►
Description
Want to view more sessions and keep the conversations going? Join us for KubeCon + CloudNativeCon North America in Seattle, December 11 - 13, 2018 (http://bit.ly/KCCNCNA18) or in Shanghai, November 14-15 (http://bit.ly/kccncchina18).
Edge Computing and Kubernetes BOF - Matthew Caulfield, Cisco (Any Skill Level)
Matt will lead an open discussion about edge computing with Kubernetes. The discussion may cover use cases, technology choices, growing pains, and what has worked for others.
If you are actively working on edge computing please attend and contribute to the discussion! If you are curious about edge computing and Kubernetes bring questions.
About Matt Caulfield
Matt is a Principal Engineer working out of Cisco's New England Development Center. He focuses on distributed systems, networking, real-time video processing, and cloud native architectures. Currently, he leads the Edge Computing team at Cisco.
Join us for KubeCon + CloudNativeCon in Barcelona May 20 - 23, Shanghai June 24 - 26, and San Diego November 18 - 21! Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy and all of the other CNCF-hosted projects.
A
A
So
we
could
start
really
digging
in
with
the
community
to
suss
out
what
are
the
real
use
cases
around
edge
computing,
and
you
know
what
are
the:
what
are
the
actual
possible
implementations
of
a
system
that
could
actually
support
edge
computing?
So
there's
going
to
be
a
pretty
open
discussion,
I,
don't
really
have
many
slides
or
anything
I
really
just
want
to
have
a
real
good
back
and
forth.
This
had
to
be
a
QA
I,
don't
you
know
it
should
be.
You
know,
multilateral
before
we
get
started.
A
You
know
if
anybody
wants
to
be
part
of
this
discussion
on
an
ongoing
basis.
Outside
of
you
know
the
bounds
of
today's
session
I
put
together
a
little
Google
form
that
you
can
fill
in.
So
if
you
can't
read
that
that's
a
lowercase
F
uppercase
m5
uppercase,
a
uppercase
Z,
lowercase
M,
you
can
go
there
and
just
fill
in
your
name
and
your
email.
Yes,
yes,
this
will
be
uploaded
on
the
site,
so
you
can
always
check
that
URL
later
but
yeah.
A
If
you
do
sound
on
that
up
on
that,
we're
not
gonna
share
your
information
or
anything
I
just
want
to
use
it
to
keep
track
of
people
and
maybe
put
together
a
an
email
list
or
slack
channel
or
something
that
we
can
start
actually
having
a
continuous
discussion
on
this
topic,
because
I
think
it's
it's
going
big
places.
So
with
that
information,
let's
get
started
so
I
have
a
few
discussion
topics
totally
open
to
keeping
this
as
an
open
mic
type
of
session.
A
If
anybody
has
anything
that
they
want
to
cover,
in
particular,
that's
not
on
the
agenda
feel
free
to
throw
it
in
there.
But
you
know
here's
what
I
was
thinking
so
starting
off
with
you
know.
What
is
that
computing?
Let's
see
if
we
can
get
on
to
like
a
common
definition
of
what
that
is
exactly
and
specifically,
where
is
the
edge
of
the
network,
and
there
might
be
multiple
answers
to
that
and
that's
totally
fine.
Let's
talk
about
real
use
cases.
A
So
if
anybody
in
here
has
a
network
that
they're
considering
deploying
many
thousands
of
kubernetes
clusters
in
or
even
you
know,
tens
of
Cabrini's
clusters,
that
would
be
really
interesting
too
here
and
specifically
not
just
how
you're
doing
the
Cooper
needs
part.
But
what
are
you
planning
to
run
on
those
clusters?
That's
part
of
the
mystery.
That's
out
there
right
now
in
terms
of
edge
computing.
What
is
the
real
use
case?
A
Is
there
a
reason
to
the
edge
computing
and
what
is
it
and
then
mythical,
fantasy
use
cases,
so
real-world
use
cases
and
then
fantasy
use
cases.
So
there
are
things
out
there
where
people
think
as
computing
might
be
necessary,
but
you
know
maybe
it's
a
bit
of
a
stretch,
maybe
their
use
cases
that
aren't
near-term
in
the
next.
You
know
four
or
five
years,
but
maybe
they're
real
in
the
next.
You
know
ten
years
so
maybe
they're
a
fantasy
now,
but
their
reality
in
the
future.
So
we
can
talk
about
those
talk
about.
A
You
know
beyond
the
use
cases.
What
do
we
think
is
the
architecture.
How
does
Cabrini's
fit
into
it?
What
other
tools
and
then,
specifically,
let's
dig
into
the
problems
with
Cabrini's.
What
can
we
solve
with
kubernetes
and
then
what
do
we
really
need
to
solve
at
the
application
layer
or
somewhere
else,
or
some
other
tools
try
to
figure
out?
You
know
what
you
know,
what
exactly
can
be
done?
Okay,
so
what
is
that
computing?
Any,
but
I
want
to
offer
a
broad
definition
on
that.
A
A
B
There
are
different
applications,
some
sorry
hard
constraint
on
the
edge
for
latency
reasons
right,
but
some
are
more
opportunistic.
The
under
sense
they're
good,
be
an
edge
if
there
is
enough
capacity
right,
right
or
they're,
sometimes
better
off
being
in
a
cloud.
That
means
it's
really
about
all
about
PC
whew,
not
really
not.
C
B
A
Yeah,
so
defining
edge
computing
is
kind
of
just
another
way
of
saying
distributed
computing
or
many
clusters,
and
it
could
be
for
any
number
of
reasons
that
you
know
that's
a
good
answer.
So
where
is
the
edge?
So
I'll
give
you
my
context:
I
mainly
work
with
service
providers,
so
mobile
operators,
for
example.
Like
you
know,
mobile
operators,
like
Cisco,
announced
a
big
partnership
with
this
company
in
India
called
Reliance
Geo,
and
they
have
180
million
subscribers
and
they
want
to
put
compute
at
their
cell
base
stations
right.
That's
obviously
an
edge
computing
use
case.
A
C
A
D
Okay,
there's
there's
absolutely
tons,
I
mean
so
so
we
had.
We
have
a
gentleman
here,
talk
about
anything
latency
sensitive,
so
we
have
people
talk
about
machine
learning
and
they
took
in
the
talk
earlier.
I
know
this
is
a
hot
topic,
but
anything
from
you
know
to
do
voice
recognition.
You
don't
be
sending
that
all
the
way
around
to
your
data
center
in
Europe
to
the
or
your
customers
in
Asia
another
another
example
might
be.
You
have
an
app
and
you
need
to
backup
the
data
somewhere.
D
D
Mean
I'm
not
saying
that
that's
it'll
be
an
example
use
case.
If
I
was
running
an
app
and
I
can't
think
of
what
and
yeah
there
was
no
net
network
congestion
back
to
my
own
data
center
somewhere.
Then
why
wouldn't
I
store
it
locally
and
then
upload
it
when
the
networks
less
congested,
because
you
know
but
then
again
it
depends
on
what
your
definition
of
edge
is
yeah.
A
C
A
G
So
we're
working
in
industrial
IOT
and
in
some
cases,
with
some
customers
you're
when
you're
manufacturing,
something
that
is
pharmaceutical,
it's
something
that
is
very
sensitive
and
there
is
not
a
possibility
for
them
or
they
don't
want
to
send
the
data
to
the
cloud.
So
in
that
case
we
need
to
do
edge
computing.
G
Mean
you
could
have
like
a
local
data
center
for
for
doing
a
compute,
but
that's
then
you're,
building
into
the
entire
cloud
infrastructure
for
one
use
case
in
this
case
that
we
have
the
IOT
with
the
data
collection
locally
and
then
in
the
normal
use
case.
We
send
it
to
the
cloud
because
that's
where
they
compute
and
the
storage
is
so.
A
G
A
F
Was
just
gonna
say
in
my
mind,
edge
computing
implies
a
bit
of
being
like
the
first
point
of
contact
for
an
end
user.
That's
their
first
access
into
the
cloud,
and
that
could
mean
colocation.
It
could
mean
specific
advertisements
being
sent
to
them,
based
on
where
they're
at
or
something
like
that.
Okay,
okay,.
A
Yeah,
so
edge
computing
isn't
really
a
well-defined
term
I
guess
then
that's
what
we're
getting
at
here.
So
I
think
we
know
it's
not
what's
in
the
public
cloud.
We
do
know
that
it
could
be
a
Nichola
facility.
It
could
be
at
a
cell
tower
or
a
little
bit
back
from
a
cell
tower
in
an
aggregation
point
of
the
network
or
a
hub
or
a
pop,
depending
on
what
type
of
telco
you're
talking
to
or
it
could
be
in
an
enterprise
network.
It
could
be
on
an
oil
rig
right
or
it
could
be.
A
H
I'm
thinking
about
this
as
a
mix
of
perspectives
like
so
you
think
about
computing
and
you
think
about
location,
and
you
think
about
computing,
really
think
about
any
of
those
so
alone.
You
need
to
combine
them
so
that
overlaps
a
lot
with
what
you
would
do
with
and
say.
If
you
were
thinking
about
software,
fine
networking
to
switch
things
back
and
forth
based
on
location.
H
A
Yeah,
so
that's
fair.
Let's
stop
trying
to
and
I
know
that
I
started
this,
but
let's
try
to
stop
defining
edge
computing
and
move
on
to
the
use
cases.
So
maybe
we
can
figure
it
out
by
recognizing
a
pattern
in
the
use
cases.
So
I'll
give
you
my
use
case
and
then
I
want
to
hear
other
use
cases,
but
my
use
case
that
I'm
familiar
with
is,
is
the
South
Tower
one
right,
and
maybe
it's
not
a
cell
tower,
but
at
an
aggregation
point
in
a
service
provider
network
and
I.
A
Don't
want
to
make
this
a
service
prior
or
a
telco
conversation
if
there
are
manufacturing
use
cases
or
other
use
cases
I
want
to
hear
those
absolutely.
My
use
case
is
service
projects.
Cisco
I
work,
for
they
work
a
lot
with
the
big
service
providers
around
the
world.
We
saw
them
lots
of
routing
equipment.
Many
of
them
are
starting
to
think
about.
A
Well,
maybe
we're
not
putting
all
these
functions
in
a
router,
maybe
we're
putting
a
lot
of
these
functions
into
general-purpose
computing
next
to
our
routers
on
our
network
and
our
routers
are
everywhere
they're
deployed
in
a
wide
footprint
close
to
people's
home
by
the
last
mile,
a
little
bit
back
from
the
last
mile.
It
doesn't
really
matter.
We
want
the
ability
to
go
and
deploy
network
functions
out
there
that
are
flexible
and
then
now
that
we're
going
and
building
out
these
little
micro
datacenters
at
the
edge.
A
We
want
the
ability
to
go
and
deploy
things
like
I'll
give
you
a
few
initial
examples,
video
caches,
and
so
that's
a
very
real
initial
use
case.
So
video,
caching
and
then
IOT
type
gateways,
so
that
today
is
something
that
actual
customers
are
actually
buying
and
that's
a
real
use
case.
But
I
do
want
to
hear
from
everybody
else.
What
what
are
the
other
real
use
cases
for
exercise
that,
yes,.
E
My
wife
works
for
a
rather
large
hospital
system
and
they're,
not
not
quite
what
I
would
define
as
edge
computing,
but
one
of
the
use
cases
for
them
is
that
there
are
a
lot
of
doctors
offices
within
the
hospital
itself
and
the
system.
Previously
they
had
one
large
data
center,
where
all
the
hospitals
were
just
connected
directly
to
that
right.
E
What
they're
starting
to
do
now
is
put
caches
in
each
of
the
smaller,
especially
the
smaller
hospitals
and
the
night
before
they
will
push
all
of
the
x-rays
files
everything
down
to
that
local
office,
so
they
don't
have
to
reach
back
to
the
main
data
center.
One
saving
them
thumb
bandwidth
costs
during
the
day,
but
also
having
that
the
data
they
are
locally
accessible.
So
if
there
is
some
type
of
outage
that
customers
aren't
impacted
by
it,
okay.
A
Yeah,
so
I
think
that
health
care
is
actually
a
big
use
case
and
I
know
that
anybody
who
went
to
the
the
edge
talk
this
morning
by
Steve
and
Megan
I
think
health
care
was
one
of
the
use
cases
that
they
talked
about
as
well.
I
think
it's
really
interesting,
pre
positioning,
so
the
two
types
of
caching
right
on
demand.
Caching
and
then
pre-positioned
caching
pre
position
cache
and
it's
really
interesting.
Especially
for
you
know,
you
know
medical
imaging
yeah
that
makes
sense
other
use
cases
yeah.
I
Hi,
so
so,
actually,
my
company
provides
like
this
concept
of
edge,
computing
and
stuff,
like
that.
We
actually
call
it
like
our
edge
networking
or
edge
sites,
but
actually
so
it
gets,
it
does
get
I'm
seeing
it
gets
confusing
like
is
it
the
cloud
or
is
it
the
edge
right?
Which
should
you
pick
at
packet?
We
actually
that's
actually
the
same
thing.
So,
for
example,
you're
saying
that
you
want
to
use
yeah,
we
definitely
the
edge
is
considered
somewhere
close
to
wherever
your
workload
is
sort
of
example,
like
in
the
healthcare
business.
I
They
like
or
processing
all
the
data
like
that
they
collect
together
like
much
closer
but
in
our
case,
actually
we're
trying
to
build
an
edge
Network
for
computing
for
networking,
for
processing
or
for
just
locally,
because
you
want
to
be
local
local
to
a
certain
site,
we're
building
that
as
still
we're.
Considering
that
the
cloud
so
I'm
not
sure
what
you
know,
everybody
has
its
own
definition
of
what
the
cloud
is
is
the
cloud
Amazon
is
the
cloud
just
you
know
these
big.
I
You
know
the
top
three
providers,
or
in
our
case
we
automate
these
data
centers
or
racks.
It
doesn't
have
to
be
a
full-blown
data
center.
You
know
we
automate,
like
you
know
like
if
it's
in
a
hospital.
Well,
you
know
we'll
put
our
automation
software
and
make
it
very
cloud
like
so
your
location
when
you
deploy
your
region.
Is
you
know
that
building
or
that
city
and
stuff
like
that?
So
that's
art.
Yes,.
A
I
That's
exactly
yeah,
yes,
so
our
customers
actually
help
us
expand.
So
we
launched
our
sites
to
get
closer.
We
don't
have
a
large
configuration
on
there,
but
we
actually
allow
customers
to
expand
on
that
and
today's
world
yeah
they
actually
have
one
over
popular
sites
like
Hong
Kong.
They,
you
know
some
customers
wanted
better
networking
and
latency
and
routing
they
also
leverage,
because
one
of
the
reasons
like
you're
saying
being
closer
to
ISP,
they
code
through
our
network
right
closer
to
the
edge.
But
then
we
leverage
on
our
network.
A
J
To
add
to
Tony's
thought
regarding
autonomous
cars
and
that
they
will
drive
the
world
basically
they're
gonna
need
lots
of
compute
sure,
and
that
depends
you
can
do
that
compute
machine
learning
and
create
the
modules
somewhere
in
the
central
datacenter
or
you
can
do
that
on
the
edge.
And
can
you
imagine
how
many
GPUs
are
crypto
mining
right
now
in
somebody's
apartment
houses
working
just
as
radiators,
basically
without.
C
J
J
A
Yeah
I
I
think
that's
actually
really
unknown
if
we
want
to
spend
a
ton
of
time
on
it,
but
interesting
use
case
of
we
talked
about
as
computing
a
lot
within
the
context
of
a
single
administrative
domain.
But
we
know
the
Internet
is
kind
of
made
up
of
multiple
administrative
domains
that
are
connected
to
each
other
through
BGP
and
other
protocols.
Do
we
see
the
same
thing
happening
for
workloads,
I
guess
could
and
I
know
that
you
know
workload.
Migration
was
like
a
common
theme
that
kind
of
petered
out
a
little
bit
with
with
cloud.
A
B
I
work
for
being
better
and
speaking
a
little
even
broader,
speaking
for
onap,
open
networking,
automation
platform
where
actually,
interestingly,
it's
from
service
providers,
several
CNC
of
vendor
members
are
part
of
it
when
I'm,
including
Cisco.
Yes,
so
really
what,
when
we
started
looking
at
this
problem,
onap
was
so
far
focused
on
network
services,
but
recently
what
we
said
was
the
opportunities
much
bigger.
We
need
to
bring
in
edge
computing
literally
and
as
we
in
the
journey,
what
we
realized
was.
We
need
to
take
a
broader
look
at
that.
B
The
whole
picture
right
I
mean
essentially
start
with
an
application
perspective.
What
are
the
requirements
right?
We
identified
several
application
categories.
Some
of
them
have
a
hard
real-time
requirement.
You
know
example,
for
example,
if
you're
doing
something
like
a
5g
mindful
function,
there
are
certain
control
functions
as
load.
Balancing
just
really
need
a
real-time
response.
Time
and
I
am
talking
like
20
milliseconds
order
right
and
they
have
to
be
have
to
be
on
the
edge.
B
There
is
no
other
way
right,
but
also,
if
you
step
into
the
next
category
broader
category,
we
see
that
applications
of
this
video
surveillance
yep.
Our
hot
candidate,
like
you,
talked
about,
but
it
is
not
that
it
is
absolute
when
it
comes
to
the
latency
requirement
there
500
millisecond
and
above
so
it's
not
a
hard,
real-time,
it's
constraint,
but
it
almost
comes
to
more
opportunistic
its
deployment.
B
But
if,
under
certain
constraints
you
know,
essentially
you
do
a
really
a
tseebo
math
I
mean
basically
an
object,
objective,
optimization
or
TCU,
and
you
figure
out
that
edge
could
be
the
light
location
or
it
could
be
somewhere
else.
Somebody's
in
the
public
cloud,
maybe
even
an
army
is
on
ec2
by
the
right.
Partnership
could
be
the
right
location
for
it.
Yeah.
A
So
I
guess:
let's
draw
a
distinction.
I
guess
I
think
there
are
network
functions
that
must
be
at
the
edge
because
that's
kind
of
where
they're
today
and
we're
virtualizing
them
and
putting
them
on
to
general
purpose
compute
and
then
there
are
functions
like
video
surveillance,
where
it's
beneficial,
but
at
the
edge,
but
the
network's
still
gonna
work
if
it's
not
at
the
edge.
So
I
think
that's
what
you're
saying
early.
B
But
sinister
and
I
just
wanted
to
add
even
the
network
functions
interesting.
If
you
break
down
the
data
plane
functions
shortly,
critical
in
the
edge,
but
even
control
plane
can
be
somewhere
else.
Okay,
so
it's
so
such
a
broad
definition.
That's
one.
We
said
when
you
look
at
the
application
landscape,
it's
better
to
call
it
more
distributed
cloud
computing
right.
So
the
third.
The
message
gets
better
and
also
sort
of
star
starting
top-down.
What
are
you
looking
from
architecture
right
got.
B
A
G
Think
I
can
just
generalize
what
I
said
before.
Is
that
the
reason
why
we're
doing
the
edge
computing
is
that
or
my
definition
is
kind
of
baggy?
The
edge
is
where
the
data
is
produced
and
and
if
you
can
work
on
the
data
where
it's
produced,
you
reduce
a
lot
of
requirements
on
the
network
and
everything
right
and
and
what
we're
doing
is
basically
what
like
we're
we're,
extending
the
intelligence
of
pre-existing
machines
or
factory
lines.
G
And
then,
when
you
are
monitoring
a
factory
line,
you
can
have
thousands
of
data
points
per
second,
but
they
are
not
really
interesting.
Okay
or
unless
something
is
deviating
from
whatever
requirements
you
set
on
them.
So,
okay,
so
the
H
is
always
monitoring,
always
computing
at
a
high
speed.
And
then,
whenever
you
get
a
deviation,
you
send
that
to
the
cloud
and
to
whoever's.
A
K
A
L
So
the
one
I
consider
is
pretty
much
you
know
using
for
distribution
of
video
yeah,
not
necessarily
for
caches.
Let's
say
we
want
to
have
you
know,
IPTV
enabled
you
know
shoe
clusters
of
users.
We
could
have
a
let's
say,
though
such
computing
devices
be
the
multicast
and
the
point
yeah
and
then
having
unicast
and
to
the
very
very
edge
of
the
the
actual
customer
right.
L
A
Yeah
CDN
type
use
cases
with
multicast
that
make
sense
okay.
So
if
you
have
additional
use
cases,
we
didn't
talk
about,
I
mean
there's
a
comment
section
on
the
forum,
so
feel
free
to
throw
them
in
there
and
I'll
try
to
keep
track
of
those.
These
are
all
really
good.
Just
going
forward
with
you
know,
I
will
skip
the
fantasy
use
cases
because
they're
not
necessarily
real
today,
but
in
terms
of
instrument
architectures.
A
So
the
way
that
I
think
about
it
is
in
each
of
these
edge
sites,
I
think
of
each
one
of
them
as
a
micro,
datacenter
and
I.
Think
of
each
one
is
a
standalone
kubernetes
cluster
I've
seen
other
architectures.
Where
somebody
wants
to
run
kubernetes,
you
know
master
centrally
and
then
teach
you
know,
treat
all
the
edges
as
workers.
So
is
there
I
guess
consensus
that
each
of
these
edge
sites
should
be
its
own
cluster,
or
do
we
see
some
kind
of
other
division
of
responsibilities
that
make
sense.
A
A
Right
yeah,
so
I
guess
is
kubernetes
deficient
in
terms
there's
multi-tenancy
today
and
do
we
expect
that
to
change
in
the
future
or
should
we
just
use
multiple
clusters?
That's
probably
the
easiest
thing
to
do
maybe
use
at
least
today,
I.
Think
using
multiple
clusters
is
the
sensible
approach
per
site.
Okay,
that
could
make
sense.
Sure.
B
Depends
on
the
size
of
the
edge
site
and
type
of
workloads
you're
trying
to
deploy
there
right,
so
one
possibility
could
be
I
mean
each
head
site
could
have
just
a
single
rack,
and
it's
not
that
you
want
to
have
one
kubernetes
cluster
per
rack
right
and
you
could
actually
I
say
that
hey
there
are
five
sites
which
are
proximal
and
Metro
distance,
and
then
they
could
be
controlled
with
the
same
kubernetes
questions
totally
possible.
It's.
A
A
A
But
in
terms
of
how
do
you
manage
that,
when
you
have
thousands
of
clusters,
is
the
question
I
really
want
to
get
at
so
does
anybody
have
any
techniques
that
they're
using
today
I
know
there's
a
lot
of
like
multi
cluster
registry
is,
you
know,
is
here
now
and
we
started
using
that,
as
you
know,
one
of
our
pox,
but
are
people
planning
to
make
use
of
that?
What
are
you
planning
to
layer
on
top
of
registry?
A
Is
there
anything
else
or
any
techniques
that
people
are
actually
deploying
today
to
solve
some
of
these
use
cases,
or
are
we
really
early?
It
might
be
that
we're
early,
okay,
okay,
maybe
we're
just
think
about
it,
so
that
that's
fine.
So
what
about
like
Multi,
cluster
load?
Balancing?
Do
you
people
so
there's
multi
cluster
in
ingress,
right
that
works
with
the
Google
load,
balancer
and
that's
kind
of
a
global
load
balancer?
A
But
when
we're
managing
things,
unlike
a
you
know,
mobile
operators
Network
sending
something
back
to
the
Google
load,
balancer
and
then
somehow
getting
it
back
to
your
edge
site
might
make
sense.
So
is
there
room
for
a
global
load?
Balancer,
that's
not
in
the
public
cloud,
but
rather
it's
it's.
You
know
on
Prem,
it's
it's
a
privately
hosted
global
load,
balancer
yeah.
I
I
G
A
G
A
Right
yeah,
so
just
to
rehash
demo
that
my
group
did
earlier
today.
Essentially,
we
wrote
a
core
DNS
plugin
and
we
rely
on
any
cast
and
you
know
BGP,
to
kind
of
get
us
to
the
closest
instance
of
that
IP
address,
which
should
be
running
at
your
closest
edge
cluster
and
then
relying
on
a
you
know:
smarts
there
to
figure
out.
Okay,
is
that
service
here
or
is
it
somewhere
else
and.
G
What
is
nice
about
that?
For
the
user
point
of
view
is
that
whenever
you're
in
the
factory
or
connecting
directly
to
your
device
that
you're,
but
then
they
frame
on
your
mobile
phone
and
then
you
walk
out
of
the
factory
you,
you
will
still
see
the
same.
There
would
be
no
changed
from
a
user
perspective,
but
it
would
just
connect
to
a
different
cloud
or
a
different
cluster.
Yeah
I.
G
D
A
D
If,
like
supplications,
for
example,
you
might
want
to
target
the
application
to
a
specific
person,
so
you
look
at
Netflix,
for
example,
they're
not
going
to
send
someone
on
Verizon
to
their
Comcast
person,
so
it's
so
some
applications.
It
might
be
that
you
need
to
talk
to
some
sort
of
central
registry,
which
then
says
you
need
to
go
to
this
exact
pipe
and
obviously
that
could
be
on
the
application
layer.
But
like
what
you
didn't
on
the
core
DNS,
you
could
extend
that
a
bit.
D
So
it
could
be
that
it's
any
cast
isn't
perfect,
but
it
could
be
that
any
cast
gets
you
to
the
nearest
DNS
server.
What
you
say
is
pretty
good
and
then
you've
got
some
sort
of
plug-in
on
your
core
DNS,
which
then
takes
all
the
information.
The
source
IP,
the
load
on
the
network.
I,
don't
know
what
clusters
are
healthy
or
not
healthy
and
then
returns.
Some
sort
of
sensible
thing-
and
that
seems
like
it's
to
me,
seems
like
it's
fairly
to
survive,
see
my
viable
to
do.
Okay,.
A
H
So
little
bouncing
means
that
you
just
direct
to
a
computer
cluster
right
now,
there's
an
alternative
to
that,
and
that
is
just
to
allocate
more
network
resources
so
that
you
get
more
access
to
whatever,
wherever
the
wherever
they
compute
resources
are.
So
those
two
are
actually
substitutes
and
if
you
can
manipulate
one,
you.
F
H
A
A
A
So
I
guess
that
brings
us
next
bullet,
which
is
kind
of
resource
scarcity
and
auto
scaling,
which
is
like.
Can
you
scale
up
and
down
the
application?
And
you
know,
can
you
deal
with
resource
scarcity?
So
one
thing:
that's
really
different
between
the
edge
versus
the
public
cloud,
I
kind
of
think
of
the
public
cloud.
Not
everybody
thinks
about
this
this
way,
but
there's,
like
a
you,
know,
bottomless
pit
of
resources
versus
the
edge,
which
is
very
much
a
finite
set
of
resources.
So
we
have
a
really
primitive
way
to
set
priority
for
workloads.
A
Today
right
you
can
say
that
this
is
a
you
know,
a
system
pod
or
whatever,
and
this
is
a
regular
prop
pod.
What
I
think
we
need
going
forward
is
probably
more
of
a
priority
for
what's
running
at
the
edge,
so
you
might
say
that
oh,
we
need
the
network
functions
to
run
there
all
the
time
with
high
priority,
but
the
video
surveillance
thing
you
know,
maybe
that's.
Okay,
if
that
gets
evicted
from
the
edge
of
the
network,
yeah.
G
The
difference
between
cloud
and
edge
is
that
often
of
the
case
that
what
you're
paying
for
usage
in
the
cloud.
So
if
you
do
more
you're
paying
more
but
right
at
the
edge
it's
like,
if
you
have
spare
capacity
for
for
doing
computing,
and
then
is
it's
not
free,
but
it's
sure
like
use
whatever
is
available.
Instead
of.
G
A
Makes
sense
and
I
guess
if
you
think
about
what
we're
talking
bout
earlier
like
who
we
connect
all
these
edges
together
into
one
big
public
edge
that
everybody
can
have
access
to
then
there's
a
question
of
well
there's
resource
scarcity.
How
do
you
decide
who
wins
like
if
I'm
gonna
pay
a
dollar
and
they're
gonna
pay
a
dollar?
Who
is
there
like
an
auctioning
weight
of
auction
off
these
edge
resources
that
are
so
scarce,
yeah.
M
M
We
are
working
on
something
that
also
allows
you
to
apply
those
policies
to
computer
resources,
because
you
can
say
the
first
one
that
ask
for
resources
gets
them
for
a
given
price.
But
then
the
resources
are
all
allocated
right
and
no
one
else
can
get
to
the
resources.
So,
as
we
were
saying
this
morning,
you
have
a
limit
to
the
access
of
resources,
which
is
actually
something
related
more
to
economic
theory.
Let's
say
yes,.
A
E
A
That's
a
good
point,
so
is
it
always
gonna
be
our
own
workloads
that
we
trust
or
would
it
be?
You
know
first
part
again,
second
party,
or
would
there
ever
be
a
place
in
time
where
we
adopt
a
third
party
and
maybe
I
think
you're
right?
The
third
party
thing
is
kind
of
a
magical
unicorn,
that's
probably
years
out
in
the
future.
The
second
party
to
use
kay
is
probably
more
real,
but
then
you
probably
know
how
many
resources
that's
gonna
require
it's
not
like.
You
need
a
balance.
Yeah.
M
It
depends
on
who
is
the
owner
of
the
infrastructure
and
how
is
difficult
to
install
new
edge
nodes
because,
for
example,
if
the
infrastructure
is
a
public
infrastructure
that
can
be
used,
for
example,
for
self-driving
cars,
I
am
a
manufacturer
of
self-driving
cars,
but
I
don't
want
to
deploy
an
edge
computing
infrastructure
on
a
whole
country.
I
will
use
the.
In
fact.
They
are
public
infrastructure.
Okay,.
A
Folks,
that's
a
good
question
yeah,
so
we're
at
time
actually
and
I
personally
have
to
run
I'm
sorry
but
feel
free
to
continue
talking.
We
probably
have
to
get
kicked
out
of
the
room,
but
if
you
can
I
sign
up
for
the
interest
list,
I'll
try
to
you
know,
get
the
people
on
there
together,
either
virtually
or
physically,
and
then
we
can
take
this
forward.
If
you
want
to
contact
me
directly,
that's
that's.
My
personal
email
feel
free
to
send
me
spam
or
whatever
you
want.