►
From YouTube: sigs.k8s.io/kind - 2019-05-06
B
A
C
C
D
C
A
C
Finish
this
is
the
part
that
I
I
don't
get
this,
who
sent
these
to
the
to
the
continuity
in
the
cube.
The
cube
Latino
right-
and
this
is
the
part
that
I'm
missing
is
what
part
of
the
code
is
sended
this
option,
but
is
introducing
this
option
in
in
the
doc
regime,
is
Dedecker
seeing
folded
the
same
force
she
arrived
or
might
have
been
because.
A
C
A
A
Right
we
should
not
have
but
I
imagine
we
can
we're
probably
just
need
to
find
like
where
it
does
this.
There
should
be
it
like.
It
should
be
doing
this
that,
like
there
should
be
compatibility.
If
not,
maybe
that's
a
we
fix
okay,
so
we
also
have
so
so
there's.
That,
though,
is
that's
the
one
thing
we
need
to
figure
out
for
kind
net
to
be
an
option
is
figuring
out
how
we
make
sure
that
the
IP
ranges
is
being
passed
well,.
C
C
D
C
Installing
a
socks
proxy
in
one
of
the
nodes
allows
you
to
to
is
a
socks
process
is
connecting
parts
is
a
switching
magnet
of
course,
so
you
stole
a
socks
proxy
in
one
of
the
nodes
you
forward
that
part
and
in
your
process
you
can
reach
everything
you
can
reach
service
IP
the
the
engineers
in
Greece,
the
ball
sports
everything.
So
that's.
D
A
C
C
A
A
So
I
spent
some
time
not
as
well
I
think
that
I
think
we
can
pretty
easily
allow
people
to
for
text
reports
like,
technically
speaking,
that's
a
pretty
simple
thing
to
support.
I
think
that
tricky
one
is
how
we
make
it
predictable,
like
which
note
the
court
is
on,
like
like,
like
after
I
run
kind
of
late
cluster.
How
do
I
know
where
that
port
is
first
single
note?
It
I
mean
it's
pretty
obvious.
A
C
A
C
Thing
is
that
that's
why
kubernetes
has
all
the
services
and
all
these
things,
and
all
these
built
or
eyepiece,
you
know
if
you,
then
you
start
to
to
redo
what
unit
is
deep
to
avoid.
At
the
same
this
way,
I
I
thought
in
another
approach
with
it,
because
you
mentioned
this
is
eighteen,
and
you
just
start
to
forward
parts.
The
cube
City
Airport
forward
is
no
it's
more
or
less
what
what
you
are
going
to
get.
What
is
the
difference.
A
The
other
one
and
so
I,
don't
know
if
there's
a
priority
in
communities
to
make
it
work
better.
The
other
one
is
that
if
we
have
a
can,
if
we
did
go
eventually
the
route
of
having
like
a
controller
doing
this
and
all
of
that
complexity,
the
upside
we
could
gain
is
that
we
could
actually
say
like
here's,
the
load,
balancer,
IP
and
everything,
and
then
tooling,
this
just
expected
to
work
with
communities,
resources
but
already
work.
A
So
if
I
have
some
tool-
and
it
like-
looks
up
the
the
load,
balancer
IP
and
then
contacts
it,
oh
if
we
can
actually
assign
an
IP,
listen
on
it
and
then
update
the
in
clustered
resource.
For
that,
then
that's
like
easiest
to
integrate
with
I,
don't
think
I
can
say
like
first
solution,
that's
a
like
that
would
be
cool
someday,
because
then
you
could
actually
implement
type
load.
Bouncer.
Oh.
C
C
A
A
Thing
that
we
were
looking
at
the
might
be
interesting
and
I'm
not
sure
how
easily
you
could
solve.
So
this
was
like
if
we
had
a
person
tied
again.
This
is
like
not
something
I
want
to
build
right
now.
This
is
a
like.
It
might
make
sense
in
the
future.
It's
significantly
more
complicated
than
just
for
the
importer
or
setting
a
socks
proxy
or
something.
C
C
C
A
C
A
A
C
A
A
F
A
Is
an
easy
problem:
we
can
run
any
of
the
existing
out-of-the-box
load
balancers.
They
should
be
fine,
but
if
one
actual
load
bouncer
that
you
can
talk
from
the
host,
which
I
think
it's
more
likely
if
you're
going
to
do
type
load
bouncer
to
begin
with,
then
you
need
to
get
that
traffic
from
the
host
into
the
cluster.
A
A
That's
where
I
would
I
don't
think
this
is
a
this,
isn't
it
today.
This
is
just
a
while
we're
talking
about
the
you
know,
just
node
port
options
or
whatever
the
this
is
solving.
This
problem
is
something
I'd
like
to
eventually
solve,
because
we
want
to.
We
want
to
emulate
as
much
as
possible.
Ideally,
we
can
get
dynamic,
very
imprisoning
and
CSI
drivers
bouncer
and
like
we
can,
we
can
mimic
as
much
as
possible
and
that
one
is
solvable.
I
just
got
it's
a
lot
of
reinventing
it's
a
bit
overly
complicated,
but
I.
A
C
A
A
C
That's
pretty
simple:
it's
the
only
announcing
routes
or
or
announcing
an
evening
with
the
art
to
fake
the
IP
address
there.
But
the
thing
is
with
the
Mac
or
the
Windows.
You
have
to
fake
the
IP
address
for
the
course
and
then
translate
that
and
turning
it
to
the
crustal
that
is
inside
the
kernel.
One
way.
A
So
so
far,
the
most
straight
for
suggestion,
I've
got
from
anyone
is
just
add
an
IP
into
an
interface.
It's
like
they
have
a
pool
of
IPs.
We
select
one,
we
edited
the
interface
and
then
instead
of
routes
you
can
actually
just
bind
a
get
to
that
IP,
and
then
we
just
need
to
get
that
traffic
into
the
cluster
and
that
that's.
D
C
A
Right
but
I
think
that's
worse
than
the
we
just
have
a
little
daemon,
listen
because
then,
as
far
as
you're
concerned
like
like,
we
can
just
make
this
look
like.
Will
you
SPECT
so,
if
I'm
running
in
the
cloud
on
kubernetes
I
expect,
when
I
create
type
load,
balancer
that
some
machine's
gonna
come
up
somewhere?
It's
not
my
concern.
It's
gonna
have
an
IP
and
when
I
hit
that
IP,
it's
gonna,
the
traffic
whines
open
the
cluster
at
my
service.
So.
A
A
But
ideally
we
could
avoid
creating
like
for
each
of
these.
The
other
option
would
be
something
like
we
try
to
create
a
container
and
bind
it
to
that
IP
address
or
we
try
to
like
get
the
like,
like
a
load.
Balancer
proxy
type
container
bound.
All
these
but
I
like
again
I
would
come
back
to
that,
because
I
think
more
people
are
clamoring
for
a
node
port
right
now
from
the
bouncer.
A
I
mean
it
would,
it
would
be
a
nice
feature,
but
it
would
also
be
one
that
I
would
probably
say
we
should
think
about
it
right
and
and
come
back
to
later.
It's
just
like
well
we're
looking
at
all
these
networking
options,
it
would
be
good
to
kind
of
keep
ideas
for
that
in
mind,
so
we
don't
more
or
less
build
it
twice
now
socks.
Perhaps
the
idea
kind
of
reminded
me
of
that,
because
that
gets
awfully
close
to
like
you.
A
A
A
Many
to
kind
of
had
something,
but
it
gets
to
cheat
a
bit
because
they
only
have
one
node,
so
it
just
it
just
they
just
make
routes
to
another
I
believe
and
and
they're
right
and
Mariette
their
control
and
they're
running
at
p.m.
so
they
have
an
IP
for
the
VM
and
they
just
like
maker
outs.
Do
it
I'm
not
aware
of
any
other
ones,
for
this,
like,
like
the
other
load
balancer
like
they're,
their
actual
criminals
load
balancer
solutions,
but
those
are
not.
A
A
Also,
don't
think
we
actually
have
to
reinvent
that
much
I
mean.
Basically,
we
like
listen
to
service
and
update
it.
That's
a
great
for
one.
We
can
use
standard
controller
stuff
for
that.
Creating
an
interface
should
alert
an
IP.
An
interface
should
be
super
complicated
and
like
listening
to
a
port,
I
mean
sure
it'll
be
pretty
similar
to
cute
cuddle
port
for
it,
but
it
like
a
cheap
way
that
we
can
wrap.
This
would
actually
be
okay,
your
IP
is
always
localhost
and
we
do
tube
cuddle
port
forward.
A
The
problem
with
that
is,
then,
your
IP
address
won't
work
in
the
cluster.
If
we
just
make
this
like
it's
a
controller
and
all
it
does
is
assign
one.
So
that's
the
other
one
that
I've
seen
and
that's
what
Kay
native
is
using
right.
Now
they
install
routes
on
hosts
to
the
mini,
cube
p.m.
and
then
there's
a
daemon
that
just
watches
for
services,
and
it
always
assigns
the
VM
IP,
but.
C
D
A
I
mean
they're
both
related
different
problems.
The
new
port
problem
could
also
be
done
by
if
we
can
figure
out
a
way
to
make
it
like,
like
just
the
UX
problem
of
like
how
do
I
know,
which
nodes
have
this
poor
veil,
like
if
I'm
scheduling
my
kerbin
any
Zappa
and
and
I
have
a
multi,
node
kind
cluster,
and
it's
listen
with
like
a
host
port
like
I'm
gonna
set
up
in
ingress
their
ingress
engine
acts.
As
my
solution,
I
need
to
make
sure
it
lands
on
the
that
have
actually
afforded
the
port
to
those.
A
A
C
C
A
C
A
A
C
C
A
C
Depe
Dukan
did
you
have
the
load
balancer?
That
is
a
matrix
of
connecting
ports.
You
know
it's
just
you
say:
I
want
to
I
suppose
this,
the
service
in
the
poor,
with
an
old
port
with
30,000
whatever.
So
you
have
the
this,
every
sport
you
have,
the
service
port
is
posted
in
the
cluster.
You
have
the
load
balancer
and
you
have
the
forward
with
pork.
You
just
need
to
connect
those
ports
and
send
to
whatever,
not
in
the
question.
C
A
C
Let's
call
the
the
port
forwarded
knows
to
him.
Do
you
have
the
configuration
with
all
the
ports
that
you
want
to
forward
there?
The
thing
is
that
you
need
to
restart
the
node
every
time
that
you
want
to
add
an
you
know,
a
new
port.
You
know,
but
you
can
keep
restarting
the
node,
because
the
IP
address
of
this
of
the
node
is
in
decimator.
C
C
The
only
way
to
do
that
thing
is
that
I
was
testing
that
and
in
work.
It
is
you
need
to
create
a
special
node
and
keep
it
through
starting.
You
want
to
add-
and
you
know
you
restarting
just
post
new
ports
and
you
configure
the
node
with
the
new
for
port
forwarding
and
that
works,
but
we
need
to
add
all
this
to
all
this
configuration.
A
C
A
It's
I
mean
I,
definitely
think
that
all
these
are
less
powerful.
Although
plants
are
thing,
but
they
can
be
done
a
lot
more
simply
and
it
might
be
like
a
step
towards
it,
especially
if
we
can
figure
out
I,
don't
think
any
of
this
loop,
too
close,
yet
you
might
be
able
to
after
creating
the
IPS.
Just
let
the
docker
container
bind
to
the
IP,
and
then
we
don't
have
to
run
it.
We
don't
we
don't
run
a
daemon
on
the
posted
darker.
A
A
C
D
C
B
C
A
Right,
so
if
we,
so
what
we
might
be
able
to
do
is
have
the
command
stateless,
Li
ish,
add
the
add,
like
a
load,
balancer
IP
from
some
pool,
the
user
specifies
with
some
default
it's
in
private
IP
space
and
then
update
the
date.
The
port
forwarding
container
to
bind
the
port
to
that
IP
and
then
updated,
and
then
we
just
might
want
to
allow
running
a
controller
so
that
you,
instead
of
invoking
this
command,
you
invoke
at
once
to
be
running,
and
then
you
create
type
load,
balancer
I.
E
A
A
Yeah
I
mean
so
so,
but
I
don't
think
it
would
be
that
big
of
a
deal
to
have
to
have
a
little
like
kind
sub-command.
That
is
a
that
is
a
controller
that
runs
against
your
cluster
kind
of
won't
have
any
problem
obtaining
a
queue
config
or
running
against
the
cluster
monitoring
for
service
changes
then-
and
you
would
only
need
to
run
this
if
you're
trying
to
have
like
load
balancer
support
so
instead
of
invoking
cube
cuddle
port
for
it.
You
would
invoke
this
thing
to
manatee
or
like
ingress,
and
it
would
manage
a
container.
B
C
A
A
It's
confusing
ascribe
it
that's
to
get
back
to
localhost
on
them,
yeah
Pacific,
to
docker
for
Mac
in
my
cell.
That's
another
thing
that
I've
intentionally
not
leverage,
because
that
solution
doesn't
work
for
Linux
like
you'll,
get
the
containers.
You
can
only
get
on
Linux,
you're
kind
of
stuck
if
you
to
talk
to
things
that
are
actually
on
the
host
loopback.
A
A
Let's,
let's
play
around
with
the
new
port
options
and
I
will
also
look
towards
a
little
bouncer
one,
but
definitely
I'm
not
trying
to
build
that
one
today,
I
think
it's
more
pressing
to
solve
the
node
port
stuff
and
I
like
I,
do
want
to
look
somewhere
and
see
what
we
can
do
about
the
complexity,
but
I'm
not
I'm,
mostly
worried
about
just
making
the
container
reliable
and
trying
to
avoid
managing
a
bunch
of
extra
state.
I'm,
not
super
worried
about
like
needing
to
run
a
controller
there,
that's
pretty
straightforward
and
like
right
now.
A
A
They're
not
typically
run
as
a
service
but
I'm
again,
I,
don't
I
hope
you're,
not
people
are
not
doing
that.
Kinda
know
we
had
somewhere
the
weekend.
That
was
doing
that,
but
the
intention
would
be
something
like
in
your
CI
job.
You
background
this
task.
That
is
doing
the
network
for
you.
If
you're,
if
you
need
that
functionality
and
if
you
don't
need
this
functionality,
you
don't
pay
for
it.
You
don't
ride.
E
Have
a
question
on
the
APR
about
container
D:
have
you
some
media
when
it
with
met
with
America
sure
what
we
cut
out
the
PR
about
container
D?
Have
you
an
idea
when
it
will
be
mad
in
will
emerge?
It
is
merged.
It
is
method,
ready.
Okay,
great
Amy
said
that
and
okay,
so
I
just
look
at
outer
Ibiza,
kinda,
roundup,
Ted
and
I.
Think
that
maybe
not
now
because
but
am
I
in
coop
economy.
We
can
talk
about.
E
E
Yes,
he
started
from
from
the
Tartar
balmy
beans,
the
idea,
but
now
it
is
much
more
flexible
because
we
can
add
more
release
of
kubernetes
into
the
same.
We
can
add
more
images,
for
instance,
I
want
to
bring
in
antiques,
because
I
always
use
and
jigs
I
want
to
bring
it
to
bring
in
some
additional
things
in
image.
So
when
the
container
start,
everything
then
goes
cause
motor
I
don't
have
to
download.
This
is
really
interesting
or
I
want
to
override
a
specific
binary,
because
I
want
to
test
the
skew
and
whatever
so.
E
A
So
we
we
have
switched
instead
of
adding
image
stars
now
we're
actually
able
to
import
them
into
the
content
store
directly,
which
is
more
space
efficient
and
time
efficient.
Yes,
I
have
to
check
this
one
well.
Well.
We
should
definitely
think
upon
that,
because
some
of
this
I've
actually
like
been
perfect
because,
like
I
the
Tarbell
build,
if
nothing
else
is
definitely
one
that
I
want
to
get
in
and
I
want
to
start
reworking
that.
A
But
a
lot
of
it's
been
blocked
on
needing
to
make
the
breaking
changes
to
the
note
image
and
I
spent
way
too
much
time
trying
to
make
that
not
breaking
and
I
got
it
working
for
like
114
Ford,
but
I
ran
into
too
many
problems
where,
before
114,
we
don't
auto,
detect
things
and
and
I've
made
it
breaking
it.
Moving
and.
E
A
The
main
concern
I
have
is
actually
more
along
the
lines
of
like
we
like.
If
we're
having
one,
that's
actually
integrated.
We
need
to
manage
things
like
that.
He's
face
it
I'm,
trying
to
avoid
us
doing
that
right
now,
so
I'm
actually
kind
of
leaning
towards
I
want
to
figure
out
the
problems
with
kind
net
and
offer
that
and
then
additionally
offer
an
option.
Where
kind
doesn't
like
you
set
some
field
and
we
don't
manage
CNI,
we
give
you
a
cluster
without
CNI
and
we
just
say
cool.
You
want
to
try
different
CNI
stuff.
A
You
know
your
way
around
see
and
I
just
installed.
This
unite
do
ever
and
it
for
most
people
that
should
just
look
like
to
cuddle,
apply,
calico,
manifest
or
whatever,
and
they
might
need
to
tweak
it,
and
we
might
have
some
like
guides
or
docs
for
that.
But
then
you
know
instead
of
shipping
it
and
shipping
in
the
binary.
We
can
just
ship
a
guide
and
you
might
continue
to
keep
the
the
surface
down.
A
I
think
if
we
ever
get
to
the
point
where
we
have
a
lot
more
people
working
on
this
whole
time,
then
like
maybe
we
should
go
full
mini
cube
and
say
we
offer
the
kitchen
sink,
but
I
think
right
now,
we're
probably
best
saying
whatever
this
is
for
people
that
are
doing
advanced
development
it.
You
know
we
offer
a
working
cluster
out
of
the
box,
but
if
you
want
to
fiddle
with
something
here's
a
switch
to
be
able
to
fiddle
with
it.
C
E
Yes,
basically,
it
is
an
alternative.
Now
it
is
an
alternative
or
a
complement
to
then
they
build
no,
the
image
that
it
isn't.
But
what
I
was
saying
so-
and
there
is
a
there-
is
some
overlap
that
I
would
like
talking
up
and
but
what
I
was
saying
it's
something
different
is
that,
while
building
this
feature,
I
I
face
the
problem
that
I
need
to
to
have
some
information
that
moves
from
the
build
phase
to
the
cluster,
create
phase,
because
I
didn't
want
to
ask
the
information
two
times
to
the
user.
E
And
this
allows
me,
for
instance,
to
to
make
decision.
I
can
take
the
tag
of
an
order
with
the
CNI.
So
if
the
notice
target
with
a
calico
I
can
create
the
count
finger
and
when
I
created
the
know
that
I
can
a
set
or
a
configure
for
calico,
if
they
noticed
a
staggered
for
Weaver
I
can
I
can
create
the
conflict.
What
we've?
If
the
notice
is
not
that
good
or
I
go
with
a
default
or
I,
don't
apply,
see
and
I
will
say.
A
Though
I
looked
at
an
approach
and
so
far
I
haven't
done
it,
because
one
of
my
concerns
that
I
need
to
write
down
summary
document
is
that
I've.
If
you
look
at
how
kind
works
today,
it's
extremely
non-branching
and
that's
something
I
want
to
continue.
It
makes
it
simple
and
reliable
and
it
makes
it
very
portable
between
like
different
environments,
but
the
goal
I
have
is
that
it
just
kind
of
spins
up
your
containers
and
other
than
the
like
minimal
options
you
mean
able
to
provide
week
or
like
error
handling.
A
We
don't
don't
do
a
lot
of
branching,
so
you
know
exactly
what
it
does
reliably,
and
so
we
also
don't
get
ourselves
into
a
quagmire,
supporting
everything
so
with
those
breaking
image,
changes.
I
we've
just
said
for
kind
of
point
three
for
the
image
format
changed.
We're
in
alpha.
All
of
the
code
for
dealing
with
the
old
type
stuff
is
just
gone
it.
A
So,
wherever
or
wherever
there's
been
variants
so
far,
I'm
kind
of
leading
towards
like
okay,
you
can
bill
your
calico
image.
You
can
have
it
pre-loaded
and
stuff
to
make
it
easier,
but
as
far
as
like
installing
the
cni
kind,
one
on
that,
instead
will
probably
even
potentially
offer
something
like
kindness,
it's
just
the
most
minimal
thing
that
perks
and
will
offer
a
knob.
That
says,
like
okay,
I
want
to
opt
out
of
the
supported
route
and
do
my
shiny
thing.
I
will
also
make
it
quicker
to
iterate
on
those
things
like.
A
Know
so
so
are
that
still
possible,
but
I
so
I
I
spent
two
weekend.
One
of
the
main
things
I
did
is
a
crack.
Of
course.
Everything
CNI
and
I
study
about
some
stuff
and
I
want
to
explore
the
kind
net
option
a
little
bit
more
because
it's
so
like
it
just
does
what
you
need
like
crank.
Alright,
I
have
a
better
everything.
I've,
like
kind
can
be
a
lot
simpler
and
lighter,
because
it
doesn't
have
a
bunch
of
problems
that
you
would
have
with
a
real
cluster.
In
this
case
we
don't
have.
A
Can
we
have
a
flat
IP
space
for
the
containers
we
know
which
can
which
nodes
are
gonna
exist?
So
we
don't?
We
don't
need
to
do
a
be
like
like
peering,
and
that
sort
of
thing,
and
a
lot
of
that
tends
to
be
the
expensive
part,
is
like
you
know,
having
an
extra
Etsy
or
whatever.
If
we
can,
what
Antonia
has
this
prototype
of
where
we
just
use
like
some
standards
he
and
I
plugins?
E
C
A
C
A
But
I
think
ya,
know
I,
think
those
are
points.
I
think
it
would
still
be
interesting
at
a
dive
metadata,
because
it's
useful
to
have
I
think
it'd,
be
interest.
I!
Think
it's
it's
going
to
be
important.
Eventually,
though,
to
let
people
do
things
like
preload,
their
extra
images
on
that's
a
nice
performance-
optimization,
oh
yeah,
but
I.
Usually
you
should.
You
should
definitely
take
a
look
at
the
breaking
image
changes
kind
of
going
along
this
philosophy.
The
other
thing
we
got
is
we
got
significantly
smaller
images,
so
we
did
things
like
tough.
A
A
On
top
of
a
nerdy,
we
shed
several
hundred
megabytes
and
I'm
increasingly
finding
out
that,
like
one
of
the
cases
where
we're
really
useful
is
like
you
know,
anybody
should
be
able
to
turn
on
like
Travis
or
circle
or
one
of
the
other
free
open-source,
the
eyes
and
spin
up
a
cluster
and
do
testing,
and
for
that
use
case,
you
know
some
of
these
changes
helped
a
lot
with
with
boot
time,
but
we
still
have
image
pull
time
and
like
shedding
and
not
so.
I
did
a
little
bit
of
work.
E
A
But
so
along
those
lines,
I
think
we
actually
have
some
more
passes
to
take
at
this.
Unlike
you
know,
this
thing
is
actually
like.
Why
are
we
still
installing
this
package
or
like
that
sort
of
thing,
I've
been
doing
some
experimentation
and
actually
what's
pretty
cool
now,
is
ignoring
the
image
bowl
problem
with
all
of
those
changes.
A
It's
actually
light
enough
that
on
docker
for
Mac
dropping
to
the
lowest
settings
that
input
on
their
VM
is
still
more
than
I
need
to
spin
up
cluster
like
I
can't
I
haven't
been
able
to
find
what's
the
lower
bound
on
resources,
because
it's
so
much
lower.
Now,
oh
and
I'm
really
hoping
we
can
get
in
that
battle
zone
of
iteration,
where,
like
you,
can
create
a
delete
cluster,
isn't
like
tens
of
seconds
or
headed
in
that
direction.
A
I,
don't
think
that's
the
absolute
priority,
but
like
everywhere
we're
not
doing
that
when
we,
when
we
say
okay,
we
want
to
spend
up
some
huge
test
matrix,
there's
gonna
be
a
lot
more
viable.
If
we
can
do
things
like
okay,
we
already
loaded
the
images
once
upfront
and
we
amortize
that
compute
and
image
size
is
one
of
the
things
that's
making
that
not
true
right
now,
because
and
most
of
the
ciq
don't
actually
have
a
poll
already.
A
A
So
I
guess
we
need
to
figure
out
today.
One
of
us
has
to
solve
the
IP
rage
thing.
I
would
look
at
that.
If
we
can
get
that
solved,
then
my
suggestion
would
be.
We
start
by
just
staging
it
into
some
directory
in
kind
just
so
that
we
can
PR
it
there.
I
don't
like
I
I
have
to
look
I'm,
not
even
sure
if
I
can
send
peers
to
that
other
repo
in.
A
B
A
C
A
It
that
worse
than
the
night
step,
would
be
if
you're,
willing
upstream
the
code
and
then
make
the
switch,
because
we'll
want
to
like
officially
publish
images
and
I'll
want
to
make
them
multi
arch,
and
maybe
some
other
tweaks
here
and
there
and
then
and
then
we'll
do
a
pair.
That
changes
the
code
in
the
talks,
but
first
I
think
we
need
to
asset
NCI
test
it
locally
and
look
at
the
IP
ranges.
You
I,
don't
know.
F
A
C
A
The
other
action
I
about
Legolas
is
actually
trying
to
get
something
in
for
being
able
to
switch
off
currency
I
own
B,
so
that
people
can
do
cilium
and
whatever
first-class
I
know.
Duffy
has
actually
died,
need
to
watch
doctor
didn't
TJ,
I
kid
so
I'm
like
here's,
how
you
can
play
with
different
C
and
I
use
and
I
think
right
now,
if
I
had
a
guest
from
my
last
time,
that's
probably
abusing
the
like
known,
manifest
location
we
should,
regardless
of
which
option
we
go
with
officially
in
kind.
A
We
should
probably
try
to
make
it
easier.
So
my
suggestion
is
to
unblock
things
we
get
that
in
and
then
we
might
be
able
to
get
ipv6
CI
working
by
having
that
leverage
turn
off
the
Dalton
CNI
while
we're
still
like.
So
we
can
do
this
in
parallel
to
kind
that
we
can
go
ahead
and
just
switch
off
the
default
CNI
and
ipv6
CI,
ad
calico
or
whatever
it
doesn't
have
to
be
perfect.
A
A
Guys
we
should,
we
should
get
in
the
other,
because
I
know
you
have
a
pair
that
adds
ipv6
to
Connie,
but
we've
been
hung
up
on
the
CNI
stuff.
So
thanks
for
my
suggestion
is
we
we
should
get
the
rest
of
that
in
and
then
the
way
we
should
enable
it
is
just
giving
an
option
to
disable
CI
and
then
the
next
step
is
like
kind
of
in
parallel
to
this.
A
We
look
at
like
kind
that
we
look
at
calico
and
we
do
eventually
make
a
switch,
but
maybe
that's
the
maybe
that's-
maybe
that's
a
point
for
up
here
or
something,
but
just
adding
the
ability
to
switch
it
on
and
off
should
be
pretty
simple
and
if
kind
net
pans
out.
That
may
be
a
pretty
straightforward
change
too,
but
I
feel
like
we
shouldn't
block
I
feel
like
we
shouldn't
block
ipv6
on
the
cni
stuff
and
I.
Just
we
can
enable
it
by
letting
it
supply
its
own.
The.
A
Shion
is
it
simple,
I
think
the
simplest
thing
to
unblock
a
whole
bunch
of
work
around
this
and
to
make
it
easier
for
users.
Just
say
like
here
is
a
networking,
no
CI
or
something
like
that
flag.
You
said
it
kind
just
skips
that
step,
and
then
you
like
do
whatever
you
want
to
install
one,
and
if
we
add
that
that
enables
Duffy's
case,
and
it
also
enables
the
like
moving
forward
on
ipv6
before
we
have
finished
moving
forward
on
like
calico
or
kind
net
or
whatever
well,.
C
A
But
I
mean
people
will
want
to
do
that,
and
it
also
enables
the
case
of
like
I
want
to
you
cilium
and
I'm
running
on
a
host
a
it
can
actually
support.
That
and
sure
you
like
sewing
is
cool,
but
like
kind,
probably
can't
ship
it
by
default,
because
we
can't
guarantee
that
everyone
has
a
new
enough
kernel
and
that
sort
of
thing
so,
instead
of
kind
trying
to
support
psyllium
in
the
code,
we
just
support
not
doing
anything.
And
then
we
have
some
Doc's
they're
like
here's,
how
you
install
psyllium
and.
C
A
Let
me
put
it
this
way
so
like
even
a
something
that
we
definitely
can
get
kind
that
in
soon
like
what
like,
why
would
I
want
to
override,
as
opposed
to
just
never
having
had
it
installed?
I
mean
I.
Think
it's
a
great
feature
to
have
I'm
just
wondering
like
why
that
would
preclude
having
the
ability
to
turn
it
off.
It
seems
like
you
can
have
both
and
the
having
the
ability
to
turn
off
seems
most
powerful,
but
it's
also
not
very
complicated
to
support.
C
A
A
Like
I,
don't
think
that's
significantly
different
from
a
lot
of
other
clusters,
and-
and
this
is
this-
is
for
the
fairly
advanced
user
that,
like
knows,
they
want
psyllium
for
the
like
typical
testing
case,
we
should
still
have
that
out
of
the
box,
so
you
know
that's
the
default
and
perhaps
that's
kinda,
but
it
looks
like
we
need
to
go
find
that
IP
range
issue.
What.
B
A
There's
a
work-in-progress
appear
on
that
I
think
we
need
to
do
some
more
figuring
out
about
the
DNS
stuff.
It
turns
out
doctor
I
think
we
all
cared
no
this,
but
just
to
just
to
clarify
docker
has
a
different
behavior.
When
you
create
a
user
defined
network,
there's
things
they
wanted
to
change,
but
for
compatibility
reasons
they
didn't
change
on
the
default
network.
So
if
you
create
a
custom
network,
they
do
stuff
like
sad,
DNS
magic,
and
we
ran
into
a
problem
with
that.
But.
C
A
C
A
And
that
sounds
like
a
recipe
for
very
confusing
situations
where,
like
the
DNS,
is
different
for
everything
and
cubelet
and
I,
like
I,
can't
figure
out.
What's
going
on
so
I
get
on
the
note
and
I
start
messing
around
and
like
all
the
DNS
looks
fine,
but
cubed
is
I'm
happy
for
some
reason
and
it's
because
they
have
complete
like
I,
think
it's
a
pretty
esoteric
feature
to
like
tell
Cuba
to
use
a
particular
resolve
call.
We
should
we.
D
A
Well,
I
think
we
just
need
a
particular
hack
to
deal
with
that
loop
thing,
but
I
haven't
had
time
at
too
much
more
I
just
know
that
I
think
we
can
make
this
work.
I
think
this
makes
sense.
There's
a
couple
things.
We
need
to
be
labeling
the
network
and
we
need
to
only
delete
networks
that
we
actually
created
and
we
need
to
give
the
users
the
option
to
select
it,
because
this
is
another
behavioral
change
and
for
things
like
Federation
I
might
need
to
spend
up.
A
D
A
I
think
what
we
can
do
is
we
can
we
can
do
like
a
check
if
the
network
exists.
If
it
doesn't
exist,
then
we
create
it
and
we
add
a
label
that
says
it's.
You
know
it's
owned
by
kind
and
then,
when
we
delete
clusters
we
can
list
for
a
one,
that's
owned
by
it,
because
you,
you
could
imagine
something
like
I
want
to
use
the
default
bridge
and
then
kind
should
be
trying
to
delete
that.
So
we
should
we
should
label
it.
A
We
should
list
and
delete
based
on
the
label,
and
we
should
have
some
option
to
pick
the
one
you
want,
because
this
is
the
kind
of
thing
that's
going
to
break
applications.
If
there's
some
totally
different
network
that
it's
on
so
being
able
to
pick
it,
we
probably
need
to
I
I'm,
not
sure
about
the
code
that
picks
the
name
and-
and
that
was
another
comment
ahead.
I
think
there's
multiple
places
that
are
like
computing,
the
network
name
right
now.
D
D
A
That's
fine!
So
like
hey,
we
surfaced
some
stuff
with
it
like
the
loop
thing
I.
Don't
that
definitely
appear
in
its
current
state,
isn't
ready
to
merge,
but
I
think
the
overall
idea
makes
sense
and
will
hopefully
solve
some
problems,
and
it
sounds
like.
Hopefully
we
can
get
this
in
and
we
can
use
host
names
then
restart
should
be
in
a
much
better
shape.
Then
the
next
thing
that
might
be
interesting
to
look
at
is
you
can
actually
tell
docker
to
restart
containers.
A
G
A
D
A
Because
that's
the
other
thing
is
like
so
after
the
network
stuff,
it
sounds
like
we
can
make
restart
work,
but
it
will
require
user
interaction.
It
might
be
nice
to
make
it
work
with
no
interaction.
I,
definitely
wasn't
an
original
concern
of
kind,
but
I
can
see
where,
like
you
know,
I'm
hacking
on
some
application
on
kubernetes
on
my
laptop
and
I
need
to
reboot
and
then
now
my
clusters
in
this,
like
broken
state
I,
would
ideally
it
would
just
come
back
up.
Yeah.
G
D
A
So
we
use
localhost
in
the
user
facing
cube
config
and
in
the
in
cluster
stuff.
We
can
probably
use
the
notes
host
names
now
for
everything
like
the
load
balancer,
you
can
use
the
host
I
think
everything
can
use
the
host
names.
Yes,
if,
if
we
can
figure
out
like
merging
the
result,
comps
in
a
reasonable
way,.
A
Not
so
we
might
have
to
go
with
something,
so
the
other
thing
that's
going
to
come
up
eventually
is
we're.
Probably
gonna
need
to
let
people
also
oh
I,
think
I
already
have
an
issue.
Drivers
we're
gonna
do
like
will
override
the
result
in
trees
a
bit
anyhow
because
of
things
like
I'm
running
it
inside
of
a
kubernetes
and
for
whatever
reason,
I
can't
modify
the
pod
that
it's
running
in.
A
So
it's
picking
up
non-desirable
dns
and
I
want
to
tell
time
to
use
some
different
DNS,
and
this
is
something
that
CRI
supports
for
creating
containers
so
like
it
seems
reasonable
that
we
could
have
some
options
for
that,
but
that
will
need
to
play
into
this
code.
That's
also
already
gonna
need
to
be
modifying
the
resolve.
Cough.
A
C
C
C
A
A
B
C
A
A
D
A
But
like
driving
it
towards
one
that
it
works,
for
that
would
probably
be
best
because
even
even
aim
is
temporary,
like
we
started,
find
as
people
do
more
things
with
it,
it's
there's
some
things
that
are
still
hard
to
make
cheap,
it's
still
hard
to
like.
Actually
do
it
strapped
your
whole
elaborate
setup,
yeah.
D
G
A
Again,
I
did
still
like
originally
nothing
like.
No
don't
do
that.
Use
like
mini
cube
or
whatever,
but
like
it
seems
like
people
are
gonna.
Do
it
anyhow,
we
if
we
can
make
it
work
better
without
too
much
for
us
I,
think,
there's
I
think
we
can
potentially
get
some
other
benefits
from
having
increased
isolation
by
using
our
own
bridge
as
well.
Anyhow,
just
like
we
essentially
get
a
private
IP
and
a
private
little
network
there,
that's
nice,
oh,
it
seems
like
an
Eveleigh.
A
A
D
A
D
A
A
You
can
start
writing
unit
tests
that
put
a
fake
and
that
will,
let
us
add,
we've
quickly
found
that,
like
when
people
try
to
Elias
and
use
pod
man,
it
doesn't
quite
pay
the
same,
so
we'll
probably
need
some
Padma
and
implementation
of
the
interface
yeah.
The
I/o
writer
is
another
one
of
those
things
like
that
that,
like
it
should
be,
and
I
can
get
it
throughout
and
yeah.
You
know
this
is
I,
think
for
like
0.5
or
something
yeah.
D
A
And
then
one
thing
that
I've
been
finding
is
that
we
keep
winding
up
meeting
more
things,
so,
like
I,
think
we're
starting
to
hit
the
point
where
we
have
all
of
the
functionality
you
need
for
the
most
part,
and
the
next
step
will
be
improving
the
actual
abstraction
and
then
switching
everything
to
only
talk
the
abstraction
like
right.
Now
we
also
pass
a
lot
of
flags
of
strings
everywhere
and
I
found
that
gets
brittle
really
fast.
A
So,
like
a
another
step
that
I
want
to
head
towards
with
that
is
like,
even
if
we're
not
using
like
the
greatest
CRI
abstractions
everywhere,
we
should
at
least
use
something
slightly
higher
level
than
like
a
docker
command
line
flag,
because
even
just
docker
can
change
on
us
yeah.
You
know
if
we
can
use
a
little
bit
more
structured
option.
So
we
have
a
little
bit
of
that,
but
not
anywhere
near
enough,
like
the
run
extraction
takes
some
things
like
mounts
as
an
action
abstraction
out,
but
it's
still
just
taking
a
lot
of
flags.
D
A
G
A
Oh
so
that's
the
other
thing
I've
considered
that,
like
so
far,
we
have
not
replaced
image
tags
except
one
14.1,
because
that's
not
in
a
kind
release
and
we've
been
making
not
breaking
changes,
but
changes
to
MA
like
enhancing
the
entry
point
to
do
the
shared
notes,
so
we
might
as
well
keep
trying
to
find
whatever
other
things
we
messed
up
in
that
trip
that
I
messed
up
in
that
change
and
iron
them
out.
Oh
tell
her
like
I'm.
A
Sorry,
I
was
thinking
we
might
as
well
just
do
a
release
a
team
con.
It
would
be
kind
of
silly
to
do
a
release
with
only
like
a
week
or
two
at
work.
So,
given
that
we
haven't
quite
gotten
point
three,
yet
we
might
as
well
just
do
it
at
Q,
Khan,
Anthon
or
right
before
at
Yukon,
and
then
we
can
start
putting
point
four
which
I
imagine
you've
got
networking
stuff.
A
We
need
to
do
more
work
on
and
then
longer-term
we've
got
like
bring
in
some
kind
of
stuff
and
do
some
internal
code
cleanup
and
start
like
as
we're
supporting
more
things.
We
can't
lean
as
heavily
on
like
pre-summit,
hits
everything
and
you've
really
gonna
start
getting
some
tests
and
stuff.
So
to
do
that,
we
have
to
not
be
exacting
a
real
line
of
code.
A
D
A
On
the
way
to
Q
Khan,
just
like
cuddly
spitting
sounds
good,
but
this
is
good
or,
like
you
know,
on
the
way
to
the
contributor
summit
or
something
like
yeah
sure.
By
the
way
we
have
a
kind
of
a
story,
or
maybe
maybe,
when
I'm
trying
to
figure
out
how
to
finish
filling
up.
These
talks
lots
it's
like
and
we're
doing
it
live.
A
C
A
A
Days
that
I
have
to
talk,
but
other
than
that,
like
I
found
the
main
thing
that's
useful
for
talk.
This
is
going
and
asking
people
questions
like
at
the
end
and
if
you're
not
planning
to
do
that,
then
catching
the
recordings
as
fine
as
the
useful
part,
is
actually
meeting
up
with
people
and
that
doesn't
actually
necessarily
have
to
be
in
the
conference
and
that
can
be
somewhere
nearby.