►
From YouTube: IETF99-RTGWG-20170717-0930
Description
RTGWG meeting session at IETF99
2017/07/17 0930
https://datatracker.ietf.org/meeting/99/proceedings/
D
F
F
F
So
our
extended
and
approved
charter
to
look
into
topic
that
might
be
outside
of
any
other
working
group.
We've
got
one
year.
She
published
thanks
AC
and
go
out
for
great
work.
They've
rewritten,
the
home
at
least
ten
times
just
before
publication
pass
also
brought
in
the
way
we
recommend
to
do
the
device
data
store,
so
great
work,
and
probably
very
good
example
how
young
model
and
ITF
government
describes
model
should
be
done.
F
F
G
G
B
F
H
Okay,
this
is
the
this.
This
is
a
really
a
simple
game
document
and
it's
an
attempt
to
call
it
to
put
the
common
types
into
for
routing
into
a
single
yang
model,
so
this
model
doesn't
have
anything
but
types
in
it
and
group
in
a
couple
groupings-
and
it's
really
a
corollary
to
what
we
have
today
for
ietf
yang
types.
This
is
IETF
routing,
yang
types
and
I
Ana,
routing
game
types,
Oh.
H
See
ya
there
we
go
so
since
actually,
actually
the
end
actor
review
is
before
the
ietf
in
Chicago,
but
we've
had
yang
doctor
review
we
and
Chicago.
We
had
some
talk.
We
had
some
discussions
about
whether
or
not
the
BGP
Saffy's
were
common
across
a
number
of
models
or
they
were
specific
to
BGP,
but
doesn't
hurt
to
put
them
in
the
common
routing
type
since
they're
well
known
and
they're
in
the
Ayane
directory,
I
mean
they're
in
and
a
registry.
We
also
got
a
comment
from
Rob.
H
He
wanted
to
new
types
that
were
in
the
open,
config
types
and
were
actually
used
by
the
current
BGP
model.
Silly
put
those
into
the
ATF
routing
types
as
well
nice
thing
about
yang,
since
you
have
the
prefix
there's
no
problem
with
duplication,
there's
no
ambiguity,
so
you
can
migrate
types
from
one
model
to
another
seamlessly.
H
H
We
did
not
take
the
security
considerations,
the
reason
you
know
the
boilerplate
security
considerations.
The
reason
we
did
not
is
these
are
only
types.
So
it's
not
really
a
yeah.
It's
not
really
a
yang
model.
You
wouldn't
use
rest
comp
or
anything
on
this
on
this
model,
and
then
we
split
the
routing
types
in
Dianna
routing
types
we
put
specifically
the
address
family
and
this
BGP
Safi
into
a
separate
module
might
so
in
the
same
model,
but
different
model
that
module
this
is
the
I
eat,
I,
Ana,
routing
types
model,
not
enough
coffee,
yet.
I
J
I
H
So
I
already
talked
about
this
along
the
lines.
Right
now
we
already
have
an
RFC
standard,
I
forget
which
it's
it's
our
as
RFC
72.
It's
not
72
23,
because
that's
the
one
for
the
interface
model,
I,
think
it's
72
24
and
that
that's
an
I
on
a
red-eye,
Ana
interface
types
that
includes
all
the
interface
types
I
Anna.
Along
those
same
lines,
we
created
a
new
model,
the
Ayana
routing
types
model
that
has
all
the
registry
I
mean
it
has
every
address
family
that
I
Ana
has
and
every
BGP
Safi
that
I
Anna
has.
H
H
And
then
we
have
actually
yeah.
This
was
okay
yeah.
So
you
see
this
was
actually
the
BGP
Saffy
snippet
that
I
showed
and
there's
about
I
think
there's
about
25,
Saffy's
or
so
that
are
defined
by
a
Ana.
Now
the
idea
is
that
the
I
Ana
would
actually
own
this
model
and
update
it
in
practice.
I,
don't
know
if
that's
worked
yet
I
know
we're
working.
If
you
go
to
the
net
mod
working
group,
there's
all
sorts
of
discussions
about
how
we
read
these
models
quicker
than
the
RFC
process.
H
H
H
Write
these
patterns
and
then
have
to
go
all
the
way
to
some
kind
of
yang
tools
where
you're,
where
you're
actually
got
a
harness,
actually
have
a
client
in
server
unit
test.
You
can
go
right
out
to
a
webpage
and
add
your
patterns
and
then
test
various
strings
against
it.
I
haven't
tested
myself.
He
he
sent
me
an
email
heart
to
do
it
I'm
gonna
plan
to
do
that
this
week
when
I
get
some
time,
but
this
is
the
route
target
one.
Oh
there's
Benoit.
C
Good
morning
so
I'm
going
right,
so
there
are
two
different
codes
to
test
the
rich
Ricci,
reg
X.
So
I
would
appreciate
some
testing
on
both
of
them.
We
put
two
outputs
to
be
able
to
compare
the
two
tools,
so
one
is
based
on
yang
Lin.
You
know
one
of
company,
so
testing
is
appreciated.
We
also
add
it
in
there
like
the
invert
match
in
case
you
need
it,
you
know,
could
you
please
send
email.
H
H
H
Actually,
she
finally
did
the
original,
but
I,
don't
know
how
I
didn't
add
this,
because
I've
made
some
modifications
to
the
to
the
to
the
route
target
and
ipv6
rod
target
from
evpn
there's
one
there's
the
es
import
rod
target.
We
missed
that
format
and
that
adds
the
type
six
route
target.
Now.
What
we
could
do
is
we
could
either
add
like
we
did
for
the
route,
our
origin.
We
could
add
this
generic
type
of
you
yang.
We
could
cover
anything
or
we
could
add
the
type
six
route
target.
H
H
B
F
D
B
I
H
They
they
I,
don't
want
them
to
be
myself,
I
mean
yeah,
you,
you
think.
I
We
should
put
them
in
aya
so
that
the
reason
they
actually
have
a
Diana
maintained
is
that
no,
for
example,
the
route
target
type
know
if
you're
going
to
have
that
be
extensible.
If
you
don't
make
a
Diana
maintain.
That
means
that
the
working
group
has
to
pick
up
a
new
working
group
document
every
single
time,
verses,
I
Anna
maintain
it's
a
lot
more
flexible
to
have
a
quickly
revved.
That's.
H
I
H
I
B
K
L
So
we
have
three
drafts
that
have
come
out
of
the
design
team
that
are
now
standards
track.
One
of
them
you
just
heard
about
two
of
them:
I'm
gonna
talk
about
in
a
moment
and
those
are
the
L&E
not
LNI.
By
the
way
LM
e
+
ni
drafts.
We
have
a
couple
of
other
drafts
that
are
headed
for
net
mod.
The
first
one
is
the
module
tags.
This
is
sort
of
hashtags
for
yang
models.
That
was
not
much
has
happened
between
the
last
ITF
and
this
one
we
had
asked
for
feedback.
L
So
the
ABS
have
issued
guidelines
on
this.
The
two
80s
who
issued
that
guidelines
are
actually
in
the
room
if
I
get
it
wrong.
They'll
correct
me:
first
thing
is:
how
are
we
going
to
deal
with
the
whole
op
state
discussion?
We
have
in
net
mod
the
network
management
data
store
architecture
document,
which
is
the
revised
data
store
for
representation
of
operational
State,
that's
happening
in
net
mod.
The
protocol
supports
happening
in
net
comp.
The
direction
is
all
models
being
developed
at
this
time
should
conform
with
the
guidelines
in
the
structure
of
an
MDA.
L
L
To
access
the
that
state
information
we'll
have
to
use
the
new
nmda
approach,
those
an
MDA
approach,
while
the
architecture
is
defined,
the
protocol
access
has
not
yet
been
defined
in
that
comp
Andres
comp.
So
we
have
a
gap
in
capability
if
we
do
follow
one
for
those
protocols
or
for
those
models
that
need
access
to
system
created,
state
or
operational
state
need
to
differentiate
between
operational
and
configured
state.
There's
a
transition
approach,
and
that's
what's
covered
in
number
two-
is
that
it
is
okay
to
have
for
a
well-structured
nmda
model.
L
Rob
Wilton,
who
I
think
he's
going
to
talk
about
talking
a
little
bit
well,
I
think
you'll
mention
that
there
for
our
nicely
structured
and
MDA
module
that
there's
some
tooling
out
there
to
take
you
from
a
nmda
compliant
structure
to
generating
this
appendix
you
can
use
that
tooling
or
you
can
do
it
by
hand
whatever
works
for
you,
just
publish
it
together.
The
base
model,
which
will
be
good
for
the
long
term,
together
with
the
transitional
support,
that's
necessary.
L
We're
gonna,
hear
quite
a
bit
about
that
I
think,
or
at
least
a
little
bit
about
it.
We're
also
going
to
if,
for
those
who
are
really
interested
in
this
topic,
come
to
net
mod
and
there's
gonna
be
even
more
discussion
and
there's
a
Q&A
session
there.
It's
basically
unstructured.
If
you
have
questions
on
how
this
all
works.
Go
there
ask
your
questions.
L
So,
in
terms
of
short
term,
if
you're
developing
a
model
in
the
routing
area,
so
this
means
is
folks
in
this
room.
You
should
restructure
your
model
to
be
an
MDA
compatible
and
the
reason
I'm
intentionally
using
the
remodel
is
you
may
end
up
with
multiple
modules,
as
so,
you
may
have
today.
One
base
module
that
you'll
have
to
break
into
two.
L
There
are
we
have
one
existing
RFC
that
really
impacts
the
area
and
that's
the
routing
config
that
RFC
will
need
to
be
changed.
It's
not
a
big
change.
I
think
we've
gone
through
when
I
say
we
collectively.
As
a
group
I've
gone
through,
it
figured
out
what
those
changes
are.
We're
expecting
a
see
as
well
as
potentially
someone
from
the
nmda
design
team
to
issue
an
update
on
that,
hopefully
not
too
long.
L
Sorry,
user
error,
so
moving
on
I'm
now
gonna
give
you
updates
on
the
two
documents
that
are
used
to
represent
logical
separation
of
a
physical
device.
So
we
have
the
logical
network
element
model.
This
is
to
give
virtual
systems.
Think
of
this
as
VMs
for
four
routers.
That's
this
one
we're
gonna
cut.
The
network
instance
is
all
about
births
and
be
a
size.
That's
the
next
one
Mike!
Please.
L
J
J
M
L
M
N
J
L
From
an
IETF
perspective,
I
would
expect
some
discussion
in
IDR
to
talk
about
what
the
plan
is
to
take
it
forward
in
a
way
that
is
functional
and
I'm,
not
talking
anything
about
the
nmda
split.
That's
that's
up
to
the
grouping
group.
That's
not
for
us
to
say,
there's
the
guidelines
out.
There,
you've
heard
it
from
the
ad.
She
says
you
know,
that's
you
take
it
on
a
case-by-case
basis.
I
think
IDR
should
discuss
it
from
a
usability
standpoint
from
the
design
team.
B
B
B
Process
has
created
a
lot
of
uncertainty
over
where
to
go
forward,
so
the
fact
that
a
group
hasn't
updated
their
model
in
the
last
few
months
to
to
keep
up
to
date.
I
don't
think
says
this
is
a
bad
model,
but
let
them
decide
when
they
decide
how
to
go
forward
on
the
nmda
stuff.
They
will
update
it,
so
it
works
and
I
think
we
can
move
on.
If
we.
L
Yeah,
so
the
conversation
in
IDR
sounds
right
as
the
this
is
an
ad
statement.
The
ad
is
stood
up
and
said
that
things
can
be
taken
on
a
case-by-case
basis,
so
I,
don't
think,
there's
anything
that
should
be
blocking
at
this
point.
We
do
have
a
transition
period
to
deal
with,
and
that
means
we
can't
do
the
ultimate
desired
solution
immediately.
So
we're
not
in
a
perfect
state.
That
is
a
true
statement.
L
Schema
mount
gives
you
that
capability.
There
was
a
slew
of
open
issues.
The
core
open
issues
have
been
resolved.
There
are
a
couple
of
small
things
out
there.
I'll
mention
one
of
them
in
a
moment.
Our
feeling
is,
is
it's
not
critical
to
resolve
those
right
now
and
we'd
prefer
to
get
the
capability
out
there
and
scheme
them
out,
and
so
we're
going
to
use
it
as
it's
currently
defined,
with
an
assumption
that
nothing
changes,
and
by
taking
that
approach
we
can
unblock
both
of
these
two
documents.
L
So
there's
a
whole
bunch
of
little
details
in
our
update
and
why
we
did
it.
We
we
got
some
good
great
feedback
from
the
routing
area
review,
as
well
as
from
yang
doctors.
This
applies
to
both
documents.
We've
updated
both
of
them
for
that
and
we've
made
a
couple
of
technical
changes.
I'm
going
to
talk
about
the
technical
changes.
Now.
Excuse
me
the
first
thing
we
did
do
we
have
a
pointer
this.
L
The
first
thing
we
did
is
we
picked
up
a
new
tree
representation,
so
there's
now
a
way
in
the
trees
document,
which
is
also
a
net
mod
to
represent
a
schema,
mount
and
mount
points.
So
the
first
thing
we
did
is:
there's
some
I
don't
see
there
we
go.
So
we
have
a
mount
point
indicating
where
the
net
logical
network
element
root
is,
and
the
other
thing
we
did
is
we
added
a
notification.
L
So
one
of
the
discussions
in
the
whole
op
State
discussion
was
the
ability
to
have
transactions
to
be
processed
asynchronously
and
being
able
to
support
that
in
a
good
way,
and
we
realized
we
needed
a
little
support
in
our
model
for
that,
because
you
might
have
a
transaction
that
creates
an
elleny
and
binds
an
interface
to
an
elleny
that
returns
immediately,
saying
sure
we're
gonna
go
work
on
it,
but
some
asynchronous
processing
happens
inside
the
router
inside
the
node.
That
makes
that
fail
later.
L
We
didn't
have
a
way
of
identifying
that,
so
we
now
have
a
notification
that
get
allows
that
information
you've
provided
it's
easily
provided,
so
you
can
have
a
transaction
that
completes
successfully.
That
says
yeah
we
found
the
Elleni,
but
in
fact
an
RPC
is
taking
place
and
then
later
on,
there's
a
failure
that
notification
can
happen.
Also,
this
can
be
used
in
normal
operations
for
internal
errors.
L
We
have
an
example
here
of
what
a
the
tree
would
look
like
in
an
actual
L&E.
So
in
the
model
the
formal
model
stops
at
the
root
when
it's
instantiated
in
an
actual
system,
and
this
is
decided
by
the
implementation.
By
the
way,
this
isn't
something
that
shows
up
a
standard.
An
implementation
decides
what
gets
represented
in
think
of
it
as
a
virtual
machine.
L
We
can,
for
example,
yang
library.
Actually
we
require
that.
That's
the
one
thing
we
require
and
then
you
can
have
whatever
modules
are
implemented
for
a
typical
node
I'll
point
out
here,
and
these
are
highlighted
in
blue-
that
there's
now
a
tree
representation
for
mounted
modules
that
show
up
with
a
slash
at
the
end.
L
O
O
Here,
of
course,
it
wouldn't
be
this
name,
but
in
the
opposite,
Rex
draft
there
was
a
desire
to
have
a
synchronous
and
asynchronous
mechanisms
for
all
of
this,
so
I
mean.
Currently
we
have
it
that
you
that
the
client
would
do
a
diff
between
the
operational
and
the
intended
to
figure
out
what
actually
got
applied,
but
there
was
a
desire
for
there
to
be
some
sort
of
notification,
a
generic
mechanism.
If
there
were
and
then
would
he
still
need
this
I.
L
So
do
we
have
a
generic
mechanism
for
failures?
In
my
opinion,
we
do
it's
called
notification,
so
we
created
a
notification
I'd
be
interested
in.
If
you
had
thoughts
on
if
there
was
a
more
generic
way
to
do
it
and
then
sure,
if
there's
a
more
generic
way
to
do
it
or
we
could,
we
could
go
that
way,
do
you
it
is
there
a
better
way
to
do
it
in
yang.
I
I
I
I
L
L
We
in
the
original
version
we
had
a
placeholder
I
thought
was
that
placeholder
could
be.
We
would
put
some
information
there,
but
provided
sort
of
a
rich
structure
to
be
augmented.
We've
simplified
that
and
we've
created
a
choice
statement
called
any
ni
type
and
the
intent
is
is
for
there
to
each
particular
model
that
supports
a
layer,
2
or
layer,
3
VPN
technology
to
augment
that
I.
Do
have
examples
on
that,
but
I'm
out
of
time
we're
going
to
talk
about
that
in
bass.
L
The
other
thing
that
we
did
is
is
we
recognize
that
just
for
scheme
amount
that
you
have
to
have
a
mount
point
for
the
information
contained
within
the
verb
or
vsi,
and
that
that
could
that
really
is
independent
of
how
that
VPN
technology
or
that
versus
is
supported
over
the
core
of
the
network.
So
this
is
common.
We
actually
came
up
with
three
types,
one
which
is
worth
actually
I,
think
I
have
it
on
the
next
slide.
L
This
was
an
example
here,
we'll
go
there,
one
is
just
for
roofs.
The
others
for
BS
is,
and
then
we
have
the
hybrid
case
where
you
have
something
like
a
bridge
router
or
something
that
does
both
l2
VPN
as
well
as
routing
within
the
context
of
the
VPN.
So
we
support
those
one
really
important
thing.
That's
confusing
seems
to
confuse
people
is
that
the
models
don't
dictate
what
goes
underneath
the
mount
point?
It's
the
implementation
that
decides
that.
L
So
an
implementation,
for
example,
can
decide
to
implement
a
particular
module
underneath
the
the
root
of
the
the
device
and
it
can
implement
a
separate
one
within
the
context
of
the
ver.
It
doesn't
have
to
implement
the
same
one
and
it
chooses
it's
an
important
concept
but
I'm
out
of
time,
so
I'm
not
going
to
spend
more
time
on
it.
L
L
We
also
had
one
open
issue,
as
I
mentioned
before.
Our
thought
is
to
just
accept
the
limitation
and
move
on,
so
we
received
one
piece
of
feedback
in
this
talk
about
the
notifications.
We
have
to
run
that
to
ground
once
that's
run
to
ground
we're
gonna
be
ready
for
last
call,
or
we
can
choose
to
resolve
that.
As
a
last
call
comment,
I
defer
to
the
chairs
on
that.
So
I.
F
L
And
if
you
do
have
comments
on
it,
please
do
send
us
comments
or
questions,
and
let's
do
it
on
the
list,
because
if
you
have
a
question
about
how
something
is
organized
I
bet,
you're,
not
the
only
person
and
well,
you
may
individually
feel
it's
a
naive
question.
There's
gonna
be
others
who
have
the
same
question,
because
this
stuff
is
a
little
confusing.
L
We
think
once
you
get
through
the
details,
that
the
models
actually
turn
out
to
be
radically
simplified
from
where,
where
the
solution
space
could
have
gone
and
where
we
started,
and
so
any
discussion
that
helps
clarify
how
things
are
work,
we
think
will
be
good
for
everyone.
So
please
bring
it
up
on
the
list
and
remember
if
you're
interested
in
the
network
instances
if
you're
interested
in
burps
or
PS
eyes,
please
come
to
Bess.
P
So
I'm,
Rob,
Walton
and
I
think
your
quick
talk
on
effectively.
They
get
some
guidelines
on
how
to
migrate
existing
iti
models
towards
this
nmda
architecture
as
I
speak
on
behalf
of
the
nmda
authors
and
this
right.
So
just
a
quick
recap
as
to
why
we're
we're
forcing
on
this
change
or
occurring
this
change.
P
What's
the
justification,
and
it
all
comes
down
to
the
opposite
thing
that
we
want
to
be
able
to
cleanly
differentiate
between
what
the
operator
is
asking
a
device
to
do
and
what
is
actually
doing
and
today
or
historically
what's
been
done,
is
often
you've
had
to
infer
what
a
device
has
been
doing
from
its
call
the
config
false
nodes.
You
don't
actually
know
what
configuration
has
necessarily
been
applied,
so
you
don't
know
the
difference
between
what
you're
asking
to
do
and
what
it's
actually
doing.
P
So
this
is
trying
to
solve
that
problem
in
an
elegant
way
and
at
the
same
time
it
also
solves
various
other
problems
have
been
coming
up
in
ITF
like
how
I
tier
I
to
our
s
should
work
and
that
sort
of
thing
so
they've
been
loads
of
discussion
velocity
last
few
years
on
this
and
the
agreed
solution
is
an
operational
data.
Store
for
to
represent
operation
of
state.
Both
net
confirm
restaurant
will
be
updated,
will
have
a
additions
to
them
to
support
this
that
be
discussed
in
net
conf.
P
It's
you're
not
binding
the
applied
configuration
of
state
together,
and
this
does
have
implications
on
what
the
models
look
like
and
that's
what's
going
through
ITF
in
the
moment,
so
I'll
call
there's
a
bit
more
details,
so
there's
three
different
sort
of
model
structures
that
we
have
floating
around
today,
or
least
three.
The
main
ones
here
are
a
split
confidence,
States
top
level
trees
so,
for
example,
interfaces
and
interfaces
States,
for
example,
they're
top
level
containers.
The
interfaces
container
has
configuration
and
interfaces.
P
States
has
upper
estate
and
you
may
have
some
duplication
of
the
configuration
nodes
into
the
state
tree,
but
that's
not
guaranteed
now,
and
the
structures
may
be
aligned
again.
That's
not
guaranteed
so
because
they're
set
the
trees
and
modeled
separately,
there's
always
a
risk
that
these
two
things
deviate.
The
second
style
that
we
come
across
is
one
open,
configure
been
pushing,
and
that
has
conflict
and
state
containers
within
a
single
tree.
P
So
you
have
one
tree
containing
a
structure,
and
then
you
have
configured
state
contains
that
underneath
and
you
have
effectively
the
configuration
under
the
config
container
and
you
have
both
the
configuration
and
the
state
and
the
state
container,
and
so
you
look
at
any
of
the
open,
config
models.
They
fairly
consistently
follow
this
structure
and
is
a
way
of
programmatically
being
able
to
look
up
both
configurational
state.
P
The
the
BGP
draft,
as
currently
expired,
effectively
follows
this
structure
today,
as
did
some
of
the
t's
drafts,
but
they've
been
updated
to
an
NBA
style,
so
the
future
or
what
we
think
is
the
future.
Is
this
combined
config
inspector
is
your
one
tree.
The
represents
both
configuration
state
together
and
I'll
cover
some
more
slides
to
illustrate
that
a
bit
more
detail.
There's
various
models
have
already
been
updated
to
this
structure.
The
ITRs
topology
module
and
the
tea
topology
model
might
be
a
couple
of
teas.
Once
I've
been
updated.
P
This
structure,
the
you
tend
I,
think
on
the
hope
is
that
everyone
will
go
this
way
so
and
to
put
forward
an
example
that
is
hopefully
really
fairly
easy
to
actually
illustrate
this.
I've
taken
a
few
leaves
from
the
beach.you
model.
Obviously
this
is
nothing
like
all
of
it.
I've
taken
for
global
leaves
two
configurable
ones:
the
a
sunburn,
the
rooster
ID,
two
global
state
leaves
total
pars
and
type
of
prefixes,
and
then
I've
got
four
neighbor
leaves
here
as
well,
which
is
under
a
list
structure,
so
you
can
see
when
it
happened.
P
What
it
looks
like
with
a
list
and
again
there's
two
configurable
leaves
their
neighbor
address
and
PAS,
and
one
set
or
two
a
two
state
leaves
as
well
under
messages
in
and
out.
Mister
mister
counts.
So
what
this
sort
of
micro
BGP
model
looks
like
with
the
traditional
IDF
split
configs
state
module
is
like
this,
so
you
can
see
that
on
the
left
in
blue
you've
got
all
the
configuration
nodes.
P
You
start
with
the
top-level
BGP
node
under
that
you've
got
the
global
configuration
and
then
you've
got
the
neighbors
list
with
two
I've
got
two
neighbor
entries
here,
the
first
and
second
neighbor
with
their
configuration.
Then
in
the
B
tree
state
tree.
If
you
were
to
replicate
all
your
conflict
into
state,
you
would
end
up
with
this
sort
of
tree.
That
looks
like
the
orange
tree
and
the
ones
that
are
in
a
shaded
dark
orange,
solid
shading
are
the
config
false
nodes.
P
Those
are
the
state
nodes
or
the
other
ones
are
the
duplication
of
the
config
nodes
and
that
allows
you
using
this
structure
to
still
represent
the
applied
State
the
existing
models,
but
you
have
the
cost
of
having
to
effectively
manage
those
two
trees
and
in
the
yang
model,
to
do
that.
You
might
have
a
lot
of
duplication
or
you
want
groupings
which
make
it
more
confusing,
and
in
many
cases
where
we've
looked
at
this,
the
state
tree
starts
to
deviate
from
the
configuration
tree.
You
can't
automatically
compare
them
the
past,
don't
necessarily
align.
P
So
what
it's
moving
to?
What
in
an
MDA
looks
like
is
effectively
like
this,
so
you
keep
that
same
conflict
tree.
We
had
before
nothing
changes
there,
but
anywhere
you
have
those
statements,
state
leaves
or
state
nodes
they
migrate
into
the
config
tree.
So
you
end
up
with
one
single
tree
and
it
contains
both
config
and
state
together
and
effectively.
You
try
and
reuse
as
much
as
possible
the
same
structures
of
the
conflict
tree.
P
So
you
should
try
and
align
there
and
then
any
way
you
want
to
have
either
leaves
or
trees
of
leaves,
underneath
that
confiture,
that's
fine
and
this
nnd
a
structure
will
work
effectively
and
we
hope
is
the
future
of
the
yang
model.
So
once
you've
got
an
operational
States
data
store,
this
tree
will
represent
the
state
in
the
operation
of
state
tree
and
the
conflict.
P
Parts
of
it
will
represent
the
configuration
in
any
of
the
configuration
data
stores
like
running
and
candidates
startup
and
intended
that
we've
added
so
I'll
just
go
back,
so
you
can
see
that's
what
we
had
before
and
you
can
just
imagine
those
trees
merging
together
and
you
end
up
with
that.
So
the
other
example
I
mentioned
was
this
sort
of
open,
config
style,
so
I
thought
again
it'd
be
useful
to
show
you
what
they
would
look
like.
P
So
you
see
a
config
and
state
container
appearing
underneath
global
and
you
see
the
same
confidence
state
containers
peering
under
first
neighbor
and
second
neighbor,
and
the
rules
that
you
used
to
build
this
style
is
that
wherever
you've
got
any
configuration
you
you
have
all
the
config
leaves
have
to
have
a
config
container
directly
above
them
and
at
the
same
time
you
create
a
state
container
with
a
duplication
of
those.
Two
config
leaves,
and
you
add
also
any
state
leaves
you
have
as
well.
P
So
all
three
of
these
represent
representations
effectively
are
modeling
the
same
fundamental
information,
so
there's
nothing
different
with
what's
being
contained
in
the
models
than
abstract
level.
All
that's
really
different
here
is
how
you
represent
them,
and
it
is
our
belief
that
creating
this
model
ends
up
to
you.
Writing
a
simpler
young
model,
there's
almost
guaranteed
to
be
more
consistent
because
we
take
away
the
inconsistencies.
P
You
take
away
the
ability
of
people
getting
it
wrong,
so
I'm
just
going
to
skip
over
what
we
think
the
advantages
are
of
having
this
combined
NMA
and
MDA
model
structure.
We
believe
it's
simpler,
shorter
to
write
and
it
sort
of
minimizes
the
use
of
groupings
and
certainly
in
the
open
conflict
models,
because
of
the
way
that
they've
structured
them.
P
It's
just
in
a
different
data
store,
and
it's
guaranteed
to
be
so
in
terms
of
the
solution
is
fully
consistent
with
all
of
the
yang
semantics
and
language
constructs.
Again,
that's
not
true
with
all
the
other
solutions
like
open
config
and
the
proposed
structure
that
we
using
here
also
takes
into
consideration
to
support
of
other
ITF
work
like
I
RS,
where
they
want
to
have
dynamic
data
stores,
and
once
you
have
that,
then
you
want
to
work
out.
How
do
you
feed
that
information
in
so
in
terms
of
the
datastore
architecture?
P
So
it
still
gives
you
a
better
structure
to
write,
even
if
you've
opened
conflict,
so
I've
made
a
suggestion
to
the
open
conflict,
people
that
perhaps
their
models
would
be
easier
to
read
if
they
structure
them
this
way
and
then
generated
they're
they're,
open
topic
stories,
so
in
terms
of
migration,
all
yang
models
produced
by
ITF
should
conform
to
the
nmda
architecture.
Personally,
I
think
there's
huge
value
if
all
the
models
produced
by
ITF
and
also
I
Triple,
E
John
Glenn
and
are
consistent
in
their
structure,
because
it
makes
it
easy
for
everyone
to
use.
P
P
You
so
and
if
you've
got
the
open
conflict
style
again,
actually
they
are
easier
to
migrate
than
than
this,
because
you
already
sought
to
have
the
same
structure
you
can
just
move.
The
existing
state
leaves
straight
across
underneath
the
confidence
state
containers,
so
nice
peek
to
see
later,
but
at
least
I
had
a
go
at
doing
this
on
the
BGP
model.
P
A
few
months
back
and
in
an
afternoon,
I
was
able
to
create
an
NMDA
compatible
version
of
the
BGP
model
that
this
currently
published,
I'm
fixing
the
sort
of
compile
errors
and
things
the
same
thing.
So
so
that's
relatively
easy
to
do
at
the
very
high
level.
You
then
have
a
question
of
whether
you
want
to
remove
the
extra
groupings
that
are
no
longer
required,
so
in
that
case
you
can't
have
been
more
effort
to
clean
up
the
models.
P
The
last
thing
it's
worth
pointing
out
is
that
now
you've
got
the
same
nodes,
represented
confidence
state
in
terms
of
those
descriptions,
whereas
it
previously
might
refer
just
to
config.
You
might
need
to
update
those
descriptions
to
be
more
generic
or
applied
both
to
config
and
state
or
explained
semantics
in
the
tune
yeah.
So
still
you
ask
okay.
M
Thank
you
for
the
excellent
BGP
example
and
the
side
comments.
I
reckon
I
have
I
guess.
I
do
have
a
question.
A
comment.
I
would
encourage
you
to.
We
have
extra
time
zippers
and
sometimes
IDR
to
present
this
in
IDR
and
hear
what
the
operators
have
to
say.
That
leads
me
to
my
second
question:
have
you
in
any
of
your
hats,
ITF
or
perhaps
your
corporate
had
asked
operators
what
they
think
about
changing
this
vgp
model
from
the
existing
deployed
ones,
so.
P
In
terms
of
presenting
these
in
idea,
yes,
if
I'm
free
I'm
happy
to
do
that
to
this
time
available
and
I'm,
also
presenting
this
in
net
mod
as
well,
I'm
I,
don't
want
to
overload
people
with
the
same
presentation
in
terms
of
the
BGP
personally
now
I,
don't
know,
and
so
I
think
that
comes
down
to
whether
people
are
implementing
the
open
conflict
models
in
the
industry
or
the
ITF
models
and,
at
the
moment,
I
think.
There's
some
hesitation
guy
tech
models,
because
they're
not
being
published
as
standard,
so
they're
not
actually
getting
passed
through.
J
M
M
Away
I
will
I
will
restate
that
you
are
incorrect
and
all
you're
stating
it,
but
let's
go
on
beyond
that
and
continue
it.
Okay,
okay,
so
you're
saying
that
you
feel
that
this
is
going
to
be
better
and
that
we
have
something
to
tell
our
operators
that
this
is
a
good
reason
to
change
I
believe
so.
Yes,
okay
and
the
reason
you're
telling
that
again
is
that
er
chair,
so
that
I
can
give
your
opinion
and
not
my
own.
M
P
A
more
consistent
model
in
many
ways
and
in
terms
of
the
implementation,
it's
more
consistent.
So
one
of
the
issues
that
open
country
models
have
today
is
that
they're
not
strictly
compliant
with
yeah,
for
example.
The
case
in
point
would
be
you
have
a
system
created
interface,
then
technically
yang
you're
not
allowed
to
have
have
that
in
there
and
company
models,
because
you
you're,
creating
config
nodes
to
represent
a
state
object
and
they
bend
the
rules
a
little
bit
with.
How
do
that
a
saying?
It's
okay?
P
If
you
only
create
the
leaves
in
the
conflict
space,
then
it's
not
really.
Configuration
is
actually
operational
State.
So,
although
all
the
open
coffee
models
aren't
quite
following
the
yang
rules
and
there's
other
things
that
they
can't
do
like
the
can't
make
use
of
choice
statements,
they
can't
make
use
of
presence
containers
because
they
don't
really
work
in
the
open
conflict
style.
So
it's
not
just
it
being
a
bit
cleaner,
I
think
there's
other
things
that
we've
noticed
and
through
its
whole
process
and
said
no.
This
is
the
best
technical
solution,
so.
M
E
H
M
H
H
H
When
we
went
to
the
NDMA
extenders
and
we
also
asked
the
implementers,
are
there
many
at
least
four
OSPF?
There
weren't
many
cases
where
the
Intendant
applied
or
there
weren't
any,
where
they
were
going
to
deliver
different
values
of
the
config,
so
it
it
was.
It
was
sort
of
the
benefit
of
a
pretty
green
field
for
OSPF.
For
us
we
ever
definitely
cut
it
down
significantly.
Q
C
P
I
can't
so
I'll
just
click.
The
answer
that
now
so
I've
been
looking
at
some
PI
am
plugins
to
help
this
process
in
various
different
ways.
So
the
easiest
one
to
come
on
I'll
come
later
is
to
create
an
extra
statuary
where
necessary.
That
can
be
done
by
hand
quite
easily,
but
I
have
to
Tony.
That
does
that
the
other
one
I've
been
using
with
some
success
is
one
that
takes
the
existing
idea
of
split
style
and
has
a
go
at
generating
a
combined
tree.
So
it's
not
perfect
in
terms
of
it
doesn't
give.
P
C
N
N
So
only
at
this,
so
one
thing
I'd
like
to
emphasize
here,
is
when
we're
talking
about
this.
This
is,
you
know,
there's
a
lot
of
good
reasons
to
do
this.
The
meat
of
the
model,
the
complexity,
the
pieces
that
folks
have
been
working
on
is
all
about
the
abstractions
and
the
information
model
and
what
you're
doing
on
some
level.
This
is
the
present
presentation
layer
impact
on
top
of
the
meat
of
the
model,
and
so
when
you
look
at
the
difference
between
the
models,
it
may
be
like
wow.
N
This
is
really
different,
but
it's
a
presentation
style
difference,
not
a
Content
style
difference.
I
think
that's
important
to
think
about
when
you're.
Looking
at
what
the
changes
is.
Rob
said
you
can
do
a
lot
of
this
very
programmatically
and
he's
working
on
tools
for
that,
and
so
I
know
the
models
may
look
like
they
change
a
fair
amount,
but
it's
its
presentation
layer.
It's
not
content.
R
Hello,
you,
girls,
can
Holly
wait
working
on
a
te
anau
model
way
to
scour
to
that
open.
Config
style
is
actually
quite
awkward
when
you
want
to
create
groupings
and
reuse
the
groupings
in
someplace
else,
for
example,
if
you
want
to
take
a
piece
of
say
three
and
use
it
as
RPC
input
and
output
having
config
state
everywhere
becomes
very
awkward,
okay,
so
the
end
MVS
die.
Actress
also
is
quite
nicely.
Thank
you.
P
So
this
is
no
more
some
questions
that
just
have
come
up
and
for
ps1
useful
to
answer
fact
to
me
and
the
question
is:
can
we
use
the
nmda
style
modules
an
existing
net?
Confer
s
cough
ie.
Do
you
have
to
wait
for
the
updated
net
command,
restaurant
and
I
think
new
sort
of
answer
this
in
his
presentation
as
well?
It's
the
same
sort
of
common
C's.
P
Yes,
you
can
but
there's
no
way
of
representing
the
applied
configuration
value
and
when
AC
spoke
about
the
OSPF
model,
no,
he
said
well,
actually
that's,
okay
for
their
model.
It
doesn't
actually
matter
the
operator
that
speaking
to
doesn't
care,
you
wouldn't
actually
have
any
differentiation
between
the
intended
configuration
value
and
what's
been
applied
and
the
other
case
where
it
doesn't
work
on
existing
devices.
P
But
well,
when
you
have
the
operational
state
datastore
is
you
can't
represent
system
created,
configurable
objects,
so
it
can't
represent
an
interface
and
exists
because
a
line
cars
been
plugged
in
and
the
system
ultimately
creates
it
or
a
default
system
interface
like
a
loopback
zero
that
aren't
actually
configured
because
effectively
you
creating
confident
erease
for
those.
So
that's
the
two
cases
where
the
an
NDA
style
model
doesn't
work
with
the
existing
net.
Con
4s
cough.
Obviously,
when
you
get
to
ones
that
support
an
operator
state,
they
store
both
of
these
supported
fine,
most
modules.
P
P
Is
you
want
to
reuse
type
there
so
effectively?
It's
very
easy
to
generate
this.
I've
got
toning,
that
has
a
gogit,
is
simple
to
do
and
that
can
be
put
into
the
draft
appendix
and
used
as
a
transition
module
effective
at
the
time
until
you've
got
an
NDA
compliance
stuff
available
and
the
expectations
that
gets
obsoleted
over
time.
But
I
would
stress
that
it's
worth
only
creating
these
extra
state
modules
when
you
need
them.
If
there's
not
a
requirement,
it's
better
not
to
do
it
rather
than
putting
them
in
just
in
case
this
sucks.
I
P
S
P
It
depends
it
depends
in
those
cases,
that's
required,
that's
fine,
so
the
final
summary
is
the
that
I
would
like,
ideally
all
ITF
modules,
to
move
to
this
style,
because
I
think
it's
the
right
answer
for
the
future
and
I
think
inconsistency
is
really
important.
Personally,
as
I've
said
here,
you
can
generate
this
extra
state
module,
even
when
it's
required,
but
we're
asking
you
not
to
put
it
in
just
automatically.
They
have
a
thing
first
as
to
whether
you
really
want
it.
P
If
you've
got
any
further
questions
this
week,
then
myself
or
any
of
the
authors
would
be
happy
to
speak
to
you
and
talk
to
you
about
it
and
there's
that
various
aliases
you
could
contact
that
and
the
talling
I'm
working
on
I'll
try
to
send
out
details
of
that
as
well
to
help
folks
as
they
migrate
towards
these
modules.
Okay,
thank.
F
You
very
much
sharper
comment
here
and
presenting
so
we'll
keep
on
guiding
and
supporting
here
on
progression.
How
to
move
from
your
existing
model
dollars
and
be
may
may
have
any
request.
Please
send
it
to
routing
working
group
as
well,
we'll
invite
drop
again
for
the
ATF
hundred
to
explain
more
details
how
to
do
it.
So
it's
not
only
theoretical
presentation
but
real
use
case
and
how
we
do.
Thank
you.
F
F
So
this
is
the
background
I
will
maybe
in
G
and
in
the
in
the
Navigator.
Just
leader
trend,
even
at
BMG
scenario
being
is
a
gateway
for
Paniagua
user
traffic.
So
they
are
a
pain
point
I.
Think
from
this,
and
you
know
being
G
scenario.
The
first
year
data
surveys,
a
lot
of
mail
penance
that
in
different
pass
and
just
Anna
its
data
in
G
as
a
post
about
forwarding
and
forwarding
function
and
controller
function,
but
poster
with
the
forwarding
furniture
and
the
controller
function
can
be
to
put
an
AK
or
PNG
device.
F
F
F
F
So
we
wanted
to
be
ng
to
be
aware
to
be
which
are
not
Asian
and
so
in
India
can
be
submitted
into
BMG
safety
under
being
DUP
here,
Eng
step.
She
is
a
user
control
and
management
component,
so
PNG
see
if
she
can
be
implemented
in
server
or
advisor
or
we
am
and
the
pin
DUP
we
rename
it
at
the
username
is
for
the
network
address
and
the
user
policy,
and
so
in
DUP
and
is
used
for
the
data
forwarding.
F
So
the
way
we
name
this
architecture
as
we
may
be
in
G
next.
So
our
try
to
answer
your
question
so
here
with
written
sales.
Data
in
in
DUP
is
usually
for
the
forwarding
for
for
the
user
of
data
for
you
and
for
the
for
the
tiny,
Daniel
trafficker
you.
You
can
see
the
retina,
the
Daniel
trafficker,
where
of
a
wonderful
Rama
being
gu
p
to
be
mg
safety,
so
so
I
think
so.
H
F
F
Yeah
yeah
so
titania
tracking
of
ARP
coming
in
the
biggest
safe
here
for
the
user
trafficker.
We
are
well
for
warrior
from
more
PGP
PNG.
You
appear
to
the
other
core
core
network
of
oil
in
the
net-
oh
yeah
so
so
have
found
yesterday
in
interface
between
GAAP
and
the
beginning,
you
peel
and
the
first
are
universally
the
service
interface
and
the
service
indicator
is
used
to
use
the
initial
restaurant,
a
no
defensive
and
AUP
here.
F
F
Thank
you!
So,
our
our
so
here
we
have
a
steroid
rafter
to
address
the
way
B&G
occur.
So
we
have
word
cancer
and
the
use
case
after
it
was
submitted.
A
2
and
a
3
are
G
so
and
also
we
have
Wesson
an
intention
for
service
interface
and
it's
a
information
model
for
in
the
IPS
work
group.
So
this
dropper
will
defend
the
momentum
that
innovation
yeah
yeah.
So
this
is
the
configuration
of
the
maybe
ng
and
the
web
in
the
spacer
on
logical
network
and
animal.
F
So
we
augment
logical
network
element
we
defined
within
the
name
and
that
enable
switch
in
in
the
Libyan
G
model
and
how
we
defend
the
interface
of
a
P&G
and
also
Java
interface
and
at
the
opens
no
channel
yeah
and
there's
nine
channel
and
also
is
there
and
OS
for
the
open
for
no
channel.
We
were
defender
of
Inferno
or
name
and
DP
idea
and
open
for
no
port
9
channel.
F
We
were
defined
West
9th,
an
idea
with
West
nantan
name
and
Tanner
society
and
Tanner
destination,
IP
underwears
now
ID,
and
for
the
SAR
and
the
chorus
we
were
augmented
against
the
jung
eunji,
the
motor
from
other
aircraft.
So
yeah
comments
are
welcome,
so
is
there,
anyone
has
the
interest
to
work
together.
F
B
I
have
a
little
bit
of
comment
on
the
terminology
that
that
AC
was
asking
about.
Maybe
something
like
subscriber
aware
or
user
aware.
Data
plane
is
what
would
sort
of
describe
the
user
plane.
Well
because
it's
it's
data,
but
it's
also
subscriber
aware
because
it
has
per
subscriber
policies
and
so
on
so
I'm,
not
sure
yeah.
F
Yes,
reducing
that
subscriber
trafficker,
you
know
what
I
did,
how
did
it
happen?
Yeah
I
think
I
think
the
data
trafficker
is
a
bit
under
our
tier
to
be
in
the
up
and
core
network
and
for
the
subscriber
travel
where,
where
messenger
from
India
to
Durban
GCP,
so
I
I'm,
not
sure
this
feather
is
be
known
to
the
you
know,
to
the
data
panel
or.
F
J
L
Burger
the
interesting
thing
about
this:
is
it
it
sort
of
parallels
the
work
that's
being
done
and
T's
for
NC
camp
with
a
CTN
we're
doing
controller
based
models
for
control
traffic
engineered
network.
The
big
difference
here
is
this
is
in
traffic
engineering,
but
it's
the
same
type
of
idea.
So
I
don't
know
if
it's
worth
talking
with.
Maybe
you
want
to
talk
with
the
actn
authors
a
little
bit
and
see
where
it
fits
in
with
their
work.
I,
don't
think
it's
right
for
T's,
but
that
doesn't
mean
it
may
not
fit
in
somewhere.
No.
L
D
D
B
D
B
D
D
B
B
Okay,
if,
if
you
know
I
guess
as
as
I
would
suggest
adding
remote
serologies,
but
we
could
handle
it
another
way.
D
D
D
Don't
want
to
cover
that
right
now,
at
least
don't
not
in
this.
In
this
version
of
the
drug
may
be
in
the
next
few
versions,
we
can
decide
otherwise,
okay,
thanks
thanks.
So
the
idea
is
to
force
the
traffic
over
the
post
convergence
path
to
make
sure
that
there
is
no
risk
of
failure
or
there
is
no
risk
of
the
packet
being
load
back.
D
D
So
how
many
segments,
if
you
have
a
symmetric
matrix
and
you
want
to
protect
links,
only
you
guarantee
to
need
at
most
two
segments
in
most
scenarios
based
on
the
simulation
that
we
did?
One
is
enough
if
you
just
if
you're
one
of
your
neighbors
or
your
new
next
hop,
is
a
loop
free
alternate.
Then
you
don't
need
any
segments,
you
don't
need
to
do
anything
just
send
it
to
the
new
next
hop
and
it
won't
bounce
the
packet
back
to
you
for
a
symmetric
links.
D
If
you
want
to
do
link
protection,
only
in
most
cases,
too,
would
be
enough.
There
are
few
cases
where
you
need
more
than
this
four
node
and
srg
protection
and
I
pointed
out
I'm
restricting
it
to
locals.
For,
for
mathematical
and
technical
reasons.
In
most
cases,
for
is
more
than
it
is
is
enough,
is
in
mostly
I'm
saying,
for
is
more
than
enough.
D
Usually
tool
is
sufficient,
but
we
can
always
cook
up
scenarios
that
you
need
more
than
this
a
bit
on
the
comments
so
important
command,
IPR
disclosure.
We
will
make
sure
that
they
are
taking
care
of
the
draft
is
specific
to
us.
Are
the
word
comments
saying:
why
didn't
you
use
other
methods?
Sure
I
mean
we're
not
saying
otherwise,
but
we're
stripping
this
two
segments
routing.
Only
the
draft
protects
shortest
paths,
which
means,
if
somebody
has
a
very
specific,
winding
path
and
to
be
very
specific
and
is
not
using
segment
routing.
D
So
if
I
am
using
a
stack
of
Sigma
trotting
an
explicit
path
that
consists
of
a
list
of
adjacency,
said
and
notes
it's
in
reality,
we
can
do
protection
for
that.
Last
comment
word
for
security
consideration,
so
the
draft
actually
says
make
sure
that
the
traffic
goes
over
the
new
post
conversations
path
very
quickly.
So
you
can
think
of
this
as
kind
of
like
a
modest
security
enhancement
rather
than
the
risk.
D
T
D
T
T
D
Would
if
I
cannot
do
link
protection?
Why
wouldn't
other
routers
be
able
to
do
Mike
real
protection?
I,
don't
see
the
relation
between
them?
Okay.
First
of
all,
this
is
a
local
behavior
and
the
fact
that
other
routers
know
or
do
not
know
that
I
can
do.
It
is
totally
irrelevant.
Do
anything
else,
I
fail
to
see
what
why
you
think
there
is
a
relation
between
them
so.
T
Instead,
it's
it
solute-free
conversions
process,
I
have
and
afterwards
they
said
where
it
becomes
interesting,
because
what
you
do
is
you
pushing
your
fast
Europe
your
repair?
If,
if
you
can
obey
the,
if
you
can
cope
with
the
constraints
that
you've
just
discovered
which
may
escalate
as
you
as
you
discover
more
information,
so,
for
example,
you
could
be
set
up
for
a
link
and
then
discover
it's
actually
a
node.
Okay,
then
I
don't
protect
it
because
I
can
only
do
links
right.
So
you
should
immediately
abandon
all
other
processing
and
go
straight
to
best-effort
convergence.
D
Okay,
maybe
there
is
a
there
is
a
miss
misconception
here.
We
do
the
best
effort,
conversions.
It
is
there,
except
because
somebody
configured
the
router
to
do
Michael
to
do
the
LFA.
He
will
do
the
TI
LFA
okay,
if,
after
a
while,
he
discovers
that
he
needs
to
do
that
that
this
was
a
note
protection
other
than
a
link.
He
can
download
the
regular
path,
which
is
already
calculated
right.
D
T
Run
no
I,
don't
think
it
is
an
implementation
detail.
You
you
have
to
describe
to
people
the
process
that
you're
going
to
to
run,
because
not
everyone
will
work
out
the
subtleties
right.
So
this
is
not
about
hiding
information.
It's
about
describing
what
you
really
want
people
to
do,
but,
secondly,
the
second
point
is
to
do
with
you,
then
go
into
loop
avoidance.
Loop
avoidance
takes
time,
and
so
you
have
to
abandon
loop
avoidance,
which
other
people
helping.
You
is
why
I
said:
okay.
D
N
Yeah,
this
no
hats
this
time,
but
I've
thought
a
little
bit
about
faster
out.
Of
course,
if
you
have
a
worse
failure
so
that
your
failure
assumption
doesn't
hold
you
get
loops
and
the
rest
of
the
network,
if
you're
trying
to
do
particularly
for
anything
like
this,
which
is
more
sophisticated
than
LFA,
what
you
finally
throw
about
mechanisms
is
you
have?
N
D
D
B
B
D
D
B
D
N
Asking
for
text
that
talks
about
the
interaction
of
it
with
micro,
14,
loop
avoidance
and
under
what
circumstances
you
need
to
bail
to
best-effort,
instead
of
because
you've
had
a
worse
failure,
and
there
are
multiple
reasons
for
that.
One
is
because
it
doesn't
have
to
be
look
just
local
behavior.
Two
is
because
we
are
trying
to
write
specifications
that
people
who
are
not
here
can
actually
go
and
implement,
and
they
may
not
have
all
of
the
understanding
and
because,
if
you
get
it
wrong,
you
loop
traffic,
so.
D
Thank
you,
the
reply
to
that
again,
as
I
mentioned,
the
failure,
the
implementation
of
this
technique
is
specific
to
certain
Philly.
If
you
have
more
failures,
then
you,
it
is
really
the
implementation
problem.
You
cannot
ask
me
if
I
say
I
am
giving
you
a
standard
force
F
what
happened
if
somebody
beside
uses
ISS
or
too
bad
I,
don't
know.
T
Disagree
and
I
guess
we'll
continue
the
discussion
as
they
put
off
I'd
like
to
ask
another
question
which
is
to
do
with
this
is
all
about
forcing
the
traffic
onto
the
post
conversions
path
from
the
PLR
right,
but
in
fact
the
a
lot
of
traffic
will
never
go
near
there
after
the
failure
and
so
I
think
I
I
keep
finding
it
hard
to
get
in
my
head,
Walt
is
actually
gonna
really
haven't
made
a
real
network
as
opposed
to
this
hypothetical
one
that
continues
to
stuff
as
to
send
stuff
by
the
PLR.
So.
D
D
F
D
D
B
D
D
B
B
B
D
D
F
B
H
N
Atlas
there
are
no
IPR
disclosures
yet
on
this
draft
it's
been
over
a
year.
Please
read
the
don'twell
and
stuff
because
the
point
is
not
to
disclose
IPR.
When
you're
asking
for
working
group
adoption.
The
point
is
to
disclose
it
beforehand
as
soon
as
possible.
Yeah
at
the
earliest
point,
not
at
the
latest
point.
Thank
you.
D
D
The
next
one
is
the
micro
loop
avoidance
again
brief
memory
refresh.
It
occurs
if
a
router
downstream
on
the
new
path
has
not
yet
converged
or
has
not
yet
programmed
its
forwarding
plane
and
may
end
up
bouncing
that
packet
back
you,
whether
directly
or
through
an
a
CMP
path,
we're
restricting
it
to
segment
routing.
D
The
objective
is
to
make
sure
that
traffic
goes
over
the
post
conversions
path,
irrespective
of
whether
routers
downstream
on
the
new
post,
converses
path
have
converged
or
not,
so
whether
they've
converge
they
are
using
some
tea,
I
left
a
mechanism
I,
don't
care,
removing
Mike,
Lupo
bodies
themselves
doesn't
matter
it
still
work
here
is
the
idea,
so
router
be
the
one
in
blue
it
previously.
It
used
to
send
traffic
down
towards
the
down
direction
towards
p2,
but
the
link
ad
has
failed.
D
So
what
it
does
is
that
it
sends
the
packet
over
the
green
path
at
the
top.
If
she
has
not
yet
conversed,
then
C
will
bounce
the
packet
back
to
be
saying.
Well,
why
are
we
sending
the
traffic
to
me
you're
the
shortest
path,
so
the
way
it
does
is
that
being
won't
tell
C,
send
the
traffic
to
e,
and
once
the
traffic
arrives
to
e
the
shortest
path
from
E
to
PE
is
via
the
post
conversion
stuff
it's
pretty
much
very
similar,
but
not
identical
to
the
2t.
I
left
a
mechanism.
D
Again,
similar
to
TI
LFA,
it
doesn't
matter
what
happens
downstream,
whether
they
converge
or
not
whether
they
using
GI,
LFA
or
not.
It
doesn't
matter
idea
is
the
same.
You
detect
a
topology
change.
You
detect
that
there
is
a
risk
of
loop,
so
you
explicitly
fork
the
traffic
over
the
forced
conversions
path
using
one
or
more
list
of
cells.
After
a
short,
while
you
basically
use
the
normal
path
without
adding
any
sets
to
it.
The
amount
of
time
that
you
wait.
D
D
How
many
segments,
if
you
can,
if
you
can
look
at
it,
it's
very
similar,
actually
Tiki
il
FAS
to
symmetric
links
maximum
to
most
often
one
in
a
symmetric,
topologies
I
can
cook
up
and
a
symmetric
links
in
topologies
with
a
symmetric
link.
There
will
be
cases
where
I
can
cook
up
topologies,
where
you
need
more
than
two
cells
for
the
link
app
scenario
it
is
guaranteed
to
be
it
doesn't
matter
whether
it
is
symmetric
as
symmetric.
It
doesn't
matter
again
update
on
the
comments
IP,
our
disclosure
would
be
made
thanks.
D
12
units,
like
I,
think
some
thanks
thanks
a
lot
like
typos
and
few
things.
What
topology
changes
are
covered
at
this
point
in
time,
we're
restricting
it
to
link
and
node
up-and-down
events
in
the
future.
We
may
restrict
it
even
more,
depending
on
the
algorithms,
how
do
to
determine
that
time
period
between
stage
1,
which
is
where
you,
the
traffic,
is
explicitly
forwarded
over
the
post
conversions
path
versus
moving
to
stage
2,
where
the
traffic
is
just
for
it
classically
just
send
it
to
the
next
hop
to
the
new
next
hop.
D
There
are
many
ways
to
do
that.
Configuration
protocol
based
simulation
is
really
off
the
topic.
Security
consideration
same
thing.
This
mechanism
ensures
faster
ways
of
sending
the
traffic
over
the
post
conversion
spot
without
risking
loops,
so
you
can
think
of
it
as
a
minor
security
enhancement
and
we're
ready
for
the
working
group
avoidance.
S
D
Again,
this
is
configuration
driven,
you
come
and
tell
me
do:
link
protection.
Okay,
so
I
will
do
link
production.
You
come
and
tell
me
do
note.
Protection
I
will
do
own
protection.
It
is
not
that
it
is
not
that
that
I
know
that
through
some
funny
mechanism,
because
it's
really
impossible,
it's
the
other
way
around
I
was.
N
D
N
D
N
B
B
B
That's
ridiculous,
so
the
local
lesser
I
mean
you've.
If
you
said
I'm
not
going
to
address
s
or
LG's
period,
then
okay,
fine,
but
you
should
state
that
as
well
I
mean
protection.
Note
protection
itself
is
a
form
of
S
or
LG,
true,
okay.
So
what
once
you
are
in
this
protection
topic?
Srl
Jesus,
obvious
you're,
talking
about
local
SLG,
an
explanation
of
why
you
restrict
this
to
local
as
opposed
to
arbitrary
a
serologies
is
completely
appropriate
for
a
document
that
you
expect
the
working
group
to
adopt.
Then.
D
T
So,
presumably
we
need
to
apply
the
same
Hollis
that
was
applied
to
those
selections
in
the
selection
of
the
alternate
route,
for
example,
if
the
original
policy
was
set
up
to
avoid
going
over
a
particular
link,
for
example,
of
security,
a
reason
you
want
to
make
sure
that
your
fast
reroute
doesn't
avoid
those
policies.
So
I
think
there
needs
to
be
some
discussion
about
policy
which
wouldn't
normally
apply
in
a
pure
best
effort.
Network,
ok,
I,
think.
D
I
think
I've
mentioned
this
many
times,
but
I
can
mention
the
more.
This
draft
protects
in
the
third
book,
protects
normal
shortest
paths
as
calculated
by
IGP,
as
well
as
Adsit
I'm,
not
sure
how
many
times
I
have
to
repeat
this
I
said
this
one,
the
main
list
multiple
times
I
wrote
it
in
a
draft
and
I
wrote
it
right
here.
D
Maybe
you're
missing
the
point:
apology,
independent,
it
doesn't
matter
how
things
are
connected
to
I
will
protect
shortest
path.
So
that's
why
it's
called
topology.
The
definition
of
topology
is
what
what
is
connectivity
right.
So
what
I'm
saying
here
it
doesn't
matter
how
your
network
is
stitched
together.
I
will
give
you
protection
yeah.
R
D
D
D
R
D
A
B
A
A
We
worked
on
some
real
world
experience,
tons
of
industry
input
and
what
happened
was
yeah.
Sorry,
what
happened
was
roughly
that
we
extended
requirements,
but
one
or
two
things
which
have
proven
important
to
some
customers.
Some
customers
insist
and
run
east-west
links,
aware
that
the
lists
east-west
link
procedures,
I,
mentioned
the
work.
A
We
had
interest
in
operational
discussion
about
internal
reach,
ability
of
nodes
within
the
fabric
in
all
kind
of
failures
in
areas.
So
we
added
considerations
for
that.
Some
people
desired
better
description
of
the
flooding
scopes,
so
we
brought
the
rules.
There
are
down
more
carefully.
The
deformation
procedures
are
non-trivial,
so
we
spelled
it
out
better.
A
We
added
BFT
interactions,
optionally,
which
are
also
desired
for
some
people.
Some
of
the
encoding
has
been
changed
for
better
implementation,
and
the
section
has
been
added
that
deals
with
a
topology
chart
which
are
not
strict
closed,
because
people
have
some
ideas.
They
don't
have
deploying
for
the
F
ideas,
all
right.
So
what
happen
in
terms
of
requirements
it
looks
like
people
are
not
overly
happy
with
lack
and
MC
lack,
no
layer-2
in
IP
fabrics
for
multiple
reasons.
I
mean
some
of
them.
There's
always
people
who
are
happy
and
some
people
are
unhappy.
A
So
we
added
mekinese
in
rift
where
it,
basically,
you
can
replace
like
for
all
practical
purposes
with
rift
so
and
you
can
run
it,
of
course,
on
links
which
have
are
known
uniquely
addressed
or
are
unnumbered.
So
we
can
support
unnumbered
parallel
links
and
basically
provide
a
full
lag
functionality.
The
MC
Lyte
discussion
are,
we
don't
want
to
touch
the
stuff
in
rift
and
see
lag
is
pretty
much
IP,
IP
multi-homing
or
otherwise.
A
It's
really
an
l2
problem
and
it's
an
ugly
problem,
because
it's
not
really
a
standardized
problem
in
terms
of
montant
multi
share,
select,
synchronization
after
discussion
looks
like
we
don't
need
any
kind
of
forwarding
adjacency
support.
We
just
use
tunnels
as
interfaces
and
to
know
there's
no
further
considerations
needed
a
next
time.
I
probably
show
something.
People
are
very
interesting
in
an
optional
mode,
where
basically,
the
whole
data
center
configuration
will
be
brought
down
to
almost
zero
touch
right.
So
it's
very
minimal
in
model
configuration
and
bring
up
the
whole
data.
Centers
will
be
configure.
A
The
hardwood
optional,
east-west
procedures
work
a
little
picture.
If
you
imagine
so,
I
show
you
an
A
and
B.
The
blue
ones
is
a
part
and
Z
is
another
part,
and
the
ten
and
twenty
are
leafs
and
alpha
is
a
spine
switch.
If
you
have
a
northbound
failure
between
a
and
alpha,
that
is
where
normally
we
would
start
black
hole
link
right,
so
we
do
the
full
dreg
regression
and
so
on
without
east-west
links,
rift
works,
but
people
desire
to
use
the
East
to
east-west
link.
A
So
we
have
optional
procedures
where
a
can
use
the
east-west
link
towards
B.
In
case
it
gets
traffic
from
ten,
so
ten
to
a
will,
go
to
B
and
then
go.
It
will
go
through
the
spine
alpha
to
go
to
twinky
the
tree.
It's
a
symmetric
because
when
20
sends
traffic
back
to
10,
it
can
obviously
not
go
through
a
so
it
will
always
go
through
B.
But
that
way
you
can
use
the
east-west
links
in
case
of
failures,
basically
to
protect
a
layer
against
northbound
failures
when
you
lose
northbound
connectivity.
A
That
is
still
completely
look
free,
which
is
very
important.
So
it's
not
generalized
routing
right.
The
east-west
will
not
be
used
in
both
directions
because
then
the
whole
algorithmically
the
whole
thing
deteriorates
into
generic
link,
state
routing
and
drift.
There's
absolutely
no
purpose.
You
can
run
eyes
eyes
because
all
kind
of
funky
topologies
getting
constructed
where
all
this
link
starts
to
forward
bi-directionally
in
the
east-west
direction.
A
Ok,
so
on
the
internal
node,
reachability
terms
of
all
kind
of
funky
failures,
right,
we
added
a
new
section
and
basically
what
we
recommend
is
that
the
internal
node
injects
its
loopback
address
into
the
north
direction,
which
basically
without
failures,
allows
you
to
reach
is
no
internal
either
from
south
and
from
north.
Ok,
the
interesting
things
happened
is
when
nodes
lose
northbound
connectivity.
A
You
you
have
to
think
that
through
what
happens,
but
basically
the
node
may
become
unreachable
from
certain
nodes.
So
what
we
recommend
is
that
in
such
scenarios,
the
notes
inject
south
PGP
ties
there,
loop
X,
which
makes
them
reachable
again,
at
least
from
the
south,
so
that
marriage
may
be
farther
a
consideration.
But
we
thought
that
through
and
there's
a
section
dealing
with
those
failures,
the
precise
D
for
regeneration
procedures
in
the
southbound
direction
has
been
written
down.
So
it's
not
entirely
trivial
I,
pretty
much
spelled
out
the
rules.
I,
don't
think
on
the
fly.
A
The
BFG
interactions
are
are
relatively
easy,
so
we
allow
or
allow
or
optionally
it
is
just
fine
to
run
BFD
on
the
same
links
as
rift
right,
and
we
just
specified
the
interaction
that
you
bring
BFD
after
the
rift
brought
the
three-way
and
the
BFD
after
it
came
up
and
goes
down,
can
break
the
stuff
down.
So
to
prevent
all
this
admin
down,
problems
that
are
protocols
have
we
aligned
the
link
ids,
which
is
basically
unique
identification
of
a
link
in
rift
with
BFD
discriminators.
A
So
implementations
can
very
easily
basically
reuse
the
values,
if
that
makes
any
sense,
and
the
BFD
may
be
used
on
subset
of
links
on
all
the
links
in
terms
of
like
having
a
micro
like
all
of
a
sudden
right
micro,
be
of
tea
running
over
a
lakh
and
ultimately
basin.
On
experience,
the
encoding
has
been
changed.
What
it
means
in
thrift
terms
is
that
a
lot
of
sets
have
been
replaced
with
maps
and
lamentation
get
much
faster.
H
A
Yeah
yeah
nothing
rift
is
a
link,
no
hope,
I,
hope,
protocol.
Okay,
there's
no
multi-hop
rift.
I
mean
we
talk
through
these
issues.
There
is
simply
no
application
for
it.
You
run
tunnels.
If
you
need
that
and
then
you
run
the
FT
over
towns
or
whatever
you
do.
Okay,
so
basically
you
make
explicit
that
this
is
only
single
Hardy,
possibly
not
what
I
shoot.
I
think
I
refer
to
the
right
drafts:
I,
don't
refer
to
the
math,
I,
hope
EFT,
but
yeah
I
I
can
look
it
up.
Yes,
but
that's
the
idea
really
trivial.
A
Basically,
it's
a
trade-off
between
having
a
control
plan
which
can
pump
the
hellos
fast
enough
to
keep
you
happy
at
a
certain
point
in
time
to
having
to
go
to
timer
resolution
where
it
doesn't
make
any
sense
to
run
the
control
plane.
So
please
invite
your
hardware,
which
for
practical
purpose,
means
BFD.
It's
a
problem
has
been
solved.
R
A
I
mean
we
went
through
the
car
for
20
years.
We
know
the
last
loss
of
light
is
not
good
enough
right.
Bfg
has.
A
purpose
has
been
built
for
a
very
specific
purpose
and
is
very
widely
used
if
loss
of
light
would
have
been
good
enough
P.
If
you
will
have
existed
right,
we
have
empty.you
problems.
We
have
single,
no
Direction
connectivity
problems,
we
have
all
kind
of
weird
stuff
and
BFD
is
doing
an
excellent
job.
A.
H
H
A
Very
good
yeah,
so
in
the
North
direction
we
recommend
you
always
inject
right,
because
that's
how
you
reach
the
node
without
any
failures
in
case
you
lost
northbound
connectivity.
You
may
not
be
able
to
get
to
it,
so
we
recommend
that
you
inject
an
s
PGP
loop,
X
prefix,
which
will
make
you
always
reachable
at
least
from
the
southbound
direction.
The
problem
is
ugly
and
the
problem
is
very
real
today,
from
what
I
understand
now
for.
H
H
A
E
You
let
me
stand
in
this
little
paint
box.
You
know
so
much.
I
hate
that
Russ
White
LinkedIn,
that
is
Tony
P's
data
center.
Sorry,
this
is
is,
is
support
for
open
fabric.
It's
pretty
much
the
same
presentation
I
gave
last
year,
except
there
are
a
couple
of
small
changes.
Oh
look,
I!
Don't
have
to
look
behind
me.
That's
amazing,.
E
Couple
of
small
changes
in
the
encoding
and
stuff
like
that,
essentially
the
idea
end
up
in
fabric,
is
to
split
the
policy
from
the
reach
ability
in
to
make
things
much
much
simpler.
This
is
trying
to
just
simplify,
is,
is
or
using
I
as
artists
as
a
base
and
simplify
things,
so
that
things
are
a
lot
easier
to
manage
simple
as
possible
distributed
links,
a
protocol,
no
policy
and
the
protocol,
no
configuration
essentially
we're
trying
to
get
to
is
having
no
configuration
on
the
boxes.
E
Everything
just
comes
up
in
boots
and
it
works
no
extra
stuff
trying
to
keep
out
of
the
feature
creep
realm
running
a
big
big
data
center
and
you
get
lots
of
features
and
basically
the
features
just
become
nerd
knobs
and
the
nerd
knobs
just
become
weeds,
and
you
end
up
spending
a
lot
of
time,
trying
to
weed
your
data
center
fabric
and
it's
really
ugly,
there's
a
loose
feature
set
right
now.
Erik
Osbourne
observed
for
me
in
a
private
email
exchange.
E
I,
don't
know
if
Erik
is
here,
I
don't
think
he
is
this
week,
but
anyway
that
there's
a
lot
of
Meru
features.
This
is
something
we
need
to
look
at
his
authors
and
contributors
and
figure
out
like.
Should
we
tighten
that
up
and
make
some
things
may
be
not
included?
Some
things
must
not
be
included
for
a
drop
ability.
This
is
just
something
that
was
brought
up
in
the
last
week.
So
it's
something
we
need
to
look
at
if
you
have
concerns
or
feedback
on
that,
I
would
like
to
hear
those.
E
Basically,
the
draft
lays
out,
in
several
small
pieces,
there's
a
fabric
location
piece,
we're
using
SPF
to
calculate
where
you
are
on
the
fabric.
This
only
works
on
a
5
stage
or
larger,
because
in
a
three
stage
fabric,
a
three
area
fabric,
you
cannot
tell
where
you
are,
because
everything
you're
symmetrical.
So
basically
all
you're
doing
is
running
SPF
with
a
hot
count
of
1.
E
We
have
changed
the
encoding.
It's
in
bold.
On
the
slide,
because
this
is
something
that's
changed,
less
and
naming
came
out
where
the
new
draft
talking
about
how
to
encode
the
tear
tear
level
in
their
draft.
It's
a
nice
change,
so
we
took
the
encoding
out
of
our
draft,
we're
just
using
theirs,
because
it's
actually
simpler
to
merge
the
encoding
schemes
in
a
way
that
makes
more
sense
for
us
is
so
that
we
don't
have
to
actually
manage
that
piece
of
it
for
ting,
optimization
or
forward
optimization
of
flooding,
essentially
we're
using
neighbors
neighbors.
E
If
you've
ever
read
the
man,
a
mobile
ad
hoc
network
work.
This
is
very
similar
to
that
in
that
you're,
using
neighbors,
neighbors
and
you're
deciding
which
neighbor
of
yours
can
reach
all
of
your
two
hops,
and
only
flooding
in
a
way
allows
one
of
your
neighbors
that
can
reach
all
of
your
two
hops
to
reflect
all
of
their
peers
or
all
of
their
intermediate
system
neighbors.
E
H
E
It
there,
but
it's
the
simplest
way
to
go
about
it.
It
seems
like
to
me
the
main
problem
you
run
into
with
eius
eius
is
there's
no
header
bits
remaining
to
do
anything
with,
so
you
can
do
different
mac
addresses
which
lubricates
it
on
lists
or
you
could
do
a
new
LSP
type,
which
is
what
link
local
flooding
basically
does
or
there's
other
games
you
can
play,
but
essentially,
since
there's
no
header
bits
to
play
with
you
know
it's
like
it's.
There
are
very
limited
options
here,
but
you.
E
E
Yes,
yes,
yes,
so,
and
if
anybody
has
any
suggestions,
please
send
them
to
me,
because
this
is
very
very
in
their
size
protocol,
because
of
the
way
the
header
bits
work
and
the
fact
that
we're
out
of
header
bits,
this
is
actually
pretty
difficult
all
right,
so
this
is
a
forwarding
up.
This
is
the
forward
flooding,
there's
also
reverse
flooding,
which
is
basically
just
don't
flood
back
to
anybody
who
is
on
your
s.
We
have
treat
towards
the
originator.
Yes,.
U
E
That's
actually
that's
actually
in
the
draft.
You
can
do
that.
The
goal
here
is
to
get
to
when
you
have
a
change.
Every
intermediate
system
only
receives
one
copy
of
the
LSP.
However,
if
you
are
concerned
with
weird
timing
problems
where
you
don't,
then
you
can
always
choose
a
backup
and
make
it
where
everybody
gets.
Two
copies,
that's!
Okay!
That's
in
the
draft!
The
other
thing
is,
is
that
in
the
draft
it
talks
about?
This
is
something
else
less
less
did
or
less
convinced
me
to
do.
E
Is
that
there's
actually
a
follow
up,
see
SPF,
so
that
you
know
that
everybody's
still
synchronized?
Okay,
because
that's
actually
simpler
than
that's
like
the
simplest
thing
in
is,
is
to
do-
is
that
you
just
send
a
si
SPF
afterwards
and
then,
if
somebody
has
a
problem,
they
can
send
a
partial
and
they
can
actually.
You
can
actually
see
the
failure
to
flood.
It's
actually
pretty
simple
mechanism.
So
there
are
two
ways
of
handling
that
in
the
draft
right
now:
yeah,
okay,
unless
it's
coming
to
them.
E
See
us
in
peace,
correct,
sorry,
not
constrained,
alright,
so
other
optimizations
there's
a
whole
section
in
the
draft
about
stuff.
We
just
don't
need
ops,
there's
some
optimized
neighbor
formation
stuff
in
there
that
Alvers
talk
to
me
about
next
steps.
Question
marks
we're
just
presenting
for
the
second
time
we
may
or
may
not
present
it
in
is,
is
this
afternoon,
depending
on
how
many
people
raise
their
hands
and
say
they
want
to
see
the
presentation
again,
and
you
know
we're
working
we're
slowly.
E
Working
on
some
implementations,
there
are
some
things
swirling
around
the
implementation
space
there,
some
people
who
want
to
implement
this
and
run
it,
but
there
are
just
some
things
going
on
as
far
as
logistics
go
that
are
making
it
difficult
for
us
to
get
there
for
that.
For
this
point,
but
questions
anything
feel
free
to
send
them
to
me.
Send
them
on
the
list
perfectly
fine,
you.
This
is
all
just
open
work,
trying
to
get
something
done.