►
From YouTube: DASH Workgroup Community Meeting 20220302
Description
March 2, 2022 Community Meeting
A
We
we
completed
that
right,
yes
and
then
so
we
have
all
the
tools
in
place
and
that's
what
the
cyrus
pipeline
you
know
the
work
was
as
high
as
pipeline
tools
and
then
the
second
one.
You
said
that
you
know
you
have.
You
have
completed
one
api
generation
of
ones.
What
was
what
was
that
part?
It's
it's
api.
B
Implementation,
so
up
till
now,
what
we
have
published
is
api
only
like
what
was
what
will
be
actually
used
in
sonic.
However,
in
parallel
to
the
sonic
work
by
the
vendors
like
each
of
the
vendors
needs
to
provide
implementation
of
those
apis
right
and
in
parallel
to
that,
we
are
working
on
simulation.
So,
as
a
back
end,
we
are
using
bmv2
simulator
and
our
goal
is
to
generate
psi
api
implementation,
which
is
translation
from
psi
to
p4
runtime,
to
be
able
to
configure
the
simulator
and
run
it
in
sonic.
B
A
A
B
Yeah
like
enhancing
p4,
compiler
and
vm,
and
both
vm
v2,
to
support,
though
this
piece
of
code
as
well,
yeah,
okay,
so
that
this
is
in
the
to-do
list
in
the
current
tool
chain.
Pr,
you
can
take
a
look
at
that.
E
D
Going,
I
know
you're
familiar
with
it,
but
the
changes
that
we're
proposing
for
fragmentation
and
how
to
end
connections
and
are
you.
B
So
in
three
weeks
it's
only
the
match:
action
portion,
which
includes
routing
plus
acls,
the
changes
with
regards
to
fragmentations
and
arcs.
They
are
related
to
connection
tracking,
which
is
not
yet
supported
by
the
simulator.
B
So
this
is
another
work
that
we
will
be
doing
after
we
have
initial
version
of
all
of
the
layers
working,
then
we
will
expand
it
for
the
connection
tracking
as
well,
and
we
will
include
that
them.
We
have,
however,
the
like
the
reference
code,
which
is
not
currently
compiled
for
connection
tracking.
B
So
it's
just
let's
say
it's
just
there
for
documentation
purposes,
so
we
need
to
do
the
work
to
actually
compile
it
and
be
able
to
run
this.
D
So
somehow
we
need
to
get
a
full
schedule
for
what
we're
getting
and
when.
Because
it's
it's
not
much
use
until
we
have
the
full
implementation
and
it
sounds
like
you're
doing
a
partial
implementation
for
three
weeks
and
then
at
some
point
and
doing
more.
And
I'm
just
wondering
if
we
need
help
here
like
this,
is.
B
Yeah,
we
definitely
do,
and
this
is
what
we
were
asking
for,
because
I
cannot
parallelize
more
so
if
someone
can
help
with
the
outstanding
items
during
those
three
weeks,
that's
something
we
could
use,
but
that's
as
much
as
we
can
do.
We
first
are
focusing
on
delivering
at
least
something
working,
and
then
we
will
be
expanding.
B
So
there
there
are
two
basically
two
projects
that
are
involved
there.
First
one
is
the
compiler,
so
we
need
to
have
to
be
able
to
to
dig
into
the
p4
compiler
front
end,
and
second,
one
is
the
simulation
tool,
which
is
basically
a
software
switch.
G
I
I
do
not,
but
I
mean
I
know,
people
who
do
and
I
host
a
like
a
public
meeting
every
couple
of
weeks
for
p4.org
for
people
that
are
interested
in
learning
internals
of
those
tools
and
how
they're.
C
Anyone
else
on
the
call
have
this
kind
of
skill
set
who
might
be
able
to
help
out,
because
I
don't
know
everyone's
background.
I
apologize.
D
Yeah
well
joining
a
group
like
this.
We
assume
the
companies
are
big
enough,
that,
yes,
the
people
are
all
pay.
I
get
paid,
that's
a
that's
a
given,
but
you
know
it's
like
in
sonic
like
there's
lots
of
people
coding
and
they're
all
paid
by
their
companies.
D
Well,
I
will
clarify
it's
just
like
sonic.
If
there's
no
contribution
from
the
member,
then
certainly
there'd
be
no,
no,
no
way
for
us
to
buy
from
that
company
right.
You
have
to
contribute,
and
so
just
like
sonic,
we
don't
buy
a
sonic
box
from
any
company
that
didn't
heavily
contribute
to
the
coding
of
sonic
right.
D
That
was
like
a
number
one
thing
in
order
to
be
even
evaluated,
because
when
you
go
to
support
it,
if
you're
not
if
you're
not
familiar
with
the
coding
practices
and
what's
in
it,
how
could
you
ever
support
it?
So
we
should
keep
that
in
mind.
If
you're
wanting
to
do
business
with
microsoft,.
C
D
Christina
because
we
have
to
have
each
company
needs
to
have
some
contributions.
Yeah
yeah.
D
D
Of
that
fine,
but
if
you
don't
understand
the
people
or
pipe
model
pipeline,
you
never
contributed
or
involved
with
it.
Then
it's
highly
unlikely
that
you've
written
the
code
properly.
H
No,
I
I
got
you
no,
no,
so
you
were
talking
in
in
general
terms,
but
but
we're
really
talking
about
room
number,
seven
for
making
use
of
p4
pipelines
yeah
with
that.
I
mean
that
is
good.
Remember
this.
D
Dash
doesn't
define
an
implementation,
so
dash
can
only
define
apis
and
behavioral
models,
and
so
that's
what
this
whole
group
is
about.
So
I
understand
we're
not
asking
people
to
contribute
their
code
because
first,
they
wouldn't
contribute
to
code.
I
think
at
least
not
today,
so
the
code
that
we're
producing
is
the
behavioral
model
which
is
built
up.
This
people
are
behavioral
of
some
sort,
whether
it's
this
bm2
or
whether
it's
the
the
next
generation
of
that
yeah.
That's
what
this
organization
is
trying
to
do.
H
Right,
which
is
okay,
which
is
why
I
spoke
up
and
asked
because
this
is
talking
about
p4,
specifically
as
opposed
to
a
behavioral
model
that
I
could
go
write.
You
know
arm
core
c
code
to
implement.
If
I
really
wanted
a
software
version,
if
I
thought
there
was
some
better
reason
for
that,
I'm
not
sure
there
is,
but
in
principle
I
could-
and
I
could
say
I'm
obeying
dash
apis,
because
I
wrote
some
software
and
I
tapped
the
packet
to
that
software.
I
do
my
thing
to
them
and
then
send
them
on
their
way.
D
H
H
Okay,
so
that's
fair,
so
christine
christina
reach
out
directly
to
us.
This
is
I
I
get.
I
get
it
now
so
general
christina
reach
out
directly
to
dell
through
me
and
we'll
have
a
we
need
to
have
a
discussion
on
part
of
this,
but
got
you
I
understand,
I'm
that
that
answers
everything
I
needed
gerald.
Thank
you
very
much
yeah.
I.
D
Also
just
want
to
make
sure
there's
no,
no
misunderstanding:
the
actual
implementation
does
not
have
to
be
p4.
We're
only
talking
about
behavioral
models
which
do
need
to
be
p
part
based.
H
Got
it
that's
what
I
say
so
let
yeah
we
yeah
yeah.
Okay!
No,
that's
enough!
That's
enough
for
the
for
this
meeting,
I
get
the
request.
Thank
you
very
much
for
the
description.
I
now
understand
what
exactly
what
you're
asking
for
here
on
number
seven-
and
I
don't
know
if
maybe
it
helped
other
people
too.
So,
thank
you
very
much.
C
I
hope
it
helped
everyone
else.
Does
anyone
else
on
the
call
have
questions
about
this.
A
A
H
Right,
why
do
you?
Why
do
okay?
So
why
is
that
important
for
dash
to
define,
because
the
compiler
and
so
on
is
a
per
vendor
based
on
their
pipeline?
So
if
we're
defining
p4.
A
H
H
A
No,
no,
I
think
that
the
behavior
model
yeah
so
the
behavior
model.
I
guess
you
know
joseph
that
that
that
we
have
defined
you
know,
especially
marianne
and
nmedia,
is
they
they
deviate
from
the
original
specification
right
like
connection
tracking
this
fragmentation
work
or
any
other
things
that
we
have
defined.
A
D
So
in
the
end,
once
we
have
those
tools,
then
we
can
just
modify
at
will
the
people
or
behavioral
model
to
any
discussion
that
we
all
agree
on
like
this
week.
We
talked
about
how
to
close
a
connection.
So
at
that
point,
once
you
have
the
tools
we
can
make
those
changes
in
p4
and
automatically
generate
those
apis
and
software
behavioral
model.
Without
somebody
trying
to
hang
crap
that
or
it'll
it'll
come
straight
from
the
p4
code
itself.
D
So
the
goal
is
to
rapidly
once
once
the
first
version
is
done
and
somebody
notices
anything
wrong.
We're
going
to
just
change
the
code,
but
we
don't
have
to
worry
about
the
tools
anymore.
So
that's
why
it's
important
to
get
to
the
stage
where
we
have
those
tools
so
that
we're
solely
focused
on
the
behaviors
and
writing
the
the
model
that
represents.
H
Behaviors
so
yeah,
so
I'm
not
going
to
fully
commit,
but
we
might,
we
might
have
someone
in
one
of
our
one
of
our
teams
who
could
actually
do
that
kind
of
work,
but
I
have
to
talk
to
there's
a
bunch
of
discussion,
so
I
said
reach
out
reach
out
to
me
privately
mm-hmm
yeah
I'll
reach
out
now
as
well.
So
again,
some
of
this
stuff.
It's
like
I
got
it
gerald.
Thank
you.
Thank
you
again,
yeah!
Well,
I
I'm
getting
this
figured
out.
Okay,.
H
I
Hey
hey
christian
now
this
is
suresh
from
punjabi.
I
don't
have
people
experts,
but
I
think
I
have
a
triple
background.
So
let
me
I
want
to
understand
what
marion
said.
So
it
is
one
thing
to
know:
understand
the
q4
as
a
language
to
define
the
pipeline
line,
but
it's
another
thing
to
know
the
internals
of
the
people
compiler.
So
are
we
looking
at
you
know
modifications
the
compiler
or
modifications
to
the
pipeline
definition.
B
For
the
reason
that
current
standard
or
of
of
the
p4
language
does
not
cover
everything
we
want
to
express
in
the
pipeline.
J
We
need
all
kind
of
helping
right,
so
we
need
to
extend
the
compiler.
This
is
one
thing
to
do,
and
maybe,
since
it
is,
it
require
more
knowledge
and
we
can
handle
that
now.
We
need
to
extend
the
bmv
to
to
support
all
the
new
functionality
like
contract,
fragmentation
and
so
on.
This
is
an
this
is
a
purely
c
plus
plus
code.
So
I
will,
I
believe
we
can.
B
B
D
Right,
yeah,
I
think
that's
that's
what
we
need
that
way.
We
can
start
building
a
schedule
around
it
and
understand
the
scope
and
who
can
help
and
start
you
know,
assigning
people
to
the
different
paths
right
now.
It's
it's
all
in
melanox,
so
there's
no
description
because
they're
doing
it
all
themselves
which
won't
scale
long-term
and
it's
not
the
purpose
so
maddie.
Do
you
think
you
can
take
a
stab
at
what
needs
to
get
done,
so
we
can
start
really
probing
who
can
help
where.
J
C
D
Yeah,
I
think
in
the
past
you
know,
I
know
that
you
know
in
sonic,
once
you
had
all
the
tasks
you
could,
that
that
is
the
key,
because
then
you
can,
you
can
ask
for
volunteers
and
they'll
volunteer
against
a
task
and
there'll
be
names
against
it
and
dates
so
yeah
perfect.
C
Yeah,
okay,
great,
then
what
was
next
this
week
we
we've
talked
at
length
about
the
fragment
paper
and
we
have
one
more
meeting
to
close
off.
Oh
marion,
we
did
have
a
question
for
you
yesterday
and
I
I
don't
want
to
put
you
on
the
spot,
but
john
carney
had
had
a
question
where
he
looked
into
the
p4
code
and
he
saw
a
double
encapsulation
somewhere.
M
I
I
was
just
not
sure
whether
there
was
an
appliance
encapsulation
and
then
another
encapsulation
or
or
not.
I
had
thought
that
I
had
seen
that,
but
I
think
then
yesterday,
when
we
looked
through
the
code,
we
couldn't
really
find
it.
So
I'm
not
sure,
but
I
mean
I
guess
if
you
could
just
answer
the
question
like
do
you
know
of
any
case
where
there's
a
double
tunnel.
B
I
think
I
know
of
one
of
them,
but
it's
not
really
a
common
case.
This
is
what
I
reviewed
with
together
would
go
on
it's
when
the
packet
is
coming.
I
guess
from
another
region
it
will
go
through
an
slb
and
it
may
have
two
layers
of
encapsulation
first,
one
being
left
over
from
the
slb
and
second
one
is
your
v-net
vxlan.
N
D
D
C
C
Okay,
you're
welcome
and
did
the
before
we
move
on
to
h.a
did.
Did
the
sdn
team
want
to
talk
about
anything
today
or
should
we
move
on.
C
Great,
so
bud
did
propose
a
aha
protocol,
pr,
not
sure
if
anyone
had
had
a
chance
to
take
a
look.
B
This
is
basically
a
description
and
a
set
of
requirements
right.
The
the
recent
pull
request
about
it's
not
really
a
proposal
of
any
protocol
yet,
but
it's
its
description.
O
Sure
so
I
started
thinking
about
the
aha
requirements
and
interactions
between
the
two
paired
devices
and
started
thinking
about
what
the
protocol
looked
like
and
the
packet
formats
would
look
like,
and
so
based
on
that,
I
mean
something
became
apparent.
Is
that
because
the
communication
path
between
the
two
devices
is
not
100
reliable?
O
O
But
if
the
communication
path
between
the
two
is
not
100
reliable,
then
there's
going
to
be
cases
where
a
remove
message
is
sent
that
from
your
peer
that
you
don't
receive,
and
so
that
will
create
like
zombie
flow
entries
and
eventually
the
flow
table
will
fill
up
with
those.
So
that's
that's
the
end.
E
O
Well,
the
the
the
dpu
that's
receiving
the
the
packets
is
constantly
updating
its
timer
right
whenever
a
new
packet
arrives
that
that
constant
updating
isn't
happening
on
the
pier.
O
So
so
say
say:
the
timeout
was
like
one
minute
or
something
like
that
for
the
fur
flow,
the
the
active
dpu,
that's
getting
that
flow
say
he's
getting
like
a
packet
every
second,
I.
O
J
O
K
Yeah
sure
yeah,
I
like
I
like
the
document
quite
a
bit.
I
think
it
made
it
so
easy
to
understand.
I
had
a
hard
time
understanding
some
of
the
earlier
discussions,
but
it
was
really
clear
and
it
seems
real,
pragmatic
yeah.
I
haven't
read
this:
I'm
I'm
anxious
to
read
it.
K
I
added
some
comments
with
regard
to
looking
how
we're
going
to
go
testing
it
because
something
that's
complicated
and
dynamic
in
real
time.
If
you
want
to
build
the
testing
capability,
I
don't
mean
conformance
testing,
even
just
observability
of
it
right
to
know
if
it's
really
working
right,
rather
than
it
being
a
mysterious
black
box
and
looking
at
the
traffic
generators
for
the
final
result,
you
really
want
to
know.
K
K
K
D
Maybe
at
some
point
let
them
they're
talking
about
implementation
here,
and
so
it's
going.
You
know,
ours
was
guidance.
Let's
see
how
this
goes.
Maybe
we
can
do
some
combination
in
the
future,
but
really
anxious
to
see
what
the
protocol
designers
are
going
to
come.
C
F
Yeah,
I'm
curious
to
see
how
the
zombie
flows
our
handle.
Is
it
going
to
be
event
based,
you
know,
do
we
need
refresh
and
yeah
quite
a
bit
of
thoughts
that
we
have
as
well
and
just
to
see
this
document
and
we'll
post
some
comments?
There.
D
Maybe
we
can
take
the
action
that
the
next
time
we
meet
that
everybody
has
read
this
document
and
and
it's
prepared
with
questions,
I'm
not
prepared
with
any
questions,
because
I
haven't
read
it,
but
certainly
will
now.
I
know
that
it's
there.
K
Like
give
me
a
count
of
how
many
active
flows
there
are
or
tell
me
the
state
of
your
state
machine
or
something
like
that,
I
don't
know
what
the
the
things
are
that
we
want
to
manage,
but
it
brings
up
that
question
to
me.
How
do
you
talk
to
this
controller?
How
do
you
observe
it?
We
want
telemetry
coming
out
of
it,
that's
configurable
to
turn
on
or
off
high
fidelity
spectrometry.
K
C
C
C
If
there's
another
way,
you'd
like
me
to
communicate,
you
know,
please
let
me
know
but
yeah
it
should
be
in
your
email
somewhere
or
you
can
come
to
the
repo
yeah.
O
Okay,
yeah,
in
fact
I
mean
we.
We
have
a
bunch
of
thoughts
with
a
lot
more
detail
than
this,
but
we
kind
of
wanted
to
propose
this
as
something
to
put
out
there
to
discuss,
because
there's
no
point
in
going
to
more
detail
which
will
have
its
own,
I'm
sure
set
of
comments
and
and
discussions
before
to
I
mean
before
getting
the
the
general
picture
agreed
upon.
O
K
F
Yes,
I
didn't
receive
the
email,
so
this
will
be
helpful.
Great.
K
A
So
one
of
the
things
in
response
to
chris's
question
on
the
api
part
for
you
know
just
in
response
to
to
this
protocol
discussion.
A
One
thing
we,
we
should
probably
understand
that
you
know
this
kind
of
elaborate
protocol
and
all
those
you
know
the
communication
that
we
are
talking
about.
J
E
D
D
Flow
state
will
be
sent
from
there.
It's
not
going
to
go
through.
You
know
every
flow
you
get
you're
not
going
to
go,
send
that
up
to
software
in
the
operating
system.
We
need
to
find
a
different
name
for
that,
because
lots
of
people,
their
hardware-
is
really
software,
but
you
don't
send
it
somewhere.
It's
automatically.
These
updates
have
to
be
sent
out
in
conjunction
with
the
data
path.
D
D
A
A
However,
you
know
the
software,
which
is
very
close
to
the
hardware,
perhaps
running
in
some
sort
of
an
embedded
core,
is
really
scanning
all
those
flows
that
are,
you
are
creating
those
those
you
know,
those
automatic
flows
that
are
created
for
the
fast
path
you
you
are
scanning
them
and
then
really
sending
them
so
keeping
all.
D
By
combining
these
record,
these
connection,
states
and
setting
them
up
the
other
side.
P
In
general,
the
first
creation
can
be
the
way
you
suggest
you
got
any
subsequent
refresh.
D
D
So
I
agree
so
be
some
combination,
but
what
I'm
saying
is
it
couldn't
be
that
it
was
all
done?
You
know
software,
just
you
just
never
keep
up.
A
D
D
Yeah,
but
you
know
what
a
lot
of
the
details,
even
though
you
know
people
made
that
argument
with
the
hey.
It
doesn't
matter
whether
we
just
process
things
fast
path,
only
or
slow
path,
only
or
combination.
And
then,
when
we
came
to
fragmentation,
it
made
a
huge
difference.
D
I
think
it's
it's
it's
better,
that
we
at
least
describe
it
in
a
in
a
manner
how
you
just
how
you
implement
it
can
be
different,
but
I
think
that
it's
pretty
obvious
here
that
there
is
a
fast
and
a
slow
path
to
this
or
a
thousand
control
channel
to
this,
and
I
think
it's
worth
outlining.
Otherwise
we
won't
have
much
to
debate
on
how
the
corner
cases
actually
exist.
D
So
behavioral
model,
main
goal
of
the
behavioral
models-
is
to
actually
allow
us
to
identify
the
corner
cases
so
that
we
can
write
tests
and
make
sure
that
they're
all
covered
up.
So
the
the
you
know
we
need
to
make
some
stand
up
like
what
that
should
look
like,
and
then
I
don't
care
how
you
implement
it
in
the
end.
But
the
behavioral
model
should
be
something
that
conceivably
could
be
implemented,
not
something
that
is
just
completely
abstract,
that
we
never
find
any
corner
cases
or
for
the
for
the
test.
D
Cases
which
we
so
dearly
need
and
aj
is
going
to
be
super
complex
to
test.
So
we're
going
to
need
to
be.
We
will
need
a
behavior
model
for
it
and
we'll
need
to
describe
that,
like
I
say
how
you
implement
implemented,
okay
sure
go
off
and
do
what
you
want,
but
the
reality
is
the
corner
cases.
The
behavioral
model
should
be
good
enough
that
you
can
identify
all
the
corner
cases.
P
Good
goal
general
on
that
thought.
I
have
a
question
about
the
interrupt
requirement
between
vendors,
because
you
know,
as
you
said,
testing
is
hard
and
even
if
the
protocol
is
public,
even
if
the
format
is
public
and
the
behavior
model
is
something
that's
standard,
there
will
be
cases
that
don't
show
up
under
performance
duress
and,
for
example,
one
implementation
might
be
perfectly
okay.
Sending
millions
of
updates
to
another
implementation
which
cannot
handle
that
and
so
there'll
be
other
differences,
like
capacities
will
be
different.
So.
D
I
think
that
I
I
agree
with
everybody
who's
saying
it's
really
like
when
you
mix
it
things
get
harder,
and
I
think
that
our
first
goal
is
to
have
one
model
that
that
specifies
the
behavior
and
originally
we're
going
to
pair
like
for
like
anyways.
D
But
what
I
don't
want
is
never
having
that
ability
and
more.
What
I
don't
want
is
having
no
ability
to
test
it
so
like
I
want
the
way
to
test
it
for
everyone
and
whether
yours
is
faster
than
somebody
else's
makes
a
little
bit
harder.
I
think
that's
fair,
that
that's,
that
might
be
a
little
bit
longer
a
goal
to
mix
it,
but
we
have
to
start
with
the
goal
of
being
able
to
mix
it
initially,
we'll
find
all
the
corner.
D
Queso
write
the
test
plans
for
it
and
we
would
test
you
know
like
for
like
in
the
beginning,
until
we
get
enough
experience
and
then,
after
that,
we
could
test
mixes
and
find
out
what
you
know.
The
potential
issues
are,
and
it's
typically
performance
right
like
there's
a
lot
of
there
could
be
a
a
big
performance
difference
between
one
vendor
and
another.
But
the
critical
point
is:
I
need
to
have
only
one
set
of
test
cases
that
everybody
goes
in.
Your
performance
is
going
to
be
different.
D
We
understand
that
and-
and
initially
you
know,
I
think
it's
quite
reasonable-
that
everybody
would
deploy
aj
pairs.
But,
as
we've
talked
in
the
past,
the
problem
comes,
you
know
later
when
we
go
and
add
a
new
more
efficient.
D
You
know
appliances.
D
Now
it's
no
longer
they're
saying
if
it's
from
the
same
company,
it
could
have
the
exact
same
problems
that
we're
talking
about,
which
is
oh,
it's
so
much
faster
and
can
update
so
much
faster
than
the
original
appliance,
and
you
could
have
issues
there
that
we're
gonna
have
to
address.
P
For
it,
you
know,
the
only
comment
that
I
would
make
is
that
the
worst
time
to
find
out
that
you
have
an
interrupt
problem
is
when
you
actually
require
the
in
a
operational
network
right.
E
P
When
you
have
all
of
these
permutations,
you
created
that
matrix
of
what
needs
to
be
tested
against
what
and
that
you
know
that,
may
be
difficult
to
do
prior
to
your
deployment.
You
will
actually
find
out
these
corner
cases
when
you
are
deploying
it
and
show
up
in
the
behavior
model.
D
You
know
that
that
is
one
difference
between
this
group
and
other
groups
is
we're
trying
to
create
behavioral
models
so
that
the
suppliers
actually
know
the
exact
behavior
that
they're
supposed
to
deploy
right.
It's
not
like
we're
just
waiting
until
you
finally
give
us
something
you
should
have
the
behavioral
model
and
your
hardware
model
should
match
the
behavior
of
the
behavioral
model
and
to
avoid
you
know
a
lot
of
what
in
the
past
would
have
been
a
lot
of
misunderstandings.
D
K
We're
already
thinking
about
system,
testing
of
h.a,
and
you
know
cooking
up
ideas,
and
you
know
we
hope
to
be
ahead
of
the
the
game
there
in
terms
of
having
some
ideas
for
everyone
to
look
at
before
we
can
get
to
that
point,
and
another
comment
is
good
old,
plug
fests
right.
That's
how
the
industry
resolves
interrupt
problems
from
a
pragmatic
point
of
view.
Everyone
does
their
best
job
of
implementing
what
they
think
the
standard
is.
K
C
I
can
say
from
experience
too,
you
know
I've
gone
from
lab
to
the
staging
environment,
to
pre-production
pilots
out
to
production,
and
I've
found
things
in
production
that
I
sh
that
I
thought
for
sure
I
would
catch
in
the
prior
three
that
because
of
the
scale
I
I
just
didn't,
I
just
couldn't
you
know,
and
so
hopefully
we
can
catch.
You
know
85
percent
of
issues.
D
Yeah,
I
think
everybody
knows
that
aj
is
hard
guys
like
there's,
no
doubt
about
it.
This
is
a
hard
topic.
It's
going
to
be
the
toughest
part,
we'll
get
that
data
path
going
and
you
know
resolved,
and
I
think
we're
coming
to
you
know
to
an
ending
on
at
least
describing
what
should
be
in
the
in
the
data
path
for
for
the
v-net
aj
will
be
much
harder.
D
I
have
a
lot
of
bright
people
work
for
me,
so
yeah
acknowledge
this
is
this
is
tough,
but
that's
if
it
wasn't
tough,
it
wouldn't
be
interesting
anyway,.
C
Right
and
I
guess
the
next
topic
is,
I
have
a
draft
in
almost
actually
it's
mostly
complete
where
we're
describing
different
customer
scenarios,
and
you
know
whether
whether
we're
doing
an
explicit
lpm
or
inserting
a
firewall
between
v-nets
or,
if
we're
you
know,
doing
default
internet
traffic
one
way
but
trusting
other
traffic
another
way
and
or
if
we're
doing
private
link
and
doing
mappings
or
routes.
And
things
like
that
and
michael
miele.
And
I
are
writing
that
up
and
maybe
we'll
be
able
to
present
that
at
some
point.
D
Yeah,
that's
a
very
important
document
and
some
people
were
on
a
call
yesterday
where
we
talked
about
it
and
she's
doing
the
inbound
stuff
now
and
has
to
do
no
outbound
stuff
now
and
and
they
need
to
add
the
inbound
stuff
but
yeah.
This
is
gonna,
be
very
important
and
that
mario's
still
on
the
line
like
this
is
not
stuff
that
you
had
access
to
before.
D
Mario,
because
we're
just
writing
it
now,
based
on
feedback
of
things
that
we're
missing,
and
this
is
going
to
definitely
change
the
p4
behavior
model,
because,
right
now
the
people
are
behavioral
models
too
simplistic.
On
the
on
the
lookup.
C
C
Creating
the
v-nets
and
the
vms
and
the
the
rules
and
what
we
want
our
environment
to
look
like
and
that's
what
we're
actually
doing
is
you
know,
put
yourself
in
the
customer's
shoes,
and
you
know
how
you,
as
a
customer,
want
to.
You
know,
configure
your
environment
and
route,
your
traffic
and
that's
what
we're
doing,
and
I
think
we
were
missing
that
piece.
A
C
To
review
it
with
michael
sigmund
to
make
sure
it's
right
or
get
some
feedback
and
then
on.
A
A
Yeah,
the
reason
I
was
asking
is
because,
if
at
least
you
know
these,
the
current
people
model
that
I
see
it,
they
are
only
going
as
far
as
basically,
the
first
use
case
is
concerned.
Right.
D
Now,
but
I
think
the
veena
tabinet
even
still
needs
to
do
this
this
this
is
this
is
required
even
in
the
basement,
the
v-net
model-
it's
not
you
you'll,
have
to
implement
it.
Within
that
model,
we
couldn't
deploy
a
peanut,
for
example,
without
this
boarding
table.
Without
these
capabilities.
A
D
That's
what
I'm
saying,
though
this
is
the
boarding
table.
This
is
the
common
use
case.
This
is
the
base
use
case
of
a
v-net,
and
so
you
know
v-net
sometimes
has
to
you
know
that
packet
has
to
go
through
a
filter
before
it
reaches
the
other
side
or
it
has
to
go
through
a
low
bouncer
before
it
reaches
the
other
side
or
it
has
to
go
to
the
internet.
You
know-
or
it
it's
an
exception
to
going
out
the
internet.
D
So
it
is
the
base
v-net
case
and
as
soon
as
christina
puts
it
out,
you'll
see
that
we
need
this
in
the
base
case,
there's
nothing.
We
can
do
in
the
base
case
without
the
boarding
table
having
these
capabilities.
C
A
Out
right
so
that
defines
another
work
stream
for
the
community
to
say
that.
Okay,
you
know
this
is
something
that
we
need
to
achieve.
We
need
to
achieve
a
base
model
to
cover
this
thing
and
we
need
to
define
the
timeline
and
then
we
need
to
say
okay,
now
this
we
have
achieved
this.
Essentially
this
behavior
model,
which
is
we
are
saying
that
it's
it's
now
complete.
A
C
C
Thanks
but
anything
else,
we
we
have
five
minutes
left
and
I
do
have
10
o'clock
today.
O
Hey,
I
have
one
question,
I
don't
know
if
this
is
pretty
simple
question
I
don't
know
if
it's
been
addressed
before,
do
we
have
an
indication
or
knowledge
of
what
the
maximum
aging
timeout
is
for
like
for,
for,
like
say,
tcp
flows.
D
We
could
take
a
medicine
action
they're
quite
long,
but
we
could
take
it
as
an
action.
I
guess
we
could
look
at
vip.
N
And
see
well
yeah,
and
this
is
also
configurable
on
the
slb,
so
for
some
flows
like
for
for
a
sequel,
slb
allows
to
configure
this
even
like
30
minutes.
I
think.
D
H
Be
anything
so
so
bud?
Could
you
explain
more
of
what
you're
thinking
about
time
out
of
tcp
flows?
I
mean
tcp
has
sort
of
natural
timeouts
in
the
in
the
specification
and
then
that
can
be
altered
all
over
the
place.
You
can
do
all
kind
of
stuff,
with
particular
implementation.
So
what
are
you
thinking
and
why
is
that
important.
N
I
would
say,
like
the
majority
like
90,
five
or
eight
percent
of
the
traffic
using
default
ones
right,
but
there
are
some
scenarios
like,
for
example,
sql
has
a
scenario
that
the
connection
to
sql
can
just
last
for
very,
very
long
and
be
idle,
and
that's
why
slb
had
to
introduce
configuration
of
timeout
on
their
flow
on
maxes
to
to
to
allow
us
to
support
us.
So
from
our
data
plane,
we
are
not
x,
expiring.
N
N
D
Well,
let's
put
this
way:
we
we
can
share
that
information
with
you.
We
found
out
the
hard
way
you
instead
of
you,
finding
out
the
hard
way
we
can
share.
But
this
is
you
know.
Definitely
you
know
a
sql
requirement.
So
that's
you
know
we
get
that
kind
of
experience
by
offering
services
right
and
it's
like
being
a
cisco
compliant
with
cisco
on
a
router.
D
O
But
that's
hey
joseph!
This
is
bud,
so
the
reason
I
was
asking
it
is
is
basically,
as
we
scope
out
internal
data
structures
like
what
kind
of
resolution
what
kind
of
bit
size
do.
We
need
to
maintain
timer
values
and
things
like
that.
C
Okay,
well
got
it.
I
have
to
run.
Thank
you
for
your
time
and
I
will
come
up
with
some
like
I
said
I'll
work
with
maureen
and
maddie
to
come
up
with
some
timelines
I'll
reach
out
to
different
companies
for
resources,
specifically
joseph
of
course
I'll
always
do
the
notes
and
you
know
feel
free
to
reach
out
to
me,
email
teams.
You
know
whatever
I
could
do
to
be
helpful
and
I'll
see
you
next
week.