►
From YouTube: IETF106-IEPG-20191117-1000
Description
IEPG meeting session at IETF106
2019/11/17 1000
https://datatracker.ietf.org/meeting/106/proceedings/
A
C
B
All
right,
well,
okay,
anyway,
I
can
only
do
this
one
hand
okay
good
morning.
This
is
the
IEP
G
at
IETF
106
in
Singapore.
If
you're
not
supposed
to
be
in
Singapore
I'm.
Sorry,
you
took
the
long
flight
if
you're
not
supposed
to
be
the
IEP
G
and
you're
supposed
to
be
at
some
other
meeting,
then
you're
in
the
wrong
room.
You
should
consider
moving
or
you
could
stay,
because
it
will
probably
be
fun
for
those
people
that
are
presenting
there's
a
microphone.
Please
don't
please
speak
into
the
microphone.
B
B
B
D
D
We
call
it
fort
because
it
made
sense
at
the
time
and
it
says
something
about
the
stronger
internet
or
something,
but
we
just
call
it
fort
and
we
try
to
not
think
very
much
about
what
it
means.
So,
there's
a
background
to
this
there's
a
fort
project
and
the
four
validator,
the
full
project
started
as
a
initiative
to
promote
route
and
security
in
the
region.
There's
there
was
a
call
from
the
open,
Technology
Fund,
which
is
a
fund
that
finances
technology
projects
having
to
do
with
freedom
on
the
internet.
D
The
idea
is
that
money,
the
tracking
tool,
will
provide
information
about
writing
incidents
for
non-technical
people
under
the
hope
that
it
will
help
people
realize
whether
their
country
government
is
PE.
Whatever
is
actually
you
know,
playing
tricks
on
them
next,
please.
So
what
I'm
going
to
talk
about
today
is
the
Arabic.
A
validator
with
a
said
is
one
of
the
two
parts
of
the
project.
Next,
please
why
another
our
PKA
validator,
you
know
lately
it
seems
to
be
that
everybody
is
developing
a
new
epoch,
a
validator.
D
There
is
at
least
four
projects
that
I
can
remember
about,
and
I
think
it's
good
having
diversity
is
good.
Having
programming
language
is.
Diversity
is
good,
and
it's
also
worth
mentioning
that
back
at
the
time
when
we
decided
to
move
forward
with
this
project.
Well,
the
options
about
RPK
validators
were
rather
limited,
so
it
may
not
send
a
lot
of
sense
at
the
time
to
create
a
new
one
sex.
Please
so
short
description
about.
Fourth,
what
is
implementing
right
now
is
repository
validation
and
the
RTL
server.
D
There
is
no
GUI,
there's
no
web
there's,
no
nothing.
So
it's
very
silly
something
you
run,
and
hopefully
you
don't
get
anything
on
the
console
right,
so
it's
rather
an
exciting.
In
that
sense,
it's
written
in
C,
which
is
you
think,
it's
great
I
mean
I,
didn't
know,
people
still
read
strong
programming
in
C,
I
mean
people,
I
thought
everybody
had
moved
to
go
and
rust
and
Python,
and
you
say
that
it
compiles
cleanly
and
runs
on
several
Linux
and
BSD
platforms.
The
license
MIT
so
basically
is
almost.
D
D
You
can
check
the
details
on
that
URL
there's
a
very
nice
table
with
all
the
LLC's
and
the
percentage
of
implementation,
but
basically
some
of
the
things
that
I
consider
interesting
in
is
validated
it
that
support
this
learn
the
support
they
say
some
files
from
the
RFC
it
supports
validation,
reconsider
validates,
goes
faster.
It
supports
policy
qualifiers
and
has
some
support
for
BGP
SEC
objects.
What
is
missing,
notably,
is
a
DB
support
and
the
HTTPS
tiles
hopefully
coming
in
version
1.2
in
a
few
months.
Next,
please,
it
has
reasonable
high
test
coverage.
D
This
is
some
of
the
platforms
where
the
tests
have
been
run.
I
haven't
tried,
compiling
it.
Obviously,
but
I
guess
it's
the
same
thing
as
always:
it's
very
easy
to
install
it
on
Divya
no.12.
Next
one,
please,
the
usual
routine
right.
You
install
some
dependencies.
You
unpack
the
Col
configure,
make
make
install
and
you're
running
I
created
a
couple
of
docker
images.
You
can
try
if
you
want
next
please,
and
there
are
two
ways
of
running
it.
One
is
what
the
unique
only
documentation
is
called
standalone.
D
What
is
basically,
it
will
run
validate
a
repository
or
a
set
of
repositories,
produce
some
CSV
output
and
exit,
no
fancy
web,
no
graphics,
nor
anything
and
the
other
way
of
running
it
is
with
a
survey
with
the
LTL
survey
enabled,
and
it
will
just
stay
there
and
the
only
thing
you'll
hear
from
it
is
where
it
finds
errors
and
or
some
inconsistencies
in
reporters
in
the
repositories.
These
are
the
basic
options
it
has
tons
of
options.
D
Some
of
them
are
quite
interesting.
For
example,
you
can
tune
a
little
bit
how
strict
the
validation
it
performs
it
performs,
which
you
think
is
kind
of
nice,
because
there
are
still
some
I
would
say,
divergences
in
the
implementations
of
the
repositories.
Next,
please
a
few
tests
they
run.
This
is
memory
consumption
running
on
a
standalone
mode.
These
are
the
M
dark.
Blue
is
a
penis
repository,
light
blue
slag
neck
and
the
sort
of
yellowish
greenish,
something
color.
D
There
is
a
Phoenix,
they
repository
it's
quite
low,
it's
quite
efficient
in
memory
usage
which
is
kind
of
nice
and
it's
I
would
say
natural
by
proud
of
it's
been
written
in
plain
C
text.
Well,
thank
you.
If
you
like
the
idea
of
having
a
new
tool
in
your
tool
chest,
please
send
us
your
box
I'm,
not
promising.
This
is
the
great
thing,
the
greatest
thing
since
sliced
bread
nor
the
disciple
free
is
probably
full
of
bugs.
D
E
Randy
Bush
our
Cassandra
EJ.
How
do
I
deal
with,
inter
instance,
slurm?
In
other
words,
I,
have
a
slurm
file
on
my
in
my
cop
in
San,
Diego
and
I.
Want
you
to
run
it
in
your
pop
in
San
Francisco?
What's
the
trust
model,
that's.
D
D
D
F
F
Good
morning
my
name
is
arianah
from
the
VRS
today,
I'd
like
to
introduce
our
research
work
to
detect
the
TMS
ik
provocation,
failure
at
the
TRD
server
monitoring
next,
please!
So,
as
you
know,
DNA
SEC
validation
failure
happens
frequently,
but
the
reliability
of
DNA
six
operation
is
a
key
of
the
nsx
deployment,
so
so
that
next,
please
we
made
a
research
question:
how
can
we
detect
validation,
failure
rapidly
and
efficiently?
F
Of
course,
active
amazement
is
an
easy
way
to
detect
DNA
segmentation
failure,
but
it
causes
high
higher
loader
for
more
DNA
SEC
available
domains
so
that
we
thought
about
how
can
we
detect
the
NSX
validation
failure
by
the
passive
measurement?
So
we
focused
on
the
to
see
the
change
of
curly
patterns
could
be
a
good
indicator
at
the
monitoring
of
power.
I'm,
sorry,
authoritative,
nameservers,
Rabia.
Next,
please
so
to
us,
detecting
validation,
failures
with
passive
measurements.
F
The
goal
of
our
research
is
depicting
chant,
detecting
change
of
clearly
patterns
at
CCTV
servers.
At
this
moment
we
are
the
GP
operators,
so
we
are
seeing
JP
DNS
servers
before
and
after
failure
when
Colin
tomahawk
is
using
both
octave
measurement
and
passive
measurement
for
the
active
measurement
we
are
using
ripe
address
for
the
DNS
client
and
the
else
worried.
F
Dietas
neighbor
to
the
iPad
rests
approximately
500
at
ISPs
on
edges,
and
also
we
prepared
local,
authoritative,
DNS
servers
which
provide
both
successful
DNS,
SEC
validation
and
failed
as
DNS
expeditions
with
different
TTL
settings
and
passive
measurement.
Vantage
point
is
JP
DNS
servers.
Next,
please,
this
figure
shows
the
measurement
ababeel.
F
The
measurement
period
for
this
research
is
equally
2018.
This
is
also
conjunction
with
the
idea
happened
on
April,
20,
80s
and
then
passive
measurement
about
age
point
is
JP
t
JP,
DNS
servers
and
active
measurement
at
the
local,
authoritative
servers
at
the
light
bottom
light
shark
is
a
local,
also
tilting
Sabah
provided
Saudi
Arabia
of
Jupiter
by
name
and
the
targeted
queries
from
light.
A
Taurus
probe
n
is
a
DNS
key
and
the
DES
records,
and
we
use
around
thousand
life
matters
probe
and
six
hundreds
of
DNS
SEC
marketers
at
home.
F
Search
for
the
Public
DNS,
because
we
do
not
know
the
concrete
behavior
of
Public
DNS
and
we
targeting
the
body
heaters
at
ISPs
on
homes
so
that
we
sent
light
probe
at
was
unique,
queries
to
our
associative
servers
and
check
the
source
address
of
the
I
curries,
which
paragate
parties
these
robots,
sent
to
Korea
to
our
local
service
and
check.
If
it
is
in
the
public,
dns
servers
IP
address
or
not
so
that
we
find,
which
type
a
terminal
node
uses
the
Public
DNS.
F
We
exclude
the
excluded
those
try
Patras
probes
from
our
measurement
next,
please,
and
we
also
measure
the
TTL
distribution
in
signed,
JP,
DNS,
JP
Zahn's
prior
to
our
measurement.
So
we
can
find
four
categories
of
the
TTL
settings
for
NS
and
DNS
keys.
This
quadrant
shots
the
classification
of
the
lengths
home,
NS
and
DNS
key
details,
so
we
don't
know
why
this
quadrant
happens
in
Japan,
but
this
is
very,
very
typical
to
some
industries
next,
please!
So,
according
to
these
TTL
settings,
we
set
up
the
eight
zones
for
this
test
next,
please
so.
F
So
this
table
shows
the
result
of
queries
that
we
observed
at
the
local,
authoritative
server,
that
is
three
laboratory
and
at
UC
when
DNA
segmentation
failed
them.
Overall
queries
increased
in
for
DNS,
SEC
related
queries,
so
the
left-hand
side
is
a
successful
validated
successful
domain
names
and
the
right-hand
side
is
a
varied
zones.
F
F
F
So
this
table
shows
the
comparison
with
local.
Also
rotated
that
is
heard
rate
was
a
third
rail
is
rabbit
and
Jade
JP
ccTLD
servers.
So
you
can
see
the
huge
nation
of
of
the
bubble
queries.
So
if
we
can
see
the
local
circuit
of
services,
we
see
much
more
queries,
but
we
still
see
some
increase
meant
on
the
TLD
servers,
but
it
is
suffering
to
say
that
the
validation
failure
in
minor
domains
are
likely
to
difficult
to
detect
next,
please
so.
F
Failure
next,
please,
so
that
we
still
continuing
this
work
because
we
have
still
some
unknown
issues,
so
that
was
one
of
the
issues
difference
in
public
and
as
a
reservist,
Public
DNS
and
as
a
reservist,
we
will
have
different
behaviors,
the
Public
DNS
of
a
very
large
and
very
good
operation
and
the
quantify
of
ccTLD
that
is
top
level
or
second
treble
was
horribly
attenuation,
because
I
am
a
TD
operator.
So
I
am
watching
over
monitoring.
F
We
want
to
have
a
victim
cause
in
real
time
detection
method
of
theory
disturb
aside,
because
we
are
monitoring
to
two
out
of
seven
JP
DNS
servers
almost
real
time,
so
that
we
can
find
which
I
used.
Those
two
servers,
as,
on
the
vantage
point,
to
find
detect
of
DNS
SEC
failure.
So
I'm
very
welcome
here
any
comments
regarding
to
research
work.
Thank
you.
G
Hi
everyone,
my
name,
is
Ray
Adams
before
I
start
on
this
topic,
a
disclaimer
I
work
for
ICANN
and
all
I.
All
opinions
expressed
here
are
my
own,
not
necessarily
I
cans.
Here's!
What
I
make
that
clear.
I'm
gonna
talk
about
private
space
names.
A
lot
of
you
are
thinking
now
not
again.
Well
again,
this
private
name
stuff.
We
talked
about
him
to
know
we
talked
about
dolls.
Can
we
just
move
on
well
next
slide.
H
G
So
if
you
look
at
the
solution
space
currently
there
are
a
lot
of
solutions
out
there.
People
suggest
that
you
register
your
own
name.
Well,
that's
not
really
private!
Before
I
go
on.
Let
me
explain
a
little
quickly
what
I
mean
with
private
use.
It's
a
it's
a
it's
a
name
that
you
can
use
on
your
own
private
network,
media
corporation,
be
it
your
home
network,
etc,
etc.
Another
argument,
just
don't
don't
don't
do
it.
G
I've
often
heard
that,
but
that
also
really
doesn't
work
used
us
internal
or
good,
local
or
dot,
alt
or
dot
home,
but
I'm
not
really
gonna
talk
about
at
home
or
home
dot.
Our
plan
and
the
problem
is,
there
is
no
defined
way
of
doing
this,
and
the
result
is
something
you
see
on
screen
here.
These
are
the
results
of.
G
Results
of
I,
thi,
I
teach
knives,
identify
technology,
health
indicators
and,
in
short,
is
the
stuff
that
we've
observed
at
the
L
root
or
iMRS
root
server
at
the
top,
the
little
13
and
names
that
do
not
exist,
that's
being
queried
for
create
for,
and
you
can
see
it
all
over
the
place.
There's
nothing
wrong
with
these
strings
I'm,
just
saying
that
we
observe
that
these
things
leak
and
they
do
not
exist
in
the
root
zone
and
next
slide.
Please
so
I'm
going
to
talk
about
a
proposed
solutions,
place
constraints.
G
So
now
that
we
understand
the
problem
space
here
are
some
constraints
in
the
solution
space.
We
need
a
simple
and
concise
BCP,
the
BCP
that
basically
says.
If
you
want
to
do
this,
this
is
your
way
of
doing
it
and
we
need
to
have
a
label
in
there
that
has
no
semantic
meaning
what
I
mean
with
semantic
meaning
is
like
it
shouldn't
have
a
literal,
meaning
like
internal
or
private
or
alt,
or
local
or
home,
or
anything
for
that
matter.
G
Because
the
moment
you
choose
private
or
internal
folks
have
an
argument
when
they
say
well
that
doesn't
really
cover
my
home
network
or
my
private
news
network
and
remember
the
X,
my
headers,
that
I
talked
about
before
right
or
in
language
text.
You
have
X,
there's
enough
semantic
headers,
they
have
no
literal
meaning,
so
we
need
something
that
has
no
semantic
meaning,
and
we
also
another
problem
with
these
restraints.
Are
they
to
angular
funny?
G
They
don't
work
in
German
or
in
Dutch,
and
it
also
smart
to
choose
a
label
that
will
never
never
be
delegated
and
we
also
need
to
choose
a
label
with
label.
I
mean
a
top-level
domain
label,
choose
a
label
that
may
not
require
our
C
67
61,
this
the
exceptionally
reserved
list
of
top-level
domains.
So
maybe
we
can
just
use
a
two
character:
ASCII
domain
next
like
this.
G
This
is
a
potential
political
minefield
to
leather
askew
domains.
So
you
can.
You
can
kill
me
later
in
the
Mike
line,
our
C
1591
talks
about
domain
name,
system,
structure
and
delegation.
There's
this
infamous
section
of
famous
section
that
talks
about
country
codes.
The
Ayana
is
nothing
the
business
of
deciding.
What
is
what
is
not
a
country,
and
the
next
section
talks
about
the
ISO
3166
list
and
keep
in
mind
that
the
next
section,
and
actually
the
entire
document
does
not
talk
about
countries
at
all.
G
It
talks
about
country
names
and
it
talks
about
entities,
but
most
countries-
and
so
I
saw
three
one
six
six
lists:
they
are
basically
the
two
character:
top-level
domains.
That's
right.
It
took
two
character
country
codes
and
if
the
next
slide
please
and
lists
them
all,
I
can
go
through
them
right
now,
it's
a
a
a
b,
a
c
and
I've
counted
them,
one
by
one,
they're,
676,
I'm,
sure,
I'm.
Sure,
there's
some
mathematical
tricks
that
you
can
use
in
order
to
do
this
faster,
but
there
are
676
next
slide.
G
Please
and
what
he
is:
oh
three
one
six
six
folks
have
done
is
they've
categorized
them.
All
of
these
individual
elements
fall
into
a
category
and
I'm
a
little
bit
of
color
blind,
so
I
hope,
I'm
gonna
get
this
right.
They
can
have
here
next
slide.
Yeah
on
the
top
left,
you
see
a
colored
squirm.
This
is
a
a
a
color
square.
It
basically
means
it's
user
assigns
and
this
one's
going
to
talk
about
in
a
minute.
I'm
gonna
go
quickly
through
the
rest.
You
have
the
white
background.
G
It's
actually
feasible,
I
know
I'm,
Valerie,
perfect,
that's
visible!
So
the
white
background,
like
a
bee,
those
haven't
been
assigned,
yet
they
may
be
assigned
in
the
future,
but
they
haven't
been
assigned
yet
so
the
categories
and
the
science
and
then
we
have
exceptionally
reserved
like
UK
this
yellow
and
EU.
G
This
yellow
small
sidestep
here
I
used
to
work
for
a
company
names
and
Namie
nets
and
I
used
to
work
it
for
nine
years
and
I
knew
UK
is
not
gb,
but
I
never
knew
the
following,
and
this
is
only
what
I
discovered
in
the
last
couple
of
months:
ISO
3166
and,
as
you
know,
and
as
UK
is
exceptionally
reserved
and
in
the
context
of
ISO
3166,
it
means
that
UK
is
not
a
country.
Another
country
name
is
not
assigned
because
ISO
3166
is
assigning
official
country
names
to
a
two-letter
code
right.
G
This
UK
is
not
a
science
to
the
United
Kingdom.
Why
not
is
because
the
ISO
3166
decided
long
ago
to
avoid
generic
terms
like
unites
or
Kingdom
or
Republic,
if
you're
gonna
stick
with
all
these
generic
names
like
United,
Kingdom
and
Republic,
you're
gonna,
soon
run
out
of
huge
case
and
ours
to
use
in
your
to
letter,
doing
so
for
the
United
Kingdom
of
Great
Britain
and
Northern
Ireland,
the
first
distinguishing
string
is
Great
Britain
and
the
official
encoding
for
the
country
code.
G
G
Yes,
so
if
you
look
at
Russia,
you
know,
let's
move
on
these.
The
other
codes
is
the
green
stuff
that's
officially
assigned,
and
then
you
have
some
transitional
e-reserves
and
in
in
in
determine
a
division.
I'm
not
gonna,
go
into
that
right
now,
but
it's
just
for
me.
It's
an
interesting
thing.
Next
slide,
please.
So
these
are
the
codes
that
I
want
to
talk
about
the
user
design
stuff
a
a
qm2
Q
Jets
XA
2
XS
jet
set
next
light.
G
So
what
does
the
ISO
standards
say
about
these
things
about
your
construction
of
the
alpha-2
codes?
The
ISO
3166
standard
uses
a
combinations
in
the
range,
a
b,
TQ
l,
ra,
t
wz,
and
why
eh-2-zed,
why
so
d,
iso
standard
talks
about
dead
range?
Of
course,
what
did
they
say
about
the
other
42,
these
user
assigned
elements?
G
They
are
not
used
in
the
standards
they
are
not
used,
they
will
never
be
assigned,
they
will
never
be
reserved,
they
are
not
used,
and
this
is
important
in
the
context
of
what
I
want
to
talk
about
in
a
minute.
So
the
standard
goes
on
in
8.1
about
special
provisions.
You
should
sometimes
need
to
extend
or
to
alter
the
use
of
country
code
elements
for
special
purposes
and
in
an
8,
1
2.
G
It
says
if
users
need
code
elements
to
present
country
names
of
countries,
country
names
not
included
in
this
part,
these
series
of
letters,
AAA,
qmq,
jet,
etc,
are
available.
This
is
guidance,
it
doesn't
mean
you
can't
use
this
for
anything
else
and
I'm
gonna.
Give
you
a
few
examples
of
how
other
standards
are
using
these
things.
Next
likely.
G
So
are
these
things
used
to
elsewhere?
These
user
sign
codes
is
o3
901
in
the
international
standard,
recording
code
and
I'm
not
going
to
pretend
I
know
anything
about
these
things.
I've
just
found
them,
so
it
reserved
said
set
for
direct
registrants,
independent
of
any
country.
Now
the
next
one
is
quite
interesting,
mi.
So
for
tu-16
codes
for
the
representation
of
currencies,
it
reserves
XA
to
exit
for
transactions
and
precious
precious
metals
as
their
country.
Independence.
I'm
gonna
talk
about
this
a
little
bit.
G
Us
dollar,
the
official
code
for
your
Zola
is
USD
and
the
US
in
USD
comes
from
the
three
one:
six
six
tenders,
GBP,
remember:
GB
UK!
You
don't
have
UK
P
if
GBP
the
GB
GB,
please
stems
from
the
ISO
three
one.
Six
extend
it.
So,
what's
all
this
XA
stuff
in
official
standards,
gold
can
be
used.
Sorry,
nothing
official
standards
in
general
gold
can
be
used
to
pay
for
stuff
right,
its
currency,
in
transactions,
etc,
etc.
G
There's
no
official
code
for
that
is
X
a
you
silver
x,
AG
Bitcoin
I'm,
not
kidding
XB
t,
so
these
are
used.
These
private
space
names
this
em.
Sorry,
these
user
sign
names
are
used
in
all
kinds
of
standards:
iso
6-1
6-2
6-2
about
excess
for
security,
clear
to
Euro,
clear,
clear
stream,
etc
away
from
the
ISO
standards.
Next
slide:
please
you
have
ICAO
International
Civil
Aviation
Organization,
it
reserves
LSATs
for
UN
travel
documents
and
WIPO
WIPO.
G
The
worth
is
lecture,
Property
Organization
uses
a
whole
range
of
these
private
space
and
things,
and
it
uses
not
for
countries
but
basically
for
their
own
offices.
So
if
you
have
a
peasant's
or
an
application
from
a
certain
con,
you
can
use
a
country
code,
but
if
it
goes
through
a
certain
Patent
Office,
you
can
actually
use
the
code
for
that
patent
on
this
closer
to
home,
to
can't
form
the
certificate
authority
in
browser
form,
it
reserves
xx
to
signify
location,
not
covered
by
ISO
3166
and
even
closer
to
her
closer
to
home
and
RFC.
G
Five.
Six,
four,
six
well
known
to
many
of
you
text
for
arson
identifying
languages.
This
is
when
you
use
action,
these
language,
taxi
your
browser,
etc,
cetera
next
slide.
This
is
the
text
lifted
from
the
standard.
From
that
sorry
from
the
BCP,
it
says,
for
example,
the
region
subtext,
a
a
set
set
and
dosing
ranges,
etc,
etc.
You
recognize
these
from
before.
They
can
be
used
to
form
a
language,
language
stack
and
here's
an
example
on
the
screen.
So
next
slide
please.
G
G
We
may
not
need
these
to
go
through
RC
67
61,
where
you,
where
you
reserved
them,
such
as
the
onion,
etc,
etc,
and
the
cool
thing
just
like
these
X
Montag's
right
using
these
two
character
costs
have
no
semantic
meaning,
which
is
a
good
thing
next,
like
so
I'm
just
gonna
use
this
as
a
suggestion.
I
don't
want
to
paint
device
yet
too
much
right
to
use
that
said.
I'm
gonna
use
that
zet
as
an
example
going
forward.
G
Luckily,
I
only
have
one
slide
left
so,
but
I'm
not
suggesting
that
we
should
use
this
now,
I'm
just
suggesting
that
that
is
available
for
this
purpose
and
next
slide.
So
what
about
D
in
a
sec?
Well,
we
all
know
that
Zed
Zed
doesn't
actually
exist
and
your
validator
will
validate
that.
That
said,
that
doesn't
exist
because
it
doesn't
validate,
can
prove
this.
That's
a
good
thing!
I,
don't
want
to
be
lied
to
right,
at
least
not
without
approval.
G
Now
some
have
suggested
that
Zed
says
or
internal
you
should
be
delegated
in
the
route
and
signs
now
I
disagree
if
I
as
a
validator
want
to
be
lied
to
I
want
to
make
that
explicit.
I
want
to
have
to
configure
it
that
people
can
lie
to
me.
So
does
the
internal
both
odds,
don't
onion
or
what
not?
If
I
want
to
allow
that
in
my
in
my
validator
space,
then
I
want
to
make
it
explicit.
G
I
want
to
have
a
truss
tank
for
that,
a
negative
trust
anchor
now
stubby
works
sure
that
bind
works
for
that
it
all
works,
basically,
so
an
even
better
if
I
wanted
resigns
his
private
space.
I
can
now
do
that
with
a
positive
trust
anchor
as
opposed
to
a
negative
trust
anchor
anyway.
This
is
my
my
rent.
My
my
little
talk
about
using
private
space
names.
I
will
take
this
to
DNA
shops
on
on
Thursday
and
I
will
bore
you
all
once
again
with
this
stuff.
D
Just
a
quick
comment:
thank
you
for
this
and
I
think
it
makes
a
lot
of
sense.
I've
been
using
dot
XY
for
a
number
of
years.
I
didn't
do
the
detailing
research
I,
just
read
the
Wikipedia
article
on
our
own
I
guess
so:
3166
I,
the
other
proposals.
I
mean
dot.
Internal
is
nice,
but
it's
just
too
long
I
mean
it's
feels
really
unnatural
to
pinger.
You
know,
TV,
you
know
the
ping
TV
dot
intimal
feels
unnatural.
It's
people
are
not
going
to
use
that.
So
thank
you
very
much
and
alright.
We
worry
thank.
I
J
Next
slide,
please
so
I
have
a
server
somewhere
and
with
a
bunch
of
people
we're
using
that
server
to
map
outs,
rpki
deployments
on
the
internet
using
a
fairly
simple
method.
We
have
two
prefixes
1
as
an
RPG,
I,
invalids,
prefix,
the
other
is
pellets
and
from
those
prefixes
we
sent
ICMP
messages
to
every
ipv4
address
on
the
internet
and
we
compare
which
of
the
source
addresses,
gets
replies.
J
Now
in
this
image
you
see,
bloom
is
IP.
Four
addresses
that
responded
to
both
the
valid
and
invalid
source
address.
This
means
they
are
not
performing.
Rpki
based,
BGP,
origin,
validation
and
yellow
is
only
responding
to
the
valid
source
address,
so
they
are
in
fact
doing
origin
validation
or
a
network
between
that
server,
and
that
IP
address
is
doing
origin
validation.
So
this
is
a
snapshot
from
August
next
place
a
month
later
next
place
and.
J
J
Next,
please,
now
a
update
on
rpki
in
context
of
a
re
are
in
the
right
community.
Roughly
a
year
ago,
the
IRR
database
was
split
into
two
parts,
almost
called
ripe
and
the
other
is
called
ripe.
Numb
off
and
the
splits
occurred,
because
the
ir
objects
in
the
ripe
source
were
authenticated
against
the
wishes
of
the
resource
holder,
but
there
also
was
a
bunch
of
data
in
the
right
eye
arm.
J
So
the
thinking
was
what,
if
we
use
this
original
addition
procedure
and
apply
it
to
IRR
objects.
In
other
words,
if
an
IRR
object
is
in
conflict,
we've
published
harpy
guy
are
ways
we
can
conclude
that
the
ir
object
is
describing
a
state
that
should
be
rejected
and
therefore
this
ir
object
can
be
deleted.
Now
the
community
came
to
consensus
on
this.
The
policy
was
ratified
last
week
and
out
of
the
65,000
ish
objects
in
this
data
source
roughly
900
are
in
conflict
with
published,
are
ways.
J
If
you
want
to
inspect
this,
you
can
download
a
small
Python
scripts,
I've
wrote
and
that
will
show
you
the
data
next
slides.
This
is
one
of
such
examples,
probably
some
time
long
before
I
was
even
walking
this
planet,
a
company
created
route
objects'
that
describe
state
that
is
in
conflict
with
our
ways
that
entity
published
an
entity
did
not
have
any
way
to
delete
these
IR
objects
in
the
right
database.
J
J
The
reason
I
already
for
exists
is
that
legacy
IR
D
was
no
longer
extensible.
There
were
some
reliability
issues
with
this
demon,
which
meant
that
we
just
had
to
restart
it
every
week,
but
the
code
was
complicated,
which
meant
that
we
could
not
add
new
features,
specifically
new
security
features
and
one
of
the
cool
things
that
is
coming
to
IR
D
is
a
mechanism
similar
to
what
is
happening
in
the
ripe
database.
J
The
IR
D
for
demon
in
its
next
release
will
have
functionality
to
consume
rpki
data
to
suppress
or
deletes
conflicting
IR
information,
and
this
will
happen
on
two
levels,
both
in
context
of
the
IRD
demon
being
a
furtive
quote-unquote
supportive
database
and
on
the
mirroring
level,
when
this
demon
consumes
data
from
other
databases
over
nrt
m.
In
both
halves
into
this
into
the
cache
of
this
demon,
origin
validation
will
be
applied,
thereby
significantly
cleaning
up
the
potential
damage
that
the
irr
can
bring
next
slide.
J
It
provides
us
operators
with
a
industry-standard
mechanism
to
get
rid
of
old
route
objects,
but
it
also
helps
protect
us
going
forward,
because
the
moment
you
put
in
our
Picaro
us
these
demons
will
reject
ir
object,
updates
that
are
in
conflict
with
the
published
arrays.
So
publishing
our
ways
becomes
more
valuable,
so
to
speak,
because
it
helps
you
clean
up
things,
but
it
also
helps
you
protect
going
forward.
J
The
timeline
for
these
things
is
roughly
spring.
2020
I
expect
the
various
evolved
parties
will
provide
more
updates,
as
these
timelines
become
more
firm,
but
the
good
news
in
all
of
this
is
that
it's
not
too
far
away.
Next
slide,
please
now
onwards
to
open
BSD,
open
BC
is
a
research
project
where
we
experiment
with
things
and
amongst
those
things
is
RPG,
I
mean.
Nowadays,
you
have
to
realize
open.
J
J
Furthermore,
we
had
to
implement
our
sync.
This
project
is
called
open.
Our
sync
and
those
three
components
have
created
a
scenario
in
which
we
can
have
original
Edition
out
of
the
box.
Next
slide.
Here's
a
small
screen
capture
the
beach
P
demon
is
running
from
Chrome
I
am
calling
the
RPG
I
clients
that
outputs
a
list
of
the
VIPs
into
a
text
file
format,
that's
suitable
for
open,
Beach
PD,
and
this
allows
the
Beach
PD
daemon
running
on
this
box
to
do
original
edition
and
show
us
fellas
were
invalid
States
now.
J
My
hope
is
that
by
implementing
our
PGI
at
this
level
in
a
network
operating
system,
we
inspire
more
commonly
used
network
operating
systems
such
as
cisco,
juniper,
others
to
consider
the
deployment
model
where
routers
extruders
themselves
can
perform
the
origin,
validation,
the
RPK
cache
validation
function
in
order
to
do
original
edition.
I
think
this
is
one
of
many
deployment
scenarios.
This
is
not
necessarily
the
recommendation
that
will
work
for
everybody,
but
I
think
there
are
some
use.
Cases
specifically
in
smaller
is
piece
where
it
could
be
beneficial
to
take
a
page
from
this
approach.
J
J
It
is
a
bit
of
an
unfortunate
situation
that
a
organization
that
is
supposed
to
operate
in
the
best
interests
of
its
community,
specifically
its
members,
that
they
are
not
publicly
distributing
their
public
crypto
keys
and
I'm,
not
sure
what
the
immediate
solution
is
other
than
to
keep
lobbying
and
keep
applying
pressure
directly
and
indirectly
to
this
organization
to
get
them
to
release
their
public
keys
to
the
public.
But
this
is
a
measurable
issue
on
the
Internet.
J
A
double-digit
percentage
of
networks
performing
RPI
region
foundation
is
not
using
the
air
Intel,
because
the
air
Intel
cannot
be
included
in
open
source
distributions,
such
as
open
BSD.
So
we
have
to
keep
an
eye
out
for
this
and
whatever
security
mechanisms
we
design.
Next,
we
should
keep
in
mind
that
we
should
make
it
hard
for
people
to
make
the
public
keys
unavailable
next
slide,
please.
So,
in
summary,
what
I
wanted
to
share
with
you
from
this
slide
deck?
Is
that
rpki
Israel?
It's
here?
J
It's
being
deployed
and
a
lot
of
people
are
using
it,
so
this
means
that
we
are
now
running
in
a
maintenance
mode
where
we
have
to
keep
our
eyes
and
ears
open
towards
the
end
users
and
the
implementers
of
RPI
suffer,
and
that's
all
for
today,
questions
comments.
Concerns
remarks
can
be
either
shared
through
the
microphone
or
emailed
to
me
directly.
K
Precise
use
of
language
sometimes
is
actually
helpful
to
avoid
confusing
people,
even
if
the
first
use
of
precise
language
males
may
called
confuse
people
who
are
used
to
imprecise
use
the
thing,
the
thing
where
I
would
want
to
do
when
it
picking
right
now
with
a
perspective
attacking
your
last
topic,
is
you
are
giving
one
map
about
rpki
deployment.
I
would
change
the
headline
for
that
to
rpki
or
Ridge
invalidation
deployment,
for
there
are.
K
K
L
J
M
M
M
So,
in
the
case
of
BGP
there's
an
event,
you
might
announce
a
prefix,
there
might
be
a
link
state
up
down,
and
then
this
sort
of
ton
of
updates
is
ripple
throughout
the
rest
of
the
network
and
at
times
trying
to
understand
why
your
router
is
kind
of
having
a
seizure
and
what
was
the
original
event
that
is
causing.
It
is
sometimes
really
hard
to
tell
now.
That's
BGP-
and
you
know,
a
fair
deal
of
study
has
gone
into
that
and
we
still
don't
understand
that
at
least
I
don't
DNS
is
about
the
same.
M
You
might
think
that
the
DNS
is
kind
of
this
transactional
thing
where
your
browser,
your
application,
is
given
some
kind
of
name.
So
it
asks
a
question
in
the
DNS.
The
DNS
goes
to
hunt
for
the
right,
authoritative,
server
and
out
pops
an
answer
right.
So
every
question:
that's
all
query
that
sitting
inside
the
dns
should
be
able
to
be
mapped
to
some
original
event
that
the
users
driven.
Nothing
could
be
further
than
the
truth.
M
The
DNS
is
like
artificial
intelligence.
It
just
talks
to
itself
and
at
least
50%
of
all
the
queries
that
we
see
inside
the
DNS
are.
Basically,
we
have
no
idea
why
they're
there
or
what's
going
on
and-
and
this
is
something
we
were
looking
at
when
we're
looking
at
an
area
called
aggressive,
insect
caching,
there's
this
kind
of
theory
that
DNS
SEC
is
really
really
good.
M
But
the
really
hard
question
is:
what
is
it
good
for
and
if
you're
not
allowed
to
answer
Dane,
it
becomes
a
really
really
hard
question,
because
if
you
can't
use
it
for
keying
in
TLS,
what's
it
good?
For,
and
one
of
the
theories
was
that
well,
if
it's
not
good
for
that
so
I,
why
don't
we
use
it
to
actually
change
the
way
negative
answers
ocation
in
the
DNS,
because
the
easiest
way
to
bring
down
the
DNS
or
parts
thereof
in
a
highly
targeted
fashion?
And
this
is
no
secret.
It's
basically
random
named
attacks.
M
You
set
up
a
largely
distributed
system.
You
set
up
random
labels
across
all
of
your
little
zombies,
and
you
just
hit
the
authoritative,
nameserver
Roy
says:
I'm
not
allowed
to
say
how
it's
really
easy.
You
just
hit
it
because
the
random
names
ensure
that
caching
doesn't
work.
All
of
those
queries
go
to
the
authoritative
servers
and
if
you
thority
of
servers,
aren't
well
defended,
they
will
melt
now.
What
you
really
wanted
was
to
get
the
recursive
x'
to
come
along
and
help
you
so
the
first
time
you
asked
even
a
random
subdomain
name.
M
It
would
go
now,
I'm,
sorry
that
doesn't
exist.
The
only
way
you
can
do
that
in
theory
was
with
signed,
domain
names
and
range
based
answers
and
whether
it's
in
sick
or
in
sex
3
makes
no
difference.
You
can
actually
manage
to
deflect
these
attacks
in
theory.
Now
that
whole
story
and
the
answer
is
a
different
presentation.
That's
not
it!
That's
not
what
we're
talking
about
today,
but
what
we
were
talking
about
and
looking
at
was
a
really
simple
thing.
M
What
happens
when
the
answer
is
No,
so
we
present
to
the
user.
A
DNS
name
go
and
resolve
this
name
and
it's
our
zone,
and
we
look
at
the
number
of
queries
and
what
queries
happen
when
the
name
does
not
exist
and
the
really
odd
thing
was,
the
person
only
got
the
name
once
only
once
it's
inside
an
online
ad
control,
this
the
user
only
exercised
that
name
once
and
only
once.
Why
are
there
more
queries?
Next,
so
we
control
the
end
user
and
we
control
the
authoritative
server.
There
are
random
labels
flying
around.
M
So
there
is
no
caching.
The
thing
is
not
signed
as
I
recall,
or
maybe
it
is
makes
no
difference
and
we're
getting
a
whole
bunch
of
queries
from
the
same
resolver
from
different
resolvers.
But
this
is
weird:
why
are
we
getting
so
many
queries
next,
on
average,
over
60
million
names
that
were
used
in
this
experiment
will
get
142
million
queries,
in
other
words,
no
has
to
be
said
at
least
twice:
no,
no
and
a
bit
more,
no
0.37
of
a
no
and
that's
the
average
amount
of
time
to
say
no.
Now
it's
really
great.
M
The
DNS
queries
are
free
cuz.
If
they
weren't,
some
will
be
making
a
massive
amount
of
money
cuz.
You
know
this
software
is
crap.
Why
is
it
doing?
This
is
sort
of
an
interesting
question
next,
so
the
naive
expectation
is
that
no
means
no
browsers
should
understand
that
asking
three
times
for
name
that
doesn't
exist
is
kind
of
two
times
to
many,
but
even
out
there
in
the
bigger
DNS,
that's
1.3
7
times
too
many.
M
So
is
this
just
the
fact
that
folk
who
write
DNS
resolvers
are
lunatics,
they
could
well
be
I,
don't
know
their
source
code
is
baroque,
I
have
no
idea.
What's
inside
it,
there
could
be
strange
and
wonderful
loops
doing
this
or
it
could
just
be.
We've
created
artificial
intelligence,
deep
in
the
DNS
and
just
haven't
realized
the
stunning
beauty
that
we
have
created,
that
the
DNS
truly
is
a
neural
network
and
it's
assumed
a
life
of
its
own
next.
M
But
there
are
some
more
prosaic
reasons
that
sort
of
begs
you
down
in
what's
going
on
and
part
of
the
I
suppose.
The
theory
is
that
at
various
parts
in
the
application
sack
no
one
gives
a
about
any
other
part
of
the
application
stack
just
do
not
care
and
happy.
Eyeballs
is
one
of
these
just
do
not
care
I'm
going
to
solve
my
problem
and
tough.
You
get
more
queries.
Well,
that's
what
you're
built
for
so,
oddly
enough
happy
eyeballs.
M
M
My
feeling
is
that
it's
serializing
not
parallelizing,
so
it
asks
for
the
quad
a
doesn't
ask
for
the
a
record
it
gets
back
in
there.
Next,
a
man
goes,
okay,
the
name
doesn't
exist.
I
will
go
away.
3%
of
the
cases
74%
only
v4,
which
is
about
the
same
as
the
amount
of
v4
and
v6
out
there
on
the
larger
Internet.
M
So
roughly
expected
next,
so
we
can
sort
of
start
dividing
things
up
a
bit
and
we
sort
of
say
well,
okay,
let's
just
group
these
a
in
quad
a
queries
together
and
if
we
group
them
together,
we
saw
74
million
seventy-three
point
five
million
dns
resolution
events,
but
we
still
saw
one
hundred
and
forty
two
thousand
four
and
three
two
million
queries.
So
we've
got
our
two
point:
three
seven
down
to
one
point:
nine
three
queries
better,
but
the
dns
is
still
weird
next.
M
Under
half
the
time
it
just
worked
with
either
one
query
now
that
might
have
been
one
A
and
one
quad
a,
but
it
was
just
one
query:
the
other
half
of
the
time,
51
percent,
that
query
got
repeated
and
the
average
was
two
point,
eight
four.
So
there's
a
lot
of
queries
going
on
next
and
the
real
question
is:
is
that
an
average
misleading?
M
Fourteen
hundred
and
thirty
queries
I'm
like
what
part
of
no
is
a
problem.
Fourteen
hundred
and
thirty
times,
whoever
wrote
that
software,
you
know
good
good
Anya
work
about
down.
On
the
other
side,
a
large
amount
of
folk
re
queried
one,
so
it
was
two
queries,
a
smaller
amount
twice.
Thirty.
Two
percent
of
queries,
though,
were
three
or
more
repeats.
So,
third
of
the
time
the
DNS
is
going
throb,
throb,
throb,
throb
throb,
and
the
answer
is
always
there.
M
N
M
The
same
query
idea
in
the
same
just
hang
on
a
second
Jared
and
you
will
start
to
see
what's
going
on
because
it's
more
than
you
think
that
was
your
yeah.
No,
so
there's
no
clues
going
on
here
does
UDP
sack.
Oh
yes,
UDP
sucks,
but
not
he
doesn't
suck
that
bad
and
the
the
response
was
not
big.
It
was
only
603
bytes
and
if
I
can't
get
a
603
but
DNS
packets
ripped
through
the
internet,
I,
don't
know
what
you're
doing
here.
M
You
should
be
fixing
your
because,
quite
frankly,
if
I
can't
get
that
through
I
can't
get
anything
through.
So
why
are
we
seeing
51
percent
of
test
generating
two
or
more
next,
so
we
started
looking
at
the
repeat
intervals
Jared
and
we
started
looking
at
the
time
between
successive
queries
from
the
original
epoch
time,
and
you
know
there
are
these
really
strong
Peaks.
You
know
next
so
who
writes
a
piece
of
DNS
resolving
software
with
a
370
millisecond
timer
Open
DNS.
M
M
Unbound,
who
writes
one
second,
bind
so
a
lot
of
what
we
see
here.
There
are
actually
signatures
of
the
resolver
folk.
Making
arbitrary
assumptions
about
the
UDP
timeout
intervals
that
have
no
relationship
to
reality
have
nothing
to
do
with
anything
else
other
than
this
number
and
oddly
enough,
they
all
use
different
numbers
and
what
they
actually
cause
is
some
degree
of
pathology
of
repeats,
because
these
time
intervals
appear
strongly
in
this
wreak.
M
Weary
signal,
what's
weird,
is:
does
UDP
suck
that
much
that
the
original
answer
never
got
there
or
is
the
resolver
simply
on
this
mindless
path
of
God
or
repeat
the
query?
Don't
care
if
I
got
an
answer
or
not
I
just
have
to
repeat
this
query:
I'm
a
driven
resolver,
no
clue
next,
so
timers
is
one
issue.
The
next
is
DNS
SEC.
It
is
a
DNS
SEC
sign
name,
so
it
has
an
insect
record.
M
Aha
is
validating
the
N
SEC
a
factor
and
don't
forget
these
are
unique
names
and
they're,
so
structured
that
they're,
even
in
unique
domains,
so
the
validation
is,
has
to
be
done
every
single
time,
so
is
DNS
ik
causing
this.
So
ok.
Let's
add
another
test:
let's
go
unsigned
next
Wow
you
have
to
get
right
or
get
right
up
here
to
slot
the
difference
between
the
blue
and
the
the
red
lines.
Now,
at
the
point
of
won't
help.
M
M
Look
I'm,
really
bored
about
this
I'm
going
to
ask
you
again:
cuz
370,
milliseconds,
says
I've
really
got
to
ask
you
again
and
again
and
again
and
again
and
again
so
signing
generates
more
load
when
the
name
doesn't
exist
because
of
validation
next,
and
we
can
actually
quantify
this
a
little
bit
when
it's
signed
the
require
ear.
Aid
is
three
point,
one
nine.
On
average
we're
unsigned
it's
two
point:
five
one.
M
M
How
many
folk
think
DNS
resolvers
are
just
standalone
boxes?
Well,
no
one
anymore,
because
if
I
ain't
there,
these
weird
things
called
load
balances
and
they
sit
inside
a
whole.
Bunch
of
you
know
single
resolver
engines,
except
when
the
load
balancers
actually
become
load
amplifiers,
as
is
a
case
from
this
Taiwanese
network,
where
the
front-end
load
balancer
decided
to
send
the
same
query
to
every
single
member
of
two
distinct
subnets.
And
you
look
at
those
query
times.
Every
sort
of
twenty
have
David
sorry.
M
Every
hundred
milliseconds
I
get
another
query
from
the
same
subnet,
but
a
different
engine
common
as
hell.
Whoever
runs
this
stuff
has
no
idea
what
they're
doing
and
and
that
a
huge
number
of
these
distributed.
Dns
resolver
farms
actually
become
DNS
resolver
amplifiers,
just
because
you
can
thank
you
very
much
next.
How
common
70%
of
the
requir
e's
come
from
these
shared
subnet
prefixes.
So
it's
not
the
same.
Ip
address
it's
a
different
IP
address,
it's
a
different
query,
but
the
original
query
is
being
forked
out
again
and
again
and
again
across
the
entire
subnet.
M
Whoever
runs
resolver
farms,
whoever
runs
yours,
fire
them,
cuz
they're.
Obviously,
at
their
job
cuz,
this
is
really
really
common.
70%
of
the
requir
e's
come
in
from
these
kind
of
odd
resolver
farms.
Next,
so
it's
kind
of
wondering
is
this
a
no
problem,
I'm
like
when
I
say
no,
do
you
not
believe
me,
whereas,
if
I
give
you
an
answer,
go:
oh
yes,
good,
fine,
there's
no
worse
than
yes!
So
next
is
the
require
rate,
lower
Wow.
M
If
I
give
you
an
answer,
you're
more
likely
to
believe
it.
But
if
you
re
query,
if
you
get
a
real
answer,
you'll
hammer
the
out
of
me
so
that
39
percent
of
the
requir
is
in
in
X
domain
signed,
become
requir
YZ,
three
point:
one:
nine
queries
per
per
name
when
it's
an
real
answer:
I
and
quad-a.
Only
13
percent
wreak
wery
much
much
better
guys,
but
your
average
is
five
point.
Eight
one
re
queries.
What
the
I'm,
like
you,
don't
even
like
a
good
answer.
It's
this
sort
of
subset
of
lunatics.
M
That
kind
of
go
look
I,
don't
care
whether
it's
a
no
or
a
yes
I'm,
just
gonna
ask
and
ask
and
ask,
and
ask,
because
you
know:
I
can
so
yeah
whether
it's
a
real
answer
or
not
a
real
answer.
There
is
this
phenomenal
wreak
weary
rate
down
there.
Next,
so
that's
the
same
graph!
You
need
to
be
right
up
there
to
see
it,
but
what
you
see
is
a
DNS.
Sec
sign
name
generates
a
lot
more
responses
than
a
real
name,
but
the
real
names
when
they
require
e
re
query
very
very
hard
next.
M
So
we
can
kind
of
put
it
out
there
inside
of
a
single
graph.
This
is
the
one
that
I
showed
you
before
42%.
Do
it
once
next
happy
eyeballs
I'd,
say
about
10%
additional
load.
Next
I
put
DNS
SEC
at
near
70%
of
the
total
load.
Next
NX
domain
itself
signed
an
X
domain,
no
vs.,
yes
about
14%.
Next,
insane,
stupidity
down
on
the
farm
about
13%
of
queries
and
the
next
wow
you
guys
are
just
inventive
I
have
no
idea.
Why
else
you're
doing
this,
but
you
know
good
on
you.
N
Yeah
heíd
shared
much
again
so
if
you
know,
if
you
can
hit
the
backspace
key
or
something
Chris
so
here,
I
don't
see
this
showing
whether
or
not
this
looks
to
be
dysfunctional,
Network
behavior
of
packet
replication
in
the
networks,
because
I've
definitely
seen
that
before
especially
with
UDP,
which
is
why
I
asked
the
prior
question
about
query,
ID
and
the
port
numbers
and
so
I'm
curious.
It
is
that
the
reasons
or
is
that
sorry.
M
It
could
will
be
I
couldn't
think
of
any
other
deliberate
pathologies
that
you
know.
I
was
smart
enough
to
understand,
but
you've
just
uncovered
one,
and-
and
thank
you,
you
know,
replicated
query
packets,
something
subtle,
though,
if
the
replication
is
happening
between
the
recursive
and
the
authoritative
I
will
see
it
if
the
replication
is
happening
earlier.
It's
actually
not
obvious
what
the
recursive
is
going
to
do
when
it
starts
to
get
back-to-back
queries
and
there's
already
a
query
in
process.
Good
engineering
says
duplicate
queries
get
held
in
occassion
till
the
answer
comes
back.
N
I
mean
in
the
case
of
a
resolver
farm
I
could
easily
see
a
subsequent
query
being
load
balance
differently,
yeah,
but
what
I
have
specifically
seen
as
I
have
seen
Network
pathologies,
because
people
are
either
configured
ports,
panning
or
mirroring,
or
have
active
or
passive
taps
in
the
middle
for
their
own
monitoring
of
the
network
that
sometimes
end
up
replicating
packets,
and
it
is
especially
harmful
for
UDP
based
protocols
like
this,
which
is
why
I
would
urge
you
to
look
at
that
further
in
part
of
them.
That's.
O
O
Is
that
being
adaptive
about
your
load
rate
is
actually
very
important,
and
the
graph
here
is
showing
about
the
various
implementations
that
we're
timing
assumptions
you
know
constants
are
almost
always
wrong.
In
some
sense,
an
interesting
line
of
inquiry
might
at
some
point
be.
Is
there
a
way
to
allow
dns
to
sort
of
actively
signal
back
into
the
system?
So
here's
how
the
cadence
should
look
like
that?
Would
get
you
a
better
response
rate?
You.
M
Know
you
know,
that's
a
great
question
and,
and
had
you
said
that
about
10
years
ago
and
the
obvious
answer
is
DNS
over
TCP.
Why
TCP?
Because,
instead
of
a
preset
timer
that
tells
you
there's
a
problem,
there's
an
acknowledgment
time.
You
actually
understand
round-trip
times,
and
you
know
when
you're
missing
a
packet,
you
wait
for
twice
the
time,
etc,
etc.
Tcp
does
a
much
better
job
of
getting
over
packet
pathologies
because
that's
its
job.
Now
it
withstand
a
loss
rate
of
up
to
about
three
or
four
percent.
M
That's
why
mobile
networks
work
as
distinct
from
don't
so.
This
is
a
good
idea,
but
ten
years
ago,
when
you
said
DNS
over
TCP
you'd
be
covered
in
arrows,
because
obviously
you're
a
heretic
you're
a
pagan
and
you
should
be
burned
at
some
local
stake
these
days.
It's
fashionable
to
say
DNS
over
TLS
and
everything
will
be
cured
and
life
will
be
wonderful
and
if
you're
doing
ultra
trendy
I'm
going
to
mention
the
magic
word,
the
magic
word
is
Doh
and
it
just
makes
everything
better.
So.
O
C
M
C
M
C
P
I,
that's
tail.
I
came
up
with
one
request,
but
you've
just
added
a
second
one
in
that
I
think
it's
probably
best
for
all
of
us
if
we
not
refer
to
the
throbbing
DNS,
but
apart
from
that
I'm
curious.
If,
if
you
happen
to
repeat
this
experiment
in
the
future,
if
you
could
also
include
the
yes/no
answer,
that
is
a
no
data
response
both
signed
and
unsigned
to
see
how.
M
We've
been
wondering
about
refused
surveil
in
X
domain,
no
answer,
because
this
phenomenal
amount
of
ghost
queries
in
the
DNS
that
just
don't
seem
to
have
an
obvious
reason
and
it's
always
been
an
issue.
What
responses
encourage,
reek,
weary
and
what
responses
kill
it
flat
and
we're
working
on
a
theory
at
one
point
that
said,
NX
domain
is
cache
no,
and
the
answer
is
well.
No,
it's
not
it
sort
of
hit
me
again,
so
yeah
I,
don't
know
where
the
no
data
refused
to
fail
or
any
other
response
code
would
have
different
response.
Q
Yeah
hi
Jeff,
Mark
Smith,
another
thought
I
had
was
that
these
persistent
queries,
even
when
you
give
them
a
good
answer,
a
load
balancers
doing
liveness
checking,
so
they
might
present
a
single
IP
address
to
their
client.
But
then
they
have
a
pool
of
upstream
servers
and
they
likes
reaching
quickly
and
so
on.
I
hide
them
for.
M
Various
reasons
and
I
think
it's
probably
a
mutual
hatred.
The
folk
who
build
load,
balancers
and
the
IETF
have
never
met
eye-to-eye.
There
is
no
standard.
There
is
no
spec,
there
is
no
behavior.
This
is
a
bit
like
the
first
weird
and
wonderful
world
of
Nats.
Everyone
builds
DNS
load,
balancers
differently.
They
screw
with
I
some
hash,
some
hash,
various
parts.
Some
do
you
know
and
so
on,
and
so
because
a
lot
of
the
recursive
resolver
families
appear
to
sit
behind
these
wacky
load
balancers.
The
end
result
is
a
set
of
queries
that
dooby-doo
behave.
M
R
M
You're,
going
down
a
completely
different
path
and
and
a
path
that
in
the
dns,
is
a
path
that
doesn't
have
very
many
answers.
The
dns
is
incredibly
opaque.
Our
attempts,
misguided
as
they
came,
of
putting
fingerprints
in
queries
otherwise
known
as
client
subnet,
have
been
loudly
met
with
accurate
accusations
of
acute
invasions
of
privacy,
and
so
the
dns
is
resistant
to
any
kind
of
realistic
measurement
of
the
things
you're
talking
about.
We
can't
see
that
clearly,
you
can
see
either
end,
but
the
middle
is
sort
of
a
mystery
yeah.
S
Robbo
Stein
couple
couple
of
comments:
one
the
time
out
issue
with
the
ridiculously
short
retry
intervals,
yeah
back
in
that
you
should
probably
expression
the
80s
we
actually
used.
Retry
intervals,
like
you
know,
two
three
seconds
and
stuff
like
that
and
as
near
as
I
can
tell
what
changed
it
was
essentially
market
pressure.
You
know
it's
basically
an
old
version
of
the
happy
eyeballs
problem,
people
whine
when
they
don't
get
an
instant
response
and
the
idea
that
your
query
global
database
it
might
take
a
little
while
well
screw
that
I
want
an
answer
now.
S
M
S
Other
thing
is,
you
know:
I've
not
actually
been
to
the
DNS
business
for
about
a
decade.
In
fact,
the
last
time
I
was
our
working
theory
about
a
lot
of
the
insane
query
that
just
like
nothing,
you
can
do
will
stop.
It
was
people
behind
stupid-looking
figured
firewalls
idiot
with
firewall,
so
they
send
lots
of
queries
and
never
get
a
response,
so
they
send
more
queries
and
they
never
turn
anything
off.
They're
like
well,
it
isn't
working
so
I'll.
Add
another
thing
and
I'll
do
this
and
I'll
do
that?
S
M
Certainly
seems
in
Chrome
in
its
tests
of
liveness,
which
it
sends
off
random
names
towards
the
root.
There
is
this
attitude
in
other
parts
that
the
DNS
is
infinite,
is
fantastically
engineered
and
and
works
exactly
as
planned.
All
the
time
and
people
could
just
throw
queries
in
it.
Whatever
magnitude
they
want
and
they'll
always
be
an
answer.
M
I
wish
it
was
so
you
know,
but
the
fact
that
it
isn't
creates
this
astonishing
sort
of
amount
of
tension
between
what
application
folk
believe
the
DNS
is
doing
and
what
operators
see
it's
doing
and
where
DNS
code
then
to
sit
in
the
middle
going.
What
the
hell
should
I
write
next
time,
how
to
defend
ourselves
and
folk
around
our
code.
It's
a
tough
one.
I
Aaron
Falk
alchemy,
thank
you
for
you
talk
Jeff
the
two
quick
points
and
a
question.
Thank
you
for
sustaining
my
belief
that
there's
a
lot
of
weirdness
in
the
internet,
I'm
glad
to
see
that
hasn't,
changed
and
I.
Think
that
you
gave
your
the
the
reason
in
I
think
your
first
couple
of
sentences,
which
is
that
it's
free
right,
there's
like
many
of
our
tragedies
of
the
Commons,
there's
really
no
reason
not
to
keep
trying.
It
would
be
interesting
if
there
was
a
way
to
quantify
the
cost
in
your
organization
from
the
the
extraneous
requests.
I
M
M
We
know
what
we're
doing
because
the
names
have
meaning,
but
your
server,
which
have
been
running
for
years,
which
has
a
whole
bunch
of
gross
queries
and
you're.
Seeing
a
thousand
queries
a
minute
is
that
nine
hundred
and
ninety
and
one
real?
How
do
you
know
no
one
does
the
suspicion
is
that
you
are
overwhelmed
with
your
legacy
and
that
real
queries
are
actually
extremely
low,
but
they're
masked
out
by
this
extraordinary
property
of
the
DNS
to
just
simply
memorize
a
repeat
and
recycle
quantifying
that
difficult
that
it
certainly
seems
bad
one.
I
M
A
lot
of
log
re
players,
so
there
is
a
massive
business
out
there
of
capturing
and
then
replaying
and
generally
they
are
relatively
benign,
but
we
also
see
a
whole
bunch
of
other
things
where
pathologies
that
are
inexplicable
there
was
a
DNS
resolver
in
Jordan
and
other
one
in
Guatemala
doing
up
to
20,000
queries
per
second
for
days
of
the
same
name,
it's
kind
of
I'm.
Sorry,
what
exactly
is
going
on
here
guys
you
know.
So
the
combination
is
just
an
amazing
pool
of
awfulness.