a
new
node
redeem
provision
state,
so
the
new
version
machine
creating
would
require
about
two
minutes.
So
we
need
with
a
wire
for
four
days
while
we're
waiting,
let's
check
how
the
components
employed
with
indeed
clutch
how
they
are
configured.
So
this
is
a
great
code.
I
have
already
eaten
started
the
communit
plugin.
So
then
we
can
interact
with
the
cluttered
directly
within
the
best
coder.
So
we
have
a
new
knowledge
hero.
A
A
Okay,
then,
let's
check
the
config
options
for
coop
controller
manager.
We
would
say
we
would
say
the
into
the
coop
controller
manager
is
running
the
club
provider
with
external,
so
this
means
separate.
The
controller
manager
is
required
to
configure
the
club
provider
back
to
agra.
So
this
is
clutter
control
man
in
pod
here,
so
it
have
a
config
to
the
cloud
provider
to
agra,
as
well
as
other
options
such
as
cloud
a
config
file.
A
A
Now,
let's
see
how
how
could
we
use
the
safe
driver,
og
creative
new
powder
with
position
volume,
so
I
also
have
a
demo
file
here
so
to
use
the
Tetra
world.
We
first
define
new
storage
class
undecided
provisional
to
disk
touch
the
SI
that
editor
come.
So
this
is
different
from
the
current
default,
a
unit
rate
or
disk
file,
a
dickster
class,
and
then
we
created
previously
referring
this
third
class
created
powder
reference,
this
persistent
volume,
so
within
this
time
powder
I
also
created
a
unique
container.
A
A
People
created
so
we
just
have
switch
to
the
namespace
she
could
system.
So
let's
switch
to
the
default
okay.
Now,
let's
create
the
dress
acid
and
we
would
fun
design
to
the
pv
say
has
been
in
prevent
prevent
data,
and
if
we
wait
a
while
a
TV
would
be
creative.
You
know
Phil
on
second
and
then
the
PV
switch
to
the
penetrator
away
switch
to
the
powder.
Then
the
powder
with
the
data
to
initialize,
so
the
initialized
is
mounted
the
L
disk
is
look
of
a
system
and
write
the
index
with
two
entire
disk.
A
A
So
we
would
fund
it
under
the
new
volume
here
is
mounted
in
this
container
and
we
also
deployed
a
surveys,
so
the
new
Papa
P
should
be
allocated.
If
we
access
this
popularity,
we
should
fund
it.
The
output
should
be
our
disk.
So
this
is
the
simple
timer
running
in
Jenks
with
air
discourteous
travel.
A
A
A
Okay,
so
for
this
node
there
are
some
times
on
this
lieutenant
Carter
club
provider
and
initialized.
This
means
the
node
have
been
created
and
have
been
joined,
the
clutter,
but
it
is
actually
not
ready
yet
because
the
cloud
provider
doesn't
initialize
the
88,
so
we
need
a
different
component
to
initialize.
This
node
98
were
to
cloud
the
node
manager
wrong.
So
let's
create
the
Cardinal
man
injured
by
cocaine.
A
A
Now
we
said
so,
it
is
adding
the
new
node
labels
with
the
intent
type.
It
also
initialized
the
node
with
the
first
domain
and
a
region
it
also
initialized,
the
node
weight
is
provided
in.
So
if
we
back
to
the
node
list
and
the
Refresh
is
manifest,
we
would
find
align.
The
new
labels
are
added
in
node
labels
and
also
the
property
had
been
initialized,
and
there
is
no
x
on
this
node.
So
now
it
is
actually
fully
ready
to
run
the
club.
A
Okay.
This
is
a
simple
demo
for
the
auto
tree
provider,
so
some
of
your
main
considering.
So
how
could
you
get
involved
if
you
have
some
Ezio's
or
have
some
proposals
or
have
some
features?
There
actually
are
several
way
to
do
this,
so
you
could
join
the
meetings,
regular
meetings,
you
could
join
the
communities
like
in
China
provided
area.
You
could
cut
involved
in
discussion
about
crafts
in
the
club
provider.
You
could
also
grab
Ezio's
from
the
project
border.
We
maintained
in
the
Kunitz
project.