►
From YouTube: Kubernetes SIG Network 2019-05-02
Description
May 5th, 2019 SIG network meeting
A
D
E
C
E
E
F
E
E
D
E
E
I
K
E
The
problem
is
for
very
large,
so
we
scale
Q
proxies
contract
requests
by
the
number
of
cores
in
the
machine,
which
seems
roughly
appropriate
for
very
large
machine.
That's
going
to
be
a
very
large
kernel
allocation
which,
if
the
colonel
can't
allocated
in
the
fast
path
that
will
fall
back
on
the
v-mail
and
we'll
complain
about
it
along
the
way,
and
the
question
here
is:
can
we
do
better
than
that.
E
E
E
E
I
J
E
C
C
L
C
C
C
E
C
F
H
C
C
G
All
right,
real,
quick
on
that
note,
the
thing
that
was
fixed
does
that
actually
prevent
where
it
is
it
is
it
just
for
logging
errors,
because
I
had
a
similar
bug
open
and
from
what
I
found
it
seemed
like
for
external
IP?
Is
it
just
logged,
but
there
was
a
problem,
but
it
didn't
prevent
you
from
kind
of
mixing
that
in
and
then
IP
tables
rules
got
screwed
up
fixing.
What,
in.
E
C
E
I
D
G
E
F
C
C
E
L
I
E
E
I
C
B
K
E
M
E
M
H
C
Doesn't
allow
the
IPS
the
balance
room
to
get
out
at
the
note,
it's
I
think
it's
a
feature
request.
D
E
Okay,
we
are
on
time
all
right.
There
we
go
2:30,
all
right
cool.
Thank
you.
Everyone
who
signed
up
for
bugs
there's
plenty
more
clearly
anyway,
who
feels
like
a
go-getter,
can
dig
into
some
of
these
and
try
to
do
some
triage
and
we'll
note
also
to
we're
all
the
way
back
to
July
now,
which
is
fantastic.
E
H
Hello
I
put
this
here
with
Cal
hello,
stick
Network,
so
really
not
not
specifically
about
the
cap
in
detail,
but
one
of
the
implementation
phases
requires
type
changes
to
CRI,
which
we
went
over
and
had
a
chat
to
the
folks
at
signo
to
kick
off
that
discussion
and
they
but
promptly
said
without
this
kept
being
in
the
right
state.
They
won't
consider
those
changes.
So
really
we
just
wanted
to
bring
that
to
the
awareness
of
the
sig
and
I
understand.
If
there's
any
week,
anything
we
can
do
because
the
cap
is
under
review.
E
H
So
yes,
and
thank
you
very
much
I
spent
time
with
Cao
and
basically
the
way
we
addressed
your
concerns
was
explicitly
calling
out
each
of
the
steps
in
each
of
the
phases
and
the
cube
proxy
stuff
was
explicitly
put
in
phase
two.
So
I
could
throw
that
I
could
show
you
what
the
phases
are
now
or
I
could
just
ping
you
afterwards,
but
I.
Think
we've
addressed
all
your
concerns.
H
E
H
A
H
G
G
E
G
E
G
E
C
So
I
presented
this
to
the
sig
last
meeting
like
wanting
to
know
if
there's
any
additional
feedback
and
if
that's
ready
to
put
into
the
cab
and
also
any
other
like
feature
requests,
I
wanted
to
pick
it
back
into
it.
Any
like
things
that
has.
The
proposer
has
not
considered
like,
for
instance,
the
one
that
just
came
up
the
ipv6
to
us.
That
looks
like
it's
gonna
change.
The
M
points
object
as
well
to
support
multiple
eyepiece
if
I'm
reading
it
correctly
yep,
so
that
probably
need
to
have
a
parody
in
the
end.
C
E
C
C
E
There's
still
there's
still
one
major
user
of
it
and
they
are
here.
They
are
now
Dan
and
Dan.
I
think
are
now
owners
of
the
user
space
proxy
as
long
as
they're
using
it
and
it's
not
causing
drag
I,
don't
mind
leaving
it.
If
you
guys
want
to
fork
it
and
move
it
out
into
separate
repo.
That
would
also
be
okay,
I
sort
of
leave
it
in
your
hands,
I.
B
Mean
we're
mostly
using
it
as
part
of
the
Nidal
Inc
proxy
and
in
theory
forking
it
out,
might
let
us
clean
things
up,
but
at
the
same
time,
if
we
manage
to
move
everybody
over
to
VN
kubernetes,
then
it
doesn't
use
cube
proxy
at
all.
It
has
its
own
proxy
implementation.
So
then
we
wouldn't
care
at
all,
so
maybe
just
leave
it
around
and
if
people
report
bugs
we'll
fix
them-
and
we
can
revisit
this
in
in
a
year-
let's
see
if
it's
time
to
kill
it
off
then
I'm.
A
Cool
I
think
that
makes
sense
mostly
curious,
because
I
was
coming
across
a
couple
issues
raised
against
it
and
wasn't
sure
how
much
time
to
invest
in
in
those
sorts
of
issues.
I
wouldn't
invest.
B
So
the
nice
property
it
has
that
the
IP
tables
one
doesn't
is
that
because
it
actually
opens
the
connection
to
the
pod
before
returning
the
connection
to
the
caller.
It
guarantees
you
get
a
live
connection,
whereas
at
the
IP
tables
proxy
can
redirect
you
to
a
dead
pod
and
it
just
hasn't
failed
its
liveness
check,
yep
and
apparently
this
is
an
issue
for
some
people.
Interesting.
E
E
Bowie
I
think
seven
changes
in
the
API
of
things
we
would
want
to
do
are
calling
nga
I.
Think
most
of
them
are
good.
There's
the
debate
about
how
they
should
exactly
look
and
what
the
implementation
should
be.
If
anybody
cares
about
ingress
in
the
ga
feature
set,
please
go.
Take
a
look
at
that
cap
and
offer
your
feedback.
A
C
E
I
E
E
H
A
J
E
H
C
H
E
N
C
N
Format,
kxz
compress
the
kernel
object,
so
this
is
probably
because.
B
C
F
M
H
D
I
E
E
Everyone
Oh
two
weeks
from
today,
I
guess
we're
gonna
meet,
but
that'll
be
the
last
one
before
team
con.
So
hopefully
we'll
have
an
outline
at
the
very
least
of
what
we
can
be
talking
about
at
Q.
Khan
I'll
try
to
share
something
before
that
Valerie.
We
need
to
circle
back
on
that
I
apologize,
I
lost
track,
yeah.