►
From YouTube: Kubernetes SIG Node 20220216
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
A
Good
morning,
good
day,
good
evening,
whatever
time
zone
you
are
it's
february,
16
2022.,
it's
a
signature,
subgroup
meeting
welcome
everybody.
We
have
one
agenda
item
today
from
imran
and
brandon
talk
about
it.
B
Fine
yeah
cool,
so
the
last
time
we
talked
about
this,
it
was
regarding
the
pr
on
the
kubernetes
repo
regarding
the
e3
test
for
law
contention
flags.
So,
after
providing
so
after
doing
addressing
your
feedback,
I
found
out
that
there
were
the
test
job
for
it
in
the
test.
Infrarepo
was
removed.
B
I
have
mentioned
the
link
to
the
comment
comment
as
well,
which
removed
that
so
I
was
not
sure
what
would
the
next
steps
be
for
this
pr,
and
especially
the
test
job
here,
someone-
I
don't
remember
the
name,
but
that
person,
I
think.
B
B
Okay
and
regarding
the
pr
on
the
kubernetes
repo
for
the
e2e
test,
I
have
addressed
your
feedback.
There
was
a
bit
of
mess-up
where
I
accidentally
pushed
an
old
version
of
the
the
changes
I
mean
so
it
over
there.
It
override
the
changes
which
I
made
addressing
your
feedback.
So
I
also
addressed
that.
A
Great,
can
you
remind
me
what
the
status
of
the
migrating
this
fla
this
arguments
to
the
conflict?
I
remember
this
was
part
of
the
bigger
work
that
you
you've
been
doing.
B
So
the
idea
is
that
we
want
these
flags
and
we
don't
want
it
to
be
deprecated,
so
we
wanted
to
move
to
public
configuration.
This
is
that
it
is
hold
up,
because
we
don't
have
an
end-to-end
test
suite
for
that.
B
So
this
right
now
this
pr
what
it
does
it
adds
the
end-to-end
test
weight
and
afterwards
later
we'll
you
know,
convert
this
add
the
test
for
for
the
same
set
of
flags
in
the
cubelet
as
a
cubelet
configuration,
so
that
there's
another
pr
for
that
which
does
the
actual
move,
the
these
flags
as
a
two
cubic
configuration
which,
which
will
then
later
be
reviewed.
B
Yeah,
so
this
does
not
require
the
serial
test
field,
because
you
know
it
is
disruptive
in
nature.
It
restarts
the
cubelet
and
also
it.
For
now
I
mean
we
add
two
extra
flags
to
the
cubelet
which
is
not
required
by
you
know
the
other
test,
suites
or
other
tests
in
the
serial
test.
Suite.
A
I
just
curious
about
long-term
plan.
Oh
it's
yeah,
it's
very!
Ideally,
we
need
to
run
most
of
the
tests
as
a
single
suit.
It's
more
efficient.
Okay!
Thank
you.
I
will
take
a
look
any
questions
on
that
topic.
A
C
Yes
sounds
great,
and
this
this
week
much
better
and
can
we
scroll
down
a
little
bit?
There
are
highlighted
signal
critical.
They
are
all
green
now
and
which
is
recovered
from
last
week,
and
the
other
parts
are
almost
the
same
from
naturally
and
from
the
signal
to
container
d.
I
think
there
are
nice
this
week
by
comparing
to
another
week
notes
and
for
signal
the
cost
it's
it's
the
same
last
week
and
for
npt.
C
C
A
Yeah,
so
pr
for
npg
from
varsha.
I
think
she
is
working
on
that,
but
there
were
some
comments
that
needs
to
be
addressed
and
great
progress
on
sigru
v2.
If
you
haven't
seen
that
the
problem
with
c
group
2
was
that
ordering
of
continuous
installation
and
configuring
of
test
was
messed
up
so
cut,
the
rd
and
test
installation
will
run
in
parallel
and
there
will
be
a
lot
of
races.
A
What
will
be
installed
first,
so
that
was
fixed
and
it
fixed
many
tests
and
the
reason
we
haven't
seen
it
before,
because
before
we've
been
sometimes
passing
containing
runtime
as
a
remote.
Sometimes
it's
contain
energy
and
in
one
case
it
would
do
some
steps
that
may
raise
with
continuity
stimulation
some
other
cases
it
wouldn't.
So
we
just
never
noticed
these
problems
before
because
of
this
messed
up
of
what
we
passes
continue.
Runtime.
A
Okay,
this
is
it
any
other
questions
or
comments
on
task
grid,
health.
A
D
We
go
and
close
that
one,
because
I
think
that
was
an
underlying
issue
with
a
bunch
of
stuff.
A
Yeah,
thank
you
for
asking.
A
D
So
it's
too
it's
it's
like.
It's
also
got
a
shruggie
on
it.
What
is
going
on
with
this
pr?
The
reason
that
there's
so
many
files
is
because
there's
a
run
c
bump
inside
of
it
like
there's
a
there's,
a
vendor
update.
D
Currently,
I
actually
just
met
with
bartek
about
this
this
morning,
so
cubelet
uses
a
lot
of
cpu
a
lot
and
like
a
huge
amount
of
it,
is
just
like
spent
on
c
advisor
like
recreating
the
labels
for
prometheus
metrics,
like
creating
the
data
structures
over
and
over
again,
even
though
they
don't
actually
change
and
then,
as
a
result
like
spending
a
ton
of
time
in
garbage
collection
for
like
the
old
labels,
which
are
the
same
as
the
new
labels.
D
But
the
cubelet
doesn't
know
that,
and
so
anyways
bartek
is
working
on
like
adding
a
new
prometheus
cached
implementation
in
order
to
collect
this
a
little
bit
more
efficiently,
something
more
like
cube,
state
metrics,
for
example,
and
so
I
this
was
my
attempt
to
like
kind
of
integrate
that,
and
so
this
builds.
But
there
are
like
issues
like
no
metrics
are
coming
out.
So
my
follow
up
on
this
is:
I
need
to
write
an
end
test
that
verifies
that
metrics
are
being
scraped
so
then
the
node
test
will
fail.
D
If,
like
that's,
not
happening,
I
can't
believe
that
we
don't
have
this
end
to
end
test,
but
we
don't.
It
would
have
prevented
at
least
one
major
regression
that
we
had
with
like
c
advisor
metrics
disappearing
in
119.
So
I'm
going
to
write
this
end
to
end
test
and
then
yeah
I've.
I've
just
been
working
with
bartek
and
damien
grusine
on
this,
so
yeah
anyways,
you
like
I
mean
needless
to
say,
you
can
dump
it
into
like
archive
it
or
something.
I
don't
think
it's
relevant
for
the
test
board,
but.
D
D
He
knows,
okay,
the
only
person
that
I
would
maybe
want
to
like
get
you
know
a
take
on
would
be
david
porter,
but
I
don't
think
that
we
need
his
feedback
quite
yet.
D
A
A
Come
with
me,
let
me
check.
A
Okay,
let's
see
here.
A
F
A
I
hope
that
that
are
irrelevant,
otherwise
yeah,
maybe
wait.
Till
tests
will
pass.
G
A
Yeah
this
minor
comment
here,
I'm
not
sure
why
it
will
fix
the
test.
Oh
this
issue,
you
mean
right
in.
A
Here
I'll
check
offline,
but
if
it's
duplicate
it's
a
big
problem
yeah.
I
don't
know
why
this
will
fix
the
test.
But
if
it
will
it's
great.
H
H
A
Okay,
let
me
put
it
out
of
scope
of
this
cersei
but
see
you
on
the
main
board.
A
Yeah
so
we
have
legit
to
we
are
that
need
reviewer,
and
we
still
have
that
misapprover.
Maybe
they
scope
beyond
what
we
own
yeah.
Maybe
we
need
to
take
it
offline
and
see
whether
we
can
make
progress
on
all
them.
I
think
this
one
is
ready
to
merge
by
the
way.
So
if
you
can
take
a
look.
A
Okay,
anything
else
for
test
part
of
the
meeting.
If
you
look
at
everything,
if
not,
we
will
go
to
backtrack
now.
A
Five
days
back,
standalone,
complete.
D
Yeah
sig
storage
might
be
missing.
If
you
look
client
is
nil,
I
mean
it's
still
kind
of
us.
I
think
we
should
triage
accept
this,
but
I
think
someone
in
six.
I
A
A
Okay,
so
it
seems
that
it's
the
same
issue
as
this
one
wonder
what
kind
of
information
we
can
ask
a
full
couplet
may
be
useful.
A
Yeah
quite
fresh
okay.
I
I
agree
it
seems
like
something
that
is
an
unpleasant
but
legit
state
of
things
I
mean
if
the
cubelet
rejects
admission.
The
job
of
the
controller
is
to
recreate
the
pod,
it's
a
very
similar
to
what
we
have
with
the
technology
measure.
It's
a
when
cubelet
rejected
mission.
You
almost
always
end
up
in
states
like
this.
I'm
not
sure
if
this
is
what
danielle
meant,
but
I
I
do
think
that,
like.
E
D
A
B
A
A
D
Oh,
maybe
your
query
just
includes
non-bugs
yeah.
That
one
was
a
feature
yeah.
F
D
Yeah,
do
we
wanna
triage,
except
that
I
mean
I
guess.
I
A
A
D
For
that
one
honestly,
I'm
not
sure
it's
possible.
There
is
yet
another
bug
in
the
cubelet
refactor
with
the
pod
worker
or
it's
possible
that
they
just
did
something
ridiculous
and,
like
the
storage
thing,
can't
retry
it's
hard
to
tell
if
this
is
a
cubelet
thing
or
a
storage
thing,
and
I
was
hoping
that
maybe
sig
storage
could
confirm
that
it's
a
cubelet
thing.
D
A
A
D
A
Oh
honey,
you
wanted
to
comment
on
this
ordering
that
complete
this
end
of
life.
D
No,
I
just
need
to
remember
to
actually
comment
on
this
one.
It's
a
sign
to
me.
I
will
try
to
remember
after
this
meeting.
G
A
I
remember
it
now,
like
we've
been
discussing
with
a
there,
is
some
race
in
taking
these
measurements.
D
G
D
Okay,
so
when
you
have
say
a
very,
I
o
intensive
workload
frequently
what
the
kernel
will
do
is
it
will
take
all
of
the
stuff
that
it
would
otherwise
be
reading
from
disk
and
caches
it
in
memory.
That
memory
is
not
like
associated
with
the
application.
As
far
as
kernel
accounting
goes.
It
goes
into
kernel
buffer,
so
like,
if
you
have
a
very
I
o
intensive
thing
on
a
system
that
doesn't
necessarily
have
a
ton
of
memory
like
or
like
you
have
an
app
that's
doing
a
bunch
of
I
o
intensive
stuff.
A
D
Well,
I
mean,
if
they're,
if
they're
running
other
workloads
on
the
node,
potentially
that
could
also
be
competing
with
kubernetes
there's
like
a
bunch
of
possibilities
here
to
me.
This
is
like
going
pretty
deep
into
like
application
level
support,
I
would
say
like
if
they're
using
oracle,
they
should
maybe
talk
to
oracle
about
what
they're
doing
here.
It
could
also
be
a
bug
in
like
oracle's
platform
right.
We
have
no
way
of
telling
we
don't
have
access
to
those
binaries,
so
those
are
some
of
the
problems
of
like
you
know.
A
D
Maybe
like
it
might
be
helpful
to
explain
to
them
that,
like
available
memory
is
not
the
same
thing
as
free
memory
and
like
the
free
memory
on
the
system
is
accurate.
Something
is
causing
a
bunch
of
memory
to
be
like
used
by
the
kernel
for
buffer
or
cache,
and
that,
like
that,
does
have
an
effect
on
available
memory
on
the
system.
F
There's
something
in
the
advisor
that
I
don't
remember
exactly,
but
but
I
remember
there,
there
is
some
special
logic
paid
to
whether
it's
available
or
whether
it's
buffer
cash.
I
believe
it's
like
subtracted
or
something
like
that.
I
don't
remember
the
exact
logic
off
my
head,
but
it
was
handled
somewhere.
D
A
It'll
be
great
to
understand
how
we
measure
that
and
like
maybe
ask
them
to
run
the
same
command
as
we
do
in
kublet,
so
to
make
sure
that
kubelet
doesn't
do
it's,
not
the
error
in
kubernetes,
it's
actually
what
is
happening
on
the
node
that
is
occupied
by
some
other
processes.