►
From YouTube: Network Plumbing WG Meeting 2018-03-01
Description
Network Plumbing WG Meeting for 2018-03-01
A
The
first
one
was
loosening
the
requirement
for
the
cube
network
object
name
to
match
the
CNI
jason
config
name,
so
I
updated
that
language
and
the
only
question
I
had
was
that
the
CNI
plugins
currently
get
the
C
and
I
Jason
network
name,
at
least
in
the
spec.
The
plugins
would
have
gotten
their
network
name
from
the
CNI
config
Jason.
And
now,
if
that
name
no
longer
matches
the
cube
network
object,
they
won't
be
able
to
get
the
cube
network
object
name.
If
that
is
a
problem,
then
we
should
figure
out
how
to
address
that.
A
B
A
There
is
no
resolution
at
the
moment
it
was
just
a
question
of
if
the
cube
object.
Network
name
is
not
the
same
as
the
CNI
config
Jason
name
that
gets
passed
to
the
plug-in.
Plugins
only
get
the
scene.
I
can
suggest,
you
know,
I
mean
they
get
C
and
I
could
fit
a
sin
when
it
gets
sent
to
the
plug-in
on
standard
in
so
if
they
no
longer
get
the
cube
network
object
name,
those
are
different.
Does
that
matter
from
a
plugins
point
of
view?
Do
we
care.
A
B
All
concerned
was
in
usage,
yep
I
think
an
implementation
that
should
be
a
non-issue
I
mean
it'd,
be
easy
input
either
way.
The
issue
I
all
along
I
thought
the
issue
was
usage
a
we
would
like
to
prevent
people
from
making
mistakes
and
B.
We
don't
want
to
have
necessarily
ty
external
network
names
to
the
kubernetes
restrictions
on
object,
names.
B
A
A
A
The
only
time
that
that
a
plug-in
would
need
to
care
is
if
it
is
constructing
the
minimal
CNI
config
Jason,
when
the
network
spec
dot
plug
in
field
is
used,
and
that
would
be
for
kind
of
like
the
thick
plugins
that
we
have
been
talking
about
in
the
past
and
there
you
probably
want
to
use
like
CNI
version,
zero,
three
zero,
because
that
enables
you
to
get
back
the
list
of
interfaces
and
things
like
that.
That
could
then
potentially
be
put
into
the
status.
If
you.
B
A
A
Everything
else
with
versioning,
I
think
or
less
as
it
does
now.
And
if
the
plugins
conformed
to
see
and
I
spec,
when
the
meta
plugin
calls
each
of
them,
they
should
do
the
right
thing,
which
would
be
if
they
don't
support
that
version
the
error
and
then
the
whole
call
fails.
Just
like
a
girlie
would.
A
Alright,
that
was
it
for
spec
changes
that
came
out
of
last
week's
meeting
or
last
times
meeting.
Were
there
any
other
thoughts
comments
on
the
spec,
it
seems
like
we
are
kind
of
running
out
of
additions
and
major
changes
to
the
spec.
So
if
nobody
has
any
other
large
comments,
there
was
also
the
open
question
of
the
alternate
network
selection,
annotation
proposal.
Maybe
we
should
find
some
closure
on
that
and
I
think
there
was
two
options
on
that:
the
first
one
was
we
either
go
with
that
alternate
proposal
or
I.
A
C
I
guess
the
one
thing
I
would
say
about
the
comma
separated
one
is
that
it
may
be
easy
for
a
user
to
type
in
if
they're,
just
trying
to
say
I'm
whipping
up
this
quick
and
I
just
want
to
say
this
comma,
that
some
of
the
other
thing,
rather
than
being
like.
Oh,
we
type
in
all
this
and
then
lint
it
or
whatever.
C
A
I
think
that
does
have
value.
You
know
if
we
want
people
to
be
able
to
do
this
quickly
and
just
kind
of
fire
off
and
test.
Even
you
know,
I
think
we
want
to
try
to
make
it
as
user
friendly
as
possible.
We
want
to
be
able
to
allow
the
flexibility
for
other
stuff,
but
I
do
think
it
has
value
to
kind
of
keep
the
non.
C
B
Yeah
I
kind
of
understand
it
I
mean
there's.
You
know
a
countervailing
thought
that
if
this
really
is
a
list
structure,
let's
just
go
ahead
and
say
that
you
know
I
mean
putting
other
syntax
into
strings.
When
we
already
have
json
syntax
at
our
service
seems
you
know
it
kind
of
offends
my
sensibilities,
but
understands
that
you
know
quoting
json.
Enamel
is
a
little
bit
painful,
but
I
mean
this
easy
way
to
do
that.
It's
not
a
big
deal
to
me.
It's
so
I'm,
not.
A
B
No,
but
my
point
is,
you
know
we're
talking
about
right
if
we're
to
offer
two
choices,
I'm
saying
the
simpler
choice:
if
it
it
logically
semantically
has
a
list
structure,
you
know
to
my
taste:
I
would
use
JSON
structure
to
express
the
list.
That's
all
I
was
saying,
but
you
know
if
you
guys
want
to
go
with
a
comma
string
with
commas.
It's
it's
not
a
huge.
C
A
B
D
A
Good,
thank
you
sure,
all
right,
so
the
next
thing
I
had
in
the
agenda
was
kind
of
like
some
decision
and
the
network
as
resources
proposal.
That
I
think
was
it
Jana
for
Dan
kenigsberg
had
kind
of
posted
in
the
doc,
maybe
a
month
ago
it
sort
of
keeps
coming
up
and
the
reason
why
I
brought
it
back
into
this
document
was
that
the
resource
management
working
group
and
Jeremy
is
sort
of
representing
that
group
here
and
he's
on
the
call,
and
also
he
did
a
short
presentation
on
some
of
that
stuff.
A
Last
time
you
know
they're
continuing
to
go
ahead
with
a
number
of
different
proposals
and
also
I
think
Fabian
Deutsch
is
on
the
call
as
well
and
he'll
be
talking
a
bit
later.
Certainly,
the
resource
management
working
group
is
going
ahead
with
number
of
proposals
that
are
sort
of
exploring
this
direction
and
I
think
in
the
future.
It
will
be
somewhat
important
to
try
to
figure
out
how
we
do
management
of
limited
resources
on
the
nodes
and
obviously
that
doesn't
really
make
a
difference
for
Sdn
type
networks.
A
But
if
your
non
Sdn
type
network
or
if
you
want
to
try
to
bring
in
some
kind
of
hardware
resource
on
the
node
for
your
networking,
we're
gonna
need
to
think
about
scheduling
and
resource
management,
and
so
anyway,
I
brought
that
back,
because
it
seems
like
something
we
probably
want
to
think
about
for
the
future.
But
my
suggestion
and
I'm
totally
open
to
you
know
what
anybody
else
thinks.
A
My
suggestion
would
be
that
we
add
it
to
the
previously
discussed
topic
section
at
the
bottom
and
decide
to
investigate
it
further
in
conjunction
with
the
resource
management
working
group,
as
kind
of
like
an
enhancement
to
this
initial
spec.
What
I
would
like
to
do
is
try
to
keep
the
one
simple
for
now
just
get
to
kind
of
like
a
v1
of
the
specification
that
people
can
actually
start
implementing
and
start
poking
holes
in
from
an
implementation
perspective,
and
that
we
can
circle
back
around
and
update
that
with
some
of
the
other
ideas.
A
B
E
A
E
Yeah
that
makes
sense
and
I
think,
but
why
we
bring
it
up
so
early,
is
because
there's
some
cooperation
needed
between
the
two
working
groups
in
order
to
get
the
device
plug-ins
into
shape.
If
we
want
to
follow
that
road
and
that's
why
I
think
it
makes
sense
to
start
looking
at
that,
even
if
it's
not
in
scope
right
now,
okay,.
A
F
A
F
E
E
Think
that's
it
can't
see.
My
slides
should
be
a
small
window.
I
can't
yeah.
Okay,
we
need
to
live
with
that
for
now.
Okay,
so
I'm,
partying
and
Petter.
Who
will
be
doing
the
demo
we're
from
from
the
Qbert
project,
which
is
about
running
VMs
on
on
on
kubernetes,
an
open
shift
actually
and
we're
bringing
with
us
a
few
requirements.
Our
networking,
which
are
also
being
discussed
in
this
group
they're
in
this
group
I
mean
they
can
never
work.
They've
been
addressed
in
the
past
year.
In
many
discussions
you
know
multiple
IPS,
multiple
networks.
E
How
can
they
represent
it
in
bots
and
we
have
the
same
problem
with
VMs.
We
want
to
provide
layer
two
now
working
two
VMs
and
we
want
to
provide
ipv6
capabilities
to
VMs
and
we
need
want
to
find
a
solution,
and
we've
been
thinking
about
that
and
also
following
the
work
here.
So
what's
the
state
today
today,
the
state
today
is
that
if
you
launch
a
VM
with
Qbert
by
default,
it
will
tie
into
the
community
it's
not
working
as
best
as
possible
in
order
to
provide
maximum
compatibility.
E
But
we
need
to
provide
more
to
our
users
because
we
expect
to
provide
them
the
solution
to
migrate
their
existing
workloads
over
to
to
Kinetics
using
Qbert,
and
because
of
that,
we
need
to
provide
them
with
the
solution.
For
example,
I
don't
know
do
virtual
virtual
functions
in
that
VM,
which
require
a
to
connectivity.
E
So
there
quite
a
few
reason
for
that,
and
what
we
also
want
to
do-
I
mean
Cuba
itself,
isn't
is
a
is
an
add-on
to
communities.
So
we
are
implementing
an
operator
pattern.
We
want
to
do
the
same
thing
for
networking,
so
we
want
to
see
that
we
can
add
our
networking
to
the
kubernetes
cluster
without
harming
any
core
parts
of
requiring
changes
on
the
node,
so
that
was
driving
our
thoughts.
E
Today,
it's
mainly
about
GPUs
but
networking
like
physical
networks,
so
InfiniBand
devices
or
virtual,
like
virtual
function,
devices,
overlay
networks,
layer,
3,
layer,
2
networks.
All
of
these
can
be
considered
to
be
resources
or
a
deep
excuse
me,
or
at
least
we
think
they
can
be
addressed
just
like
other
resources
on
the
host
and
be
brought
into
the
pot
in
order
to
be
consumed
by
the
workload.
E
Actually,
we
also
think
that
thick
or
managed
networking
you
know
like
with
what
is
happening
with
cobia
or
with
other
Becht
plugins
could
also
be
solved
in
this
approach.
So
what
if
yeah,
okay
and
now
small
exchange
device,
plugins
I'm,
not
sure
how
familiar
everybody
is
with
device
plugins
and
not
so
much
dan
and
Jenny,
set
about
that
in
the
past
so
connects.
E
Resource
all
resource
yeah,
that's
so
that's
you
know
they
use
the
term
resource
and
Jeremy.
Please
correct
me
if
I'm
wrong,
but
the
currently
the
term
resource
is
used,
but
the
proposal
came
from
the
GPU
direction
to
make
virtual
tour
GPUs
usable
in
pods
for
computing
reasons.
So
today
it's
pretty
much
tied
to
GPUs
and
some
I
would
say
primarily
to
GPUs
what
they
use
the
term
resource
and
to
me
and
that's
why
I
was
following
this
approaches
that
the
pattern
they
take
expose
a
quantity
of
of
a
specific
resource
for
a
node.
E
That
is
what
device
plugins.
Do
you
start
a
device
bug
in
that
device?
Dragon
is
looking
on.
Your
note
is
that
resource
which
I'm
I'm
I
was
developed
for
is
that
present
on
the
node?
If
it
is
then
it
in
it,
then
it
is
exposing
the
quantity,
the
capability
on
the
known
API
to
the
cluster
level
and
the
resource.
B
E
It
depends
on
the
interpretation
for
for
overlay
networks,
who
could
say
it's
a
host
provided
resource,
so
you
need
some
kind
of
host
setup
in
order
to
provide
it
to
to
whatever
whoever
is
consuming
it.
So
community,
isn't
that
example
or
CNI
whoever.
So
it's
a
resource
provided
by
the
host
because
view,
and
also
that
overlay
network
will
also
go
over
a
physical
device
out
of
their
host.
So
there
is
room
for
interpretation,
I
agree,
but
on
the
other
end
it
is
the
resource
which
the
host
is
providing,
which
you
can
conceive
with
Apollo.
So.
F
F
F
That's
connected
to
a
fixed
set
of
Nick's
I
could
be
accessing
a
specific
network
connected
to
a
fungible
set
of
Nick's
I
could
be
accessing,
I
could
just
be
getting
access
to
a
network,
but
not
necessarily
sole
access
like
not
a
dedicated
device
to
a
fixed
or
fungible
network
or
I
could
be
asking
for
an
essentially
an
unbounded
thing
like
the
X
LAN
case
and
whatever
model
we
pursue
here
sort
of
has
to
comprehend
that
and
address
it
yeah.
So.
B
I
say
I
would
say
also
a
little
bit
more
fundamentally
even
right
in
some
sense.
In
plain
English,
you
know
everything
is
a
resource
it,
but
you
know
when
you
make
an
API,
you
have
to
have
some
semantics
right.
I
mean
objects,
are
a
resource.
You
know,
the
ability
to
create
X
is
a
resource.
The
CPU
know,
electricity
is
a
resource.
Clean
air
is
a
resource,
I
mean
just
because
even
plane
that
you
can
call
some
of
your
resource
doesn't
mean
you
should
shoot
horn
it
into
the
kubernetes
resource.
Api
right.
E
So,
on
the
other
end,
I
wonder
what
makes
networking
so
specific.
So
first
I
agree:
there
are
different
ways
how
you
can
consume
networks
and
how
you
can
actually
attach
them.
I
mean
the
best
obvious
example
is,
for
example,
for
VMS
pci
passed
through
in
order
to
directly
attach
a
NIC
into
a
VM
using
the
PC
iPad
through
mechanism
or
a
bridge
it
onto
that
device,
and
so
there
are
different
ways
that
we
can
consume
networks.
E
I
don't
disagree
with
that,
but
that
that
is
different
ways
of
consuming
the
resource
and
you
could
expose
them
in
different
ways
how
how
the
resource
you
can
exchange
with
resources
yeah.
So
that
is
what
I
wanted
to
get
to
sorry
for
not
finding
the
right
words
right
away,
but
yes,
you
express
them
as
different
resources.
That's
actually
what
we'll
see
in
the
example
later
on
that
you
could
express
the
same:
the
same
physical
device,
for
example.
E
Let's
say
it's
a
physical
device:
you
can
express
it
as
a
as
a
layer,
2
device
if
you
want
to
get
access
to
the
network
interface,
but
you
can
express
it
as
a
PCI
device.
In
case
you
want
to
do
pass-through,
so
it's
a
different
resource
depending
on
how
you
want
to
consume
it.
The
point
why
so
we
can
discuss
about
that
if
it
is
feasible
to
send
networking
fits
into
that
a
class
or
not.
E
But
the
reason
why
we
did
it
is
that
the
device
plugins
can
kubernetes
actually
have
some
things
which
and
I'm
sorry
about
that.
They
have
some
things
which
make
the
peening
and
which
we
see
was
is
emerging
here
as
well.
For
example,
they
already
have
a
way
to
for
deployment
and
out
to
speak
to
the
qubit.
E
They
have
scheduled
awareness
so
on
the
fly
they
can
detect
if
the
resources
present
or
not,
that
is
really
handy
for,
for
example,
manage
networks
and
they
have
API
integration
and
I
mean
that
is
what
we
currently
seeing
with
the
Sun
News
ation
of
the
CR
D,
which
is
not
a
conflicting
by
the
way
which,
with
this
proposal
here,
it's
rather
a
more
more
fine
granular
addition
to
it,
but
it
has
API
integration.
So
today
you
can
specify
that
in
container
spec
and
what
resource
you
want
to
consume.
B
That's
all
you!
It's!
It's
really
important.
Two
things
are
really
important
and
one
that's
not-
and
this
is
really
my
criticism
right
so
I
understand
why
the
fact
that
you've
got
something
that's
going
to
do
your
easy
deployment
and
it's
already
integrating
the
API.
Those
are
great
features
right
and
the
scheduler
awareness.
E
F
They
the
device,
sorry
they're,
the
resources
in
the
device
plug-in
API
seems
to
fit
a
model
where
you
have
discrete
things
that
don't
really
have
relationships
to
each
other
right.
If
I
ask
for
a
GPU
I,
don't
really
care
which
GPU
I
get
because
then
they
have
no
relationship
to
each
other
they're
not
connected
to
each
other.
If
I
ask
for
a
NIC
that
is
sort
of
insufficient,
because
I'm
actually
asking
for
a
connection
into
a
larger
system
right,
a
NIC
on
its
own
isn't
useful
and
I
think
some
would
argue
with
you.
F
But
your
general
point
definitely
I
mean
what
are
you
gonna
do
with
a
NIC
if
you
can't
specify
to
which
network
it
is
it
connected
right?
Either.
I
have
literally
flight
hardware
NICs,
where
they
have
wires,
pre
routed
to
particular
networks
or
I.
Have
s
RI
of
e
NICs,
which
are
going
to
be
hooked
up
somehow
to
virtual
overlays
or
virtual
channels
or
I.
Have
these
software-defined
things
which
are
connected
through
software
mechanisms
in
the
same
way,
but
in
all
cases
I
need
this?
Actually,
what
I'm?
E
The
box
yeah,
that's
perfect,
I
mean
that's
absolutely
right,
and
that
is
actually
so.
Maybe
maybe
the
problem
here
is
that
the
device
prong
is
really
just
a
mechanism
to
model
all
this,
and
you
know
the
resource.
I
should
have
written
down
a
few
examples
of
how
there
is
two
resources.
Look
because
effectively
depends
on
the
rate
on
the
plug-in
itself,
how
it
exposes
a
resource,
so
you
don't
have
to
expose
at
all
a
device
itself,
for
example
for
any
open.
E
If
you,
if
we
imagine
the
plug-in,
which
is
speaking
to
Neutron,
for
example,
you
wouldn't
imagine
that
this
plug-in
is
exposing
any
physical
devices.
What
rather
expose
the
number
of
networks
known
by
that
neutral
provider
so
and
then
on
the
fly
this
indeed,
the
pot
can
be
connected
to
those
networks,
so
the
plug-in
is
in
reso
yeah.
So
the
plug-in
is
advertising
as
resources,
the
known
networks
and
not
any
physical
connection.
But
that's
that's
the
also
the
charming
side
of
this
approach
that
that
you
can
expose
different
or
networking
as
different
resources.
E
So
you
can
expose
them
as
really
logical
networks.
You
can
expose
them
as
physical
devices.
It
depends
on
the
use
case,
for
example,
if
you
want
to
do
I,
don't
know
some
peeking
into
network
traffic,
then
you
might
want
to
really
you
know,
bridge
to
a
specific
device
or
put
a
specific
device
into
permissions
mode
and
that
can
be
modeled
as
a
device
plug-in
actually
on
a
slide
later
or
in
the
end.
We
accept
a
specific
scene,
I
scene
I
slide
and
we
can
get
to
that
later
on
our.
E
E
We
want
to
get
to
touch
too,
and
all
of
this
can
be-
and
this
is
rough
guess
we
have
not
done
it-
we'll
be
working
on
that
or
we'll
take
a
look
at
that,
at
least
until
you
see
now
actually
to
perform
this
network
plumbing
work
because
I
said
DP
is
just
a
way
to
leverage
the
API
and
make
the
schedule
awareness,
but
you
see
and
I
and
all
the
plugins
which
exists
there
to
do
actually
the
network
plumbing.
So
it's
it's.
It's
a
it's
a
pretty
big
hammer
which
might
be
used
or
yeah
it.
E
F
Interesting
part
here
is
that
there's
a
non-obvious
relationship
between
the
network
and
the
device
plug-in
that
isn't
portable
and
that
users
shouldn't
really
be
exposed
to
if
we
can
avoid
it.
The
user
wants
to
say,
connect
me
to
this
network.
Sometimes
that's
fine
and
there's
no
device
implications
there.
It's
just
in
via
the
X
LAN
thing
and
other
times
it
means
allocate
me
a
specific
NIC
on
a
specific
pool
of
machines
so
that
I
can
get
access
to
this
physically
guarded
network.
F
But
the
user
shouldn't
be
aware
of
that
right,
and
so
what
we
end
up
with
is
a
sort
of
backside
relationship
between
the
users,
expression
of
connect
me
to
this
network
and
some
other
entities.
Understanding
of
oh.
That
means
you
need
an
SR
Iove
device,
or
that
means
you
need
to
run
on
machines
that
are
labeled
connected
to
secure
net
right
right.
E
So
I
guess
I
I
don't
want
to
get
into
that
rat
hole
or
let
users
decide
to
know
about
all
these
details.
So
this
is
actually
the
the
best
example.
I
have
one
in
the
one
of
the
documents
to
express
that
so
here.
This
is
how
we
would
imagine
an
ipv6
connection
to
look
in
upon
definition.
So
here
the
important
part
is
actually
down
here
and
that
is
encoding
how
the
resources
working
so
you've
got
your
bender
string,
which
is
the
first
part
with
all
the
slash
and
and
the
resource
to
consume.
E
And
here
is
that
abstraction
which
looking
for
we're
effectively
saying
I,
want
to
consume
net
a
from
the
given
from
the
given
vendor
or
the
given
provider,
and
this
provider
is
encoding.
That's
an
ipv6
address
which
we
want
to
consume
in
that
phone
and
you
can
model
many
things
along
that
lines.
So
if
we
take
the
example
above
now
here
we
explicitly
provide
a
provider
for
what
have
been
or
providing
string
which,
for
which
the
terms
it
is
to
to
expose
a
specific
physical
device
to
the
body.
E
F
E
C
Wouldn't
I
would
think
that
you
know
this
probably
is
intrinsic
to
a
lot
of
hardware
devices
and
I
feel
like
could
cheap.
You
is
even
a
fairly
decent
example
to
is
and
relative
to
what
we're
trying
to
do
here,
but
with
a
GPU,
for
example,
you
might
care
about
what
kind
of
chipset
is
on
that
GPU
of,
depending
on
the
you
know,
type
of
algorithms
you're
gonna
run
on
it.
F
Of
a
minute,
but
there's
a
difference
in
that
there's
really
not
such
a
thing
that
I'm
aware
of
as
a
software
emulated
GPU,
whereas
there
is
such
a
thing
as
a
software
and
related
virtual
NIC,
and
if
we
expose
to
an
end
user,
hey
you
have
to
request.
Even
if
we
go
through
the
there's,
there's
a
secondary
proposal
that
isn't
been
covered
here.
That's
called
the
resource
class,
which
provides
something
like
storage
class,
which
gives
you
an
abstraction
in
which
the
actual
device
and
the
the
abstract
name
that
the
user
uses.
F
E
B
Solely
see
if
I
understand,
the
idea
here
is
that
for
each
distinct
Network
class
there
would
be
a
different
limit,
so
I
see
on
the
screen
here
to
network
classes.
One
is
five
network
Cupertino
and
the
other
is
six
thought
networked
up.
Cooper,
tayo
tayo
right
so,
depending
what
network
class
I
use
a
different
key
in
my
limit.
E
Yes,
you
so
that
is
that's
obviously
a
gap
in
the
API.
So
today
you
need
to
specify
the
quantity
of
resources
you
want
to
get
more
than
one
might
make
sense,
but
it
often
doesn't
make
sense.
So
here
we
would
probably
you
here's
some
what
to
do
with
the
device
plugin
or
at
least
to
evaluate.
In
my
opinion,
you.
B
E
E
B
B
E
Actually,
here's
here's
actually,
where
I
see
the
the
connection
to
the
network
acidity,
which
is
taking
place
currently
I
mean
this
can
be
a
reference
to
those
CR.
Ds
I
mean
the
red
or
blue,
which
are
now
added,
could
be
really
I
mean
there
would
be
it's
a
far
fetched.
That
is
why
I
basically
agreed
with
Dan.
That
is
something
for
me
too,
but
to
me
it's
still
charming
and
the
detail.
It
needs
to
have
a
closer
look
and
I
actually
think
a
good
discussion
about
they
lose
the
the
ends
of
it.
To
understand.
D
D
E
A
E
That's
the
vendor
of
the
plug-in
the
resource
is
actually
red.
So
if
we
imagine
that
network
KS
IO
is,
is
the
reference
seen
I
plugin
from
kubernetes
blessed
by
by
you,
for
example,
and
and
that
is
watching
for
all
Network
definitions
on
the
on
the
cluster.
So
looking
for
network
see
are
these
and
whenever
such
a
network
is
defined,
then
that
pot
can
be
connected,
and
then
this
red
network
resource
is
available
on
the
host
or
any
host.
Where
I
don't
know.
E
E
H
Network
is
everywhere,
we
have
these
demon,
I
mean
at
least
in
the
in
the
Nvidia
case.
We
have
these
demon
says
you
know
that
run
device
plugins
and
next
are
rare
and
write,
write
everything
out
to
the
update
the
notice
that
capacity
and
why
not,
as
the
allocating
the
allocate
calls
occur
to
the
device
plugin,
so
we're
putting
a
lot
of
the
magic,
and
this
is
where
the
vendors
do
their
value.
Add
to
the
extension
point:
we've
added
the
device,
plugin
extension
point,
yeah.
E
Actually,
we
will
see
it
I
mean
Peter
will
show
it
in
a
minute
how
available
resources
will
pop
up
on
the
fly.
So
your
device
plugin,
is
watching
the
cluster
and
it
will
watch
for
new
newly
defined
networks
as
it
sees
them.
It
can
do
any
set
of
locally
on
the
node
once
it
discovers
these
new
newly
defined
networks.
Whatever
that
back
and
then
is
it
depends
on
the
vendor.
You
know
you
might
want
to
do.
Everything
will
be
explained.
B
E
D
F
E
Right,
that's
up,
I
mean
that's
a
lot
of
I
mean
I
mean
that's
not
impossible.
Is
it
because
with
especially
because
we're
using
the
vendor
string?
That
would
be
unique
to
the
device
plugin
vendor
and
in
this
case
we've
red
and
blue,
that
vendor
would
then
set
up
a
redirect
to
to
an
open
API
specification
for
that,
given
network
definition
or
never
c
or
d
or
whatever
it
will
keep
it
it
doesn't.
It
doesn't
quite
feel
right
to
me,
but.
F
F
If
you
presume
for
a
minute
that
that
the
network
c
rd
is,
is
a
thing
that
the
idea
of
multiple
networks
is
going
to
proceed
right,
then
what
you
really
want
the
user
to
express
is.
I
want
to
join
network
red
and-
and
I
want
to
join
network
blue
right
and
that's
all
the
users
should
have
to
move
privacy.
F
Then
in
looking
at
the
definition
of
red
and
blue,
there
perhaps
is
an
optional
link
to
a
resource
class
that
says,
if
you
want
to
join
red,
you
have
to
get
a
resource
from
this
resource
class
and
the
resource
class
is
a
is,
maybe
your
sr
iove
thing
and
blue.
Well,
that's
just
the
X
LAN,
so
there's
no
resource
class
needed
and
then
some
admission
control
process
says
you
know.
F
B
Sort
of
I
think
there's
still
level
of
indirection
here
within
a
this
kind
of
CRD
concept.
We're
talking
about
there's
like
there's
still
room
for
a
class
concept
right.
We
want
to
be
able
to
make
multiple
networks
of
the
same
sort,
they're
just
different
individuals,
but
they
all
are
of
the
same
class
and
it's
at
the
class
level.
You
say:
oh
well,
this
need,
you
know
a
piece
of
detail.
F
They
shouldn't
have
to
know
anything
about
the
network
or
how
its
implemented
and
the
abstractions
the
abstractions
should
cover
the
rest,
and
that
may
mean
that
we
added
like
we
literally,
have
to
add
an
admission
controller
for
network
device.
Linkage
and
that's
okay,
I
think
that's
completely
a
reasonable
thing
to
do.
I
mean
that
said
in.
A
The
example
that
fabien
has
here
I'm
feel
like
it's
getting
a
little
bit
more
complicated
as
we
speak,
but
the
user
would
only
really
have
to
specify
you
know
that
they
want
to
connect
to
network
red
blue
and
in
turn
in
this
pod
spec
right
here.
So
does
that
not
satisfy
that
requirement?
Tim
I
mean
there's
a
little
bit
more
typing
with
resources
and
limits,
and
the
network's
well.
F
I
think
it
resources
is
the
wrong
place
to
represent
it
here,
because
I
think
the
quantity
here
part
of
it
is
not
appropriate
and
I.
Think
the
use
of
the
resource
name
structure
to
sort
of
embed
the
the
user-defined
name
isn't
really
appropriate.
I
think
it
would
make
more
sense
in
either
an
annotation
or
in
a
like
a
real
field.
E
I
wonder:
how
is
that
the
definition
I
I
I'm,
actually
I
I,
don't
know
checked
with
anything
I
I
think
it's
a
good
point
to
say:
the
user
should
should
have
a
few
barriers
to
express
what
he
needs
and
all
the
prefixing
with
the
vendor
string
should
should
be
not
necessary
to
be
done
by
the
user,
but
I
actually
wonder:
why
should
it
be
different
for
GPUs?
Why
should
the?
Why
should
a
user
specify
oh
well
for
GPS
might
make
more
sense
right?
F
B
A
I
was
actually
gonna
ask
about
that
because
we
have
talked
in
various
places
about
quas
and
bandwidth,
and
you
know
what
does
that
look
like
here?
You
know,
for
example,
if
you
have
a
40
gig
Nick
and
a
pod
wants
10
gig
of
VAT.
How
do
we
specify
that
kind
of
thing
and
it
seems
like
resources,
are
actually
be
involved,
but
I
don't
really
know
what
that
looks
like.
D
E
A
A
G
Hello,
I'm,
better
and
before
I
jump
to
the
general
I,
just
a
brief
introduction
to
to
the
device
plugin.
So
what
it
does,
it
exposes
connections
to
note
bridges
as
a
resource
I
mean
it's,
maybe
not
the
best
usage
of
the
device
plug
in
for
networks,
but
it
was
for
POC
and
I.
Think
it's
ostrich
illustrates
the
case.
Somehow
it
is
deployed
as
a
daemon
set
and
it
accepts
less
white
list
of
bridges
as
a
parameter.
G
We
have
to
find
some
problems
with
missing
ipython
points
and
the
device
plugin,
and
that
is
that
we
have
only
allocated
for
goal,
which
is
it's
called
before
that
pod
is
actually
created.
So
we
don't
have
access
to
the
network
name
space
on
the
bottom,
but
we
solve
it
with
a
heck
and
there
is
ongoing
discussion
with
the
resource
where
group
about
extending
of
the
API.
So
we
can
have
it
a
hand,
light
kind
of
90
loyal
like
having
at
and
delete
call
to
handle
it
by
the
C&I.
G
G
Thanks
so
I
prepared
the
mini
cube
instance
and
only
extra
Frank
I
did
there
was
that
I've
enabled
to
device
plug-in
feature
yet
so,
let's
jump
into
it.
First
thing
you
need
to
do
is
to
create
a
config
map
and
that
it
this
is
the
parameter
for
the
device
spoken
on
this
bridges
here
is
a
whitelist
of
bridges
that
should
be
exposed,
so
the
device
plugin.
The
next
step
is
to
deploy
it
demons
that
it
contains
device
plug-in
itself.
We
can
wait
for
it.
G
E
G
It's
running
and
I
can
ask
kubernetes
to
give
me
like
the
resources
on
on
the
mini
cube,
node
and
just
so,
as
you
can
see,
there
is
a
bridge
resource,
exposing
the
library
to
zero
and
there
is
zero
connection.
So
I
really
know
what
that
mean.
Is
that
the
bridge
it
was
not
created
under
not
yet
so
in
case,
I
asked
I,
create
a
pot
and
I
ask
for
this
bridge.
There
won't
be
in
place
to
schedule
it
on.
So
let's
try
to
create
the
pot.
G
G
Now,
let's
look
into
the
available
resources
on
the
node,
and
now
there
is
100
connections
to
the
British
available
debt.
That's
because
there
is
no
way
how
to
express
infinite
resources
for
now.
So
it's
just
an
arbitrary
number
and
if
I
get
information
about
the
pot,
you
can
see
it's
not
spending
anymore.
It's
actually
running.
G
E
A
A
All
right
will
do
all
right,
so
we
are
at
the
top
of
the
year
by
yeah
top
of
the
hour.
So
thanks
for
the
demo
and
great
discussion
and
let's
continue
discussion,
the
mailing
list
and
comments
on
Fabian's
document,
which
I
believe
he
linked
to
from
the
agenda
doc
and
we
will
meet
in
two
weeks
on
March
15th.