►
From YouTube: Kubernetes SIG Network 2018-03-22
Description
Kubernetes SIG network meeting from March 22nd, 2018
A
B
Don't
know
me,
my
name
is
Paris.
I
am
a
co-chair
for
the
special
interest
group
4
contributor
experience.
What
we're
trying
to
do
is
do
a
little
road
show
of
sorts
and
what
I
make
all
my
notifications
here
reason.
Why
is
we
know
that
this
is
in
a
very
large
and
complex
project
now
and
getting
messages
out
can
be
difficult
in
this
setting,
since
email
is
so
rigorous,
sometimes
so
what
we're
gonna
do
is
clearly
go
around
the
SIG's
and
let
folks
know
what
we're
doing
all
of
your
contributors
awesome
contributors,
to
say
the
least.
B
So
we
just
want
to
make
sure
that
everybody
is
on
the
same
page.
I
did
include
in
the
notes
a
very
lengthy
document
about
like
what
we're
doing
and
why
so
I
just
kind
of
want
to
go
through
that
at
a
very
high
level
and
also
leave
it
for
you
to
do,
leave
it
for
you
to
read
and
give
us
feedback
at
a
later
date.
B
So,
first
and
foremost,
I
just
wanted
to
kick
it
off
and
say
that
one
of
the
main
goals
this
year
for
contributor
experience
is
truly
a
automate
and
non
automates
and
document,
and
also
to
work
on
contributors.
Having
the
same
experience
across
repos
we've
been
getting
a
ton
of
feedback
from
both
new
and
current
contributors
that
it's
hard
for
them
to
jump
back
and
forth
in
different
repos
when
there's
different
contribution,
workflows
and
BOTS,
and
people's
need,
if
ur
ins
and
certain
places.
B
How
you're,
using
certain
labeling
things
like
that,
and
we're
working
on
label
definitions
right
now,
also
better,
defining
bot
commands
and
things
like
that?
There
are
links
and
the
in
the
document
as
well
and
where
you
can
comment
on
these
things
and
where
we
can
have
open
discussions
without
that.
Also,
we've
been
working
on
a
lot
with
issue
triage
guidelines
and
we
want
to
see
what
your
thoughts
are
on.
Applying
all
current
issued
on
the
issue,
triage
guidelines
that
we
have
across
all
covered
any
github
repositories
and
then
also
with
the
labeling.
B
We're
really
trying
hard
to
get
this
right,
because
one
of
the
things
that
we
are
trying
to
do
is
link
a
lot
of
labels
to
some
of
the
dev
stats
dashboards.
If
you're
not
familiar
with
dev
stats,
definitely
check
it
out.
It's
20-plus
dashboards
that
show
the
health
and
the
velocity
of
the
project
and
by
adding
labels
in
there.
That
would
be
another
filter
for
us
to
have,
which
will
be
great
for
us
to
define
the
health
and
metrics
of
the
project,
and
the
most
important
piece
to
this
grant
is
that
we
have
defined
out.
B
B
B
I
mentioned
charter.
We
did
just
complete
our
Charter,
but
we
actually
are
going
to
be
going
back
today
and
doing
updates
to
that,
because
we
have
renamed
the
signe
units
into
chairs
and
then
we've
also
adopted
the
technical
lead
model
that
the
steering
committee
has
given
governance
on.
So
if
you
want
to
check
ours
out
it's
there,
but
it
is
going
to
change
today,
we're
also
including
sub
project
owners,
and
if
this
crew
needs
any
help
with
charter
related
things,
we're
happy
to
help
as
well.
I
also
included
a
projects
list.
B
This
is
all
the
projects
that
are
currently
underway
with
contributor
experience
and
we'd
love
to
have
more
involvement
from
other
devs
on
the
project.
For
some
of
this
stuff
and
I
wanted
to
include
that,
so
you
know
and
have
an
idea
of
what
it
is,
we're
actually
working
on
and
then
we've
also
been.
We've
also
pumped
out
a
new
version
of
a
contributor
guide
and
we
would
love
for
you
to
poke
holes
at
it
and
review
and
make
any
comments
that
you
think
we
should
that
we
should
do
and
make
better.
B
We
are
doing
a
developer
guide
as
well,
and
that
is
underway.
Currently,
the
link
is
in
the
link
for
the
issue
is
in
there
as
well,
and
we
would
love
to
have
comments
on
this.
We
know
this
is
definitely
going
to
be
a
lot
of
work,
but
the
more
we
can
get
comments
on
the
better,
because
then
we
can
either
make
new
documentation.
That's
referencing,
the
material
that
you
think
is
missing
or
update
the
information
that
you
think
we
already
have
that
just
needs
to
be
updated.
B
So
that's
that's
about
workflow
and
documentation,
and
things
like
that.
The
next
section
that
I
wanted
to
discuss
is
mentoring.
So
how
does
your
group
currently
do?
Mentoring
and
this
isn't
necessarily
for
just
new
contributors
either
I'm
talking
about
all
ladders
of
our
wrong,
so
members,
reviewers
rulers,
etc.
We've
created
a
ton
of
different
kinds
of
mentoring
programs
and
address
things
like
time,
burdens
and
one
that
was
one
of
the
main.
The
main
things
that
I've
been
thinking
about
in
planning
these
things
is:
what
can
we
do
that
releases?
B
A
lot
of
the
time
burdens
from
traditional
mentoring
program
that
would
still
help
and
build
success
for
people.
One
of
those
things
is
this
concept
of
mentors
on
demand,
which
is
very
similar
to
use
our
office
hours,
but
this
is
contributor
office
hours.
If
you
will,
we
are
dubbing,
it
need
our
contributors,
so
this
is
targeting
new
and
current
contributors
is
a
live
livestream
session
every
month,
first
Wednesday
of
the
month,
we
would
love
to
have
you
come
on
if,
if
you
want
their
on
YouTube,
you
can
go
and
see
our
previous
sessions.
B
It's
just
us
talking
on
zoom'
and
we
collect
questions
and
slack
and
Twitter,
and
the
questions
range
from.
What's
your
favorite
color
to
and
test
flaking,
and
we
help
them
out.
However,
we
can
we
also
do
a
new
concept
called
group
mentoring.
This
is
something
that's
very
new
g-force
in
general,
so
this
would
be
a
four
to
one
mentoring
relationship
and
it's
AIT's
like
a
two
to
three
month
process,
depending
on
where
you're
trying
to
go
right.
B
Now
we
have
a
test
with
cops
workload,
API
and
QA
DM,
where
we're
taking
a
group
of
ten
people
who
are
all
current
numbers
into
code,
reviewer
ship
and
we
have
bi-weekly
stand-ups.
We
do
different
kinds
of
workshops.
Tim
came
on
and
did
a
code
review
session
that
was
very
highly
rated,
so
we
would
love
to
have
involvement
from
this
crew
and
the
leads
and
chairs.
If
you
want
to
reach
out
to
me
and
how
we
can
get
you
involved,
that
would
be
really
awesome
and
then
we're
also
creating
a
buddy
program.
B
B
So
if
you
want
to
pair
program
with
someone
that
you
don't
get
the
time
to
do,
if
you
want
to
do
a
code
review
session
for
an
hour,
if
you
want
to
ask
them
anything,
this
is
this
is
where
you
would
do
it
we're
hoping
to
have
a
test
out
within
the
next
two
to
three
weeks
on
this.
So
more
soon
than
that
information
and
then
really
the
last
thing
I
wanted
to
talk
about
is
slack.
B
We
now
have
almost
thirty
four
thousand
members
and
slack,
so
it
would
be
cool
if
we
could
have
more
things,
pindy
or
channels
like
important
documents
like
your
charter
and
how
people
can
get
involved
in
your
group,
because
slack
is
sort
of
this
portal
right
now
into
entrance
of
continent
of
being
a
contributor
or
a
user
of
kubernetes.
So
just
a
very
cool
way
of
getting
someone
up
to
speed.
It's
just
through
pinning
documents
and
being
helpful
and
slack.
We
also
have
a
slack
admins
group.
Now,
if
you
need
anything,
we
have
slack
guidelines.
B
We
actually
just
created
an
announcements
channel
today
and
that's
gonna
be
really
cool
because
and
we'll
have
things
like
code
freeze
and
other
really
important
or
quite
announcements
come
through
on
that
and
then
last
but
not
least,
use
our
office
hours.
We
would
love
to
have
more
people
talking
to
our
users
through
their
office
hours.
Yesterday
was
user
was
the
office
hour
sessions
and
they
go
really
well
they're,
viewed
by
hundreds
of
people
and
the
questions
are
pretty
awesome,
but
that's
it
for
me.
B
I
just
wanted
to
kind
of
bridge
this
gap
between
contributor
experience
and
other
SIG's.
So
if
you
have
any
feedback
for
us
as
far
as
what
we're
doing,
why
we're
doing
it,
how
we're
doing
it
or
what
we
could
do
better,
please
visit
our
select
channel,
is
that
our
mailing
list,
then
don't
pigeon
al
whatever
awesome.
A
C
B
C
A
A
Alright,
next
on
today's
list
is
a
cig
test,
review
and
I
think
we
can
keep
this
relatively
quick
I
looked
in
advance
and
there
are
kind
of
few
areas
of
test
that
are
in
trouble
at
the
moment.
The
first
is
the
cube
ATM
tests,
which
Bowie
also
pointed
out
and
I
think
I've
been
in
trouble
for
a
while
and
then
the
second
are
the
IP
vs
tests,
which
are
also
showing
lots
of
failures.
D
E
E
E
D
A
G
A
A
E
Burnett,
could
that
suppose
applying
basic
traffic
shipping
during
Hodor
travel
is
the
way
I
happen
if
bender,
what
is
being
I
didn't,
who
their
policy
data,
horror,
Q,
polite
thing,
I
could
a
dozen
dispositor,
because
there
was
no
traffic
shipping
packing
butter,
but
we
can
replicate
the
behavior
we
have.
We
out,
we
have
stay
in
community
affairs,
is,
is
in
putting
the
feature.
In
fact,
Tina
has
already
support
the
traffic
of
shipping
packaging
and
the
other
process
proposal,
and
some
people
have
dated.
E
A
H
There's
a
little
bit
more
work
that
has
to
be
done.
Port
map
uses
these
capability
arguments,
these
sort
of
dynamic
configuration
of
substitutions
and
we
haven't
merged
those
into
the
CNI,
a
traffic
shaping
plugin,
yet
right,
okay,
so
it
can
be
merged
essentially
now
into
cube
net
sense.
Cube
net
generates
its
own
config,
but
it
is
not
yet
available
for
generic
C&I
in
cube.
Thank
you
for
dating.
C
So
it
kind
of
sounds
like
this
is
a
two-step
process.
I
mean
merging
into
cube
net
would
allow
us
to
remove
all
of
the
cube
net
code
that
does
bandwidth
shaping
on
it
by
itself
and
get
a
bunch
of
code
out
and
then,
as
a
second
step
as
part
of
the
cube
net
deprecation.
Add
the
capabilities
to
a
CNI
driver
to
pass
that
through
right
and
also.
I
E
J
E
J
And
then
I
have
a
common
for
this,
so
so
this
feature
auto.
It
exists
for
quite
some
time,
but
it's
never
I
graduate
to
beta
Rho
GA,
because
because
it's
like
actually
there's
a
lot
more
requirement
to
form
form
only
support
bandwidth,
shaping
traffic,
shaping
and
bandwidth
based
scheduling
and
whatnot.
So
basically
I
would
expect
it
to
like
to
stay
in
alpha.
Even
after
this,
even
after
the
change
to
see
and
I
never
plug
in
okay,.
E
Agree
so
another
question
is:
shall
we
graduate
has
teacher
from
an
occasion
to
speaker
or
not
awesome
traffic
shipping?
My
country
behavior,
is
just
increase
the
bandwidth
and
the
equalizer,
but
the
thing
singer:
Hagen
has
a
supporter
birthday
increases
pairs
and
first
egress,
so
so
Mike
so
I.
What
I
mean
who
repress
the
traffic
shipping
from
pkg
by
the
ways
to
therapy
is
Singapore?
Has
the
pen
do
I
do
features
in
shipping
and
if
we
want
to
tab
repeater
colada
in
the
future.
A
Is
the
question
whether
we
should
move
this
from
annotations
to
spec
fields?
I
think
the
answer
is
for
the
first
step.
We
would
leave
it
how
it
is,
and
we
would
need
to
have
that
larger
discussion
about
what's
required
to
move
this
from
alpha
to
beta,
but
that
could
be
a
follow-on
after
Lee.
We
do
the
work
to
support
this
in
CNI
yeah.
F
I
think
there's
maybe
don't
get
me
wrong.
I
hear
this
from
several
people,
but
I
think
there's
a
lot
of
concerns
around
using
this
as
a
general
mechanism.
I'm,
not
convinced
that
every
C&I
plugin
out
there
can
reasonably
a
blend
banks
shaping
in
the
same
way
and
I'm
concerned
about
whether
it
becomes
a
like
a
schedule
resource
or
whether
it's
just
a
constraint,
but
we
don't
actually
schedule
on
it
and
therefore
over
commutable
I
think
we
would
need
to
compose
pretty
carefully
and
and
I'm
more
concerned.
Even
about
with
that
that
the.
F
B
A
K
K
A
F
E
E
F
The
right
place
to
discuss
this
isn't
in
a
pull
request
over
the
code.
It's
in
a
pull
request
over
a
design
proposal,
because
it's
a
much
bigger
topic
like
honestly,
the
annotations
that
we
have
are
sort
of
a
vestigial
leftover
piece
of
an
older
era,
and
they
probably
should
have
been
deleted
a
long
time
ago,
but
they
were
really
just
sort
of
no
cost.
So
we
let
them
ride,
but
they
don't
apply
to
all
network
plugins.
F
They
don't
work
well
with
the
plug-in
model
and
they're,
not
I,
don't
believe
that
they're
scalable
in
the
way
that
we
want
the
system
to
work,
and
so
we
shouldn't,
even
though
the
barter
is
pretty
low
for
alpha
features.
We
shouldn't
alpha
things
that
we
have
no
plans
of
taking
further.
So
we
need
to
discuss
that.
Okay,.
E
F
I,
like
I,
said,
I
welcome
the
conversation.
I
I've
heard
it
from
a
customer
again
just
this
week,
like
literally
this
week,
so
I'm
happy
to
discuss
it.
I,
I
and
many
others
have
thoughts
about
how
we
would
actually
want
to
express
it
and
how
we
want
to
implement
it,
and
they
don't
look
like
what
we've
currently
got
with
those
annotations.
F
A
A
H
A
I
I
I
Yes,
okay,
so
actually
I
am
with
us
from
Red,
Hat
and
I'm
going
to
present
this
idea
where
it
it
enables
to
be
able
to
use
a
specialized
network
hardware
for
network
heavy
user
application
boats,
and
this
solution,
I
ties
together,
device
plug-in
and
Sen
I
plug-in
and
also
this
solution
is
based
on
CR
DPS
to
multi
network
de-facto
standard
was
adding
a
specialized
network.
Hardware
is
essentially
another.
I
So
here
is
how
the
story
starts.
First
of
all,
admin
creates
a
I,
am
assuming
elf
you
a
few
details.
I
am
just
assuming
that
folks
are
aware
about
about
about
the
facts
which
are
already
there
in
the
network:
CRD
based
multi
network
de
facto
standard,
so
just
to
save
some
time.
I
am
just
trying
to
keep
explaining
those,
but
please
feel
free
to
interrupt
whenever
you
feel
that
I
might
be
going
too
fast.
So
so
story
starts
with
the
admin
creating
a
CR
DPS
network
object.
I
So,
as
we
see
in
this
slide,
this
network
objects
are
created
for,
for
example,
control
network
and
the
data
network.
Now
in
this
solution,
because
we
are
using
device,
plug-ins
admin
is
expected
to
to
know
which,
how
the
device
plug-in
is
going
to
advertise
by
what,
by
what
extended
resource
name
device
plugin
is
going
to
advertise
that
network.
I
It
is
expected
that
this
plug-in
will
advert
networks
to
be
consumed
using
device
plug-in
api's
in
this
format.
So
admin
is
expected
to
notice.
In
future,
this
could
be
a
resource
class
Lane.
So
essentially,
this
is
the
mapping
between
the
network's.
They
are
do
you
object
and
resource
with
which
this
CRD
object
is
going
to
be
mapped
so
today,
because
we
don't
have
a
source
classes.
Today,
it's
going
to
be
mapped
with
with
a
resource
name
which
is
going
to
be
advertised
from
device.
Plugin.
I
The
idea,
yet
the
daemon
building
invoke
the
a
pipe
provider
API
or
provision
that
network
at
the
host
level
and
then,
after
once
it
has
created
successfully
the
network
at
the
host
level
or
at
the
cluster
level.
Then
first
structure,
the
availability
will
be
advertised
through
the
DBA,
be
a
list
and
watch
api
to
the
tubular,
and
then
culet
will
show
the
availability
of
that
network
in
the
noted
status.
I
So
this
is
the
network
creation
flow,
so
admin
creates
CRT
object.
Daemon
figures
this
out
that
watches
this
and
after
figuring
out,
okay,
it's
this
network
object
is
targeting
me
with
the
plug-in
name.
Then
this
daemon
CNI
daemon
in
provisions
that
network
at
the
host
level
and
then
advertises
it
to
the
cubelet
using
the
DPP
is,
and
then
this
network
becomes
available
in
the
node
status
as
a
as
an
eligible
resource.
I
Extended
resources
like
any
other
extended
resource
which
today
device
per
gain,
is
equal
to
advertise,
for
example,
GPUs.
So
now
how
how
the
port
is
going
to
request
attachment
to
these
networks.
So
this
is
going
to
be
the
same
as
mentioned
in
the
crt-based
multi
network
de
facto
standard.
Using
the
this
standard
annotation,
the
ODE
will
request
the
will
mention
the
network.
Cid
object
names
like
control
network,
so
these
CTR
a
date
I
net,
are
the
names
of
the
CR
D
Network
objects
that
this
port
wants
to.
I
So
user
needs
to
just
mention
the
CR
T
Network
objectives,
which
the
admin
has
created.
Now
there
will
be
a
admission
controller,
Network
resource
and
just
I
am
just
pulling
it
in
network
resource
admission
controller.
So
this
network,
this
admission
controller,
will
figure
out.
Ok,
we'll
we'll
figure
out
what
3
Rd
objects
will
will
fetch
the
CHC
network
cid
objects
which
were
requested
in
the
portal
check
and
then,
after
in
those
in
the
network
of
check.
I
It
will
look
for
the
look
for
the
annotation
resource
name,
which
is
actually
which
we
discussed
in
the
previous
slide,
which
is
the
mapping
between
the
network
object
and
the
resource
which
this
network
Seattle
jet
is
going
to
consume,
and
then
admission
plug
in
this
animation
controller
will
mutate.
The
odd
request
to
to
add
in
the
limit
section
in
a
way
in
which
the
scheduler
can
understand
that
the
very
it
the
schedule
that
has
to
share
dual
to
which
on
which
notes
a
jeweler
has
to
share,
do
in
this
port.
I
So
so
admission
controller
mutates
the
resources
section
in
the
poor,
spec,
and
so
so
from
the
annotation
in
the
pod
object,
admission,
controller
ads
the
ads
the
limits
section
in
the
it
is
emitters,
the
resources
section
in
the
forest
bag
and
another
task
with
another
responsibility
which
admission
controller
has
to
perform,
is
to
add
a
as
a
context
UID,
which
is
a
UID
generated
by
admission
and
controller
in
the
poor
annotations.
So
we'll
see
why?
What
is
the
use
of
this
context?
Why
do
why
we
need
this?
So.
I
I
After
that
now
this
is
the
this:
is
the
complete
port
creation
flow?
So
this
we
have
discussed
now
till
the
point
number
three
that
poor
gets
created
at
the
API
server
and
because
it
has
extended
resources
and
format
in
the
limits
and
in
the
resources
section
scheduler
can
figure
out
like
in
a
negative
way
which
nodes
are
able
to
support
this
poor
vision
with
nodes.
I
So
when,
in
the
admit
phase
this
allocate
RPC
call
is
invoked
at
the
device
plugin,
we
at
this
combined
CNI
in
plus
device
plug
in
a
demon
and
now
in
context
of
this
specialized
nick
in
the
response
in
the
olympia
response,
whatever
mouthparts
and
the
environment
variables
are
required
to
be
present
in
the
container
those
are
sent
in.
They
look
at
response
to
the
giblet.
I
Also
like
this
is
asked
from
the
DPA
API
that
in
the
allocate
we
want
to
extend,
allocate
to
passport
and
notations
also
to
the
device
plugin.
So
the
context
to
ID,
which
we
saw
in
the
previous
slide,
which
the
admission
controller
was
adding
in
the
poor
object
that
needs
to
get
passed
to
the
see
analytical
plugin
and
that
that
context
ID
is
used
to
track
the
track.
This
request,
like
while
the
request,
goes
to
the
CNI
as
well.
I
So
when
we
have
passed
these
mount
paths
to
the
cubelet
and
the
but
and
cubelet
has
allocated
those
won't
mount
parts
and
these
device
for
a
pod
CNI.
If
today's
he
and
I
this
this
device
plug-in
will
not
be
able
to
know
which
port
is
using
that
allocated
interface,
so
when
the
request
comes
to
the
CNI
for
attaching
the
network,
then
CNI
will
not
know
which
to
which
interface
was
allocated
to
this
port.
So
for
that
we
are
this
proposal.
I
Is
this
proposal
is
asking
for
adding
context
to
ID
by
a
missional
controller
and
that
will
be
used
as
a
cookie
to
figure
out
as
a
as
a
as
a
magic
glue
to
figure
out
like
which
port
is
using,
which
interface?
So
when
the,
when
the
request
reaches
at
the
CNI
and
the
CNI
invokes
the
CNI
ere
call,
we
are
expecting
like
meta
plug-in,
which
is
described
in
detail
in
the
CR
D
based
a
de
facto
standard
document.
So
one
example
is
Moltres
of
that.
Then
it
will
after
providing
a
default
Network.
I
This
meta
plugin
will
look
for
this
annotation
networks.
Penetration
is
string
and
see
what
are
the
networks?
What
are
the
network?
Are?
The
networks
are
the
objects
corresponding
which
are
requested
if
the
poorab
check
and
then
after
fetching,
the
networks
here
to
object.
Its
figures
out
what
is
the
CNI
binary
binary
name,
which
I,
which
this
meta
planning
needs
to
actually
invoke
to
get
this
additional
network
attached
to
the
port
so
using
this
mechanism.
I
This
is
the
same
as
the
multi
networks
here
device
proposal
no
change
there,
so
Mehta
plugin
invokes
the
C&I
static
binary
and
then
this
binary
talks
to
the
CNI
DB
demon
over
a
UNIX
socket.
This
CNI
binary
is
stateless
binary
but
CNI
deeply
daemon
stores
a
mapping
of
context
UID
to
the
device
allocated
device.
So
this
CNI
static,
binary
figure,
solves
the
context
UID
from
the
poor
and
rotations
and
pass
and
and
and
giving
this
context
to
ID
to
the
daemon.
I
It
requests
the
interface
which
the
binary
needs
to
pass
into
the
container
namespace
and
these
steps
of
Eid,
which
I
just
explained,
are
mentioned
here.
That
CNI
invokes
the
ad
to
get
network
configuration
done
for
the
port
and
then
meta
plug-in
figures
out
using
the
network
objects.
Crd
objects:
how
to
invoke
what
is
the
name
of
the
executable
and
then
meta
plug-in
invokes
that
executable,
then
executable.
This
binary
executable
passed
a
context
to
ID
to
figure
out
a
located
device
to
the
daemon
and
because
Damon
is
stored.
I
So-
and
this
is
the
last
slide,
this
is
the
port
deletion
flow.
This
will
be
the
same
as
the
addition,
so
Dell
will
being
walked
at
this
aesthetic,
binary
and
then
again
from
the
context
to
ID.
It
will
be
figured
out
like
which
interface
was
allocated
to
the
poor,
and
then
deallocate
will
happen
of
that
of
that
interface,
so
that
that
was
all
I
wanted
to
share
questions.
I
think.
I
So
I
think
Fabian
can
answer
it
after
me,
but
as
far
as
my
understanding
goes,
Fabian
is
asking
some
changes
in
the
DB
API,
which
will
enable
two
to
pass
the
interface
into
the
or
network
namespace
from
the
device.
Plugin
allocate
API
itself
and
like
device
plugin
will
invoke
internally
CNI
and
we
will
not
wait
for
the
cubelet
to
invoke
the
CNI
and
all
additionally,
it
requires
some
changes
in
the
in
the
CNI
and
CRI
as
well,
so
in
detail.
I
think
Fabian
can
explain.
I
L
Yeah
thanks
for
pressing
this
button,
we
see
you
asked
on
trying
to
action.
The
device
plug-in
API
user
journey
seems
really
complicated,
consume
this
in
production
is
this
and
on
one
hand
there
me
are
the
reasons
to
expand
the
device
plug-in
API.
So
that
might
happen
over
time.
But
for
this
specific
use
case,
are
you
like
really
seeking
feedback
from
the
community
on
on,
like
figuring
out
a
better
strategy
for
handling
Network
Devices
I?
L
F
My
my
initial
feedback
is
sort
of
the
same
as
vish
that
this
is
a
pretty
complicated
baroque
way
of
making
things
happen
where
it
really
screams
to
me
that
there's
some
other
primitives
missing
or
some
other
linkages
missing,
and
given
that
we
have
ten
whole
minutes
left
Fabien
I,
don't
know
if
you're
here,
but
I,
think
your
slot
is
in
jeopardy.
But
I
think
that
that
this
applies
to
both
topics,
I'm
all
for
trying
to
figure
out
how
to
make
this
work
and,
in
fact,
I
believe
very
strongly.
F
I
F
So
I
think,
there's
I,
think
there's
a
nugget
of
interesting
here
and
I
think
that
it's
a
neat
way
to
sort
of
do
the
prototype,
but
I
think
that
it's
it
feels
like
a
prototype.
You
really
want
is
to
figure
out
how
we
would
actually
turn
this
into
a
real
living
thing
and
I
think
it's
going
to
need
deeper
changes,
which
you
know,
unfortunately,
means
slower,
but
I
think
we
can.
We
can
do
it,
but
I
have
some
like
really
fundamental
use
case,
questions
that
weren't
answered
by
this
or
by
even
Fabian's
Doc's
do
so.
F
How
do
we
want
to
proceed
with
this?
Like?
We
only
have
eight
minutes
left
I
can
go
into
them
now.
I
literally
have
two
pages
of
notes
from
there
from
the
doc
pre-read
or
we
can
take
it
back
to
email
or
like
what
else
do
people
want
other
more
fundamental
questions?
People
have
I
feel
like
we
should
probably
do.
F
F
K
C
M
K
Maybe
notes
oh
hey,
I'm,
Tom,
Ian,
and
so
you
know,
I
was
planning
to
speak
about
the
broader
thing
honestly,
I
think
because
they're
so
so
it's
great
to
hear
it
that
we're
looking
for
Lyman's
I.
Think
that's
that's
the
key
here
to
me
and
so
I
think
you
know
from
from
our
perspective,
I
mean
well,
but
because
this
was
preparing
why
it
looks
so
complicated
it's
just
because
that
is
something
that
was
today
with
future
heavens.
K
Well,
what
my
proposals
more
about
and
that's
the
difference
between
the
two
is,
you
know
I
did
not
you
know
to
care
too
much
about
you
know.
How
did
the
api's
look
today,
but
rather
how
could
I
look
in
future?
You
know
how
could
we,
how
could
we
mang
or
change
device
plugins?
How
could
we
eventually
do
smaller
changes
to
see
and
I
in
order
to
fit
all
this
better
together,
I
think
a
faced
of
that
little
scone.
K
N
F
There's
I
think
there's
a
this
is
part
of
the
problem.
I
know,
I,
read
fabien
stock
and
there's
a
couple
of
use
cases
in
there,
but
some
of
the
cases
were
things
like
I
have
a
like
a
physical
topology
that
I
need
to
represent
through
committees
and
be
able
to
schedule
against
part
of
it
was
I
have
effectively
fungible
resources
like
s
RI,
o
V
functions
that
are
all
effectively
switched
into
a
fabric
and
any
function
can
be
any
network,
but
there's
a
limited
number
of
them,
and
then
there's
also
this.
F
This
concern
of
we
have
an
increasing
presence
of
devices
that
are
actually
multi-purpose
devices,
so
you
know
it's
a
GPU
with
a
network
card
slapped
onto
it
and
those
sorts
of
things
and
our
plug-in
model
right
now
between
networking
and
devices
is
just
completely
not
coherent
if
they
don't
interoperate
at
all.
In
fact,
the
thing
that
came
to
my
attention,
through
the
conversations
was
or
through
the
reading,
was
they
don't
even
use
the
same
life
cycle.
Folks,
Mike's,
plug-in
API
operates
in
a
very
fundamentally
different
way
than
the
network
plug-in
API.
Okay,.
N
So
this
is
yes,
so
you
just
listed
three
things:
Tim
the
first
two
sounded
kind
of
similar
and
that
they're
about
getting
scheduling,
constraints
in
quickly
inferring
the
scheduling
constraints
and
slide
by
network
attachments,
and
the
second
one
was
this
thing
about
a
device?
That's
both
networking
and
something
else.
So
there
it
I'm.
What
I'm
hearing
is
two
different
sort
of
things
were
trying
to
accomplish
here:
one
is
getting
scheduling,
constraints
injected
and
the
other
is
unifying
these
two
different
kinds
of
plug-in
models.
That's.
F
N
K
The
slide
tech
I
shared
on
the
email
list,
I
think
that's
a
good
introduction.
I
specifically
prepared
that
for
the
zig
Network
audience,
so
you
know
brief
introduction
into
device
plugins
and
why
we
want
to
use
them
for
for
networking
as
well,
and
maybe
that's
a
good
garden.
It
has
some
links
to
the
you
know:
kubernetes
Doc's,
you
can
go
from
there.
I
guess.
N
K
That's
the
thing
I
mean
I
will
put
it
the
other
way
around
network
devices,
even
overlay
networks
are
just
another
kind
of
device
and-
and
that
is
I-
think
where
the
point
of
is
they?
If
you
take
that
device
class,
they
share
a
lot
of
requirements.
We
have
another
devices
as
well
like
scheduling,
awareness,
and
you
know:
lifecycle
management,
so
setting
them
up
and
delivering
drivers
all
that
kind
of
stuff,
so
I
think
there's
a
lot
of
overlap
which
actually
drove
that
idea.
F
In
in
trying
to
understand
all
of
this
in
my
head,
I
kept
coming
up
with
more
questions
about
well
what,
if
this
and
what,
if
that
and
so
I,
don't
even
think
that
I
have
a
clear
understanding
of
the
actual
requirements.
I
hope
that
a
lot
of
these,
what
ifs
we
can
just
say,
yeah,
don't
do
that,
but,
like
there's
physical
constraints
that
aren't
really
captured
by
the
model,
at
least
as
I
understand
it.
F
You
know
if
I
have
physical
wiring,
that
that
is
in
two
different
networks:
I'm,
not
sure
how
to
represent
that
right
and
I'm,
not
sure
I'm,
not
sure
that
we
can
say
we're
not
gonna
handle
that
and
I'm,
not
sure
that
we
can
say
that
we
don't
have
that.
We
have
to
either
I'm
not
sure
which
side
of
that
fence
we
have
to
follow
on.
So
one.
K
Note
here
is
that
I
think
this
whole
approach
is
when
the
one
has
is
to
some
degree,
just
a
mechanism.
I
think
it's
it's
you
know
in
in
in
large
areas.
It's
just
a
mechanism
to
formalize
how
we
get
devices
into
parts
and
into
containers
depends
on
the
namespace.
Obviously,
and
it
doesn't
solve
some
of
the
problems
like
you
said.
Physical
wiring,
but
I
think
that
can
even
key
I
mean
that
that
is
an
operation
problem.
K
F
F
What
I'm
proposing
I
have
like
I
said:
I
have
notes
and
we
didn't
get
to
talk
much
about
them
today,
I
held
off
on
sending
them
to
you,
because
they're
sort
of
very
hard
to
digest
unless
you're
me
I
will
actually
turn
them
into
full
sentences
and
send
them
to
the
mailing
list
in
response
to
your
slide
decks,
and
we
can
go
with
that
for
now.
I
see
from
the
chat
that
Dan
is
willing
to
volunteer
his
network
plumbing
call
next
week
for
those
people
who
are
at
the
resource
management
face-to-face.
L
F
P
Technically,
the
agenda
ends
three,
which
is
one
hour
after
I,
think
the
Signet
work
slop
Pacific
time
I
may
be
messing
with
the
time
to
thumbs
and
then
that
afternoon
is
a
super
interesting
topic.
We're
trying
to
get
through
the
resource.
Api
stuff.
Is
there
any
chance?
We
can
move
the
call
one
hour
later
on
next
Thursday
so
that
we
could
not
have
that
overlooked.
That's
you
know
spreading
of
folks.
I
could.
C
I
F
P
F
P
F
To
drop
off
right
now,
I'm
late
for
my
next
one
I
will
see
you
guys
in
a
week
and
I
will
send
my
notes
up
before
then.