►
From YouTube: Kubernetes SIG Node 20210708
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
A
A
A
So
yeah
it
was
a
short
week
for
us
because
we
had
like
independence
day,
but
something
was
happening
plus
we
have
a
quad
freeze
tomorrow,
so
I
would
assume
that
many
people
were
reviewing
prs
and
trying
to
merge
big
ticket
items
into
122..
A
B
Yeah
the
first
two
items
are
both
me
so
because
it's
code
freeze
today,
I
have
an
item
in
here
for
the
stuff
that
the
release
team
has
been
tracking
for
ci
signal.
We
should
probably
make
sure
that
we
look
at
all
four
of
those
tests
which
they
are
saying,
are
potentially
release
impacting.
B
But
the
first
item
was
a
report
on
the
node
conformance
tests
from
bringing
it
up
at
sig
arch
at
the
last
sig
arch
meetings.
That
was
last
week,
and
so
we
had
an
item
on
our
backlog
to
discuss.
B
Basically
like
should
we
rename
node
conformance
to
cri
conformance?
How
do
we
define
the
scope?
How
can
we
better
document
this
and
previously
from
like
a
slap
conversation,
dims
has
said
well,
this
is
mostly
about
you
know
ensuring
that
the
various
different
cri
implementations
all
work
with
node
things,
and
so
I
basically
went
to
a
sig
arch
meeting
and
because
a
lot
of
the
conformance
folks
were
there,
I
asked
them.
You
know
like.
Does
this
sound
accurate?
Is
this
reasonable?
B
Should
we
go
ahead
and
rename
them
and
a
lot
of
the
sort
of
like
original
people
who
worked
on
the
node
conformist
tests
were
at
this
meeting
and
they
were
like?
No?
No,
it
is
not
merely
a
matter
of
cri
conformance.
B
So
specifically,
everybody
agreed
that
we
should
probably
take
conformance
out
of
the
name
of
these
tests,
because
conformance
has
a
very
specific
meaning
and
the
fact
that
they
have
conformance
in
the
name
is
confusing,
but
it
is
not
merely
limited
to
the
cri,
and
so
I
linked
that
old
issue
there,
which
was
basically
the
original
issue
tracking
when
we
added
node
conformance
tests
to
begin
with,
and
one
of
the
things
that
was
sort
of
discussed
when
we
went
through
this,
was
that,
like
right
now,
running
conformance
suites
require
you
to
spin
up
an
entire
cluster
and
there
was
a
goal
for
the
node
conformance
test
that,
like
basically,
we
would
be
able
to
run
conformance
tests
that
don't
require
a
full
cluster
that
just
require
a
cubelet.
B
And
so,
given
that
context,
I
think
it's
a
little
bit
more
complicated
than
just
saying.
Oh,
you
know,
we've
got
these.
These
node
conformance
tests,
they
will
check
the
cri,
maybe
we
can
call
them
cri
conformance
or
something
like
that.
It
is
a
little
bit
more
complicated
than
that,
and
so
I
think
basically,
we
need
to
sit
down
as
a
sig
and
discuss
what
we
want.
A
Kind
of
when
you
don't
want
to
bring
up
entire
cluster
and
still
want
to
test
what's
happening.
B
Yeah
and
yeah,
I
know
that
the
the
conformance
name
has
been
very
confusing
because
from
what
I
can
tell
like,
it's
they're,
certainly
not
like
conformance
as
in
the
kubernetes
conformance
project,
because
that
is
really
only
focusing
on
like
user-facing
apis
and
a
lot
of
the
cubelet
stuff
is
not
that
the
cubelet
conformance
tests
are
like
checking.
You
know
basically
behaviors
on
a
cubelet
which
may
or
may
not
actually
have
anything
to
do
with
like
outward-facing
like
api
things.
So.
B
A
Yeah
one
thing
about
node
conformance,
node
conformance
testaron,
as
the
p
submits.
So
if
you
want
to,
I
mean
this
is
one
of
the
reason
to
have
a
note
conformance.
So
it's
a
basic
set
of
tests
that
you
want
to
have
to
be
stable.
B
Yeah,
so
I
think
that
we
need
to
think
of
probably
a
better
name
for
node
conformance,
and
I
know
that
we're
also
right
now
sort
of
going
through
all
of
our
weirdly
named
node
specific
test
selectors.
So
I
think,
like
now
is
probably
the
time
when
we
go
through.
We
say:
okay,
is
this
test
in
or
out.
B
There
were
not,
I
mostly
just
got
a
bunch
of
historical
context
to
take
back.
That
was
different
than
what
initially
I
had
heard
on
slack
from
dims,
so
I
mean
I
think
we
still
have
the
item
going
forward
to.
We
need
to
improve
the
documentation
around
this,
but
I
think
the
reason
the
documentation
hasn't
been
improved
is
because
nobody's
really
in
agreement
on
in
terms
of
what
node
conformance
means
right
now.
So.
A
It's
from
what
I
understand
like
most
of
the
tests
are
marked
as
not
performance,
most
of
the
like
primary
tests
of
the
probe
test
or
the
config
tests.
A
Okay,
I
suggest
action
item.
I
think
so
it's
still
on
me.
I
I'm
the
item
on
me,
so
I
I
will
put
something
down
and
we
can
discuss.
B
Sounds
good
yeah.
I
think
that
that's
good,
and
mostly
I
just
wanted
to
like
bring
you
more
information,
because
I
think
that
we
were
offering.
We
were
working
on
incomplete
information
before
so
now.
I
think
this
is
kind
of
gone
from
a
oh.
This
is
like
an
easy
thing.
We
just
need
to
change
the
name
and
like
update
a
few
docs
to
this
is
a
bit
of
a
refactoring
project
and
sort
of
like.
I
think
it
aligns
well
with
some
of
the
the
work
that
we're
already
doing
so.
B
Yeah
and
the
next
item-
I
don't
think
you
need
me
for
any
of
these.
It's
just
mostly.
I
think
these
are
the
four
like
flaky
tests
that
are
being
tracked
right
now
for
ci
signal
and
just
making
sure
that
we
get
an
update
on
all
of
them.
I
think
that
the
test
freeze
is
next
wednesday,
so
we've
still
got
a
little
bit
more
time,
even
after
code
freeze
today
to
keep
looking
at
these.
D
B
The
reason
I
ask
is
because
that
was
what
I
did
for
when
I
put
together
the
probe
tests
like
for
the
liveness
probes.
I
basically
like,
went
and
ran
a
loop
for
an
hour
and
like
did
the
math
on
you
know
the
things
that
were
potentially
flaking
and
whatnot,
and
I
picked
the
threshold
such
that
they
still
demonstrated
correctness,
but
the
test
basically
never
flaked,
like
even
in
the
worst
case
scenario,
if,
like
things
raced
in
a
way
that
it
was
the
slowest
possible
thing,
it
still
demonstrated
the
correct
behavior.
B
But
the
test
wasn't
like
flaking,
because
things
were
too
slow.
So
if,
if
there's
no
risk
of
like
a
correctness
issue,
I
would
say
yeah
increase
the
the
thresholds.
We
may
have
just
not
written
the
test
in
a
very
good
way,
which
is
fine.
D
D
I
still
think
if,
if
the
the
thing
from
clayton
merge,
it
will
reduce
the
time
to
to
propagate
the
the
status
but
yeah.
I
suggest
we
just.
B
Do
that
both-
and
I
think
that
I
mean
I
hope
that
clayton's
thing
merges-
I
don't
know
what
the
I
don't
know.
What
the
definitive
answer
is
on
that
just
because
there
are
so
many
bugs
like
this
one
that
one
is
going
to
address,
but
I
think
that
it's
also
probably
a
good
thing
to
just
increase
the
threshold,
because
you
know
there
could
still
be
in
the
weird
race
condition
like
happens,
and
we
get
it
really
slow
this
one
time
and
the
test
fails.
A
Yeah,
I
wonder
what
does
mean
for
the
product
like
if
we
cannot
write
a
test
that
will
ensure
that
one
startup
probe
succeeds
like
containers
ready?
Is
it
actually
not
ready
in
in
many
cases
in
production
or
it
just
like
our
test
environment
is
so
bad
that
it's
too
slow
and
no
production
environments
will
ever
experience.
This
problem.
D
The
the
thing
is
locally,
as
with
the
current
master,
I
cannot
make
it
fail,
so
I
I
just
have
to
to
rely
on
the
on
the
ci
and
and
then
try
to
understand.
What's
going
on
because
locally
on
kind,
it
never
fails.
D
Yeah,
I
think
it's
it's
the
the
propagation
of
the
the
status
between
the
you
have
the
container
stitches
and
then
the
pot
status,
and
then
you
know
we
propagate
everything,
so
my
guess
is
is
maybe
we
run
like
many
tests
in
parallel
and
we
we
have
like.
I
don't
know
the
plague
that
is
like
over
overloaded
or
something,
and
it
takes
more
time
to
to
propagate.
B
I
haven't
looked
at
this
one
specifically,
but
I
know
with
like
the
probe
termination
grace
period
seconds.
I
mean
some.
Sometimes
it's
just
a
matter
of
for
the
startup
program.
B
I
think
it's
a
little
bit
different,
since
that
one
is
only
really
probing
at
the
beginning
of
the
container's
life
cycle,
but
at
least
for
the
termination
grace
period
seconds
like
it
was
basically
possible
for
if
you
timed
it
really
really
poorly
in
terms
of
the
interval,
depending
on
when
the
jitter
hit
and
then
also
like,
basically
another
random
startup
for
the
worker
and
then,
like
you
know,
when
the
first
failure
hit
it
could
potentially
sometimes
take
up
to,
like,
I
think,
basically
2x
the
test
period
to
fail
or
something
like
that
for
given
the
threshold.
B
B
Okay,
he
he
gave
a
bunch
of
updates.
I
don't
know
if
he's
busy,
we
might
want
to
reassign
that
one.
A
Sorry
yeah
does
anybody
else
want
to
take
a
look.
A
D
D
D
You
can
see
better
if
you
ignore
white
spaces,
because
I
have
like
a
bunch
of
of
course
that
was
just
put
inside
the
yeah
in
the
gear
yeah
in.
D
A
D
C
Yeah,
I
think
I
think
lana
tried
to
pink
someone
if
you
will
check
below,
but
we
still
don't
don't
have
any
answer
so
again.
From
my
perspective,
it's
like
looks
like
some
c
advisor
problem,
because
currently
our
ci
is
running
with
the
c
advisor,
not
his
like
runtime.
Oh.
A
I
think
david.
A
A
Okay,
I
will
pink
david
in
case
he
has
some
ideas,
yeah
and
finally,
on
the
board,
I
already
reached
a
few
items
that
just
added
them
to
the
board.
B
So
that
was
a
failing
test,
but
it
doesn't
affect
any
of
the
pre-submits.
B
Yeah,
oh
apparently
needs
a
reading
base
already.
I
I
submitted
this
one
yesterday,
mostly,
I
was
just
looking
to
add
periodic
jobs
for
cubelet
swap
stuff
now
that
it's
merged.
A
So
yeah,
I
hope
you
guys
would
anybody
else
wants
to
take
a
look.
I
think
it's
some
all
right.
It's.
B
I
basically,
I
just
renamed
the
jobs
to
all
be
consistently
named
and.
A
Yeah
I'll
take
both
as
well,
but
I
don't
really
want
that
mic
to
take
at
all
node
and
then
that's
for
female
containers.
A
C
A
Let's
just
quickly
take
a
look
at
what's
what
needs
approver,
yeah,
zero,
still,
okay,
it
seems
to
lose
lgtm.
A
So
this
is
what
this
one
is
interesting
like
parker,
following
up
on
some
pull
requests
I
used
yesterday
and
in
this
past
it
used
to
wait
for
events
to
hear
it
used
to
wait
for
events
and
expect
no
events.
Why
couldn't
where?
Is
it?
Oh
yeah,
this
one
wait
for
events
and
expect
no
events
and
then
check
for
both
being
for
this
gun.
A
So
it's
changed
to
wait
for
events
and
now
event
supposed
to
come,
but
the
problem
is
that
we
have
an
issue
that
saying
that
events
are
highly
unreliable
on
tests
and
there
are
many
cases
when
events
just
wouldn't
show
up.
So
that's
not
supposed
to
fully
rely
on
events
ever
and
that's
a
comment
I
made
and
yeah.
A
Yeah,
I
don't
know
whether
we
want
to
take
it
and
see
whether
it
flakes,
but
I
think
we
know
already
that
many
tests
that
rely
on
events
or
will
eventually
start
breaking
just
because
they
weren't
highly
unreliable.
D
A
D
It
the
case
in
like
a
real
world
clusters
that
we
lose
events.
Oh.
A
Yeah
some
highly
high
throughput
systems
like
when
we
have
a
lot
of
pods
and
many
events
they
start
like.
There
are
a
whole
bunch
of
throttlings
being
kicked
on
and
you
can
start
losing
events.
Okay,.
B
Yeah
events
are
lossy,
so
I
mean
you
may
have
seen.
For
example,
the
cubelet
tries
to
like
it
has
an
event
manager
and
it
tries
to
like
send
events
and
like,
for
example,
if
there's
a
network
partition
or
something
like
that,
the
cubelet
may
retry
a
few
times
but
like
it,
may
give
up
trying
to
sync
an
event
if
you've
been
working
on
particular
cubelet
issues
now
granted
like.
B
D
A
D
A
A
And
the
interesting
thing
that
the
test
used
to
rely
on,
no
events
will
come,
and
even
if
events
are
unreliable
and
just
didn't
show
up,
the
test
will
still
go
forward
right.
So
the
change
of
test
that
start
a
line
on
existing
movements
is
troublesome.
A
Okay,
I
will
put
it
on
me
reviewer,
but
I
think
something
going
on
there.
So
soon
it
will
go
back
to
mesa
program
yeah.
I
think
we
have
plenty
of
things
that
oh
wait
needs
approval.
C
And
probably
we
have
some
like
pretty
urgent
bug
fix
for
the
dynamic
kubrick
configuration
that,
like
it
was
set
to
duplicate
it
and
that,
and
so,
like
a
lot
of
our
serial
links,
started
to
fail.
Like
memory
manager,
cpu
measure
the
block,
they
all
started
to
fail
because
it
failed
to
pass
the
duplicated
version
for
the
for
the
feature.
A
A
Or
interesting
yeah
I
will
be
interested
to
take
a
look,
but
I
cannot
approve
so.
Can
you
pass
the
link
in
the
chat.
C
C
A
C
A
Was
here
for
a
long
time,
but
dedication
dynamic,
config
was
happened
yesterday
it
was
immersed
yesterday.
A
A
B
And
I
guess
that
those
end
to
ends
didn't
get
exercised
in
the
pr
where
they
made
the
change.
So
I
think
that
one's
a
separate
thing
mm-hmm.
A
A
So
last
item
here
is
two
more:
this
one's
we
just
discussed
francesca.
Can
you
join.
A
He
wants
some
help.
Okay,
if
anybody
will
have
time,
maybe
I
don't
know
whether
you'll
get
it
this
week.
A
Okay,
any
more
agenda
items
for
today.
B
B
Yeah,
I
feel
like
the
problem
right
now
is
not
like
reviewers.
The
problem
is
like
blocking
on
approvers,
but
here
let
me
very
quickly
share
my
screen.
I
think
we
should
keep
this
on
the
recording
and
let
me
just
take
a
quick
look
at
the
I.
D
Did
the
plus
one
on
your
pr
to
get
one
more
approval,
but.
D
And
I
I
spoke
to
a
bob
killen
as
well
about
this.
He
he
knows,
and
he
spoke
to
don
the
other
day.
So.
B
B
And
hopefully
this
is
the
right
tab.
I
have
two
of
them
open,
okay,
good.
So,
as
you
can
see,
we've
got
like
42
things.
I
think
right
now
that
need
approver,
and
so
let
me
make
sure
that
those
are
all
accurate
yeah.
It
looks
that
way.
B
43
things
that
need
approver
so,
like
all
of
the
ones
out
here,
never
mind,
here's
another
one.
B
So,
as
you
can
see,
we
got
a
lot
of
things
with
the
approvers
right
now
and
there's
some
stuff
that
still
needs
review.
I
think
I've
just
been
focusing
in
terms
of
like
anything
else
that
needed
review,
I'm
only
looking
at
stuff
that
are
like
critical,
urgent
or
important
soon,
and
this
one
is
a
windows
thing,
so
I
looked
at
it,
but
I
couldn't
really
do
anything
with
it
like.
B
I
think
the
windows
people
need
to
lgtm
it
and
if
they
need
an
approval
from
node,
they
should
last
us
in
terms
of
importance
and
stuff
there's
also
not
a
lot
of
stuff
here.
Paco
mentioned
this
one.
This
disable
the
cube
read-only
port
by
default.
I
think
we
need
an
approver
on
that
one.
So
if
somebody
wants
to
take
a
look
at
that,
that
would
probably
be
good.
B
Yeah,
I
think
that
was
part
of
the
concern,
and
I
know
I
think
a
couple
of
these
are
related
to
caps
that
we're
trying
to
land.
So
if
we
get
eyes
on
those
as
well,
that
would
be
great,
but
other
than
that.
I
don't
think
that
anything
else
on
this
list
is
like.
B
I
think
all
of
these
are
just
for
the
most
part,
I'm
not
sure
what's
going
on
with
this
one,
but
this
one
is
controversial
and
I
think
the
rest
of
these
are
all
related
to
caps,
so
they
definitely
need
eyes,
but
we're
really
just
blocked
out
of
proverbs
right
now.
So.
B
A
Yeah,
I
noticed
that
even
even
one
with
lgtm,
if
you
interested
to
take
a
look,
it
helps
to
have
second
look
on
some
prs,
because
you
you
leave
some
comments
and
then
it
will
be
easier
for
approver
just
to
make
a
final
decision.
B
B
I
can
do
feature
approvals,
so
I've
been
trying
to
like
to
make
sure
that,
for
a
lot
of
our
caps
that
we
want
like
another
set
of
eyes
on
for
approval,
I've
been
going
through
and
double
checking,
basically
like
the
stuff
related
to
anything
with
package
features
which
is
actually
a
lot
of
work,
because
you
have
to
make
sure
that
they've
met
all
of
the
cap
criteria
in
order
to
make
changes
to
feature
flags.
B
So
you've
got
to
go
back
and
read
all
the
graduation
criteria
and
then
I
just
kind
of
go
and
check
off
my
list
but
yeah
as
far
as
like
stuff
that
needs
note
or
stuff
that
needs
api
reviews,
there's
still
quite
a
lot
of
stuff
on
the
backlog,
so
just
make
sure
that
folks
are
aware
make
sure
that
they're
looking
at
stuff,
that's
more
marked
important,
soon,
critical,
urgent,
especially
and
hopefully
we'll
get
as
much
stuff
merged
as
possible.
A
Is
contributing
country
backs
seek
on
misery.
C
A
To
be
applied
automatically
fast,
without
the
need
for
comment.
B
Oh
yeah,
so
I
did.
I
did
follow
up
on
that
one
that
was
actually
a
sick
testing
thing
and
it
was
supposed
to
be
fixed
quite
a
long
time
ago
there
was
basically
an
issue
with
like
github
api
throttling,
and
so
they
made
like
some
sort
of
change
that
should
have
fixed
it
and
for
a
while
it
was
relatively
fixed.
B
So
if
the
label
is
not
being
applied
right
now
without
having
to
comment
on
things
to
like,
basically
kick
the
bot
then
feel
free
to
reopen
that
issue,
because
that
was
quite
a
long
time
ago.
I
think
it
was
yeah.
A
I
noticed
it
yesterday
that,
after
my
comments,
it
applied
need
to
review
a
new
metric
based
contact,
so
I
I
thought,
that's
not
fixed
at
times
like
that.
It's
extremely
important
because
it
helps
to
detect
something
faster.
B
I'm
trying
to
see
oh
here
we
go
yeah
I'll
show
you
the
I'll
link
the
issue
in
the
chat
for
the
notes,
and
if,
if
you
want
to
reopen
this
and
say
oh,
it's
still,
a
problem
again
feel
free,
but
it.
B
And
the
issue
was
that
there
was
basically
that
the
github
search
size
limit
was
too
big
and
they
were
having
some
issues
with
pagination
and
api
throttling.
So.
A
A
Yeah,
unfortunately,
nobody
from
ap
apac
showed
up,
let's
keep
it
doing
once
a
month
and
if
it
wouldn't
work,
we
can
try
something
new.
B
Yeah,
I
don't
know,
maybe
it's
just
because
it's
like
you
know
the
day
before
code
freeze
there,
so
they
were
all
pooped,
because
it's
the
end
of
their
day
or
I'm
not
entirely
sure
but
yeah.
Let's
maybe
the
other
thing
too.
We
might
be
able
to
set
up
like
a
reminder
or
something
like
that.
That
might
help.