►
From YouTube: Kubernetes SIG Node 20200609
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
A
Yeah,
okay,
let's
do
and
I
think
the
nits
first
to
start
last
time,
because
two
weeks
ago
we
can
solve
the
the
last
the
previous
ones.
So
the
last
time
we
have
met
the
victor
and
talk
about
signal,
the
testing
and
because
we
ran
half
time,
and
so
we
didn't,
we
didn't
give
you
time.
So
we
thought.
Maybe
you
want
to
start
today.
Yeah.
B
Sure
thank
you
dawn,
so
I
put
some
stuff
on
here
signal
test
enhancement,
there's
a
link
for
the
dot
I
just
want
to
start
out
and
say:
there's
been
some
great
efforts
from
the
team
to
sort
of
jump
in
look
at
these
tests
to
understand.
What's
going
on
and
fix
some
issues,
one
of
the
comments
this
is
directly
out
of
our
meeting,
says:
hey.
B
We
had
an
we
made
an
entire
test
grid,
tab
blue
with
one
PR
and
that's
when
you
look
at
the
test
grid
tabs-
and
you
know
there
was
a
lot
of
red,
so
blue
is
definitely
a
step
in
the
right
direction.
So
that's
good
and
that's
what
we
want
to
give
a
shout
out
to
everybody
say:
hey
thanks
for
all
the
hard
work,
all
the
PRS.
It's
really
good
know
at
the
same
time,
what
at
least
for
me
and
what
I'm
seeing
is
a
lot
of
activity
and
a
lot
of
review
requests.
B
I
was
like
how
can
I
keep
up
with
all
this,
so
I
did
a
little
project
board
I'm
experimenting
with
that
to
sort
of
keep
track
of
these
PRS
and
just
to
give
everybody
sort
of
an
idea
of
how
much
work
and
effort
is
going
on.
There's
10
PR
has
already
completed
emerged
on
just
the
signal
testing
piece,
and
these
are
include
things
like
some
dark
updates.
How
do
you
run
these
tests
if
you
just
want
to
run
a
particular
test?
B
Some
cost
image
fixes
a
little
test
config
and
what
are
the
big
things
we
had
before?
That
was
causing
us
some
issues
as
if
there
was
a
test
that
said,
hey
use
this
image
and
the
test
wasn't
there.
It
was
a
solid
failure.
So
now,
if
the
image
is
missing,
there's
an
explicit
failure,
so
that's
good
and
there's
13
PRS
still
in
progress,
so
we've
got
10,
merge,
there's
30
more
that
are
currently
in
progress
being
reviewed.
There's
some
dock
updates,
more
image
cleanup
and
then
there's
some
things
with
the
benchmark
test.
B
So
I
know
probably
ed
Morgan
on
I
want
to
talk
just
briefly
about
this
benchmark
test.
What
what
happens
is
there's
some
benchmark
tests
that
run
the
benchmark.
Tests
were
failing
without
a
memory,
and
so
there
were
a
couple
of
approaches
to
fix
it.
The
one
we
sort
of
I
think
agreed
on
and
that
we're
pushing
forward
is,
let's
lower
the
number
of
pods
from
105
to
90.
Now
this
is
just
really
a
temporary
workaround
to
get
the
test
of
running
again.
B
The
analysis
so
far
looks
like
container
D
was
split
up
to
a
separate
process,
I
guess,
but
that
one
seems
to
be
what's
using
more
than
memory
and
so
just
to
make
sure
that
we
didn't
lose
that
I
open
the
issue
to
track
that
and
say
hey.
You
know
we're
gonna,
probably
lower.
This
number
pause
from
105
to
90
get
the
test
going,
but
we've
got
this
underlying
issue
where
at
some
point
you
know
it
seemed
like
the
history
was
way
back
few
years
ago.
This
was
a
test
says:
hey.
B
How
many
pots
can
we
safely
spin
up
on
a
note
and
Don
I?
Think
there
was
a
big
issue.
I
think
you
had
put
back
in
their
comment
way
before
so
that
was
sort
of
what
this
test
was
about
and
that's
sort
of.
What's
going
on
so
I
think
there's
if
anybody's
interested
in
issue
interest
in
container
D,
helping
debug
I
know,
there's
been
some
great
help
from
Roy
and
kneeing
and
and
others
to
help
figure
this
out.
So
that's
all
appreciated.
B
C
B
A
Huge
progress,
actually
we
used
to
have
the
family
media
and
for
the
eat
not
eat,
eat
has
and
but
because
the
not
all
the
imagery
have
not
the
image
owner,
so
they
either
ones
and
enlightenment.
And
so
we,
then
we
have
the
policy
if
you
are
making
a
snack.
The
know
that
et
ties
to
particularly
fair
for
one
at
the
one
month
that
time
and
we
have
to
retire
that
image.
It's
giving
kinda
speakers,
the
other
image
issue
details.
B
A
So
then
we
also
have
the
policy.
Next
we
want
to
have
like
the
get
rid
of
the
image
like
like
we
used
to
have
the
Debian
and
Ubuntu
you,
and
also
we
asked
her
to
have
like
that.
We
also
have
the
chorus
and
we
also
asked
a
dryer
but
I
believe
the
retina
actually
meant
a
different
side
is
the
separate
one
is
not
really
here,
because
a
lot
of
work
to
end
animal
not
know
the
et
tags,
so
they'll
ever
get
around
to
end
up
with
that
as
the
package
it
together.
A
B
A
D
I
also
wanted
also
wanted
to
draw
something
else
out
there
for
all
the
all
the
testing
related
work.
There
is
a
as
it
was
already
mentioned.
There
is
a
project
port
in
the
kubernetes
all
and
a
lot
of
working
there.
The
only
the
only
people
who
are
a
perverse
are
the
are
the
chairs
for
signal.
I
was
also
if
we
could
get
like
some
a
periodic
time
slot,
where
people
could
walk
through
those
issues,
PRS
and
prioritize
them.
That
will
be.
That
will
be
also
really
helpful.
A
We
yeah,
we
should
fix
this
yeah
thanks,
yeah
and
I.
Think
there's
a
couple.
Pr
already
start
to
add
more
people
to
maintain
those
kind
of
things,
and
we
also
need
to
define
more
clearly
define
those
policies
and
yeah
another
things.
I
think
the
Victor
already
mentioned
that
benchmark
tests
yeah
we
have
the
benchmark
tests
and
we
even
have
like
the
performance
I
think
that
people
propose
a
couple
weeks
ago
to
those
so
about
the
know:
the
performance
things
like
the
skin
being
attacked
using
our
mechanism.
A
So
we
build
those
the
performance
tests
for
the
note,
the
powder
density
or
you
could
say
the
know,
the
skill
behind
his
skill
and
then
the
mechanics
actually
that
the
people
build.
That
whole
thing
is
my
intent.
That's
the
three
or
four
years
ago
and
doing
the
stair
I
development
during
the
continuity
development
and
the
crowd
development.
Actually,
using
that
a
lot
to
measure
the
continuity
and
accumulate
the
resource
usage.
A
So
then
each
release
we
basically
based
on
those
data
so
and
to
make
sure
I
understand
how
much
do
we
want
a
set
for
our
benchmark
tests.
So
we
maybe
want
to
be
someone
want
to
pick
up
that
work,
also
maintain
that
one
and
and
also
for
that
work,
they're
more
work,
I,
happy
new.
My
intern
also
put
that
like
that
to
do
list
mission
is
there
and
they
want
to
evolve
because
we
talked
about
initially
when
I
hire
him
as
the
intern.
A
D
D
A
Used
to
have
the
issues
there,
but
I
think
that,
because
nobody
unity
kappa
so
movie,
maybe
those
issues
are
in
route
him
and
being
odd
being
the
auto
closed
by
the
bob.
But
I
can't
pick
up
those
things
because
I
remember
the
the
dog
I
revealed
he
wrote
about
know
the
performance
perform
static,
understa
to
do
things.
We
can
pick
it
up.
We
can
start
from
there
to
be
fine
issues
and
we
used
to
have
the
leader
help
he'll
pleaded
tiger
label
in
the
Cuban
ideas
and
a
signal
order.
A
A
Thanks
for
everyone
warned
here
and
a
half
because
you
used
to
be
knackered,
not
many
people
look
at
the
note
the
III
test.
We
are
the
only
one,
and
so
the
issue
is
not
bob
up
too
much
to
the
community.
So
much
because
internal
the
team
is
already
tightest
and
already
stopped.
So
now
we
connect
the
reefs
attention
and
there
is
the
awareness
to
the
community
so
I'm
glad
that
more
people
want
to
help
you.
B
Yes,
thank
you
don,
so
you
know
we're
still
meeting
once
or
twice
a
week,
and
this
is
just
to
go
over
some
really
kept
enhancements
and
stuff
that
sort
of
allow
dedicated
focused
on
outside
of
signal
just
because
there's
not
enough
time
to
do
everything
in
an
hour.
So
we
met
today
and
you
know,
there's
one
kept
that
was
approved,
and
this
was
really
just
adding
a
slight
improvement
for
sort
of
topology
manager
stuff
for
really
getting
better
alignment
between
GPU
devices
and
improving
performance,
which
is
an
optional
extension,
but
there's
the
link.
B
It
was
approved
and
they'll.
Just
one
quick
update,
you
know,
Kevin
I,
don't
know
if
he's
on
the
call
clues
from
Nvidia
they've
done
a
lot
of
work
there
and
you
know
it's
sort
of
late
in
the
enhancer
process
for
19,
so
it
was
approved.
It's
not
a
huge
change
thing
to
I
said
a
cig
note,
but
it's
not
likely
to
be
implemented
until
you
know,
1.20.
C
B
E
B
B
B
E
Common
I
would
add,
was
BG
had
presented
a
proposal
to
City
odd
level
alignment
as
a
technology
manager
policy,
where
we
today
only
have
container
level
alignment
I
know
both
Kevin
and
myself,
four
plus
one
on
that
proposal.
Right
now
and
I
think
we
were
waiting
to
see
if
Alex
from
Intel
had
time
to
offer
additional
context.
But
given
that
our
current
topology
alignment
policy
as
a
per
node
policy
and
that
it
seems
like
any
future,
evolution
might
become
a
per
pod
policy.
E
But
we
lacked
a
clear
action
next
step
to
do
that,
we
wanted
to
bias
to
enabling
more
flexibility
on
the
per
node
version
of
that
policy
and
support
Pablo
alignments.
So
I
wanted
to
thank
VG
and
the
team
from
Samsung
I'm
playing
together
like
a
very
clear
description
on
where
and
how
they
would
use
this
policy
in
a
classic
5g
deployment,
where
you
might
have
a
set
of
notes
that
are
only
running
UPF
functions
and
other
nodes
that
are
doing
different
work,
which
might
not
run
with
this
policy.
E
So
one
thing
I
just
want
to
make
note
of
that
is
like.
It
was
really
really
helpful
when
evaluating
that
kept
to
understand
like
the
overall
topology.
The
cluster
would
be
deployed
in
to
to
understand
the
system
and
if
it
would
actually
be
used,
but
it
may
be
abstract,
very
concrete
and
so
maybe
going
forward.
One
of
the
lessons
I
would
say
it's
come
out
of
the
the
discussions
we've
had
and
Victor
up.
E
Maybe
you
could
second,
this
was
it's
really
helpful
when
people
or
more
concrete
on
their
use
case,
to
drive
understanding
so
like
decisions
on
things
being
per
pod
versus
per
node
are
not
as
easy
like
and
just
being
clear,
I'm
saying
well
up
heard.
No,
the
thing
makes
sense
for
me
because
my
use
case
is
around
dedicated
GPU
nodes
right.
E
That's
really
helpful
for
us
as
a
community
understand
and
and
biases
towards
doing
something,
whereas
absent
like
a
concrete
deployment,
topology,
we're
kind
of
biased
to
just
do
nothing,
and
that's
not
good
for
anybody
so
anyway,
if
folks
want
to
review
BG's
proposal
on
pod
level.
Topology,
please
make
those
comments
today,
because
I
wanted
to
unblock
that
this
week.
It's.
E
B
E
Recording
is
I
need
to
check
if
I
had
uploaded
that
one
I'm
like
a
week
behind
on
my
uploads
but
there's
also
a
a
blink.
Maybe
we
can
put
adhere
to
the
slide
deck
on.
Why
and
where
it
would
be
used,
so
it
had
a
5
node
deployment
for
we're
user
plain
functions
would
go
and
stuff
and
it
was
the
most
insightful
useful
piece
of
slides,
I've,
seen
related
to
5g
in
a
while.
So
I
just.
A
You
know:
can
you
upload
after
review,
if
possible
and
at
least
think
of
the
slack
back
here,
the
back
of
a
lot
of
our
us
and
I
heard
a
lot
of
us
for
the
west
coast
cannot
attend
it's
the
six
o'clock
it's
to
earlier
for
many
of
us
here
and
but
I.
Think
many
awful
I
want
to
know
more
about
those
things
and
expansion.
This
is
well
be
change
after
our
call
our
neck,
the
resource
management.
So
so
one
things
what
I
my
comment!
A
It
is
I
wanted
to
say
that
it's
harder
for
us
to
find
a
solution.
One
size
fit
all
and,
and
also
so
I
mean
I,
didn't
see
the
teenagers
of
a
proposal
from
his
problem,
so
I
can
see
that
clearly
wired
they
came
from
and
to
support
the
mobile
application
to
support
the
Falchi
did
I
also
have
other
people
and
how
other
used
pieces,
not
all
it
is
the
tiny
cutting
kisses.
So
so
so
when
things
I
really
like
what
Alex
proposed
last
time,
I
think
it's
more
than
a
month
ago.
A
He
gives
the
proposal
what
in
doubt
doing
that's
really
aligned
with
the
three
years
ago
the
decision
we
made
it
so
next,
the
more.
This
is
also
why
we
have
that
continental
interface
and
all
those
plaque
or
extensible
API.
We
defined
we
try
to
say
oh
here
it
is
the
identified
common
core
for
all
the
general
graph
that
Walker
node
and
then
we
have
like
the
the
runtime
cars
and
another
label
or
those
kind
of
things
and
Adam
II
seen
control
all
those
kind
of
things
concept.
So
you
can't
customize
D
for
the
know.
A
Customize
the
different
of
the
they'll
also
have
the
sky
donut
extensible
framework
and
you
can
plug
in
your
scheduler
and
you
can
plug
in
your
content
of
runtime
and
for
different.
So
that's
not
last
many
years
we
try
to
push
forward
and
the
annex
the
demo,
actually
I'm
meaningly
next.
He
give
all
this
kind
of
the
scenario
different
the
hardware
and
how
we
are
going
to
do,
and
so
so
that's
that's.
The
I
just
wanted.
A
E
I
think
I
think
that
exact
sentiment
will
be
probably
captured
in
the
recording.
So
there's
no
disagreement,
they're
doing
and
I
think
I
think
the
issue
is
a
where
there's
no
clear
consensus
on
how
to
move
forward
even
an
analysis
POC
that
he
presented.
E
E
G
Actually
I
fought
about
with
us
forward
and
we've
had
some
internal
discussions
with
an
Intel
and
I
put
already
in
a
web
of
what
future
topics
for
this
topology
discussion
making.
So
we
have
some
ideas:
how
to
implement
the
things
we
can
share
all
the
items
we
have
it's
a
question
of
like
where
we
want
to
go
and
I
will
try
to
summarize
all
the
current
painting
proposals
and
outline
like
few
options
like
where
we
can
put
with
things
and
when,
based
on
on
words,
we
can
discuss
like
what
is
for
future
steps.
Yes,.
E
G
It's
it's
a
good
incremental
step.
I
understand
for
reason.
I
understand
like
what
asks
what
we
tries
to
solve.
I
just
want
to
verify
like
few
minor
things,
to
make
sure
what
like
what
future
things
will
not
be
prevented
like
we
will
not
end
up
this
something
what
we
need
to
support
for
multiple
years,
what
we
cannot
change.
A
Successful
for
tracking
of
the
time
so
thanks
for
David
to
remind
us,
because
we
just
want
the
hell
out
after
summary,
about
resource
management,
work
group
and
now,
let's
move
back
to
our
target
our
agenda.
So
so
second,
a
container
set
the
car
water.
So
it's
back
to
the
first
away.
So
second
container.
H
So
far,
so
good
also
I
wanted
to
mention
that
we
found
some
edge
cases
on
termination.
Like
termination,
was
the
biggest
concern
it
was
raised
here
about
this
about
this
cycle
cap,
and
we
found
some
some
corner
cases
when
like
if
the
some
current
cases
does
murder
the
details,
and
we
have
some
ideas
on
how
to
improve
it.
Regarding
the
the.
H
I
wasn't
sure
if
that
like,
if
we
cannot
rely
on
that
feel
being
said
of
if
that
was
added
just
because
when
they
field
was
added,
it
needed
the
support
when
it
was
not
site
for
to
interoperate
with
all
the
releases,
and
it
was
just
not
removed,
or
something
like
that
does
anyone
here
happen
to
know,
because
that
will
be
useful
for
the
Hewlett
graceful
shutdown
or
maybe,
if
it's
useful
they
are,
we
can
use
it
for
that.
I've
get.
E
F
I,
looked
at
the
gate,
blame
super
quick.
It
looks
like
it
may
be
added
I
to
handle
pods
that
are
missing
the
spec
entirely,
and
that
may
happen
when
we
load
pods
from
checkpoint
or
something.
But
it
looks
like
we've
written
the
code
or
it
was
added
with
the
intention
that
the
cubelet
could
continue
running
even
when
certain,
even
when
we
didn't
have
a
pod
spec
at
all,.
H
H
Because,
like
the
bodies
being
restored
from
from
continual
labels,
so
we
like
docker
container
or
the
container
runtime
label,
so
at
that
point
the
pods
should
be
but
yeah.
Maybe
when
you
restore
it
from
level,
it's
not
said
yeah,
maybe
it's
something
like
that.
Make
sense.
Sorry,
but
yeah.
That
was
the
quick
update
regarding
inside
cars.
H
J
A
E
So
this
effort
was
initiated
originally
by
myself
and
because
Chaudhary,
who
is
no
longer
with
Red
Hat
I.
So
he
did
a
lot
of
great
work
to
help
push
it
forward,
but
he's
not
presently
working
in
this
place
anymore.
So
it's
it's
great
to
hear.
Numeron
all
had
been
chatting
with
folks
at
Netflix
that
had
an
interest
and
it's
good
to
see
the
renewed
interest.
E
Curious
if
the
unicorn
reappears
for
that,
but
the
challenge
we
had
there
was
trying
to
figure
out,
probably
how
to
handle
upgrade
safely
and
I
would
I
would
bring
in
Jordan
Liggett,
who
I
remember
being
a
very
prominent
reviewer
on
this
particular
PR
as
well,
since
I
think
at
the
time
it
was
myself
Vegas
and
Jordan
that
were
trying
to
iterate
on
this.
There
were
also
some
areas
that
were
maybe
under
defined
and
so
it'd
be
good.
Ranade
know
we
haven't
chatted
like
if
the
Netflix
use
case
was
was
a
node
wide
remapping
yeah.
J
It
doesn't
know
like
remapping,
yeah
and
I
think
yeah.
They
are
they're
interested
in
a
node
wide
remapping,
the
phase,
one
that
you
are
trying
to
tackle.
You
and
I
think.
One
of
the
issues
we
had
here
was
with
docker,
where,
if
you
have
a
privileged
part,
I
think
with
cryo
and
continuity,
you
can
switch
off
the
user
name
space,
but
docker
did
not
have
support
for
that.
I.
Remember
that
as
one
of
the
excuse
but
I'm
not
allowed
to
check
well
what
we
agreed
to
do.
A
We
talked
about
this
plant
at
a
signal
and
I.
Remember
that
I'm
nantao
and
my
Cobra
is
here
today
we
talked
about
this
one.
Then
they
they
try
to
take
back
those
issue
to
the
continuity
community
and
they
continue
discussing
I.
Don't
remember,
are
we
put
back,
but
we
they
actually
take
that
task.
I,
hope,
I'm,
not
sure
micro
is
here.
Yeah.
K
K
A
A
My
purpose
or
something
Eric
can
you,
you
have
invited
the
Jordan
and
give
make
attend
today's
meeting,
maybe
sometimes
next
week
or
maybe
even
fall.
So
we
can't
have
like
the
recap
on
the
status
and
so
far
I
heard
a
couple
issue
when
it
is
figure
out
the
upgrade
and
we
can
focus
on
those
things
another
one.
It
is
the
continuity
and,
if
I
remember
correctly,
they
take
that
job
back
to
the
continuity
we
upgrade
and
also
Sarah
I
change
you
and
we
agree.
A
A
So
we
can
lay
out
after
one
by
one
obviously
user
name
in
space
actually
is
pretty
common
popular
topic,
but
all
many
people
with
the
hands
they
want
that
one
to
enhance
the
there's
that
creative
management
and
policy
and
but
is
actually
is
more
from
the
Cooper
90s
community,
leading
that
effort,
because
you
knew
so
maybe
we
should
have
focus
on
those
things.
Yeah
yeah.
E
No
and
I
don't
know
if
he
ever
discuss
this
in
the
past,
but
like
I'll,
just
caution
that
this
is
a
very
difficult
to
reason
through
when
thinking
about
all
the
permutations
and
so
for
sure,
the
the
I
would
hope
that
we
could
continue
working
from
we
kisses
PR
and
that's
now
closed
63,
sick,
this
five
horn,
but
even
just
point
out
myself
I'm
trying
to
remember.
Oh
all,
the
challenges
that
we
ran
into.
But
what
I'm
curious
about
is
is
the
use
case
that
we
were
driving
for
universally
shared
among
folks.
E
Or
is
anyone
exploring
this
as
an
attempt
to
try
to
get
a
rootless
cubelet
like
I,
want
to
I'm
curious
if
people
are
exploring
this
as
a
means
to
support
better
security,
isolation
of
end-user
pods
or
as
an
indirect
mechanism
to
try
to
support
a
rootless
qubit
and
if
we
can
maybe
align
on
like
what
success
criteria
is?
That
would
be
helpful
because
I
know
this
has
also
come
up
in
the
past.
E
So,
like
heroes,
suta
in
the
past
had
raised
this
as
a
issue
tied
to
rootless
qubit
and
so,
but
definitely
for
the
original
cap.
I
think
it's
still
solid
and
I
think
the
PR
that
PICUs
had
was
really
close,
but
still
had
some
issues
that
we
could
talk
through
and
they're.
Just
all
of
my
memory
because
they're
two
years
old,
I.
A
Is
the
first
attack
nothing
for
the
lutenist
cubed
and
so
routine
X
cubed
and
the
Melissa
for
the
former?
What
you
you
list
them
and
we
the
problem
it
is
we
never.
This
is
being
going
on
the
user's
name
space.
It's
been
going
around
for
a
couple
years,
and
now
we
I
just
naturally
say
that
week
has
already
made
the
majorly
things
down,
but
it
are
several
open
question.
We
need
closer
on
those
open
question
and
move
forward.
E
E
E
Some
of
the
CRR
pieces
I
think
well
I.
Don't
even
think
that
as
much
as
like
other
like
it
definitely
predates
the
ephemeral
containers
and
yes,
and
so
we
need
to
map
these
features
to
each
other
and
I.
Think
it
also
predates
pit
namespace
sharing
and
so
yeah
probably
deserves
a
an
updated
cap.
And
so
if
there
is
a
group
within
the
community,
whether
that's
a
set
of
folks,
ideally
from
multiple
interests,
whether
that's
Netflix
or
vincit,
and
your
team
to
just
kind
of
to
pick
up
a
new
cap.
That's
probably
what
I
would
recommend.
E
E
K
M
E
E
A
F
Actually
talk
about
that
a
little
bit
so
we
had
to
revert
his
PR
because
it
caused
some
performance
regressions.
It
doesn't
so
normally
when
we
delete
containers.
All
of
that
is
sort
of
synchronous
through
the
life
cycle
event
generator,
it
doesn't
seem
like
sandbox
deletion
is
done
the
same
way,
and
so
it
added
like
10
ish
seconds
to
pod
deletion,
which
was
unacceptable,
so
he's
investigating
how
to
do
that.
He
may
have
had
questions,
but
I,
don't
know
what
those
are,
but
that's.
B
F
B
F
M
Sure
this
is
a
quick
update,
so
the
core
implementation
I
am
pretty
much
close
to
completion
on
that
I
got
some
tired
reviews
from
David
looked
into
it
and
left
some
comments,
I
addressed
them.
Thank
you,
David
and
I
believe
the
API
has
been
partially
reviewed,
at
least
by
DeMorgan
and
I've,
addressed
the
concerns
and
change
the
code,
but
I
don't
know
if
there's
a
second
pass.
M
That's
happened
yet
I'm
planning
to
there
are
a
couple
of
small
items
left,
which
is
the
resource
quota
limit
range
I,
don't
think,
there's
much
work
to
do
in
resource
quota.
We
agreed
to
use
just
limits
for
now.
I
do
want
to
go
in
and
take
a
doubt
that
we
have
an
implementation.
I
wanna
take
a
look
at
if
there
is
any
real
problems
with
corner
cases
there
shouldn't
be,
and
then
there
is
a
limit
range
which
is
which
should
be
fairly
simple.
M
To
just
ensure
that
resource
code,
update
is
not
blocking
resizing
and
limit
range
arrow
keeps
the
limits
within
bounds.
Then
the
next
big
task
is
to
do
the
e.
Do
we
test
so
the
goals
for
alpha
release
409?
He
needs
to
have
the
basic
e
to
e,
and
then
he
do
it
for
multiple
containers.
So
far,
I've
been
testing
them
manually
and
David.
I
did
check
that
concurrency.
To
can
two
concurrent
updates
the
mutex
mechanism
that
I
have
there.
M
F
M
F
F
We
decided
we
needed
to
only
ensure
that
we
don't
exceed
that
request.
Don't
exceed
limits
internal
to
the
note
or
if
we
need
to
make
sure
that
that
can't
ever
happen
external
to
the
node,
meaning
that
I'm
not
sure
if
we
decided
it's
fine
for
pod
assigned
resources
to
exceed
the
node
capacity
or
if
we
just
need
to
make
sure
that
those
aren't
ever
applied
so
I,
don't
remember
what
we
decided,
but
I
think
there
may
be
races
there
still
but
I'll
take
another
look:
is
this
for
teachers
quota?
F
M
M
M
A
N
A
very
quick
description
so
in
the
past
qubit
was
the
advisor
would
report
GPO
metrics.
That
was
a
change
that
was
made
two
or
three
years
ago,
I.
Think
since
then,
we've
decided
to
move
towards
a
different
model
for
reporting.
Metrics
we've
noticed
more
recently
is
because
qubit
actually
reports
GPU
metrics,
specifically
NBA
GPU
metrics.
N
It
now
has
an
open
handle
on
the
Nvidia
driver,
and
this
is
a
problem,
because
what
that
means
is
that,
if
from
a
cluster
and
main
perspective,
I
want
to
update
the
Nvidia
driver
or
remove
the
Nvidia
driver,
I
need
to
kill
qubit,
so
I
have
a
pure
that
is
in
flight
right
now,
where
it's
still
working
progress,
but
I'm,
adding
a
a
flag
to
Cuba
to
disable.
This
reporting
of
GPO
metrics,
which
is
done
by
the
way
to
Missy,
advisor
and
I,
think
the
discussion
I
want
to
have
is
well.
N
Is
it
possible
in
the
future
to
enable
this
by
default?
Right
now,
the
PR
is
just
if
you
are
close
to
Edmund,
you
can
just
label
this
flag
to
disabled
do
but
to
collect
these
metrics.
There's
some
dependency
on
the
advisor
to
get
real,
ender
and
I
think
these
are
your
generally.
Your
question
here
is:
is
this
something
that
we
think
is
desirable
and
if
so,
what
would
be
the
process?
Do
we
need
to
just
open
it?
You
okay,
is
that
something
that
we
want
to
start
advertising.
N
E
N
N
My
experience,
people
don't
really
use
these
metrics,
because
the
metrics
are
really
collected
are
a
bit
very
reliable
or
were
very
reliable,
especially
GPS
ization.
So
it
will
mean
I,
think
people
there's
probably
people
out
there.
There
use
it,
but
I
definitely
agree
with
the
fact
that
at
least
we
should
be
trying
to
remove
this.
You.
A
Know
so
we've
learned
of
the
path,
but
I
want
to
say
that
we
definitely
crowd
this
one,
because
we
used
to
have
like
a
lot
of
debate
in
the
signal.
We
don't
want
this
plan,
but
do
we
accept
it?
It's
a
to
move
forward,
move
the
ball
forward
right
that
time.
So
we
definitely
support
this
one
and
the
better.
Still
we
don't
know
we
we're
not.
We
cannot
cover
all
the
use
cases
here
sown
it's
just
a
move
on
with
your
PR
and
I.
A
Do
this
disable
this
one,
the
flag
and
the
two
people
in
a
cylinder
and
s-trans
possible
and
make
that
announcement
through
the
release,
notes
and
all
those
other
thing
and
all
the
other
cheek
and
nose?
And
then,
if
there's
the
no
concern
no
complaining
from
the
community
and
from
them
all
the
visitors
that
is
winter,
so
we
just
next
ring
is
all
maybe
follow
you
meaning,
so
we
could
just
completely
make
this
default
this
happening.
A
E
N
N
F
F
I
think
GPU
metrics
seem
to
be
different
in
that
a
one
renaud
has
said
that
they
have
problems
with
opening
with
using
the
GPU
driver,
which
is
specific
to
them
as
well
as
we
have
already,
as
the
sake
agreed
on
and
out
of
tree
replacement
for
these
metrics
GPUs
metrics,
specifically
I
do
think
that
we
shouldn't
remove
this
flag
and
remove
GPU
metrics
entirely
without
deprecating.
The
summary
api,
as
we
have
in
the
past,
tried
to
maintain
backwards.
F
Compatibility
so
I
agree
with
your
earlier
point,
Derek
that
we
should
still
tie
this
to
an
overall
effort
to
remove
our
or
to
remove
any
legacy
API,
as
we
have
that
dependency
advisor,
but
I
think
in
this
case
adding
the
flag
so
that
users
that
have
issues
with
cubelet
using
the
driver.
It
is
the
right
solution
here.
N
N
A
You
there's
the
comment
from
the
Dames
and
Alex,
so
I
hope
you
miss
your
Alex
and
yeah
yeah
so
say,
but
those
GPU
metrics
that
have
I
think
they're
also
signal
to
make
the
conscious
decision.
It's
not
like
the
that
hand,
because
at
least
I
strongly
against
to
those
add
into
the
cuban
NSA
and
said
over
there,
and
but
we
have
to
because
we
have
many
customer
to
support
a
new
side
of
the
workload.
That's
that
time.
So
we
connect
the
make
the
neck
the
conscious
decision.
So
so
I
also
agree
with
the
themes.
A
They
say:
okay,
too
many
of
the
flags
and
but
to
take
this
one
with
the
deprecated.
After
summary,
and
all
those
kind
of
things
that's
been
like
the
three
and
a
four
years
ago
that
I
fir
to
ask
that
even
before
CRI
and
we
didn't
make
much
progress,
so
I
agree
need
to
think
about
the
retire
as
much
as
if
we
could
coming
up
a
little
bit
of
the
code
base
and
remove
some
of
the
one
of
you
when
there's
code
and
cleaning
up
the
code
base.
That's
the
good
choice.
A
O
Understand
that
done,
it's
just
that
it's
frustrating
to
see
that
things
are
stuck
and
we
are
adding
flags
just
to
let
somebody
disabled
something
for
their
own
use
case,
rather
than
actually
making
a
complete
effort
to
to
clean
things
up-
and
you
know
it's
since
we
are
doing
this,
we
have
less
incentive
to
go,
do
the
actual
work
that
is
needed.
That's
what
I
see
it.
Yes,.