►
From YouTube: Kubernetes SIG Node 20180911
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
A
Daemon
set
that
queries.
The
cubelet
for
its
device
gets
that
information
and
then
uses
that
to
attach
metadata
to
metrics
about
these
devices
on
the
node.
So,
for
example,
for
GPUs
the
monitoring
game
and
set
wouldn't
query.
The
devices
endpoint
find
out
that
this
workload
pod
is
using
two
GPUs
and
which
GPUs
and
which
containers
are
using
them
and
then
would
be
able
to
export
metrics
that
have
information
about
the
workload
attached
to
them,
and
it
would
adhere
to
the
node
monitoring
guidelines
that
Saleh
is
working
on.
C
A
The
cubelet,
when
it
gets
a
upon
during
admission,
will
first
call
allocate
on
a
device,
and
only
then
will
it
actually
assign
it
to.
It
only
assigns
that
device
to
a
container
if
the
allocate
call
succeeds.
So
if
the
allocate
call
fails
or
the
device
plugin
isn't
ready
for
some
reason,
then
it
doesn't
allocate
that
that
device
to
the
container.
C
A
D
A
B
E
So,
instead
of
like
having
devices
as
the
rest
resource,
would
it
make
sense
to
have
some
song
like
global
scoping,
I'm,
not
sure
I
understand
here
the
REST
API
that
you're
proposing
now,
if
I
understand
correctly,
is
like
there
will
be
a
cubelet
endpoint
called
slash
devices
so
I'm
asking
if
instead
it
makes
sense
to
have
a
higher
level
resource
like
save
pods,
pods,
slash,
pod,
name
or
pod.
You
ID
followed
by
devices,
because
if
we
have
that
scoping,
then
you
can
add
more
API
than
there
in
the
future.
E
A
A
So
I
don't
expect
this
to
be
used
purely
by
device
plugins
themselves.
The
idea
is
that
this
theming
set
can
be
separate
from
device
plugins
and
doesn't
necessarily
have
to
be
the
same.
The
only
thing
that
so
I'll
try
and
flush
out,
what's
what's
in
my
mind,
a
little
bit
more
to
hopefully
add
some
clarity.
A
If
I'm
a
device
vendor
I
publish
a
daemon
for
a
device,
plugging
daemon
set
that
does
allocation
the
LP
or
allocation
and
device
health
and
such
and
then
I
also
publish
a
separate
game,
except
that
those
monitoring
and
because
I
get
to
choose
the
resource
name
under
which
I
advertise
my
devices
say
who
I.
Also,
then,
when
I
run,
my
monitoring
daemon
set
only
query
for
the
Foo
devices.
E
One
of
the
things
that
you
would
want
we
could
want
see
advisor
to
do
is
like
we
could
say,
go
query
this
board
endpoint
and
then
we
can
have
see
advisor,
get
a
list
of
all
the
things
that
are
attached
to
a
pod,
and
then
it
can
deterministically
continue
monitoring
right
rather
than
having
to
query
suffered
api's
and
having
to
know
that
multiple
API
is
exists,
each
for
different
entities
that
get
attached
to
a
pod.
It's
just
that
it's
it's
like
this
one
place
discovery
for
all
introspection
on
xtext,
but.
E
A
E
I
gotta,
just
clarify
I,
didn't
mean
when
I
meant
pods
I
didn't
mean
like
attached
to
the
pod
itself,
but
like
attaches,
that
should
logically
be
amazing.
The
question
is
like:
which
way
do
we
go?
Do
we
from
part
to
its
different
resources,
which
is
the
natural
way
of
looking
at
how
resources
are
being
managed,
or
do
we
start
from
the
resource?
I
see
the
need
for
possible
need
for
exposing
an
API
that
starts
from
the
resource
itself,
but
it's
one
I'm
just
like
asking
you
to
think
through
the
other
scenario
as
well
yeah.
E
A
Yeah
I
I
think
that's
a
good
point
and
I'll
try
and
I
I
think
it
at
the
core.
The
reason
why
I'm
choosing
this
route
or
I
think
we
should
choose
this
route,
it's
that
were.
This
is
a
discovery
API
as
much
as
anything
else,
and
you
can't
discover
pods
that
you
don't
or
you
can't
query
for
pods
necessarily
that
you're
trying
to
discover
at
the
same
time
so
it's
hard
to
have
a
an
endpoint
that
is
like
querying
for
specific
pods.
That
would
meet
this
use
case,
but
I
can
think
of.
E
Yeah,
fair
enough
I
mean
maybe
you
might
have
to
like,
create
all
the
pods
and
then
start
building
a
reverse
lookup
based
on
your
local
cache.
Such
an
API
might
help
other
agents,
while
not
breaking
the
you
state
primary
use
case
that
you're
tackling
I
feel
that
it's
at
least
worth
thinking
through
yeah.
A
E
D
A
G
Yeah,
the
color
containers
itself
is
trying
to
implement
the
possum
box
with
the
lightweight
VM
and
working
on
the
integration
through
the
Shillito
api's
to
integrate
with
kubernetes
and
country
the
so
far
so
good.
We
have
the
code,
that's
passed
most
of
the
government's
equally
test
cases
and,
except
for
cases,
failed
and
all
of
the
path
forward
related
best
one
of
the
the
cra-cra
acai,
and
we
investigate
investigating
on
this
on
this
feature,
and
we
found
the
Kanpur
forest
which,
if
I'm
finding
say
alright,
yes
and
all
of
them,
I
implemented
in
the
ER.
G
G
G
Simple
simplest
to
be
simply
that's
run
what
this
command
I
just
enter,
two
coordinates
magnums
face
and
that
mantra
so
cat
in
the
in
the
namespace
to
connect
to
the
thing
apart
and
in
this
command
line
in
AI,
Center
and
so
kind
of
both
from
the
coastal
ecosystems.
And
you
know
if
the
the
possum
box
is
implemented
by
is
the
namespace
that
works.
Fine.
However,
if
the,
if
the
container,
if
the
possum
box
is
implemented
by
other
other
tools,
such
as
by
the
we
are
like
Kara
containers,
this
this
world
works
anymore.
G
G
The
second
possible
solution
is
at
the
socket
in
the
working
dinner.
This
is
this
is
imagined
in
the
in
a
in
basically
this
issue
and
that's
proposed
two
years
ago
and
decided
to
want
to
put
some
honey
in
for
a
container
and
then
it
may
execute
the
so
cat
decided
to
do
the
to
the
perform
perform
from.
G
From
my
my
opinion,
I
think
the
infra
container
is
not
designed
to
be
to
do.
Anya
asked
the
RTC
inside,
so
this
method
maybe
come.
They
can
cannot
be
violently
and
really
well.
Now
we
have
the
another
solution
that
is
by
Sunday.
You
can't
see
our
CRE
interface
and
we
are
low
to
either
remove
the
container
insider
running
hot,
so
we
we
may
have
remains
implemented
by
this
way.
You
can't
create
a
when
we
got
a
/
forward
requesting
and
but
Kenobi
itself.
D
H
Shouldn't
this
impulses,
unlike
the
complex
it
here,
how
couponing
to
management
to
lifecycle
of
the
container,
so
the
basically
Cuban
I,
don't
have
no
idea
about
this-
is
so
hot
container
crate
here
and
there
are
secretly
created
by
that's
their
I
and
there.
But
then
they
don't
know
the
other
Staters
about
this
type
container.
If
this
time
container
it
is
died
and
then
there's
the
no
visibility
to
what's
going
on.
Who
is
that
functionally.
G
H
G
H
G
I
G
G
If
we,
if
we
like
to
do,
we
have
you
know
my
my
my
name
is
through
deployment
days.
Personally,
we
can,
we
can
have
a,
we
can
become
four
imager
and
the
try
to
keep
it
not
not
not
be
garbage
collected
and
secondary.
We
may
try
to
thing
cost,
because
the
so
cat
and
the
Sun
Hat
is
just
like
in
the
house
of
systems.
G
J
H
Kinda,
my
question
is
the
rabbit
ears.
I
feel.
What
do
you
describe
here?
No
matter
is
the
image
handling
orders
like
the
left
stack
of
this
time
container.
It
is
work
maybe
for
the
particular
gaming
production,
but
Takumi
is
not
a
general
solution
for
the
signal.
So
next
any
so,
basically
there's
a
lot
of
things.
People
could
do
to
have
this
a
feature
work
in
their
production
and
if
it
is
not
the
communistic,
it's
not
like
the
predictable
behavior
for
any
gaming
solution,
so
that
can
make
me
a
little
bit
concerned.
I
also
I.
H
G
Actually,
the
the
number
three
is
the
used
to
implemented
and
the
chain
to
do
pretty
local
and
but
for
the
for
the
first
for
salon.
We
have
to
change
both
the
and
can
heavy
part
and
that
they
be
the
other
machine
machine,
machine,
API
part-
and
you
know
the
this-
will
change
too
many
protocol
for
the
for
the
for
the
for
the
farm.
Not
only
that
you
find
a
patient,
but
also
the
support
of
those
the
shimma
shimma
api
is
a
is
a
is
a
api
even
face
for
the
pub
and
there's
other
other
protocols.
G
G
But
the
number
three
this
this
just
work,
see
in
the
in
the
pan
to
sear
CRA
semantics.
This
work
for
anything
it's
just
a
very
local
change,
and
actually
it
could
be
done
just
in
the
in
the
powerful
implementation
itself
in
well
functions,
so
so
from
the
from
the
impairment
implementation
consider
region
in
this.
G
G
J
It's
a
minor
implementation
detail
from
a
container,
D
and
cryo
perspective.
To
do
this
launch,
however,
the
you
know
it
had,
it
does
have
larger
repercussions
up
stream
insofar
as
if
you
were
to
ask
how
many
containers
give
me
a
list
of
containers,
this
was
now
gonna
show
up
in
the
would
would
be
no,
you
know
I
guess
we
could
filter
it,
but
it
would
still
need
to
be
hosted
and
you
know
be
long-lived
right.
J
J
F
H
K
K
G
L
G
H
H
G
K
H
J
J
H
So
so
so
she
about
you
propose
so
I
can't
understand
why
you
want
limited,
changing
and
the
local
area
and
other
things,
and
actually
the
even
for
you
for
your
situation.
If
you,
you
want
to
implement
this
kind
of
things,
so
basically
right
now,
I
think
need
to
revisit
this
from
the
from
the
bottom
and
how
to
increment
as
well
ready
for
your
particularly
you
just
mentioned
sorry
I.
Take
it
back
because
you
miss
and
I
to
your
use
cases.
You
don't
need
a
dependable
and
you
basically
just
run
the
pass
off
the
component
answer.
H
Okay,
I
would
just
want
to
suggest
you
for
short
period
harm.
If
you
have
use
cases
depend
on
as
well.
I
do
suggest
you
to
do
some
other
way
and
make
that
visible.
Even
you
want
implement
of
the
time
container,
which
you
have
those
so
cut,
all
those
can
offer
to
us
invaded
and
don't
just
put
those
implementation
in
c
RI.
You
could
do
those
cannot
think
through
some
other
way.
Naca,
for
example,
I
denise
in
control,
and
you
could
have
said.
H
Oh,
I
needed
this
kind
of
particular
feature
format,
giving
production
support
and
for
the
cutter-
and
you
could
add
objectives-
are
certain
things
to
handle
this
or
so
you
do
have
the
leads
opinion
here.
This
isn't
just
my
suggestion
for
the
production
who
want
to
run
kata
and
also
depend
on
the
port
forwarding.
So
then
you
have
the
recipe
area
for
all
those
kind
of
things,
but
do
not
put
such
of
the
things
to
the
user
user
work
people
so
anyway,
but
you
just
mention
ID
you'll
use
kisses
thank
people
now
israel.
H
G
G
G
My
certain
support,
such
as
the
perform
I
just
I,
just
try
to
try
to
his
tower
on
discover
the
potential
potential
cavity
of
the
CI
in
a
very
brief
class
biggest
part
of
our
wait,
and
oh
maybe
I
I
can
say
because
we
have
the
yes,
you
re
CRE
in
their
face
by
now.
So
if
we
can,
we
have
a
brand
Hemi,
that's
a
promise
yeah,
so
we
can
because
a
part
of
it
the
things
like
to
do
a
do
a
power
forward
without
another
another.
J
H
Seminar
to
the
debacle
so
which
it
is
thus
the
twist
you
want
to
have
the
we
subpoenaed
here
energy,
we
don't
want
like
okay,
there's
the
every
single
and
like
the
integration,
extensive
height
in
the
Kuban
areas,
the
people
people
I
deserve-
make
a
different
container
there.
So
then,
there's
no
way
subpoena
to
customer
there's
no
way
for
for
the
for
the
item
into
control
to
manage
me
those
kind
of
things
so
so,
basically
you're
right
the
reason
we
design
content
a
random
interface.
That
many
reason
one
of
the
reason
it
is.
H
We
want
to
have
that
extensive
in
it.
We
want
to
have
that
flexibility
in
the
future.
We
can
I,
but
I
still
need
the
extra
api
cue
to
do.
That
is
like
okay,
we
had
in
here
and
then
we
act
like
tempered
container
short
time.
We
need
a
single
power
to
those
those
snaps
cycle
management
of
those
kind
of
things.
If
there's
the
problem
that
comes
in
a
guy
who
is
going
to
manage
those
kind
of
things,
you
did
propose
a
lot
of
things
way,
but
we
want
up
through
the
90s
way,
yep.
G
I
I
To
do
this,
yes,
and
also
the
fact
that
this
works
for
traditional
Linux
containers,
and
now
we
have
to
change
our
implementation
for
the
VM
I
can
recipes
pause,
so
I
think
it's
a
bigger
change
that
way,
and
it
also
pushes
some
concept
it
back
into
cubed
again,
as
in
like
what
kind
of
container
to
launch
how
we
are
going
to
manage
this,
we
have
to
figure
out
which
image
to
use.
So
it's
not
in
a
way
it's
better
than
the
original.
We
propose
because
there's
more
possibility
for
more
visibility,
but
it's
also
more
change.
G
Sorry,
I
I,
don't
see
what
what
you
what
you
mentioned,
the
bigger
change
in
the
in
the
coop
side
and
I'd
like
to
summary
the
comments
from
from
the
based
medium,
the
first
one
is
by
by
the
method
of
three
they.
It
will
introduce
a
container,
that's
invisible
to
the
users,
but
it
may
be
found
by
the
public
places,
but
the
life
cycle
is
not
my
recommended
by
Kubrick.
This
is
the
biggest
question
here
right.
H
So
all
the
issue
you
are
going
to
have
it
is
actually
we
already
been
discussed
is
in
signal
when
we
talk
about
the
bathtub,
sir
yeah,
so
so
so,
basically
a
lot
of
things
we
discussed
in
a
different
part.
We
also
have
a
lot
of
this
throughout,
not
just
not
not
team
and
also
sick
architecture.
Discussing
about
the
back
part,
so
we
basically
made
decision
on
the
back
part
and
I
think
we
could
make
that
do.
H
If
that
is,
can
go
smoothly,
we
could
make
that
increase,
revisited,
well,
increase
data
scope
and
info
all
to
those
kind
of
the
things
like
like,
though,
what
we
are
discussed
here,
what
so,
but
that
will
actually
like
the
invisible
of
those
can
become
cleaner
to
Cuba
night.
It
has
been
discussed
before
and
also
cumin.
It
is
self
to
create
of
the
tongue,
for
the
back
always
been
discussed
before,
and
and
also
how
to
change.
J
We
I
think
the
conclusion
is
on
the
container
D
side
and
the
cryo
side
we're
going
to
investigate
you.
If
you
want,
you
can
open
up
issues
that
we'd
already
have
them
for
the
elite
to
to
do
this
port
forwarding
without
launching
spoke
at
by
BSC
and
I,
both
both
cryo
and
Kennedy
are
using
seen
on
you
know,
can
we
can
configure
those
correctly?
We
shouldn't
have
to
do
this.
J
Well,
today,
in
container
D
we're
handling
the
port
forward
via
the
so
cat
executable
and
that's
what
closed
closure
problem
and
we
needed
to
do
an
investigation
to
you
know
both
in
cryo
and
in
container
D,
to
see
if
that
there's
a
better
way.
We
can
do
that
port
forwarding
or
you
so
that
it's
just
it's
just
config
via
sinon,
if
that's
possible,.
J
We
so
we
just
need
to
verify
it,
that's
still
the
case,
and
then
we
can,
just
you
know,
initiate
port
forwarding
with
with
CI
networking
today,
and
if
that
is
still
the
case,
then
what
is
the
best
way
to
do
this?
For
you
either
in
the
shim
layer
were
in
a
cryo,
you
know
API
layer
or
or
to
do
it
this
this
other
way,
which
is
running
a
temporary
container.
That's
if
that's
what
needs
to
be
done
possible
solution.
Number
three,
we're
just
saying:
we
need
to
do
more
investigation,
I
do.
D
I
do
think
if
the
issue
around
image,
garbage
collection
and
the
killed'
not
being
aware
I
do
think
it'd
be
generally
useful
if
I
could
provide
a
whitelist
or
some
type
of
hint
to
the
cubelet.
That
says,
ignore
these
images
from
cleanup,
but
I
agree
with.
Basically
the
sentiment
here
that
let's
try
to
explore
all
the
things
that
can
be
done
behind
the
container
engine
before
trying.
J
G
I
I
didn't
think
they
personally
I,
don't
think
the
networking
so
the
bathroom
in
four
days.
Actually,
the
GFCI
that's
person
itself
is
trying
to
define
how
to
park
the
interface
to
the
to
the
to
the
pot
and
and
not
specifying
who,
since
if
you
leave
the
space
to
the
to
the
funnies,
and
if
you
try
to
move
this
tree
to
see
a
CI
base
that
that
means
you
have
to
make
this
determination
for
for
each
different,
CI
parties
best
best
would
be
will
be
terrible.
H
She
sorry
I
have
to
cut
you
here
so
I
think
we
have
focus
a
bro
of
any
way
to
look
in
here
and
let's
just
take
action
from
there.
So
then
we
can.
We
can
continue
discuss
next
week
because
we
don't
have
the
time
here
and
if
you
want,
you
can
send
the
email
and
the
to
the
kubernetes
is
signaled
and
we
can
carry
discuss
in
there.
Also
there's
also
okay
I'll
come
back
next
week,
thanks
thanks
everyone
and
talk
to
you
guys
next
week.