►
From YouTube: Kubernetes SIG Network meeting for 20230803
Description
Kubernetes SIG Network meeting for 20230803
A
Thank
you,
hello,
everyone
and
welcome
to
the
August
3rd
edition
of
the
Sig
Network
meeting,
just
a
reminder
that
this
meeting
is
under
the
kubernetes
code
of
conduct,
as
always
boiling
down
to
basically
the
nice
to
one
another.
So
please
be
nice
to
one
another.
Please
try
to
use
the
raise
hand
feature
in
Zoom,
as
that
can
help
us
to
organize
interjections
comments
and
stuff
like
that.
We
only
have
a
couple.
Well,
we
have.
A
A
I
am
feeling
a
little
bit
under
the
weather
today.
So
please
bear
with
me
I'm,
probably
going
to
be
a
little
bit
slow,
Carry,
Me
I'm,
going
to
start
sharing
my
screen.
A
A
First,
we're
going
to
start
with
triage
I've
already
opened
up
all
the
unassigned
triage
items,
so
we
can
start
going
through
that
the
first
one
here
is
this
Ingress
slash,
Health,
Z
404
not
found
I,
don't
know
if
anybody's
had
a
chance
to
look
at
this
one
yet
because
Erna
was
three
weeks
ago.
Actually,
so
we
put
in
this
Ingress
on
Azure
AKs.
D
E
A
I
think
I
agree
with
that.
I
can
take
this
one
and
kind
of
move
it
that
way.
I
don't
think
I
can
actually
move
it.
Maybe
I
can
I'll
figure
that
out
later.
A
Next,
one
is
from
Jay:
should
Google
proxy
delete
its
rules
when
being
gracefully
deleted?
F
You
yes,
so
in
the
past
we've
talked
about,
should
Cube
proxy
do
its
own
cleanup
on
termination
and
the
the
general
problem.
Is
it
doesn't
really
know
whether
it's
going
down
in
order
to
come
right
back
up,
in
which
case
tearing
down
its
rules
would
be
disastrous
or
whether
it's
going
down
to
stay
down?
We
had
this
like
Cube
proxy
dash
dash
cleanup
flag.
That's.
A
F
Here,
yeah,
we
were
never
able
to
like
actually
keep
maintained
and
like
it
was
never
consistent
enough
and
we've
I
think
more
or
less
said
we're
not
really
doing
that
anymore.
So,
unless
somebody
like
Dan
tell
me
if
you
disagree
so.
G
F
Oh
okay,
all
right
so
maybe
I
misremembering
that
aspect
of
it
yeah
anyway.
The
the
point
of
cleanup
and
exit
was
to
make
the
I'm
going
down
in
tour
to
stay
down
an
explicit
thing.
Obviously,
that
doesn't
map
well
to
like
a
demon
set
being
turned
off.
A
Okay,
so
should
I
respond
to
this
one.
He
left
oh
wait.
He
already
Antonio
already
kind
of
made
it
clear.
So
should
we
I
guess
I
could
take
this
one
and
just
kind
of
try
to
move
it
to
close
with
Jay,
because
he
seems
to
be
kind
of
getting
that
basically
I'll
just
check
in
with
him
and
see
if
he
accepts
this
or
if
he
has
any
more
comments,
and
if
he
does
accept
this,
then
we
can
probably
consider
this
one
closed.
Okay,.
F
H
Says
use
case
only
building
like
using
kubernetes
Cube
ADM
to
set
up
a
cluster
and
he
just
doesn't
want
coup
proxy
to
run
or
be
installed
because
Kube
ADM
supports
that
use
case.
So
then
he
doesn't
even
need
to
worry
about
this
problem.
Let.
I
Right
because,
once
once
you
clean
there
are
no
rules,
no
service
for
this
node.
So
you
don't
want
that
because
even
today,
if
we'll
start
a
node,
sorry
not
restart
anode.
If
we
restart
proxy
everything
works
all
right.
At
least
the
existing
rules
will
work
right.
The
new
rules
until
proxy
comes
up
and
does
the
cash
and
so
on
everything
works.
So
I
don't
want
to
I,
don't
want
to
like,
assuming
that
we're
dying
forever.
F
H
A
I
added
your
comment
here,
basically
to
kind
of
check
in
with
him,
because
I
think
that
I
mean
yeah.
If
you
can
just
not
because
right,
you
can
flag
it
and
say:
please
do
not
you.
H
A
Your
comment
makes
perfect
sense
to
me,
so
I'll
I'll
go
for
that
unless
you
want
to
make
the
comment
on
here.
I'll
just
say
that
and
just
kind
of
try.
H
The
problem
by
yeah
allowing
the
normal
Cube
ADM
process
to
go
and
then
he's
unwinding
it
when
he
could
just
prevent
the
issue
in
the
first
place.
A
A
Cool
who's
working
on
that.
K
J
C
C
K
C
K
G
The
the
big
problem
that's
getting
fixed
here
is
is
that
we
never
really
dealt
with
dual
stack
for
for
this,
and
and
so,
if
you're
on
dual
stack,
you,
you
could
only
pass
one
bind
address
and
if
we
have
to
Auto
detect,
we
intentionally
ignore
the
second
IP
in
in
the
node
object,
and
only
look
at
the
first
one
and
we're
sort
of
trying
to
figure
out.
If
you
know
we're
going
to
break
anyone.
If
we
fix
that-
and
it
looks
like
no
so,
except
maybe
on
Windows.
E
C
A
Yeah
no
worries
all
right,
so
this
annotation,
not
usable
with
external
name
and.
C
F
I'm
trying
to
figure
out
what
they
expect
it
to
do.
Yeah.
J
A
B
B
F
J
A
Last
one
also
from
Lars
can't
restrict
access
from
kubernetes
nodes
with
a
load.
Balancer
Source
ranges
for
proxy
mode
equals.
Oh,
is
this
related
to
the
other
one?
Oh,
no,.
K
K
Go
ahead
but
the
load
Source
ranges
cannot
be
used
because
the
source
address
will
be
the
load
balancer
address
in
in
the
ipvs
case,
so
it
you
can
do
it.
If
you
write
a
load,
balancer
address
s
well,
I
I
cannot
see,
it
can
be
done
in
any
simple
way,
but
it
is
a
problem
and
it
differs
from
IP
tables
proxy.
F
F
Logo
I
see
what
you
mean
now,
because
right
of
course,
because
that's
a
local
address.
H
F
G
F
A
G
F
K
E
If
you
put
the
source
on
them,
the
kernel
will
pick
it
if
there
is
no
existing
Source
address
in
the
packet.
You.
B
K
J
K
F
F
Go
ahead,
so
this
came
up
in
like
in
all
the
testing
around
combinations
of
Antonio's
laughing,
because
he
knows
the
combinations
of
fields
that
service
has
all
the
reasons
that
we
dislove
Service
for
the
various
combinations
of
type
and
other
fields
that
are
allowed.
F
We
find
that
there
was
an
e
to
e
test
which
was
tickling
this,
which
escaped
notice
for
a
long
time,
which
was
that
if
you
have
a
service
of
type
cluster
IP
and
anything,
that's
not
load
balancer
type,
you
can
patch
the
status
to
have
a
load.
Balancer,
Ingress
I
think
this
yeah,
the
says
spec,
but
it
should
be
status.
F
The
there's
code.
We
have
code
that
specifically
says
when
you
can,
when
you
update
a
service
and
the
type
is
not
load
balancer,
we
wipe
the
load
balancer
status
and
I.
Remember
writing
that
and
that's
great
if
you're,
converting
something
from
a
load
balancer
to
a
cluster
IP.
But
if
I
have
a
cluster
IP
service
and
then
I'm
allowed
to
patch
the
status
now
I
have
a
cluster
IP.
That
indicates
there's
a
load
balancer
Ingress,
which
is
totally
meaningless.
I,
don't
know
if
what
I
have
an
audited
like
Cube
proxy
I.
F
Think
Cube
proxy
would
ignore
that.
But
you
know
it's
not
inconceivable
that
it
might
slip
through
and
actually
be
trying
to
set
up
routes
and
stuff
for
it.
But
it's
super
volatile
because
the
next
time
that
service
spec
is
updated
for
any
reason,
including
changing
a
label
or
something
then
that
field
will
get
wiped
again.
So
the
question
that
I
posed
to
the
group
is
a:
does
anybody
consider
this
to
be
a
breaking
change?
If
we
were
to
fix
this
and
B?
F
How
breaking
are
we
willing
to
tolerate
and
I
wrote
here
the
options
we
could?
We
could
fix
validation
as
we
should
have
done
and
just
say:
no,
no.
You
can't
set
this
field
unless
type
is
load.
Balancer,
that's
a
really
breaking
change.
So
if
anybody
is
actually
doing
this,
it
will
stop
them
from
working
completely.
F
F
I
thought
we
could
maybe
throw
a
warning,
but
it
turns
out
the
the
pre
prepare
for
update
code
is
called
before
the
warnings
code,
so
that's
actually
hard
or
we
could
just
throw
a
warning
and
leave
it
the
way
it
is
and
let
it
be
volatile
anybody
have
thoughts
on
this.
One.
I
C
F
F
Status
deprecated,
that's
a
great
Point
Antonio.
Okay,
so
we
add
a
feature
gate
that
is
on
by
default,
which
says:
don't
allow
it
and
then,
if
anybody
experiences
a
problem,
they
can
go
flip
that
feature
gate,
and
then
we
can
talk
about
how
to
fix
it.
For
real
I
forgot
that
that
was
there.
Yes,
that's
a
great
answer.
F
A
Satisfied
Antonio:
do
you
mind
throwing
your
idea
on
to
the
comments
there
for
that
one,
please
yeah.
C
D
Okay,
yeah
I,
don't
want
to
monopolize
too
much
time
here,
but
Gateway
API
we're
going
to
be
trying
to
have
updates
every
Sig
Network
meeting
between
now
and
our
plan
GA
release
in
October
trying
to
avoid
any
surprises
in
any
direction
between
now
and
then
and
we've
got
a
pretty
tight
timeline
before
we
want
to
hit
1.0
before
we
get
there.
We
are
planning
to
release
0.8
this
month.
D
That
includes
three
main
things:
one
that
includes
gamma.
Our
kind
of
mesh
inter
initiative
is
going
to
experimental,
which
basically
means
it's
part
of
the
release.
It's
something
like
you
can
use.
Then
we've
got
cell
validation
at
long
last,
we're
embracing
built-in
Advanced
crd
validation.
D
It
will
work
for
125
and
above
but
Gateway
API
promises
to
support
120
and
the
trailing
five
versions.
So
we
will
still
need
to
some.
Some
users
will
still
need
to
install
the
web
hook
and
then
finally,
there's
some
version
bumps
in
this
release.
D
That
means
you
know
changing,
which
version
is
stored,
served
Etc
just
again
trying
to
align
everything
up,
so
1.0
goes
smoothly
in
October
and
yeah,
as
I
mentioned,
we're
absolutely
trying
to
make
sure
we
have
check-ins
as
frequently
as
possible
in
these
meetings
and
anything
that
comes
up
in
our
more
regular
Gateway
API
meetings.
That
feels
like
it
might
be
a
little
controversial,
we're
going
to
try
and
bring
up
to
here
so
there's
again,
less
surprises
as
we
go,
and
first
up
Flynn
had
one
thing
about
timeout
formatting
in
timeout.
J
D
Maybe
I'll
hand
it
off
to
Flynn
and
you
can
probably
provide
a
bit
more
context.
B
So,
there's
a
very
long
comment
at
the
end
of
this
thing,
but
for
our
purposes
immediately
the
tldr
there,
which
is
that
Shane.
Why
did
you
make
that
smaller,
that's
very
weird,
the
tldr
is
that
if
you
look
at
what
cell
does
and
you
look
at
other
things
that
are
already
in
the
kubernetes
world,
like
Meadowview,
1.
duration
and
things
like
that,
all
of
them
boil
down
to
goes
timed
out:
parse
duration,
under
the
hood.
If
you
look
at
the
source
code,
so.
B
A
we
should
probably
just
use
that
to
reduce
scope
going
into
0.8,
but
B.
That
format
also
has
some
horrible
things
like
floating
Point
numbers
and
very,
very,
very
open-ended
things
that
it
will
allow.
So
Rob
came
up
with
a
regex
that
he
and
I
both
think
should
do
a
reasonable
job
of
letting
us
put
some
bounds
on
what
the
input
will
actually
allow,
while
still
leaving
the
door
open
for
discussing
things
like
well.
F
Think
your
regex
is
wrong.
You'll
allow
octal
seeming
numbers.
What
does
it
mean
if
I
specify
zero
seven?
Seven,
seven
minutes
is
that
octal
or
integer.
J
B
F
B
Well
also,
note
that
also
note
that
the
unit
at
the
end
is
required.
So
there
isn't
a
way
for
yaml
to
interpret
this
as
a
number.
It
will
always
be
interpreted
as
a
string,
so
we
don't
have
so
we
don't
have
to
worry
about
yaml
itself,
giving
you
an
octal
value.
When
you
didn't
mean
one,
we
just
have
to
worry
about
people,
thinking
that
they
wrote
an
octal
value
and
not
getting
it.
F
I
guess
the
real
question
is
is
what
is
go?
Parse
duration.
Do
if
I
specify
zero,
seven,
seven,
seven.
F
While
we're
talking
the
so
I
read
I
read
over
this
proposal
yesterday,
sorry
did
you
want
to
keep
going
or
do
you
want
to
talk
about
it
now.
B
F
I
mean
my
hand
was,
was
still
up,
but
I
I
read
it
over
I
other
than
the
possible
little
wart
with
the
regex.
It
seems
okay
for
for
anybody
for
following
along
the
big
issue.
Here
is
when
we
embed
a
format
into
a
string
in
our
API.
We
require,
then
clients
be
able
to
parse
that
to
do
anything
with
it.
F
If
you
want
to
consume
one
of
these
types-
and
you
want
to
compare
two
different
durations
for
some
reason,
if
you're
a
controller-
and
you
want
to
say-
is
this
duration,
less
than
zero?
You
need
to
have
a
you
need
to
parse
it
right,
which
means
we
need
to
write
a
grammar
for
for
such
a
string,
formatted
field,
which
means
we
can
never
ever
change
it,
because
it'll
go
out
into
client
code
and
that
client
code
will
potentially
never
be
updated.
So
we
get
one
shot
to
Define
this.
F
We
can't
expand
it
and
we
can't
constrict
it.
We
get
it
right
or
we
get
it
wrong,
which
is
why
we've
had
such
a
huge
debate
about
what
we
want
to
put
here
right,
like
C,
for
example,
parsing
of
IP
addresses,
which
you
think
has
a
pretty
well
understood
grammar,
but
it
turns
out
that
we
were
just
using
goes
embedded
parser
and
its
parser
treats
leading
zeros
differently
than
perhaps
other
parsers
do.
So.
F
F
So
we
would
need
to
specifically
document
that
right
for
those
following
along
this
is
why
we're
like-
hey-
let's
not
use
floating
point,
because
then
we
have
to
describe
what
the
accuracy
of
our
floating
Point
parsing
is
right
and
let's,
let's
not
do
complex
compound
values
like
1m30s,
which
go
parsing
allows,
but
then
our
clients
all
have
to
be
able
to
parse
that.
B
B
B
The
kubr,
S
crate
in
Rust
just
got
support
for
this
added
recently,
in
fact,
for
this
particular
Gap
that
mimics
the
behavior
of
timed
parse
duration,
so
I
will
get
to
go
to
my
colleagues
and
at
buoyant
and
say,
let's
rip
the
floating
Point
part
out
of
that,
and
that
will
probably
make
them
happy
and
python
has
a
pip
installable
package.
That
also
does
the
right
thing:
the
boys
business,
the
implementations
that
I
have
looked
at,
do
support
out
of
order,
durations
and
repeated
durations.
F
Yeah,
so
whatever
we
decide
here,
we
we
need
to
aggressively
document
what
we
claim
to
support
and
we
need
to
aggressively
validate
that.
We
only
get
the
things
that
we
claim
to
support
and
we
need
to
have
a
like
a
published
set
of
test
cases
that
if
you
want
to
implement
a
parser
for
this
you'd,
better
parse,
all
of
these
test
cases
correctly
yeah.
B
F
B
J
B
F
The
other,
the
other
part
of
this
that
had
me
worried,
is
about
serialization
and
deserialization,
and
this
is
why
I
didn't
like
the
meta,
V1
duration
as
a
field
type,
because.
F
It
deserializes
as
part
of
the
if
I
read
it
correctly,
it
deserializes
it
has
like
a
unmarshalled
Json
and
if
that
unmarshall
Json
fails,
basically
they
have
time.
Parse
duration
fails
it's
going
to
throw
a
500
something
internal
error,
as
opposed
to
actually
returning
a
useful
like
this
is
not
a
valid
duration
string
and
that's
true,
if
you
feed
like
a
like
an
invalid,
any
string,
formatted
thing
that
is
invalid.
This
is
why
we
don't
in
apis.
F
We
don't
have
IP
addresses
received
into
a
net.ip
or
an
IP
what
net
Adder
field,
because
the
parsing
gets
done
at
on
marshalling
time,
as
opposed
to
at
validation
time.
So
instead
we
receive
it
as
a
string
and
then
we
later
parse
it,
which
is
also
how
we
miss
things.
So
what
I?
What
I
would
like
to
see
here
is
that
we
Define
a
kubernetes
type
that
we
can
stick
somewhere,
meta,
V1
or
something
like.
Maybe
we
replace
duration
as
far
as
I
can
tell
meta.
F
I
see
that
Chad
Antonio's
saying
probably
something
interesting
about
duration,
serialization.
Of
course
there
was.
C
Because
Jordan,
so
somebody
related
to
this
I
forget
the
details,
but
what's
related
to
that,
they
couldn't
round
trip.
Okay,
I
I,
don't
remember.
F
Okay,
so
shocking,
most
of
the
config
apis
are
not
GA
apis.
So
if
you
need
to
go
make
some
minor
changes
to
the
spec,
that's
probably
doable
but
I'd
like
to
see
us
Define
it
as
a
type
which
has
a
allows
garbage
strings
in,
but
has
a
well-known
validation
function
for
handwritten
validation
and
which
has
a
cell
validation.
That
does
the
right
thing.
Basically,
what
I
don't
want
to
see
is
everybody
who
embeds?
One
of
these
types
has
to
run
this
regex
validation
themselves.
B
B
I
would
agree
with
that
too.
Rob
go
ahead
since
you're,
playing
with
your
mute
button.
D
Yeah
I
just
trying
to
understand,
if
we
add
this
as
a
format
on
meta
V1
as
a
kubernetes,
lib
I,
think
that
means
we
need
to
wait
for
129,
then
to
be.
B
F
And
yeah
you're
gonna
need
a
consult
from
the
cell
folks
because
I
don't
know
enough
about
how
to
make
a
type
which
has
automatic
cell
validation
right
like
in
the
fullness
of
time.
In
the
future,
everything
will
get
first
pass
validated
by
built-in
validation
rules,
so
we'll
be
able
to
say
you
have
a
cube,
duration
string
and
that
will
be
automatically
validated
and
people
don't
need
to
manually
run
that
regex
yeah.
D
So
Joe
Joe
has
already
been
helpful
on
this
specific
PR
issue,
whatever
I
I
think
so.
So
the
good
thing
is
that
cell
already
supports
this,
for
the
sake
of
validation
in
the
sense
of
like
min
max
values,
in
that
their
duration,
parsing
would
work
for
this.
It
would
work
for
a
lot.
You
know
a
broader,
a
superset
of
this
format,
but
it
also
works
for
this
format,
so
you
can
do
min
max
validation
with
cell
already
using
this
format.
B
I
think
we
need
to
well
I
personally
need
to
make
certain
of
exactly
what
happens
when
the
cell
parsing
fails.
I
yeah
I
think
it
does
something
reasonable,
but
we
need
to
double
check
that
yeah.
D
So
I
think
we
have
pretty
clear
AIS
for
this.
It
is
remarkably
more
complicated
than
you
would
ever
think
going
into
this,
but
yeah.
Thank
you
for
the
discussion
and.
B
B
Do
we
also
have
broad
consensus
that
replacing
the
metav1.
duration
type
is
the
right
way
to
go
as
opposed
to
creating
a
new
type.
F
I
I
think
I
would
approach
this
as
create
a
new
type,
prove
that
it
does
all
the
things
that
we
want
it
to
do
and
then
say:
okay,
there's,
you
know
eight
places
where
we
use
meta,
V1
duration.
Can
we
switch
them
all
to
this?
And
if
so,
we
can
reappropriate
the
name.
I,
don't
know
enough
about
those
contexts
to
say
up
front
whether
it's
going
to
work
or
not.
Okay,.
B
I
was
kind
of
thinking
of
metaview,
1.,
timestamp
or
dot
dot,
timeout
or
dot
expiration,
or
something
like
that.
But
we
can
there's,
there's
definitely
a
certain
poetry
declaration
V2.
What
was
that
shape
across.
A
I'll
spoke
people
would
be
good
all
right,
we're
eeking
up
on
the
last
10
minutes
here.
Let's
give
Dave
a
chance
to
talk
about
Gateway
back
end
protocol
selection
gut.
M
Yeah,
this
is
just
a
sanity
check,
so
the
problem
is
give
API
doesn't
have
a
way
to
specify
back-end
protocols
or
hint
back-end
protocols
right
now,
so
one
of
the
options
is
because
kubernetes
has
a
standard
application
protocols
defined
in
the
kept
enhancement
right
now.
There's
HTC,
websocket
and
websocket
secure.
M
This
Gap
service
proposes
hey,
let's
use
these
constants
as
sort
of
the
hint,
so
this
would
then
mean
Gateway
implementation
should
look
at
the
service
app
protocol
to
kind
of
know
that
hey
you're
intending
to
for
it
to
be
this
back-end
protocol.
This
then
allows,
for
example,
hey
you
can
have
an
HP
route,
do
traffic
splitting
and
it
can
handle
like
Multiplex
over
like
like
HTTP
and
one
and
also
like
websockets,
because
there's
many
servers
that
do
that
type
of
stuff.
M
So
the
question
I
have
is
like.
Is
there
any
objection
to
that
and
then
from
this
group,
and
also
there
are
some
sort
of
like
open-ended
questions?
I
wouldn't
say
that
these
questions
preclude
the
Gap
from
being
like
accepted
or
merged.
One
of
them
is
sort
of
this
multiplexing
TCP
EDP
can
be
Multiplex
can
ICP.
J
M
It
a
CTP
and
also
like
more
complex
protocols.
I
know,
Dior
I
couldn't
find
your
issue,
but
there
is
this
concept
of
like
grpc
was
a
standard
app
protocol,
but
because
it
has
different
transports
like
how
do
you
present
that
and
likewise
like
you,
can
also
have
websocket
over
HTTP,
2
and
hp3
as
well?
So
like
right
now,
the
existing
protocols
in
the
standard
cap
reference
specific
rfcs,
so
those
are
HP
one.
L
Go
ahead
there
yeah
thanks
Dave
I,
wanted
to
raise
one
thing:
maybe
I'm
mixing
with
something
but
I
watched
the
go,
API
design
review
and
there
was
a
discussion
whether
we
want
change
gamma
in
the
future
for
cluster
apis
or
something
like
this
and
then
how
would
this
coexist
if
we
accept
so
it
just
came
up
now
if
we
accept
like
looking
at
a
protocol.
So
if
we're
like
changing
from
referencing
to
service,
maybe
it's
not
related
so
just,
but
this
is
something
I
just
thought
about.
M
D
Yeah
I
mean
it
seems
like
so
much
of
Gateway
API,
including
this
discussion
of
app
protocol,
where
we're
trying
to
Define
the
connection
between
the
Gateway
and
the
back
end
and
assumes
that
there's
some
kind
of
proxy
and
not
pass
through
where
currently
cluster
IP.
It
does
not
have
that
assumption
so
again,
I,
don't
know
what
that
means
for
cluster
IP
gateways.
We
haven't
really
gone.
D
I
so
far,
every
every
Gateway
implementation
I'm
familiar
with
there's
like
it's
like
a
proxy
implementation
right
where
there
there
is
a
unique
connection
from
client
to
Gateway
and
then
gateway
to
back
end
where,
like
service
cluster
IP,
whatever
would
be
so
far
just
passed
pass
through
right
is
that
a.
F
D
D
I
I
yes
I,
agree,
but
that
so
far
has
not
happened
with
Gateway
API
I
I
agree
that
that's
a
gap
and
kind
of
what
I'm
saying
is
at
protocol
itself
is
something
that
seems
to
rely
on
this
idea
that
there's
a
distinct
thing
between
client
and
Gateway
and
then
Gateway
and
back
end
and
app
protocol
is
defining
the
protocol
used
between
Gateway
and
back
end,
which
doesn't
really
seem
to
make
sense
in
a
pass-through
world.
L
About
what
would
be
correct
for
the
gamma
future
implementation,
where,
in
gamma,
you
probably
wouldn't
go
through
the
Gateway
or,
like
you
know,
So.
B
B
J
L
L
L
So
we
basically
reference
the
service
for
it.
But
if
we
gonna
move
the
Direction
with,
like
you
know,
not
referencing
the
service,
how?
How
would
we
derive?
This
are
protocol.
J
B
D
Yeah
I
mean
I,
think
I
think
it
can
only
possibly
be
the
latter
one,
because
we're
we're
seeing
more
and
more
that
there
are
examples
where
this
cannot
be
supported
and
I
think
that
cluster
IP
gateways
is
potentially
one
of
those
in
the
future.
I
get
talking
about
something,
that's
basically
undefined
at
this
point
and
just
trying
to
use
imagination
but
yeah.
B
Well,
a
use
case
that
we
definitely
see
with
existing
Ingress
stuff
is
a
client
out
somewhere
in
the
world,
makes
an
HTTP
2
connection
to
the
napi
Gateway
and
the
backend
service
only
speaks
HTTP
one,
and
so
the
the
Gateway
needs
to
do
that
translation.
So
I'm
curious,
what's
going
to
happen
with
the
the
gatewayless
Gateway
that
you're
describing
in
that
world,
would
you
simply
require
that
the
backend
service
must
be
able
to
speak
H2?
I
I
I
just
want
to
add
that
a
few
years
back,
we
had
this
HTTP
one
to
proxies
for
upgrading
and
downgrading
traffic
all
over
the
place
and
we're
about
to
go
into
another
phase
with
HTTP
three
to
two
on
one.
So
protocol
upgrades
and
downgrades:
that's
something
we're
gonna
have
to
live
with
right,
every
every
or
so
often
we
get
a
lot
of
it
and
then
it
dies
out,
but
it's
just
featured
it's
always
there.
B
C
But
there
is
one
thing
that
I
don't
understand
so
you're
kind
of
great
for
native
tp2
to
attp3
or
or
whatever,
but
you
cannot
agree
want
to
use
grpcs.
You
are
tied
to
http
2,
for
example,
so
there
are
things
that
you
can
upgrade,
but
there
are
others
that
you
cannot.
You
cannot
move
from
HTTP
to
http
3
to
grpc
for
example.
C
M
The
the
intent
of
this
and
also
there's
sort
of
a
table
and
I
get
that
says
like
if
you're
using
this
route
and
you're
using
this
protocol
like
this,
is
compatible
and
like
so
there's
sort
of
like
a
recommendation.
So,
for
example,
websocket
should
work
with
HTTP
route
as
a
front.
End
I
think
for
like
some
of
this
stuff,
like,
for
example,
like
grpc
and
websocket
over
like
H2
and
H3.
M
Those
to
me
are
sort
of
like
hey.
We
don't
need
it
at
the
moment,
especially
with
like
each
HTTP
3
being
over
UDP.
C
M
Like
that
changes
a
lot
of
like
people's
fundamentals,
so
for
now
like
what
this
enables
is
HTC
and
websocket
HP
one,
and
as
we
figure
out,
maybe
this
like
hp3
stuff
in
the
past,
then
we
can
extend
this
Gap
to
handle
more
and
potentially
advise
like
what
you're
describing
are
people
talking
about
like
supported,
upgrades
and
things
like
that.
M
It
doesn't
make
opinions
on
actually
I
guess
something
like
it
doesn't
make
opinions
on
upgrading
and
stuff
like
that,
but
there
are
sort
of
like
suggested
hints
on
like
this
is
clearly
going
to
be
incompatible
and
if,
for
example,
HTTP
3
wasn't
a
thing,
then
I
would
say
like
setting
your
kubernetes
service,
CDP
and
and
pointing
each
free
route
to
it
should
be
marked
and
compatible.
But
that's
clearly
not
the
case
anymore.
M
So
so
I
kind
of
you
view
the
discussion
about
like
protocols
and
upgrades
to
kind
of
be
like
separate
from
what
is
the
mechanic
to
hint.
What
the
backing
protocol
is,
if,
if
I'm
being
clear,.
A
We
are
unfortunately
running
out
of
time
or
in
our
last
minute
sounds
like
you
have
a
need
for
even
more
feedback,
Dave
and
discussion.
A
Everybody
who's
been
in
this
discussion.
If
you
would
go
jump
onto
the
pull
request
and
do
some
async
discussion
to
continue
this
here,
we
would
appreciate
it
so
that
we
can
move
forward
with
it
and
figure
out
what
the
right
thing
to
do
is
I
agree
with
Flynn's
statement
personally
that
having
the
ability
to
at
least
describe
it,
makes
sense
to
me.
A
And
also
Dave,
if
you
want
to
put
this
one
on
the
agenda
for
the
next
meeting
too,
so
we
can
continue
the
discussion.
That
would
be
good
too.
J
M
And
then
just
solely
decide
on.
Is
this
mechanic
useful
and
does
everyone
agree
that
the
standard
labels
in
this
kubernetes
cap
be
a
reasonable
thing
to
use
so
yeah
yeah
sounds
good,
so.