►
From YouTube: Kubernetes Resource Management WG 20180815
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
B
Service
mesh
and
one
effectively
network
service
faster.
You
can
think
of
this
way.
You
know
we're
all
familiar
with
this
teo
and
the
stuff
that
it
does
for
normal
sort
of
application,
networking
at
layer
4
through
layer
7.
You
know,
tcp
connections,
HTTP
connections,
all
of
that
stuff
and
that's
awesome
and
we've
got
a
whole
set
of
use
cases
that
people
have
around
and
a
fee.
You
know
multi
network
physical
mix,
sr
io,
v,
a
whole
laundry
list
of
things
where
it
turns
out
that
you
can
apply.
B
This
seems
smart
things
that
are
being
done
in
service
mesh.
You
can
apply
a
logs
to
them
to
your
l-2
and
l-3
networking
traffic,
and
so
we
call
that
concept,
network
service
mesh
and
one
of
the
things
that
it
does
that's
really
nice
is.
It
allows
you
to
use
the
existing
device,
plugin
API,
with
no
alterations
in
order
to
achieve
your
ends,
so
we
don't
need
to
do
all
kinds
of
weird
things:
to
change
the
device
plug-in
API
in
order
to
properly
support
partner,
remix
Messer
Iove.
All
of
that
stuff.
B
So
this
presentation-
I'm
not
gonna,
do
the
whole
deep
dive
into
network
service
mesh
because
most
of
it
is
not
necessarily
directly
relevant
nice.
The
resource
management
working
group,
although
I'd,
be
happy
to
do
that.
If
you
guys
would
like
what
I'm
gonna
do
is
start
with
a
very
simple
narrative
story
that
talks
about
how
we
would
handle
in
you
know
physical
mix
with
no
wiring
physical
mix
of
the
pods
you're,
using
the
existing
device
plug-in
API
without
having
to
make
changes
to
it.
B
B
So
this
is
this:
we
have
several
stories
like
this,
so
this
is
a
chapter
in
our
overall
narrative
deck,
so
yeah
hannah
is
writing
a
kubernetes
app
to
be
deployed
on
her
on-prem
kubernetes
cluster
and
one
of
the
pods
in
her
app
needs
to
use
some
hardware
Nick.
Now
some
of
the
nodes
in
Hannah's
kubernetes
cluster
has
special
Hardware
Nick's,
some
don't
so.
This
intrinsically
becomes
a
resource
and
scheduling
problem,
but
it's
not
as
simple
as
I
want
a
hardware
Nick,
because
not
all
the
special
hardware
it
makes
are
the
same.
B
Not
all
them
are
connected
to
the
same
network
right.
Some
of
them
are
connected
to
various
different
networks,
and
even
when
you
talk
about
Nick's
connecting
to
the
same
network,
some
of
them
will
have
different
speeds,
so
tended
versus
100
Gig,
and
so
those
aren't
really
giving
her
the
same
resource
in
this
situation
and
not
all
of
those
mix
are
being
treated
the
same
by
the
network.
They
connect
to
right
so
on
note.
B
You
can't
think
about
it
in
terms
of
the
network
than
it
connects
to,
because
that
doesn't
actually
describe
the
network
service,
that's
being
provided
either.
Nor
can
you
think
of
it
solely
in
terms
of
the
network
plus
speed
of
the
neck.
It
turns
out
that
actually,
what
you're,
really
getting
when
you're
looking
for
this
physical
hardware
resource
is
a
particular
kind
of
network
service
that
could
encompass
all
kinds
of
properties
so,
but
this
complication
is
not
actually
how
Hannah
sees
her
world
right.
B
From
her
point
of
view,
the
problem
looks
like
this:
she
has
a
pod
that
has
to
be
scheduled
so
that
it
can
connect
to
somewhere,
where
there's
a
hardware
date
that
she
needs
and
that
Nick
has
to
be
injected
into
her
pod.
So
this
is
a
very
simple
thing
from
Hannah's
point
of
view,
but
this
is
a
problem
because
you've
got
all
these
harbour
nicks,
and
so
how
did
she
has
to
figure
out
how
she
is
going
to
get
scheduled
to
them
right?
B
You
know
what
services
are
provided
to
the
various
hardware
Nick's.
You
know
how
to
pick
to
the
node
and
hardware
in
which
is
something
you
really
want
to
scheduler
to
do.
You
know
avoiding
conflict
with
people
who
are
already
using
that
node
that
a
hardware
neck,
and
so
this
is
where
we
sort
of
aventurine,
with
our
little
cartoon
character
for
memory
super
smash.
This
is
Ariane.
May
the
NSM
the
resource
mesh
spider,
so
she's
sort
of
her
mascot.
B
So
really
quickly,
just
another
brief
introduction
here.
So
what
is
network
service
mesh,
it's
sort
of
like
Sto,
but
for
l2
and
l3
protocols
right
so
when
you're
dealing
with
this
do
your
payloads
are
HTTP
messages
or
TCP
streams.
Here?
Are
the
payloads
really
are
IP
e
and
Ethernet,
and
other
things
like
that?
So
if
we
go
back
to
look
at
Heather
and
Hannah's
problem
right,
she
just
wants
to
be
able
to
play
a
pod
and
get
one
of
the
hardware
interfaces
that
she
needs
injected
into
it.
B
She
doesn't
want
to
have
to
think
about
which
card
were
in
her
fit
neck
to
make
a
particular
one
in
the
cluster.
She
doesn't
want
to
have
to
think
about
what
note
it's
on
she
doesn't
want
to
have
to
think
about.
You
know
wait
up
with
some
hardware
Nick
that
doesn't
do
the
thing
that
she
needs
and,
of
course
she
also
still
wants
to
get
her
normal
kubernetes
networking
through
the
we'll
kubernetes
interface
that
gets
injected
into
her
pot
right.
B
Unfortunately,
it
doesn't
quite
get
us
there
and
the
reason
it
doesn't
quite
get
us.
There
is
basically
fundamentally
because,
at
the
point
that
vise
plugin
is
being
asked
to
handle
the
allocation
of
devices,
we
don't
actually
have
the
information
we
need
to
handle
the
networking
macness
of
Nyx
and,
at
the
point
we
get
to
something
like
seeing.
I
we're
not
getting
the
information
to
handle
the
hardware
Ness
in
order
to
configure
the
network
enos
of
Nyx.
So
there's
a
disconnect
there.
B
The
other
thing
to
note
is
that
the
the
network
enos
that
the
henna
is
seeking
is
not
actually
kubernetes
network
heinous
that
she
gets
from
CNI
right,
she's,
literally
looking
for
a
hardware
NIC
that
hardware
Nick
is
going
to
provide
a
radically
different
network
service,
then
what
you're
getting
from
kubernetes
intrinsically
that's
part
of
why
you're
asking
from
a
heart
for
a
hardware
neck,
and
you
can't
just
again
be
a
generic
hardware.
B
Nick,
you
can't
say:
I've
got
20
hardware
Nick's
in
my
cluster,
give
me
the
hardware
Nick,
because
what
that
Nick
is
plugged
into
and
the
characteristics
of
the
service
being
provided
for
it,
but
build
by
being
plugged
into
something
is
very
really
realistically
different.
So
you
can't
just
say
something
like
well
give
me
a
hardware
Nick
or
give
me
an
SSRI
of
Enoch
and
expect
to
get
any
meaningful
use
for
response
and,
of
course,
this.
This
is
this
results
inside
Penda
any
questions
so
far.
B
Good
always
happy
to
know
that
it
makes
sense.
So
this
is
where
we
sort
of
get
to
help
me.
Your
network
service
mesh
could
start
helping
right,
so
Network
circus
finish
thinks
about
things
in
terms
of
three
basic
concepts.
The
first
is
a
network
service.
It's
the
abstract
thing
that
you
wish
would
happen
when
you
shove
packets
or
frames
on
that
particular
interface.
It
could
include
things
about
Kwas,
ecl's
connectivity
to
special
networks.
You
know
a
whole
variety
of
things:
security,
firewall,
IDs,
I,
don't
care
right
from
network
service
missions.
Point
of
view.
B
It
is
a
thing
that
you
want
in
the
abstract
right,
so
this
free
for
Hannah's
problem.
This
is
the
stuff
that's
provided
by
the
hardware
mix
in
her
cluster.
That
do
do
the
thing
that
she
needs
right.
So
if
there
are
20
Hardware
necks,
maybe
4
of
them
actually
can
provide
the
network
service
she's.
Looking
for
and
then
the
second
concept
is
a
network
service
endpoint,
it's
a
concrete
thing
that
provides
an
instance
of
that
network
service
right.
B
B
B
B
So
you
know
hemisphere
right,
so
she
immediately
gets
this
concept
and
physically,
it
looks
sort
of
like
this.
You've
got
the
port
in
the
top
of
rack
switched
it's
provide!
That's
the
network
service.
Endpoint
you've
got
the
hardware
NIC
that
is
essentially
giving
her
in
the
l2
l3
connection
to
that
that
she's
looking
for
everything's
still
making
good
sense
to
folks.
A
B
Well,
we
can
definitely
do
that
when
we
sort
of
sent
them
through
this
a
little
bit.
So
how
do
you
use
this
as
sort
of
the
central
question?
We
got
you
of
these
three
abstract
concepts
right
so,
in
network
service
mesh
we
have
a
C
or
D
that
we
call
a
network
service,
you
give
it
a
name
and
it
exposes
channels.
You
could
these
think
of
these,
this
as
being
an
analogue
to
services
and
kubernetes
and
think
of
channels
as
being
an
analogue
to
ports.
B
The
big
difference
here
is
that
we
focus
on
the
payload
rather
than
necessarily
on
the
transport.
Please
don't
network
service
does
a
lot
more
than
handling
hardware
NICs,
so
that
there
are
a
few
more
handles
here
than
you
would
do
if
you
were
doing
a
completely
bespoke
solution
for
our
drain
X
and,
of
course,
with
this,
never
services
really
doing
as
an
abstract
is
representing
that
class
of
hardware
NICs.
They
do
the
thing
that
had
that
Hannah
needs
so
in
the
spec,
you
just
treat
it
like
any
other
resource
right.
B
So
you
you
have
network
service
metrics
at
I/o,
slash
enhanced
corporate
connectivity.
If
enhanced
corporate
connectivity
is
the
name
of
your
network
service-
and
this
is
right
and
appropriate,
because
you
really
are
asking
for
a
physical
resource
you're
asking
for
some
hardware
NIC-
that
is
able
to
provide
you,
this
network
service
in
the
cluster
and
then
that
particular
resource
gets
advertised
as
a
device
plug-in
in
the
normal
sort
of
way.
So
all
the
normal
scheduling
that
works
well
for
device
plug-in
API.
B
B
So
when
Hanna
runs
her
pod,
it
gets
scheduled
of
the
normal
way
by
the
device.
Plugin
API
and
the
normal
scheduling,
Corelli's,
and
she
would
need
to
run
it
with
a
network
service
mesh
in
it
container
and
a
config
map.
That
basically
says
look.
This
is
the
kind
of
network
service
that
I
want
so
just
a
question.
There
is
this:
basically,
the
equipment
like
a
sidecar
container
in
Sto
and
then
it's
really
conceptually
closed.
So,
if
the
time
being
you
shorthand
it
that
way
in
your
head,
you
are
not
too
far
off,
which.
A
B
B
A
B
The
normal
container,
this
is
not
being
a
container
I'm
discussing
here,
okay
and
then
there
is
a
little
bit
of
funkiness
there,
because
you're
talking
about
networking
us.
Typically,
we
scope
the
network
enos
to
pods
and
not
to
container
so
there's
a
little
bit
of
fuzziness.
There
I
absolutely
agree
so
effectively.
Here
when
you
run
your
you
run
it
within
a
container
and
a
config
map
and
network.
So
the
Edit
container
is
not
rocket.
B
Science,
there's
no
reason
to
write
more
than
one
of
them,
so
network
service
mesh
is
literally
writing
one
and
it
just
takes
a
config
map.
You
can
basically
say:
hey
I,
want
this
network
service
and
by
the
way,
I'd
like
it
as
a
hardware,
Nick
and
since
you've
already
been
scheduled
and
allocated
the
hardware
Nick,
it's
a
perfectly
valid
thing
for
you
to
be
able
to
add
to
the
request
connection
that
you
send
to
the
network
service
manager.
B
That
would
be
set
by
an
environment
variable
indicating
that
you
have
in
fact
been
allocated
this
hardware
Nick
and
then
based
upon
that
it
will
inject
that
hardware
Nick.
Let
the
init
container
know
that
this
is
work,
has
been
done
and
will
basically
finish
the
unit
container
exits
and
the
pod
proceeds
normally
so
fairly
straightforward
and
simple,
and
no
actual
changes
to
the
device
plug-in
API
in
the
scheduling
it
works
exactly
the
way
it
always
has.
This
is
just
dealing
with
the
networking
as
part
of
the
problem
which.
C
B
The
network
service
manager
is
handling
that
directly.
We
do
not.
You
see
a
tie
in
this,
because
this
is
actually
not
a
problem
that
CNI
is
terribly
well
scopes
for.
Ci
works
really
really
well
for
handling
the
communities
networking
stuff,
but
here
we're
literally
just
talking
about
injecting
a
hardware
Nick
into
the
system,
and
we
do
this
for
a
lot
of
other
things
as
well,
where
you
explicitly
aren't
looking
for
kubernetes
network
services
when
you're
trying
to
wire
up
network
service
specialist,
so
yeah
this
is
not
going
through
CNI
at
all.
B
This
is
not
part
of
CNI
lifecycle,
and
this
is
part
of
how
we
escaped
the
trap
that
you
have
that
I
pointed
out
earlier,
where
the
Weisse
plug-in
doesn't
know
enough
to
be
helpful
with
network
enos,
because
that's
not
knowable
at
that
point
and
the
CNI
doesn't
get
the
necessary
information
to
be
able
to
do
anything.
So
we
literally
don't
have
that
problem,
because
we
don't
actually
fall
into
that
trap.
B
D
D
B
So
keep
in
mind
here
is
effectively
what
we've
done
here
is.
So
there
are
two
ways
you
can
look
at
this
problem.
A
lot
of
the
situations
I've
seen
the
original
ways
that
the
recent
the
name
of
a
resource
was
being
used.
Was
this
defining
a
class
of
hardware
right,
and
so
you
see
things
like
the
SR
Iove
device,
plugin
now
an
SSRI
to
aviod
isolation.
It's
not
all
that
interesting.
Now.
B
There
are
all
kinds
of
interesting
discussions
you
can
make
about
adding
various
properties
to
devices,
and
some
of
those
move
proved
to
be
entirely
fruitful.
But
the
thing
that
we've
realized
here
is
at
least
when
you
talk
about
networks
and
network
services,
because
we're
dealing
with
network
services
in
the
generic
and
they're
simply
mean
things
that
you
don't
have
to
burden
the
user
with
figuring
out
a
broad
collection
of
properties.
B
D
B
Yep
but
no
excellent
question,
though
cool
so
then
you
know,
of
course,
you've
just
used
the
device
plug-in
mechanisms
of
the
normal
way
to
schedule
the
resource
for
the
hardware
Nick.
So
nothing
has
to
change
here
about
the
device
plug-in
API.
So.
B
So
one
of
the
things
that
work
that
I'm
presuming
here
is
that
a
pod
has
a
single
network
namespace,
and
so,
if
I'm
talking
about
injecting
a
Nick
into
a
pod,
I
am
talking
intrinsically
about
about
injecting
that
hardware
Nick
into
a
network
namespace,
which
means
that
it
is
intrinsically
shared
across
the
entire
pod.
That's
just
the
nature
of
the
way
positive
Network
namespaces
work
it
in
kubernetes
today,
but.
B
Is
true,
so
there
is
a
fuzzy
side
effects
here,
which
is
that
if
a
container
makes
a
request
for
a
hardware
Nick
for
that
container
in
that
PI,
not
Nick,
by
virtue
of
the
way
that
no
fundamentally
handled
in
kubernetes,
that
Nick
will
be
available
to
every
container
in
that
lot,
but
I
think
if
I'm
and
if
I
make
a
request
for
tunics,
you
will
get
tunics
in
that
pot
right.
You
can
anything
you
want
as
long
as
the
reasons
are
available.
A
B
A
B
B
My
colleague
Sergey
is
here
to
talk
about
some
of
the
other
things.
Cuz
he's
done
some
of
the
work
on
this
in
code,
and
he
has
a
laundry
list
of
this
would
make
life
better
but,
like
I
said,
there's
a
big
difference
between
we
have
to
do
a
major
reorder
to
make
sure
it
work,
and
how
do
you
monitor
the
health
of
these
Nick's,
the
health
of
the
next?
We
have
monitoring
of
the
health
of
the
Nick's.
B
At
this
stage,
it's
important
to
understand
the
network
service
mesh
is
still
under
development
right,
so
we
literally
have
been
moving
this
concept
forward
for
the
last
four
months
and
that
it
does
quite
a
lot
more
than
just
managing
hardware
decks.
We
also
deal
with
the
problem
of
you
know,
problems
associated
with
multi
network
and,
if
the
whole
laundry
list
of
things
that
turn
out
to
map
it
turns
out
a
very,
very
large
and
complex
set
of
use
cases
mapped
to
a
relatively
simple
constructs,
which
is
always
a
good
sign.
A
Okay,
I'm
just
thinking
for
the
audience
of
folks
on
this
chat
right
now
for
the
areas
that
have
been
under
discussion
for
the
device
plug-in
API
that
we
feel
like
would
need
to
change.
It
was
around
monitoring
and
potentially
health.
So
just
curious.
If
you
had
any
insight
on
what
you
would
like,
as
a
user
could.
A
A
B
Okay,
so
you
know
from
a
monitoring
standpoint:
that's
really
on
what
you
want
to
do
in
terms
of
health
of
the
neck.
In
networking
you
can
treat
the
health
of
a
neck
from
something
extremely
simple,
like
those
the
NIC
report
is
being
off,
but.
A
B
E
Yes
and
also
I,
don't
know
whether
you
followed
David,
I,
suppose
devices
I
mean
proposal
and
I
think
well.
His
workers
trend
provides
some
API
for
third
party
that
needs
to
discover
the
device
assignment
information
so
that,
because
associated
allocated
device
with
the
product,
you
know
that's
done.
You're.
A
E
B
E
E
B
Network
service
mission,
I
Oh,
slash,
the
name
of
the
service,
is
effectively
what
we're
looking
at,
or
something
some
rough
variation
of
that
sorry
about
video
movie
network
service
effectively.
The
name
of
the
network
service
represents
a
class
of
things
that
are
considered
to
be
equivalently
good
for
this
purpose.
B
So
that's
an
interesting
problem.
The
there's
a
whole
interesting
problem
here
of
trying
to
sort
out
and
decide
how
you're
going
to
name
the
services
being
provided
by
a
physical,
Nick
and
effectively
I
think
the
situation
you're
we
determine
is,
let's
say,
for
example,
I
have
three
Knicks
and
two
of
them
can
provide
service
speed
and
two
of
them
can
provide
service
a,
but
only
one
of
them
can
provide
service
a
or
service
B.
Is
that
more
or
less
the
thing
you're
gonna
get
so
there's
there's
this
overlap
between
them.
Yeah.
B
A
set
of
speed
characteristics,
a
set
of
things
that
you
care
about
right,
and
you
know
the
things
the
set
of
things
that
you
care
about,
and
some
of
the
things
you
care
about.
It
may
be
as
simple
as
this
NIC
is
actually
going
to
something.
That's
running
you
on
the
shortest
possible
fibre
between
Chicago
and
New.
York
right
I
mean
there
are
a
lots
of
things
that
it
could
be,
but
effectively.
B
The
developer
only
really
cares
about
one
set
of
characteristics,
and
that
is
does
it
do
this
thing
that
I
need
to
do,
and
so
yes,
we
do
basically
have
to
pegging
NIC
to
a
particular
choice
of
name
for
that
class
of
things.
So
a
resource
in
it
belongs
to
exactly
one
resource
name,
but
many
mix
can
provide
the
same
resource
name.
F
C
Well
not
entirely
you
had
in
you
one
of
your
first
pictures.
You
had
like
network
one
which
had
10
gig
connection
and
the
hundred
gig
connection
from
user
perspective.
I
would
say:
I
won't
network
one
and
I
don't
really
care
what
is
with
speed.
However,
it
might
be
a
second
class
of
users
which
says
I
want
network
one
and
I
won't
hide
a
hundred
geek
in
until
until
reelect
with
hundred
geek
Nanak
is
able
to
provide
both
like
network
one
and
network
100
geek.
B
This
is:
do
you
want
to
go
through
and
enumerate
and
have
the
developer,
deploy
their
app
have
to
enumerate
the
complete
set
of
these
specs,
and
then
how
do
you
do
the
fuzzy
matching
around
them,
which
could
be
very
semantic
and
context
sensitive?
You
know,
and
that's
one
way
of
approaching
the
problem
and
I'm
sure
there
are
use
cases
for
which
it's
a
valuable
way
to
approach
the
problem.
B
But,
but
if
you
try
and
enumerate
all
those
things
in
the
semantics
around
them,
because
those
semantics
can
be
complicated,
you
can
wind
up
with
a
very
a
very,
very
complicated
solution
and
a
very
complicated
to
use
solution
in
many
cases.
So
in
some
cases,
it'll
be
worth
it
in
some
cases
it
won't.
So
we
give
you
this
sort
of
easy
to
use
version
of
this
problem.
E
B
B
It
really
depends
on
how
you
want
to
define
the
network
services
that
are
being
provided,
and
the
main
point
I
make
with
this
is
that
the
things
that
go
into
that
are
enormous,
Li,
more
complex
than
just
big
speed
or
or
network
you're
attached
to
or
any
of
the
rest
of
it,
and
so
that's
going
to
be
a
decision
on
the
part
of
the
person
deploying
the
cluster
okay.
B
One
network
service
manager
per
node
that
can
handle
all
the
network
services
that
all
the
network
service
stuff
going
on.
On
that
note,
so
you
only
wind
up
with
the
one
daemon
set
you
know,
so
you
end
up
with
one
point:
network
service
manager
per
node,
so
you're
gonna
have
as
many
network
services
as
you
would
like,
including
as
many
network
service
resources.
B
Here
you
would
just
have
the
network
service
manager
acting
as
multiple
device,
plugin
api's
for
that
case,
their
vice
plug-ins
for
that
case,
because,
obviously
it's
going
to
advertise
multiple
resources,
but
you
don't
have
to
run
entirely
separate
daemon
sets
for
each
one.
E
B
A
purely
secure
numeration
at
this
time,
so
it's
literally
mapping
the
hardware
nicks
on
that
node
to
the
network
services
as
an
enumeration,
so
think
of
it
as
a
a
map
that
has
a
basically
they
think
of
it.
As
a
map
structure,
we,
the
key,
is
the
particular
hardware
neck
and
the
value
is
the
network
service
that
it
is
being
okay,.
B
Are
some
tricky
corners
there
around
dynamically,
changing
that
config
map
that
we're
still
trying
to
work
through
so
I'll
give
you
like
one
classic
example
say
you
want
to
change
dynamically,
the
config
map
for
a
service,
but
you
want
to
change
it
to
take
a
NIC.
That
is
the
sign
to
service
a
and
you
want
to
assign
that
NIC
to
service
feet,
network
service
B,
but
somebody's
already,
using
that
Nick
for
network
service
a.
How
do
you
actually
handle
that
problem?
How
do
you
handle
life
cycling
of
it?
E
Yes,
it's
actually
also
some
of
you
have
discussed
well,
they
are
discussing
the
new
recipe
I,
like,
as
you
mentioned,
lifecycle
management,
but
they're.
Changing
the
my
opinion
of
the
parts
already
run
you
or
not,
I
think
we
actually
have
some
very
similar
discussions
and
I
guess.
The
question
is
whether
the
logic
should
be
handled
in
the
video
like
resource
management
are
like.
We
want
some
generic
way
to
handle
that
so
I
will
also
add
a
link
of
some
discussions,
and
maybe
we
can
take
a
look
massively.
B
So
you
know,
and
then
also
noting
in
here,
that
you
essentially
get
connected
to
the
right
hardware
neck,
because
you're
asking
not
for
a
particular
device,
ID
you're
asking
for
the
network
service
that
is
provided
and
then
you
know,
the
network
service
mesh
mechanism
is
handling
all
the
network
enos
there
for
you.
So
this
approach
has
a
few
nice
things
in
terms
of
deployability,
so
we
don't
need
any
changes
for
this
approach
to
the
device
plug-in
API.
There
are
things
that
would
certainly
be
nice
and
helpful,
but
it
works
out
of
the
box.
B
We
also
don't
need
any
special
C&I
plugins
because
we're
basically
completely
orthogonal
to
CI
here,
because
you're
not
actually
asking
for
kubernetes
network
services
that
you're
asking
for
our
Riddick
you're,
asking
for
a
more
general
network
service,
cool
I,
think
that's
the
end
of
the
neck.
At
this
point,
so
that's
I,
guess
where.
B
Project,
let
me
go
ahead
and
grab
that
really
quickly
and
I
will
stick
it
into
actually
I
think
I
did
stick
it
into
the
know,
I'm
a
complete.
So
let
me
stick
that
into
the
needy
minutes.
B
So
we
have
a
you
know,
connection
to
all
of
our
in
re
media
that
our
repo,
which
is
now
linked
in
the
meeting
minutes.
We
have
the
connection
to
all
the
resources
that
we
have,
including
we
have
weekly
meetings
at
8
a.m.
Friday
morning
on
zoom'.
We've
got
a
mailing
list,
IRC
channels,
all
the
rest
of
that
as
well
and
and
as
I
said,
we
do
quite
a
bit
more
than
just
handle
hardware
NICs.
If
you
guys
are
interested
in
me
of
that,
you
will
find
that
Dec.
If
you
go
a
little
bit
further.
B
A
Don't
know
if
there
any
other
following
questions,
but
I
guess
a
big.
Thank
you
for
talking
about
the
exciting
things
you
guys
are
doing
with
device
plugins.
It's
always
good
to
see
the
work
that
we
as
a
convenient
on
get
used
and
eliminated.
So
thank
you
and
if
our
concrete
changes
that
you'd
like
to
see
I
know,
you
alluded
to
the
list
of
things
that
you
think
we'd
all
looked
at
to
see
those
as
well
so
make
the
experience
less
fuzzy
and
verify
things
where
they
could
be
clarified.
So,
thank
you.
A
B
Talk
a
little
bit
about
that,
so
my
colleague
Sergei
actually
has
done
some
early
implementation
for
this
stuff
around
us,
our
iove
and
he
put
together
somewhat
of
a
wish
list
of
some
of
the
things
that
he
encountered.
That
might
be
potentially
nice,
it's
nice
to
have,
and
it
might
be
very
productive
to
talk
through
that,
particularly
since
I
know,
part
of
my
past
experience
dealing
with
you
guys
is
half
the
time
when
I've
come
in
and
said
things
like
I
think
I
need
deallocate.
What
I
discover
is
no
no
I,
don't
need
to
allocate.
B
F
F
F
So
one
of
the
one
of
the
interesting
discoveries
for
myself
was
that
the
life
cycle
and
then
when
I
started
looking
at
it,
it
seemed
even
though
it's
working
as
ed
mentioned
perfectly
now,
it
could
be
a
I
think
it
could
be
optimized
a
little
bit,
especially
since
in
the
networking
world
the
resources
or
the
network
services
are
a
bit
more
dynamic
than
the
GPU
example.
You
install
GPU
in
the
server
in
most
of
the
cases.
F
Basically,
you
need
to
power
off
the
server
to
switch
the
GPU
to
replace
it
to
add
something
in
the
networking
world.
The
services
could
be
deep.
Provisioned
reprovision
changed
modified
in
certain
ways
so,
and
if
we
had
a
little
bit
better
at
the
lifecycle
management,
it
would
help
us
a
lot
example.
I
have
a
network
service
for
a
villain
I
like
the
long
time,
and
currently
it
has
10
interfaces
which
are
VF
and
each
backed
up
by
10
gigabit
interface.
F
So
we
upgrade
the
we
we
could
add
or
move
could
change
this
disservice
for
VF
provided
by
another
node,
which
has
a
40
gig
network
card
or
a
hundred
gig.
So
basically
changing
that
network
service
and
replacing
this
VF
by
a
different
VF.
It's
it's
very
possible
and
I
don't
need
to
basically
I
don't
need
to
restart
Trello
at
the
server
it
can.
It
can
be
done
pretty
much
on
fly.
How
would
I
inform
the
couplet
that
this
instances
are
gone
or
being
replaced?
F
E
It
should
yeah,
could
you
try
it
and
if
you
see
any
problem,
let
us
know
so
we
to
have
some
improvements.
We
need
to
do
further
research,
lifecycle
management
sent
you
the
document,
I
shared
a
racist
latest
in
the.
Why
it
it's
in
the
current
status,
but
basically,
if
some
resources
no
longer
other
what
highest
we
set
its
capacity
to
zero,
that
they
removing
eight
but
I,
actually
curious
to
know
like
how
it
affects
your
yeah
use
case
like
most.
A
A
F
A
An
abstraction
of
what
you
could
think
it
was
like
an
Amazon
is
G
or
as
our
resource
template
or
something
where
you
can
address
a
set
of
machines
and
modulus
Li,
and
one
of
these
cases
would
be
to
have
a
bare
metal,
actuator
or
the
actuator.
So
you
can
treat
bare
metal
machines,
they
can
come
and
go
and
scale
dynamically.
And
so,
if
you
were
targeting
a
bare
metal
use
case,
I
would
be
curious,
like
if
I
was
to
pair
your
work
here
with
that
other
project.
A
F
F
It's
again,
it's
mostly
for
the
REE
REE
provisioning.
Let's
say
we
have
a
port
which
is
with
a
bunch
of
SR
AV
devices
attached
to
it.
What
happened
if
this?
If
there
is
a
network
change,
so
the
network
operator,
they
changed
the
trunk
or
they
change
the
interface
which
was
mapping
this
VF
instances.
F
Currently,
we
don't
have
any
feedback
which
port
uses
this
specific
VF,
since
there
is
no
put
ID
name
space
or
the
namespace
name
or
the
port
name
in
the
allocate
request.
If
we
had
that,
then
we
could
think
about
some
logic.
To
informing
put
your
VF
is
no
longer
available,
I
mean
just
exit
or
some
sort
of
reaction,
or
even
even
just
pushing
some
look
or
event.
Something
like
that
could
be
extremely
useful
to
the
end-user.
E
F
F
F
E
We
currently
don't
do
anything
well
Roger
and
the
continuous
data
already
run.
You
very
simply
like
deduct.
The
database
count
from
the
allocated
pub
and
it's
kinda
still
reflected
in
capacity.
So
the
idea
is,
it's
a
bit
hard
to
differentiate
like
a
simple,
yet
device,
failure
versus
a
device,
plug-in
failure
and
the
way
don't
want
to
cause
a
dramatic
failures
difference
the
device
plug
in
failure
which
can
happen
currently
you're.
The
best
braking
upgrade.
B
Because
it
strikes
me
as
a
very
hard
problem
to
decide
what
is
the
right
thing
to
do
in
that
situation,
and
probably
the
least
harmful
answer
for
that
question
is
nothing
until
someone
has
a
really
good
idea
about
what
we
should
do.
That's
reliable
and
I
I'm
not
smart
enough
to
have
an
idea.
E
Mm-Hmm
yeah
I
think
yeah
generally,
we
expect,
like
thirty
hundred
in
it's
the
hand,
order
outside
the
like
a
couplet,
but
we
can
sing
about,
or
rather
we
need
some
general
approach,
but
I
don't
know
whether
I
think
there
are
some
missing
hooks
that
can
check
some
house
needs,
so
certain
things
and
I
don't
know
whether
you
can
use
them.
What
did
she
was
kissing.
F
Locked
memory
amount
and
it
was
crushing
so
we
had
to
use
some
work
around
with
some
doing
to
be
able
to
extend
that
locked
memory
size.
So
the
application
would
be
running
success
successfully
and
the
second,
since
SRV
in
most
of
the
cases,
are
used
in
the
NF
VI
scenario,
with
the
high
performance
data
plane.
Things
like
the
CPU
ping
and
the
Numa
allocation
or
being
aware
is
very
important
for
a
good
performance.
So
that's
pretty
much
what
I
have.
E
I
think
I
assume
people
have
discussed
that
a
lot,
but
quite
a
lot
work,
and
so,
if
you
are
interested
I,
think
some
folks
in
this
group
may
have
better
knowledge
for
the
yogi
mate.
That
seems
a
generous
ignore
the
future.
So.
A
A
B
C
B
B
There
was
a
some
conversation
for
some
time
about
making
sure
that
was
actually
true,
because
it
was
realized
that
privileged
container
privileged
containers
for
general
use
are
probably
not
a
great
plan,
but
you
know
I
can
go
check
in
on
see
the
see
what
the
state
of
the
work
is
around
that,
but
I
think
we
are
either
there
or
close
to
the
point
where,
if
you
have
a
high
performance,
CNF
cloud
native
network
function
that
needs
real
access
to
that
Nick
V
is
something
like
DVD
K.
You
can
get
it
without
having
to
be
privileged.