►
From YouTube: IETF106-RIFT-20191121-1330
Description
RIFT meeting session at IETF106
2019/11/21 1330
https://datatracker.ietf.org/meeting/106/proceedings/
B
C
C
It's
it
they
kind
of
part
of
Sandy's
presentation,
because
she's
driving
the
applicability
draft.
So
we
talked
through
all
these
things,
but
not
everything
made
it
made
in
you
know
such
kind
of
lucky
Luger
hat,
have
a
receive
the
site.
Yeah
I
sent
your
slides
I
mean
we
have
time.
If
we
run
out
of
time,
I'm
cool
I
think.
A
We'll
have
time
so
zero
nine
has
been
posted.
We've
issued
the
working
group
consensus
and
waiting
her
write
up
reviews
her
ongoing
the
document
is
long,
so
it
takes
a
lot
of
time.
Unfortunately,
so
we
expect
to
finalize
sometime
between
Singapore
Vancouver.
Hopefully
we
also
plan
to
do
land
which
I
wouldn't
call
it
fixes,
but
simplification,
so
we're
still
I'm
still
looking
for
someone
who
is
native
speaker-
and
you
know,
leisure
person
who
could
really
simplify
the
text.
A
We
have
adopted
an
specification
after
it
has
been
updated
to
match
0-9
it's
pretty
much
ready
for
working
with
last
call
more
rocks
me,
there's
a
multi
custom
beer
and
listen
PGP
and
segment
routing,
so
it's
all
being
planned.
We
also
plan
to
recharge
there
before
one
cover.
Since
main
document
is
done.
We
are
going
to
focus
on
progress,
ink
and
finalizing
the
other
documents
of
specifically
PGP
segment
routing
the
multicast.
D
Hi
Rodney,
so
since
this
went
through
working
west
call,
I
am
read
the
document
and
I'm,
not
a
native
speaker
in
case.
You
want
me
to
read
it.
But
how
are
we
in
this
question?
Maybe
for
you
and
the
Shepherd
and
the
author's?
How
are
we
reviewing
the
model
itself
and
in
then
we
make
it
a
multi-part
question.
You
mean
3v
part
of
it.
Yes,
so
I
noticed
that
when
we
talk
about,
you
know
doing
the
modeling
thrift,
which
is
fine
yeah,
that's
perfect!
D
It's
no
problem,
but
since
we're
the
first
ones
to
do
it,
we're
probably
gonna
have
to
go
through
any
growing
pains
that
we
have
to
go
straight.
So
I
noticed
several
things.
The
appendix
is
normative
right,
which
means
that
it's
fine,
but
we
should.
You
know,
make
sure
that
the
model
is
reviewed.
Somehow
I,
don't
know
that
we
have
anyone
who,
beyond
the
authors
who
know
thrift
I
mean
we
don't
have
a
thrift
doctor
or
yang
dr.
sorry,
we
don't
have
a
thrift
doctor
that
could
check
the
syntax
and
everything
else.
D
D
D
A
Pians
very
valid
concern,
so
I
try
to
find
people
outside
of
ITF
fox
general
drift,
but
the
answer
solves.
Why
should
they
so
I
assume
we
are
going
to
iterate
same
way?
We
do
with
young
models
after
we
have
published
and
if
you
found
by
building
new
implementation
that
something
doesn't
work,
we
can
always
risk.
Pin
and
you
know,
do
quick
update
yes,
but
that
doesn't
I
understand
but
I
mean
obviously
Tony
knows
what's
in
it
right,
but
asking
him
to
review
it.
It's
like
drinking
so
kool-aid,
so
it
doesn't
really
help
run.
C
20P
juniper,
so
finding
someone
to
review
the
thrift
is
more
up
to
the
chairs.
Okay,
the
stones
of
thrift
in
the
industry
are
not
necessarily
you
know
in
our
field
now.
Why
should
they
I,
don't
know
up
to
the
chairs?
To
like
do
the
salesman's
job
as
to
the
thrift
compiler?
Well,
we
will
be
facing
the
same
problems:
GRP
c2
and
the
GRP
c3
right.
If
they
move
major
version,
is
like
good
luck
to
you.
Thrift
has
been
very,
very
stable
because
Apache,
you
know,
put
that
stuff
away
back
and
the
compiler
has
been
stable.
C
D
D
You
said
if
Apache
goes
and
changes
a
compiler
to
some
other
major
version
or
something
there
may
be
issues,
meaning
that
if
you
did
your
implementation
and
it
did
mine
and
I,
compile
it
with
something
else,
we
might
not
be
able
to
control
it.
Is
that
true,
or
am
I
understanding
so
I
will
just
for
a
second
Bruin.
C
Bruno
I
I
jump
in
an
approver
can
talk
so
first
thing:
it's
more
subtle
because
the
compilers
deliver
actually
the
code
in
different
target
languages.
Umesh
issues
between
target
languages,
welcome
to
the
beautiful
world
of
modeling.
That's
why
I'm
so
happy.
We
have
actually
three
programming
languages,
implementations
and
each
of
them
have
a
different
back-end
code
come
from
the
compiler
and
they
all
interrupt.
Second,
a
bola
will
probably
talk
about
it.
C
We
very
carefully
checked
whether
what
thrift
was
promised
is
actually
valid
like
when
we
omit
elements
and
the
optional
without
default,
and
whether
the
stuff
really
shows
up-
and
we
can-
we
can
decode
something
we
added
an
element
and
so
on.
So
we
did
pretty
I
would
say:
annal
checking
with
Bruno
whether
the
stuff
holds
up,
because
otherwise
you
knew
we
couldn't
progress.
The
model
without
like
doing
go
going
through
the
major
versions,
but
yes,
I
mean
Industries
is
doing
tons
of
ideal.
C
A
I
can
hear,
but
cannot
speak,
he's
saying
that
when
he
tried
compiled
for
another
language
he
had
some
problems
with
handling
and
a
number
soon
we'll
see
some
minor
issues
between
languages,
but
at
least
the
songs
that
have
been
test
should
be
probably,
and
the
interdependency
is
not
really
idea.
Domain
right
well,.
D
Is
that
all
the
first
two?
You
know
some
other
time
that
you
want
to
argue
that
with
someone
else,
but
what
I
was
going
to
say
is
that
we
should
at
least
include
I
just
look
at
the
draft
and
there's
some
there's
an
implementation
looks
like
maybe
considerations
section.
Maybe
you
wanted
to
include
an
implementation
status
section
that
talks
about
you
know
a
little
couple.
Paragraphs
of
what
implementations
have
already
been
done,
yeah
this
compiled
to
this
language
that
compiled
that
other
language.
D
F
D
Whatever
there's
an
RFC
that
talks
about
best
practices
in
indicate
implementations,
so
it's
basically
a
section
that
talks
about
the
current
state
and
in
fact,
if
is
removed
when
the
RC
is
published.
So
it
is
mostly
for
information
of
the
reviewers,
and
you
know
everyone
else,
and
so,
when
everyone
starts
saying,
we
don't
have
know
where
the
heck
this
and
you
know
who
reviewed
it
and
whatever
at
least
we
can
say
well,
but
look
it's
compiled
in
different
languages
and
you
know
whatever
there's
some
inference.
A
C
So
yeah,
okay,
no
problem,
let's
put
an
implementation
section
in
that
get
basically
nuked
before
the
stuff
gets.
You
know
RFC
as
proposed
standard
as
a
meta
observation.
This
is
not
different
than
having
an
implementation
in
the
C
compiler
bug,
but
just
a
different
level.
I
mean
it's
not
like.
We
fundamentally
changing
the
world
right,
we're
getting
better
tools
and
each
of
the
tools
we'll
have
it
quirks
and
the
quirks
can
affect
that
now,
the
resulting
output
and
these
quirks
will
be
fixed
if
enough
people
have
interest
and
run
this
stuff.
A
B
G
I
want
to
discuss
controlled
desegregation
and
multihoming
in
data
center
networks.
It's
not
specifically
about
reefed,
but
I
want
to
go
through
the
slides
to
show
problems
which
are
difficult
to
solve
without
reached,
so
essentially
why
we
want
to
do
this
creation
and
aggregation,
of
course,
some
problems
which
are
difficult
to
solve
if
we
are
using
BGP,
which
is
commonplace
in
a
really
large
data
center
networks.
Now
that
basically
boils
down
to
following
things
in
factory.
G
If
we
consider
any
particular
level,
it's
completely
disjoint
nodes,
don't
know
about
each
other,
and
some
decisions
cannot
be
really
made
using
only
local
information,
but
it
can
be
solved
with
reefed
and
quite
elegantly
to
limit
the
scope
of
discussion.
We
are
talking
about
layer,
three
data
center
networks,
only
considering
only
enriched
ability,
but
not
considering
the
asymmetrical
bandwidth,
and
we
are
not
talking
about
details
of
rift
implementation
because
it
was
discussed
many
times
before,
just
to
remind
the
topology
of
the
network.
We
are
talking
about
in
examples.
G
I
will
use
these
two
spine
levels
and
one
Li
for
totally
level
all
the
directions.
South
north
east
west,
and
essentially
we
are
following
quasi
quasi
core
closed
building
procedure.
Recursively,
we
are
taking
some
elements
with
some
radix
combining
them
into
Samson,
which
we
can
consider
switch
an
element
with
higher
Alex
and
if
that's
not
enough,
very
repeated,
that's
it
and
basically
different
ways
to
scales
and
data
center
networks.
G
But
there
is
a
problem
potential
problem.
When
we
aggregate
in
multi
pass
routes
what
kind
of
multiple
routes
we
have
and
how
they
propagated
I'm
talking
about
propagation
of
routes,
to
contrast,
what
we
have
this
distance
reactor
and
BGP
to
what
potentially
can
be
done
with
drift.
So
what
we
have
first
is
default
or
entire
data
center
network
aggregate,
which
Kyocera
seen
it
it's
usually
not.
Usually
it's
always
owned
north
toward
top
of
fabric
and
it
propagates
south
from
top
of
fabric.
G
It
shouldn't
propagate
north
from
some
node
on
the
intermediate
level,
and
there
could
be
small
aggregates
situational.
They
appear
in
foreign
cases
when
we
do,
for
example,
host
multihoming
or
when
we
introduce
extra
mesh
between
spine
levels
or,
if
you're,
trying
to
do
something
like
redundancy
on
the
top
of
Port
lado,
which
some
hyper
scalars
do.
It
introduces
valley
routing,
so
basically,
a
traffic
could
go
down
to
the
leaf
levels
and
reflects
back
to
spine
one
level
and
then
reflects
back
to
correct
leaf
and
it's
problematic
because
it
has
its
own
corner
cases.
G
Those
small
aggregates
raise
a
point.
It
depends
on
the
place
in
the
network,
say
my
points
house.
They
might
point.
North
really
depends
on
where
we
are
in
the
network
and
then
normally
propagates
from
the
origin.
Up
to
the
top
of
fabric
then
reflected
and
propagates
nose
or
south
to
the
rest
of
the
network.
G
Okay,
there
are
some
positive
things
with
aggregation.
Why
we
do
disagree
eight.
Basically,
the
reason
is
when
we
combine
multipass
and
there
is
an
aggregate.
This
aggregate
covers
multiple
destinations
and
we
have
favors
somewhere,
not
on
the
dirac
next-hop,
but
remote
favors
reach
ability
becomes
a
symmetrical.
Basically,
not
all
the
intermediate.
Next
hops
are
valid
for
all
the
remote
destinations
and
if
we
try
to
use
only
aggregate
Road,
what
eventually
can
happen
is
that
we
come
to
some
node
which
originates
that
aggregate,
but
there
is
no
way
to
go
further
from
that.
G
And
of
course
there
is
a
fallback
scenario.
Basically
the
simple
thing
to
do.
We
can
just
start
with
disagree.
Gated
everything
never
greg
8,
it
works,
but
it
doesn't
scale
and
essentially
we
are
leaving
with
the
worst
case
scenario.
All
the
time
and
really
forwarding
state,
especially
multiple
Fortin
state,
can
be
a
problem
and
I
will
show
that
later.
G
G
Because
do
we
need
to
propagate
that
specific?
Because,
as
notes
on
our
level
advertise
aggregate,
but
they
don't
have
rich
ability,
don't
have
routes
to
that
specific
destination
and
we
really
need
to
inject
or
propagate
specific
or-
and
another
question
is:
do
we
have
a
full
set
of
all
possible
valid
valid
routes
in
the
network
is
a
current
moment,
because
if
we
don't,
then
we
are
potentially
broke
hall
in
Samson.
G
That
means
that
we
cannot
go
from
this
node
to
this
node,
and
this
error
just
shows
a
way
we
want
to
send
traffic,
so
this
particular
form
means
that
this
node
and
these
links,
which,
with
the
dashed
lines,
they're
ineligible
they
invalid.
If
you
want
to
get
to
that
particular
destination-
and
that
means
as
three
of
the
spine
1
notes
these
days,
and
these
cannot
really
use
the
default
is
they
need
to
have
more
specific
routes
because
they
cannot
send
traffic?
G
To
this
note,
via
this
top
of
fabric
note,
what's
interesting
is
if
we
have
ferrous
lower
in
the
topology
like
anti
note
or
link
from
spine
1
down
to
tour.
Note
then
affected
part
of
the
topology
is
much
larger,
so
boa
Stratos
is
quite
large,
and
that
means
that
we
need
to
disaggregate
not
on
the
spinal
arrows,
but
we
really
need
to
disaggregate
on
the
leaf
layers.
All
those
leaves
cannot
select
any
note
from
this
plane
to
go
to
that
destinations.
They
all
need
to
go
through
the
hero
plane.
G
That's
also
a
property
of
PLAs
topology,
because
basically
the
higher
up
you
go
in
this
topology,
then
narrower
and
the
smaller
subset
of
remain
Interpol
Japan
you
can
go
and
once
we
select
plane
as
it
is
so
this
we
cannot
switch
to
another
plane.
We
went
to
this
note.
We
are
confine
it
to
this
plane
and
if
it's
a
problem
with
the
connectivity
from
this
plane
to
the
destination,
Belloc
just
be
code.
G
And
well,
that's
how
it
looks
with
2
spine
layers.
Situation
becomes
a
little
bit
more
interesting
if
you
have
more
than
two
spine
layers,
but
essentially
the
same.
The
lore
is
a
little.
We,
we
have
failed,
the
largest
by
Australian
small
notes
need
to
disaggregate.
More
notes
cannot
use
the
default,
but.
G
Really
most
of
the
deployments
are
fine,
with
one
or
two
spine
levels
very
few
deployments.
They
really
need
more
than
two
spine
nails
and
another
thing
is
favors
and
forward,
and
state
problem
is
that
in
modern
switches,
especially
if
we
are
talking
about
commodity
hardware,
S&P
fortune
state
is
a
scarce
resource
and
it's
not
vendor
specific.
It's
just
how
workups
lookups
work
essentially
once
we're
dealing
with
ECP
after
a
longest
prefix
match
lookup,
you
have
a
couple
of
additional
tables.
G
We
have
a
simple
group
table
which
in
turn,
points
to
the
block
of
next
hops
and
those
resources
are
limited,
so
we
should
be
careful
and
they
can
be
consumed
quickly
because
usually
it's
okay
to
have
moderate
number
of
groups,
but
especially
of
intermediate
spines,
which
have
very
widely
seen
P
looking
north
the
number
of
consumed
next
hop
entries
is
basically
a
number
of
unique
next
hop
groups,
multiplied
by
the
weeds
or
PCP
and
Boozer
PCP
could
be
32
or
64
on
modern
switches.
So
it's
really
scarce
resource.
G
All
the
Leon
leaves
on
top
of
rack
switches,
number
of
potentially
unique.
Next
subgroups
is
usually
limited
by
the
number
of
combination
of
next
hop,
for
example.
Typically,
if
you
have
some
slight
8
up
links,
there
are
only
2000
5256
potential
combinations
of
our
valid
and
invalid
next
hop,
so
that's
256.
G
Next
four
groups
marks
another
situation
where
we
have
to
deal
with
aggregation
and
control.
The
disaggregation
is
if
we
want
to
do,
host
multiform
and
and
specifically
a
layer,
3
host,
multi
forming
so
setup
is
4
and
we
have
a
pair
of
switches
and
a
number
of
hosts
our
door
home
to
those
switches.
Each
switch
is
a
bear,
originates
the
same
obligate
and
in
this
situation,
leaf
or
Torah,
don't
have
enough
local
information
to
figure
out
foreign
things.
First,
if
the
host
is
dead
or
it's
just
the
link
towards
earth
crust
is
dead
locally.
G
We
cannot
really
decide
that
means
that
only
horse
decides
and
host
makes
a
decision
to
inject
more
specific
routes
in
case
of
one
of
the
airplane
sparrow,
and
another
thing
is
that
leaf
cannot
decide
locally
if
another
leaf
in
the
pair
is
alive
or
dead,
and
if
you
have
a
bunch
of
specifics-
and
we
are
sure
that
another
leaf
is
dead
and
not
an
advertising.
The
aggregate,
it's
okay,
to
suppress
that
specifics,
those
specifics,
but
if
we
don't,
we
need
to
propagate
them
and
that
decision
cannot
be
made
locally.
G
That
could
be
a
really
serious
problem,
because
if
you
consider
a
scenario
with
a
power
cycle
or
powerup
for
pod
or
even
entire
data,
centers
switches
in
the
prayer
never
comes
up
at
the
same
time
and
it
can
be
a
real
massive
injection
of
specific
routes.
So
how
it
looks
first,
consider
that
pair
and
just
couple
hosts
at
such
are
patched
and
whatever
for
on
one
of
the
links,
so
both
switches
but
originated
aggregate
for
the
pair
they
agree.
G
G
South's
reflection
allows
to
distribute
information
about
neighbors
on
the
same
level,
so
we
know
that
if
the
neighbor
is
alive
or
dead
and
what
kind
of
support
information
and
what
kind
of
reach
ability
it
has
and
it
allows
to
solve
a
lot
of
those
problems
without
some
extra
mechanism
or
trying
to
attach
finite
States
machine
to
the
BGP
or
whatever,
without
some
external
decision-making
logic.
That's
it.
C
Yeah
and
actually
it's
a
practical
output
from
having
people
lab
that
staff
and
looking
like
this
aggregation,
behavior
right
in
the
cost
of
routes,
is
that
some
people
are
asking
actually
for
knobs
were
at
between
the
leaf
and
the
pod
level,
especially
the
run
server.
They
would
like
to
have
the
desegregation
all
the
time,
because
that
singing
3
speeds
up,
you
know,
link
failure
to
the
to
the
leaf
without
actually
being
a
big
burden.
C
So
if
you
have
a
rack
with
40
servers,
it's
like
you
know
hundred
addresses,
so
they
would
like
to
have
to
Don
the
server
always
pretty
for
desegregated.
They
don't
want
to
deal
with
the
first
link
failure,
whereas
at
higher
levels,
of
course,
you
can
cause
a
lot
of
these
aggregation
all
right.
So
I
basically
tell
you
what
was
going
on
in
nothing
too
exciting
specs
done,
which
is
like
going
through
the
motions
documenting
more
and
more
so
the
village
sketches
is
another
thing.
C
So
I
just
tell
you
what's
going
on
the
review
status,
so
that's
like
the
second
round.
We
had
actually
already
a
first
round
on
the
force
stuff
done,
and
now
we
progressed
to
the
age.
So,
from
the
security
side
we
kind
of
cleared
everything
up,
which
I
mean
those
are
the
people
most
interesting
questions
so
that
that
stuff
is
all
wrapped
up
and
all
from
the
ops
inform
the
routing
year
we
had.
Basically,
you
know
issues
to
clean
up
and
on
the
Jen
art
from
sparks.
We
got
like
not
ready
and
somewhat
a
larger
discussion.
C
I
talk
about
a
lot
of,
although
I
improved
on
the
on
the
truss,
but
basically
the
beak
I'll
talk
talk
also
about
what
is
outstanding,
but
would
I
not
mention
is
that
from
one
reviewer
came
well,
instead,
the
spec
that
will
work
right,
like
it's
under
specified
by
my
feeling,
like
yeah,
so
the
argument
was
okay
we
have
to,
and
you
know
the
spec
is
not
precise
enough.
People
will
do
weird
stuff.
So
the
best
answer
we
had
was:
okay,
you
have
to
interoperable
implementations,
fully
document
that
you
know
carefully.
It
really
lasts
two
years.
C
Look
it
up.
We
have
a
circuit
implementation,
come
going
at
least
I
know
over
fourth
and
all
the
questions
these
people
ask
can
be
basically
pointed
at
the
spec,
so
I
asked
for
well,
please
please
give
me
a
measuring
stick
right.
What
you
mean
like
the
spec
is
ready,
except
your
gut
feeling,
and
nothing
came
back
so
I
assume
that
it
will
take
care
of
itself.
C
So
let's
look
what
I
cleaned
up
and
what
is
still
outstanding
and
where
I
asked
to
like
some
participation,
so
the
Yaak
of
way,
whatever
is
to
contest
that
you
just
remove
it
and
the
life
goes
on
so
the
requirements
section.
It
was
not
as
much
as
discussion
of
requirements
but
like.
Why
is
a
requirement
section
in
there
right
so
like?
Well,
we
removed
it
fine.
Why
should
we
care
is
that
extra
came
from
two
reviewers
like
one
reviewer
and
more
had
discussion?
C
If
two
reviewers
said
like
you
know,
why
is
the
requirement
section
there,
though
requirement
section
is
going?
The
SR
extension
section
is
gone
because
that's
in
a
different
draft,
so
they
asked
why
sits
stuff
overlapping
and
it's
not
complete,
so
that's
removed,
and
that
was
actually
fairly
persistent
that
it's
very
very
much
surprising,
because
it
is
very
contrary
to
what
I
heard
originally
from
people,
which
is
that
they,
like
a
narrative
right,
explain
it
better
and
please
push
the
implementation
super
detail
somewhere
off.
So
there
was
only
in
appendices.
C
C
Then
there
was
a
lot
of
very
good
actually
like
three
reviewers
people
really
chewed
through
this
one
reviewer
kind
of
you
know
ran
out
of
time
of
3/4
of
the
document,
but
two
reviewers
really
chewed
through
the
document
like
really
really
careful
reading
and
the
glossary
section
has
so
no
dag
and
ztp
I
think
there
was
a
little
bit
of
arrow,
but
then
they
hammered
on
the
three-way.
Well,
this
is
a
3-way
which
is
a
bi-directional
idea.
Sincere
I
this
is
loose
so
that
has
been
specified
out
precisely.
C
There
was
actually
a
SPF
north
and
south
split
was
not
specified,
so
that
is
now
in
glossary
properly.
There
was
something
more
so
security
and
the
law
right.
They
said.
Basically,
when
you
read
it
with
with
the
time
it
becomes
clear,
but
when
you
heat
it,
you
don't
know
what
it
is.
So
it
makes
just
lowers
readability
of
the
draft.
So
I
thought
it
was.
It
was
really
excellent.
I
mean
there
really
must
have
gone
carefully
through
his
stuff
and
thought
about
it.
Then.
C
Another
very
good
comment
was
that
with
the
anti
and
estée
people
were
confusing
anti
with
the
node
thai,
because
we
have
no
north
I
note,
South
Thai
and
the
one
writer
we
started
to
throw
this
entire
nest
ice
and
in
some
examples
people
got
confused.
So
the
whole
spec
of
the
entire
estée
has
basically
has
been
expanded
to
north
or
south
Thai.
C
It
was
it's
a
small
thing,
but
two
people
said
that
you
know
that
is
where
the
stumbled
in
a
know,
mentally
the
general
topology
they
cross,
Bart
bar
cross
bars,
which
you
know,
is
very
key,
but
obviously
not
an
easy,
read,
no
matter
how
well
we
try
to
put
it
out
there.
That
has
been
what
people
nailed
on
it.
There
was
no
terminology
section,
so
we
were
kind
of
like
introducing
the
terminology
as
we
went
along
and
a
bunch
of
variables
here
and
there.
So
that
has
been
pulled
up
to
make
for
a
simple
reading.
C
And
then
you
know
these
pictures,
ASCII
art
right
people
were
stumbling
because
it's
really
complex
stuff.
You
remember
I,
told
you
the
HH
will
confuse
the
hell
out
of
people.
Let's
make
it
an
O
right,
so
everything
is
going
to
replace
with
a
little
dot
because
people
we
have
these
2h
h
for
an
rj45
jacks
and
people
stumble
like
other
one.
Is
this
or
in
our
pictures
change?
It
sounds
trivial,
but
you
get
from
two
people
means
like
yeah.
It's
a
mental
hurdle
right
to
understand
this
back
then.
C
We
had
a
lot
of
things
where
things
were
still
called
spines,
which
is
like
a
very,
very
old
staff
which
were
really
top
of
fabrics
so
that
improve
you
know
just
properly
named
things.
The
way
we
introduced
them
on
readability,
of
example,
there
was
I
mean
that's
an
is,
is
flavor
respect,
so
we
have
these
deep
clauses
and
there
were
1
point
1
point,
2
point,
3,
/
4,
and
you
know,
after
discussion
with
a
reviewer
eat
much
better
to
like
change
the
numbering
which
to
XML
to
RFC,
allows
you.
C
So
you
have
like
you
know,
number
alphabet
Roman
something
right,
so
it
makes
for
like
easier
discussion.
It
doesn't
needs,
but
I
mean
ultimately
this
thing.
If
it
comes
multiple
times
it
will
help
people
I
mean
not
a
big
price
and
then
lots
of
stuff
on
the
normative
so
stuff,
like
UI,
6,
T
4.
You
know
those
discussion,
I
mean
the
spec
I
agree.
C
C
What
fell
out
was
an
interesting
discussion
and
there's
more
like
open
to
the
group.
I
mean
we're
very
far
where,
like
in
the
last
things
but
taste,
so
we're
not
an
immovable
think
the
question
was
brought
up
or
do
we
should
introduce
something
into
the
schema
to
support
the
BFD
strict
mode.
So
I
put
a
clarification
in
how
the
PFT
is
supposed
to
work,
because
that
wasn't
clear
it
wasn't
breaking
the
spec,
but
people
could
do
different
things,
so
there
would
be
mildly
unexpected.
C
Top
I
don't
know
it's
a
good
I,
don't
even
know
what
the
strict
mode
has
to
be
normative
right
so,
like
you
can
safely
ignore
it.
Everything
will
work.
But
if
you
don't
have
to
indication
you
don't
know
whether
the
other
guy
desires
and
strict.
So
like
you
have
no
signaling
you,
you
may
ignore
the
signal
by
you.
Don't
have
to
signal.
You
cannot
know
whether
you're
willing
to
use
it
or
ignore
it.
Just.
C
A
separate
draft
on
strict
BD
in
read
yeah,
but
you
unique
the
information
element
to
indicate
that
you
want
it
strict,
which
would
basically
necessitate
a
minor
schema
revision
and
since
we
know
that
RFC
da
to
do
with
just
one
little
field,
so
what's
the
big
deal,
I,
don't
know
just
a
question
to
the
room
and
after
you
know,
the
reviewers
went
through
the
stuff.
You
know
with
pincers
then,
and
that's
a
very
interesting
discussion
for
Alvaro
now
I.
C
Basically,
the
spec
has
somehow
to
pour
the
schema
into
registries
right
because
that's
what
we
have
and
our
registry
are
only
built
for
maybe
one
small
space
with
200
numbers.
This
schema
is
generating
pages
and
pages
pages
of
requests
right,
because
every
schema
element,
if
you
want,
is
its
own
namespace
and
it
has
elements
inside
and
each
of
them
have
a
reserved
value,
possibly
right.
So
the
first
thing
is,
you
have
the
schema
element
I
wish
I
would
have,
and
so
now
you
have
a
structure.
C
Something
and
you
like
GRP
see
every
element
is
numbered,
so
you
have
to
reserve
those
elements.
You
have
say:
I
need
a
namespace
for
whatever
tie
an
entire
I
have
four
different
elements
and
they
have
numbering
one
two,
three
four.
So
we
have
to
reserve
that
and
then
within
that
their
structures
again
right.
So
you
start
to
generate
pages.
I
mean
knowing
me
I,
just
wrote
a
small
schema
compiled
on
bank.
How
he
comes
right
here
is
all
my
requests
now.
He
said
the
right
thing
to
do.
C
A
C
So
here's
observation
brew
noise
is
always
almost
right.
He
can
talk
and
that's
good.
So
there
is
a
big
value
to
the
registry
I
see,
and
you
will
see
that
when
so,
let's
say
we
have
this
schema
version
two
and
now
we
go
to
version
3
so
like
yeah
okay,
so
we
have
new
set
of
registry
for
version
3
right,
but
mostly
we
go
miners
like
let's
add
a
field.
C
It's
comes
with
two
ones:
oh
I
know:
oh
I
can
go
and
look
at
this
field
and
it
comes
with
two
zero
and
now
I
know
I
don't
have
to.
But
now,
if
you
just
have
the
previous
document
with
your
schema
to
zero
and
then
another
documents
get
probably
with
two
one,
you
have
to
run
a
diff
and
hope
for
the
best,
whereas
the
registry
is
very
normative.
It's
very.
You
can
look
at
the
whole
thing
and
see
which
version
supports.
You
know
which
field
so
I
think
there's
a
lot
potentially
lot
of
value.
C
Having
registries
running
this
way,
no
because
the
histo
histo
Racal
track
is
then
very
normative,
very
well
controlled,
and
what
also
that
can
of
course
happen.
So,
let's
say
you
have
the
schema
version.
2
I'd
really
are
completely
independent
of
riffed.
If
we
go
into
now
schema
stuff,
that's
all
just
price
of
living
right
for
luxury,
of
avoiding
90%
of
the
dumb
work,
we're
doing
manually
that
the
schema
compiler
does.
For
us.
This
is
the
price.
C
So
the
other
thing
that
the
registry
will
avoid
is
that
we
have
a
document
published
to
zero
right
and
now
two
extensions
come
up
like
two
eyes:
I
steal
these
same
thing
and
they
basically
start
to
extend
the
schema
and
they
go
like
to
one
and
the
other.
One
is
to
one
or
even
two
too,
but
they
can
collide
by
taking
a
code
point
on
the
structure
right.
C
So
if
you
don't
have
anything,
you
may
just
happily
progress
those
documents
and
extend
the
schema
and
everything
hits
the
fan,
even
publish
them
and
everything
hits
the
fan
where
people
named
implemented,
they
say:
do
documents,
look
the
same
code
point
right,
even
if
the
schema
they,
even
if
they
talk
to
each
other
and
they
realize.
Are
you
thinking
two
one?
Oh
you
better
take
two
right
now,
the
two
does
it
imply
that
one
is
in
it?
Well,
maybe
not
even
right.
C
So
now
you
have
to
synchronize
the
guys
who
gonna
do
that
they
may
talk
to
each
other
if
you're
lucky,
but
if
they
meet
on
the
registry,
and
they
have
to
ask
okay
I'm
coming
with
the
one
with
this
field
and
the
other
people
tell
them
no,
not
first,
two
ones
already
taken
and
second,
this
field
is
possibly
the
same
value.
Then
you
cannot
devolved
its
utilizes
the
schema
evolution
right
because
the
schema
the
beauties
in
like
eyes,
I
steal
these.
C
Multiple
people
can
progress
it,
but
ultimately,
when
multiple
people
progress,
eyes,
eyes
extension,
they
have
to
go
to
the
registry
and
ask
for
the
number
and
if
the
number
collides
it's
a
no-go
and
here
the
registry
would
play
the
same
serialization
right.
It
will
make
sure
that
the
schema
version,
progress
in
a
linear
fashion
and,
second,
that
you
will
not
have
a
collision
of
the
code
points.
So
I
think
there's
value
I.
A
D
C
C
A
C
C
Alright,
so
I
just
set
it
out
there
in
the
room.
That
was
what
was
going
on
then
a
very
good
comment:
someone
read
the
math
section
of
the
sequence
number
and
realized
that
I
was
full
of
it
when
I
called
in
all
the
relations
symmetric,
so
they
needed
fixing.
So
there
is
the
two
to
do
store
out
there.
So
one
observation
was
that
it's
very
cute
that
I'm
giving
them
a
dot
thing
and
you
can
run
into
a
dot
compiler
and
look
at
the
FSM.
C
But
it's
not
like
you
know
the
most
beautiful
thing
in
the
world.
We
have
the
version
3
and
exactly
one
RFC
published.
So
it's
like
going
because
I
think
we
looked
at
it
six
months
ago.
It
wasn't
there
yet
so
I
have
to
put
the
SVG
in
which
we
have
right
off
the
FSM
and
publishes
as
as
version
3
in
PDF
with
the
pictures.
So
that's
something
isn't
it.
The
second
one
is
basically
simplify
the
language
right.
C
You
know
that
the
famous
thing
like
so
the
same,
who
said
the
speck
it
maybe
underspecified
heads,
said
that
it's
too
much
text.
Maybe
it
should
be
simpler.
So
it's
like
the
the
Mozart
movie.
You
know
where,
after
the
opera,
the
Emperor
tells
Mozart.
You
know
too
many
notes.
A
great
thing
was
just
too
many
notes
right,
so
we,
but
we
have
to
tie
it
in
the
language.
I
mean
I'm
sympathetic
with
that,
which
means
probably
shorter
sentences.
Smoothing
over
the
declaration.
C
You
know
the
the
some
of
the
descriptive
text,
volunteers,
highly
desire
right
and
it's
basically
just
reading
and
just
smoothing
the
language
into
more
readable
text.
It
wasn't
like
it's
factually
wrong
or
you
know
it's
very
unclear.
Some
of
the
sound
like
can
you
at
this
and
the
other,
but
basically
smooth
this
thing
for
easier
reading
sentence
is
too
long,
maybe
no
little
bit
on
complexity.
Sometimes
you
know
context
may
be
better
clarified.
So
that's
just
no!
H
D
Editorially
true,
the
RCO
does
a
wonderful
job
all
the
time.
However,
we
need
document
to
be
readable
and
reviewable
by
a
bunch
of
other
people,
so
I
would
prefer
if
it's
Jeff's
just
before,
if
this
was
done
before.
In
other
words,
we
don't
want
to
just
leave
everything
to
the
editor,
because
you
know
they
have
a
lot
of
work
done.
Hundred
and
fifty
four-page
spec
writes
and.
C
I
also
think
it's
non-productive
right,
because
what
you
want
is
sit
there
and
have
someone
read
and
ask
question
and
smooth
it:
it's
not
like
you
throw
it
over
to
the
editor.
It
comes
with
a
bunch
of
question.
He's
like
hungry,
140
pages
right
so
I
mean
I,
got
these
comments
and
it
wasn't
like
it's
unusable,
it's
more
like
well,
no,
it
could
be
smoother
all
right
sentences,
sometimes
too
long,
edging
on
here
get
on
the
grammar
look
at
stuff.
C
C
A
C
C
Look,
for
example,
I
realized
this
one
RFC
published
in
version
3
right
the
one
and
he
doesn't
have
any
picture
so
like
yeah
yeah.
Exactly
so
then
I
started
to
look
at
Wade's
document
and
how
you
put
a
picture
in
and
it
could
be
the
first.
Yes,
it
looked
fairly
arcane,
but
you
know
my
mana.
If
you
know
good
well,
a
well
meant
assumption
was
that
the
stuff
actually
works
well,
I
think
I'll
be
pushing
the
curve
and
Bruno
does
the
ASCII
art
we
see
where
we
end.
C
Ok
and
don't
fine-tooth
comb
on
the
normative
that
more
stuff
needs
to
be
normative.
I
got
one
two
examples
which
were
correct
right,
but
it's
again
reading
and
thinking
through
the
stuff:
that's
probably
less,
for
people
who
just
enjoy
you
know
and
English
and
smooth
English,
that's
more
the
people
who
are
actually
well
in
the
protocol
for
a
bit
because
it
can
lead
to
discussion,
because
the
protocol
is
because
it's
very
free
routing
is
extremely
permissive
compared
to
normal
stuff.
We
allow
tons
of
stuff
because
well
do
what
you
want.
You
can
break
it
like.
C
You
know,
replace
your
algorithm.
What
do
we
care
right?
Which
normal
protocols
don't,
but
there
were
a
little
bit
of
discussions,
but
that's
pretty
much
the
output
from
them
from
the
reviewers,
so
basically
I
think
that
I
addressed
except
the
Jen
art,
which
was
Sparks
and
possibly
I.
Think
the
ops
dear
gave
me
some
of
these
readability
comments
which
I
have
to
address
yeah
John.
F
You
Tony
Jones
cutter.
It
occurs
to
me
that
since
you
spent
so
much
time
intimidating
people
by
mentioning
the
page
count
on
your
document
and
then
saying
that
you
want
detailed
review
lots
of
these
appendix
don't
worry
and
since
you're
so
good
at
parallelism.
Maybe
you
know
it
might
work
I
mean
it's.
F
You
get
a
certain
quality
of
review
this
way,
but
to
the
extent
that
what
you're
trying
to
do
is
smooth
out
sentence
as
it
could
work
as
you
could
farm
out,
you
know
sections
to
different
reviewers,
so
maybe
I
don't
have
time
to
do.
150
pages
for
you,
but
maybe
20
pages
would
be
pretty
easy
to
a
certain
extent.
I.
C
Think
the
introductory
I
mean
so
I
have
been
misled
by
this
group.
Who
was
originally
very.
You
know,
insistent
that
they
like
in
their
ativ
rather
than
the
drawers
back
and
then
their
ratings
do
not
paralyse
that.
Well,
but
you
know
they're,
the
narrative'
is
the
front,
so
I
don't
think
that
can
be
practically
split
up,
especially
it
has
to
like
lead
one
into
the
next.
C
You
have
to
understand
why
you
desegregate,
and
you
understand
what
is
a
Fallen
Leaf
and
why
you
need
negative,
but
from
their
own
yeah
like
ztp,
could
be
probably
just
completely
ripped
out
and
reviewed
separately.
John.
Have
you
just
volunteered
to
to
review
it
hugely
without
a
commitment
to
a
date.
C
C
I
A
G
G
It
stopped
working
okay,
so
now
going
to
the
applicability
draft.
Basically,
we
are
talking
about
updates
since
the
previous,
a
Jeff
and
next
step.
So
what
was
changed
and
mostly
not
changed?
Most
landed
as
well
escaping
examples
please,
especially
with
miss
cabling,
between
layers
ipv4
over
ipv6
fabric,
no
home
in
summers
fabric
with
a
controller
we
scrapped
off
options
where
to
attach
it,
submit
mismatch
and
addressed
feminists
and
any
cost
consideration
next
light.
Please.
G
So
here
we
mostly
have
diagrams
from
the
applicability
draft,
and
from
is
cabling
we
consider
interaction
between
ztp
in
particular,
and
not
only
but
also
ztp
and
mascara
in
between
levels,
which
is
in
rate
apology.
So
what
kind
of
miss
cabling
can
we
have
and
how
pruning
works
in
that
case,
because
if
it's
not
part
of
well,
it
apologies
that
link
should
be
pruned
next
place
before
will
be
six.
G
So
basic
idea
is
drifted
in
advertising
before
prefixes
and
using
v6
network
same
as
our
C
55
for
49,
for
example,
so
using
online
v6
topology
and
using
it
as
next.
Hopefully,
for
and
it's
expected
that
Intel
fabric,
you
supports
at
the
same
time
of
just
families,
but
it's
possible
to
indicate
what
kind
of
rotors
families
no
supports.
So
it
is
possible
to
have
two
different
topologies
for
different
addressed
families.
On
top
of
the
same
physical
topology,.
G
G
Also
was
that
it
subnet
mismatched
and
addressed
families
and
next
slide,
please
so
any
cost
considerations
and
the
reference
to
the
policy
guided
practice
prefixes,
because
normally,
if
you
have
any
caste
and
look
from
the
top
of
fabric,
it's
essentially
multi
pass
route.
There
is
no
way
to
sit
in
a
particular
node.
We
have
no
way
one
for
them
to
disk
in
Russia
if
it's
just
multi
pass
or
single
destination
or
E
or
any
caste,
but
there
can
be
well
it
situation
and
we
might
want
to
specifically
guide
traffic.
G
For
example,
if
we
inject
traffic
from
some
leaf
nodes-
and
we
have
any
kind,
any
cars
destinations
behind
some
other
leaf
nodes,
we
essentially
want
to
distribute
it
to
do
Rio
anycast,
and
if
there
is
another
leaf
with
any
cost
destination
attached
within
the
same
port,
we
will
never
reach.
They
remain
in
nodes
normally
because
it
will
be
shortcutted
inside
the
port.
We
don't
want
to
that
and
to
avoid
it
there
is
an
option
to
use
policy,
guided
practices
and
I
guess
that's
it.
Yeah
want
comments.
G
H
G
H
Yeah
and
that's
because
we
actually
put
in
the
rib
the
the
information
from
the
pod
itself,
and
we
had
the
discussion
about
tunneling
a
long
time
ago,
and
maybe
you
will
resurrect
those
ideas
in
a
separate
draft,
but
there
might
be
cases
where
are
you
in
the
in
the
top
of
that?
You
don't
want.
You
learn
this
thing
from
the
flooding,
but
you
actually
don't
put
it
in
your
writing
table.
H
You
just
pass
it
up
sort
of
yeah,
and
if
you
do
that,
then
you
get
the
kind
of
results
you
want
right,
because
because
you
don't
have
the
rot
self
to
the
unicast
address,
there
are
cases.
For
me,
the
case
was
more
mobility.
If
things
move
very
rapidly,
you
might
want
send
them
to
the
tough,
because
maybe
I
need
a
tough
as
a
real
state
of
where
things
are
yeah.
I
G
C
G
C
All
right
I'll,
if
you
know
the
cage,
so
that
shows
you
some
of
the
configuration.
So
what
is
the
v4
of
a
v6
thingy?
So
basically,
what
happens
is
that
if
you
come
up
and
you
only
have
v6
and
D
right,
you
only
have
link
local
addresses
and
that's
it
and
what
riff
does
for
you.
It
indicates
on
the
link
whether
the
link
is
v4
forward
incapable,
because
otherwise,
how
do
you
know
what
they
can
do
v4
over
v6
or
you
can't
alright.
C
So
in
case
you
know
it
is
v4
capable
and
you
have
to
link
locals
and
nothing
else.
But
it's
of
course,
how
you
optimize
your
implementation.
You
know
you
may
be
very
smart
about
what
you
have
v4
next,
so
v6
link
law
calls.
But
let's
say
you
know
you
have
a
link
loco
and
you
won't
before
and
nothing
else.
Then
what
you'll
have
on
your
side
of
the
implementation?
You
will
have
v6
table
which
is
resolving
Linley
a
link-local.
You
know
the
B
to
the
Mac
and
that's
really
everything
that
you
need.
C
C
So
that
is
what
section
of
the
black
ability
document
talks
about
now.
Do
you
have
to
do
it?
That
way,
no
I
mean
the
implementation
is
fairly
lengthy,
ain't
right,
but
it's
like
a
technique
that
you
can
use
that
makes,
especially
since
we've
forcing
CTP
so
much
on
the
protocol.
That
is
such
high
demand.
C
C
So
that's
well
here:
that's
what
I
omitted
right!
You
get
from
node
B
when
you
get
the
lie
with
the
v4,
capable.
That's
also
how
you
know
what
his
address
is
right,
source
address
and
then
you
can
resolve
it
to
the
Mac
that
you
use
it
as
his
gateway
to
point
as
to
him
as
a
gateway.
But
in
this
case
something
funky
can
have
happen,
people
can
change
configuration
and
all
of
a
sudden,
the
v6
address.
C
So
if
you
really
want
to
rectify
this
condition,
you
really
have
to
clear
digestion,
see
I
mean
you
can
just
give
an
alarm
and
say:
look
the
v6
got
an
address
mismatch
and
the
other
thing
that
next
one
Jeffries
sorry
I
really
have
to
be
in
front
and
point
to
it.
So
this
is
some
of
the
stuff
that
we
talk
about
the
loot
bags
for
controllers
and
controllers
consideration,
and
basically
it
all
boils
down
to
how
much
of
internal
reach
ability
do.
C
You
want
to
hit
a
specific
point
on
the
fabric.
Somehow
so,
if
you
start
to
look
at
the
loopback
reachability,
so
when
you
go
from
leaf
to
leaf
the
green
stuff
link,
failures
present
no
problems,
you
just
you,
you
pop
up
the
default
right
and
even
if
it
isn't
desegregated,
it
means
the
default
works,
because
otherwise
referal
desegregate
and
guide
you
the
right
way
and
once
you
hit
the
top
of
the
fabric,
the
guy
knows
precisely
which
ways
are
feasible
to
get
you
to
the
leaf
so
leave
to
leave.
C
No
special
consideration
leave
to
leave
loop
X,
always
work
right
without
actually
on
a
seeing
B
loopback,
it's
unnecessary,
you
just
send
towards
B,
and
the
fabric
will
figure
it
out.
But
if
you
want
to
go
from
bottom
to
top,
things
starts
to
get
interesting
right.
So
that's
kind
of
the
North
normal
case
and
some
people
like
that
for
bazillion
reasons.
Now
imagine
someone
wants
to
do.
Egress
engineering
on
the
fabric
because
they
go
out
at
the
top
of
people.
Does
that
about
half?
C
C
So,
if
you
want
that
to
work,
then
you
have
to
push
the
loop
back
of
the
node
one
South,
which
is
no
big
deal
right
with
just
like
you,
advertise
D
for
out
you
just
ad
is
more
specific,
all
right.
It's
not
desegregation,
because
we
have
special
ties
for
this
aggregation
right
for
computation
purposes.
It's
just
you
D
for
doesn't
have
to
be
a
default.
Actually
the
applicability
trough
talks.
When
is
it
that
you
really
don't
want
to
advertise
the
D,
for
you
want
to
advertise
the
internal
fabric
default
right
and
it
goes
to
its.
C
For
this
stuff
really
to
work
with
link
failures,
I
mean
that's,
that's
just
an
example.
What
goes
into
applicability
draft
in
details
right
and
how
you
achieve
that?
You
can
do
PGP.
You
can
do
multiple
mechanism,
but
that's
basically
how
you
deal
with
the
internal
rich
ability
right.
Do
you
want
to
come
from
north
to
the
south?
You
know.
No.
Do
you
want
to
go
from
the
bottom?
To
top,
do
you
will
need
loop,
X
and
so
on?
So
that's
just
something
to
flash
it
out.
A
little
bit
makes
it
work.
C
G
Really
a
question,
but
a
male
comment
when
we
inject
in
specific,
are
basically
horse
trout's
on
the
back
routes
for
top
of
fabric
notes
and
propagating
them
down
it's
much
easier
on
the
hardware
resources,
because
all
those
routes
are
essentially
unipass.
It's
they're
not
consuming
it
simply
resources.
So
it's
certainly
not
difficult
to
handle.
Well.
C
G
I
C
Is
as
wide
as
you
fan
out
in
the
port
right,
yeah,
okay,
yeah,
also
that
you
concerned
yeah
yeah.
It
would
be
like
port
file,
okay,
so
yeah,
so
that's
I,
assume
this
is
applicability.
Dropped
will
actually
take
another
at
least
two
revisions,
because
more
and
more
stuff
comes
up
the
more
people
bang
on
it
and
do
lap
things
and
think
through
the
stuff
you
know
just
which
I
think
is
exactly
what
Alvaro
desired.
You
know,
because
we
have
been
called
the
marketing
document,
but
you'd
still
this.