►
From YouTube: SIG Cloud Provider 2022-05-11
Description
Agenda: https://docs.google.com/document/d/1OZE-ub-v6B8y-GuaWejL-vU_f9jsjBbrim4LtTfxssw/edit#
[joelspeed] extraction/migration schedule
[steve wong] We (Nick and Steve Wong) need “lightning talk” presentations from individual cloud providers now for KubeCon Europe
event link https://sched.co/ytow
draft of deck is here - https://docs.google.com/presentation/d/1x6igOAIJmNN7xlEiiX98TTufsqztXbbu/edit?usp=sharing&ouid=118341252518163971765&rtpof=true&sd=true
A
All
right
apologies
in
advance.
If
my
dog
barks,
I
am
gonna,
move
my
office
downstairs,
so
it
shouldn't
be
a
problem
in
the
future,
but
we'll
have
to
get
through
one
more
one
more
meeting
so
welcome
everyone.
It
is
may
11th.
Today
this
is
sig
cloud
provider.
We
looks
like
we
have
a
couple
items
on
the
agenda,
so
let's
go
ahead
and
get
started.
A
C
A
B
B
A
A
A
A
B
Think
that
the
next
paragraph,
too,
is
where
they
were
hitting
a
lot
of
problems
like
upgrading.
Where
the
you
know,
the
kcm
you
know
is
not,
or
you
know
the
kubelet
when
it
has
cloud
provider
equal,
openstack,
it's
not
adding
that
annotation
and
so
the
transit
between
from
the
kcm
to
the
ccm.
Then
you
know
again,
this
flapping
can
occur.
A
A
B
B
A
B
A
A
B
A
Yeah
and-
and
we
I
think,
we
really
need
to
decide
what
our
support
model
is
and,
like
you
know,
if
we
have
load
balancer
code
that
we're
not
supporting
in
in
the
external
and
or
entry
there's
really
three
places
now
yeah.
We
should
be
much
more
explicit
in
our
documentation
about
like
what
we're
supporting
and
what
we're
not.
A
B
B
A
B
D
A
D
So.
The
problem
with
that
situation
is
that
both
the
cubelet
and
the
node
controller
running
on
the
ccm
will
manage
node
addresses
on
the
node.
So
if
there
is
any
disparity
between
what
the
ccm
thinks
the
node
addresses
should
be
and
what
the
entry
cloud
provider
thinks
the
node
addresses
should
be,
then
they
are
going
to
flap.
D
D
Oh
sorry,
keebler
then
cubelet
will
annotate
its
node
with
the
with
the
node
ip
that
it
was
provided,
and
this
means
that
the
cloud
provider
can
filter
the
node
addresses
that
it
is
that
it
is
adding
to
the
node,
and
in
fact
this
is
cubelet.
Does
this
itself
so
keyblock
running
with
entry
gets
a
bunch
of
node
addresses
from
the
entry
cloud
provider
and
then
it
filters
them
by
node
id,
and
this
was
a
change
in
124.
D
D
That
means
that,
in
this
upgrade
situation,
cubelet
is
not
configured
with
external
cloud
provider.
So
it's
not
providing.
So
when
the
the
node
controller,
running
and
ccn
runs,
it
can't
see
the
node
ip
and
therefore
it
node
controller
is
writing.
The
unfiltered
addresses
and
cubelet
is.
Writing
writing
the
filtered
addresses,
so
they
flap
between
two
potentially
very
different
sets
of
node
addresses
and
this
causes
breakage
in
practice.
D
That
means
that,
if
keebler
is,
is
running
yeah,
whether
it's
configured
for
entry
or
out
of
tree,
it
will
always
add
that
annotation
so
that
when
we're
in
this
dual
controller
situation,
the
ccm
can
always
see
that
annotation
and
therefore
we
flap
between
a
consistent
set
of
addresses
so
still
not
a
great
situation.
But
it's
a
small
fix
rather
than
a
big
fix.
A
Is
that
no
yeah?
It's?
I
think
it's
it's
definitely
clear
to
me,
although
I
we
did
encounter
this
this
problem,
so
I
have
some
context
and
I
agree
that
the
ideal
fix
would
involve
not
having
both
cubelet
and
the
the
node
controller
control
the
address
fields,
but
I
think
we
can
merge
this
even
if
we
were
to
fix
it.
You
know,
even
if
we
were
to
fix
that
issue
as
well
like.
I
don't
think
this
would
cause
any
issues
right,
we're
just.
C
D
First
now,
to
be
honest,
I
don't
think
that's
actually
documented
anyway,
that
that's
a
thing
but
or
if
it
is,
I've
never
found
it,
but
yeah.
We
have
these.
We
have
these
existing
configurations
where
the
non-primary
network
is
listed
on
the
machine
first
and
and
we
also
specify
node
ip
on
the
on
the
keyblade.
So
when
cubelet
comes
up,
it
knows
which
of
its
node
addresses
is
actually
the
one
that
can
route
to
the
to
the
cube
api
server.
D
C
That's
fair,
I
the
reason
I
ask
and
my
information
is
woefully
out
of
date.
The
last
time
I
looked
at
this
was
about
four
years
ago,
but
when
I
looked
at
this
and
I
went
to
sig
node
and
chatted
with
them,
the
general
impression
of
the
community
was
that-
and
this
had
been
a
recent
change
at
that
point-
was
that
I
p
main
the
the
ip
address
in
the
node
object
should
either
be
controlled
by
the
controller
manager
or
the
cubelet,
but
not
both
and
that
having
the
no
the
cubelet
control
it.
C
C
Well,
there's
also
there's
a
difference
between
the
node
ipflag
and
writing
to
the
node
object
and
in
fact
the
node
ip
flag
is
not
deprecated.
It's
all
the
rest
of
those
flags
and
again
I
haven't
looked
in
the
code
in
four
years,
so
I
I'm
not
the
best,
I'm
just
trying
to
understand
and
just
wanted
to
sort
of
yeah.
B
D
D
A
change
we
made
recently
was
yeah.
Previously
cubelet
was
was
annotating.
The
known
object
with
with
its
node
ip,
but
ccm
wasn't
doing
anything
with
it,
which
meant
that
when
you,
when
you
switched
from
ccm
to
external,
then
your
your
node
behavior
changed.
So
all
we
did
was
we
moved
the
the
node
filtering
code
into
cloud
provider
and
now
both
cubelet
and
ccm
use
the
same
code,
but
it's
yeah.
It's
just
yeah
the
in
the
external
case.
It's
only
the
filtering
is
only
done
by
by
ccm.
D
B
C
Yeah
I
mean
I
just
I
wanted
to
understand
and
wow
and
also
exactly
to
the
point
you
just
made.
I
wanted
to
understand
how
we
ended
up
with
two
and
that
you
know
that
wasn't
meant
to
you
know.
Obviously,
that's
not
meant
to
be
the
case,
and
I
just
wanted
to
make
sure
that
we
were,
you
know
properly
dealing
with
it
and
it
sounds
good.
A
D
A
C
B
B
C
A
A
D
D
But
it's
it's
kind
of
ccm
related.
I
I
wonder
if
it
fits
at
this
point
in
the
agenda,
essentially
it
it
makes
a
small
change
to
conflict
parsing
in
the
legacy
cloud
provider
so
that
it
doesn't
bath.
If
you
send
it
confident
directives
that
it
doesn't
understand,
for
example,
because
they're
only
present
in
external
cloud
provider.
A
D
Upgrade
thing:
it
just
means
that
for
the
next
two
cube
releases
before
we
can
get
rid
of
a
cloud
provider
for
kcm,
it
means
we
don't
need
to
maintain
two
different
cloud
conflicts.
We
can
just
use
one
and
and
kcm
that
entry
cloud
provider
and
casein
will
ignore
anything
that
it
doesn't
understand,
as
as
it
is
as
it
is.
If,
if
the
user
adds
external
ccm
config
to
their
cloud
comp,
then
their
cluster
dies
because
kcm
doesn't
come
up.
A
A
E
Yeah
thanks,
so
this
is
probably
more
a
question
for
the
extraction
meeting,
but
that's
a
little
bit
late
in
the
day
for
me,
so
I've
not
managed
to
attend
it.
So
far
I
had
heard
on
the
grapevine
that
there
was
some
delays
with
the
webhooks
project,
so
I
was
kind
of
wondering
what
the
schedule
with
the
extraction
migration
stuff
is
and
when
we're
thinking
of
clicking
the
feature
gates
so
they're
on
by
default
for
the
saving
cloud
provider
stuff
and
also
where's
best
to
track
that
in
future.
A
C
C
C
That
explains
all
of
the
problems
that
we
have
with
testing.
The
issue
is
that
a
huge
number
of
our
tests
are
google
specific,
I
mean
I
say
huge.
It's
probably
less
than
100
of
the
tests
are
google
specific,
but
while
we
have
a
working
ccm
distribution
in
sig
cloud
provider,
gce,
that's
outside
of
kk
and
inside
of
kk,
with
the
cube
up
scripts
and
no
ccm,
we
don't
actually
get
a
completely
healthy
cluster.
C
When
you
turn
off
all
the
cloud
provider
code
and
as
a
result,
quite
a
few
of
the
tests
fail,
and
so
we
need
to
get
people
on
board.
With
fixing
the
tests,
we
need
to
off
basically
say
we're
okay
with
not
running
those
tests
or
we
need
to
build
a
test
system
which
would
allow
us
to
push
off
running
those
tests
to
a
separate
repo
that
then
takes.
C
The
current
code
builds
the
distribution
on
top
of
it
runs
those
tests
there
and
then
pulls
the
test
results
back,
and
that's
that
document
that
I
said,
I
need
to
write,
there's
an
earlier
version
of
it.
That's
actually
already
out
for
cl
sig
cloud
provider,
which
explains
that
last
option
I
just
talked
about
it
was
written
by
joe
betts
and
it's
called
last
known
good.
So
if,
if
what
I
said,
you
seemed
kind
of
complicated
and
you
want
to
better
understand
it.
E
C
That
that
is
oddly
a
kind
of
a
call
for
for
the
storage
team,
so
I
mean
the
best
call
for
that
is
going
to
be
michelle
ow,
but
specifically
that
that
has
to
do
with
a
very
spectacular
legacy
use
case
where
I
cut
a
customer.
So
I
mean
this
is
generally
going
to
be
like
an
azure,
an
amazon,
a
google
customer
has,
let's
say
a
database
with
a
bunch
of
storage
and
they
want
to
migrate
to
kubernetes,
and
they
want
to
bring
that
that
disk
with
them
the
ccm
web
hook.
C
Work
has
to
do
with
extracting
the
little
bit
of
cube
api
server
code
that
allows
that
to
happen
today.
It's
basically
an
admission
controller,
not
an
interesting
web
hook,
controller
but
just
an
admission
controller
that
calls
to
the
cloud
provider
code,
and
so
we
want
to
pull
that
out
and
replace
it
with
a
web
hook.
That
does
the
same
thing
so
that
we
no
longer
need
to
link
the
the
cloud
provider
code
into
the
cube
api
server.
C
A
A
A
E
E
A
F
So
and
with
andrew
gone,
I
guess
maybe
walter
has
volunteered
for
a
role
there,
if
I'm
not
mistaken
too,
but
that
doesn't
appear
to
have
been
updated
in
the
deck.
It's
still
the
stuff
that
I
cut
and
pasted
over
from
the
last
time,
then
ibm
has
submitted
their
little
lightning
talk
for
everybody
else.
F
It's
due.
If
you
can
get
that
to
me,
I'm
flying
out
thursday.
So
I'd
really
like
to
have
that.
If
you
intend
to
give
me
a
video
I'd
like
to
have
it
before
I
get
on
the
plane,
because
if
I
have
to
do
any
video
editing
or
transcoding
to
get
it
in
the
deck,
I
really
would
prefer
to
do
it
on
my
home
system,
rather
than
the
laptop
in
a
hotel
room
with
lousy,
wi-fi.
A
F
A
F
A
F
A
F
F
A
F
F
Now,
that's
monday,
so
it's
possible
that
people
won't
be
there
by
monday,
but
anyway,
I
think
that
was
one
procedure
to
get
yourself
a
table
in
some
room
and
if
we
wanted
to
make
it
later
in
the
week
things
I've
seen
before
you
could
just
the
conference,
I
think,
provides
lunch.
So
we
could
just
make
a
declaration
that
we'll
find
each
other
in
a
table
at
a
lunch
at
the
convention
center
or
we
could
try
to
meet
during
one
of
the
evenings.