►
From YouTube: Kubernetes WG IoT Edge 20230503
Description
May 3, 2023 meeting of the CNCF IoT Edge Working Group. No agenda items submitted so this meeting is an open "birds of a feather" discussion on IoT Edge
A
Hi
welcome
to
the
May
3rd
meeting
of
the
cncf
iot
edge
working
group
on
today's
agenda.
I
put
a
number
of
very
brief
things
that
are
more
in
the
recap:
interesting
item
territory,
but,
as
always
group
members
are
free
to
late,
add
anything
that
you
like.
If
we
don't
get
time
for
at
this
meeting,
we'll
just
carry
it
on
to
the
next.
A
Also,
if
we
run
out
of
agenda
items
we'll
just
go
into
freeform
birds
of
a
feather
open,
Forum
discussion
on
whatever
members
are
interested
in
chatting
about
you're
free
to
bring
up
questions
or
anything
interesting
in
the
field
of
iot
and
Edge
that
you've
come
across.
You
know
in
in
the
last
few
weeks.
Let
me
quickly
go
over
the
things
that
I
added
into
there.
A
The
first
one
is
that
the
Eclipse
Foundation
sponsors,
50,
plus
Edge,
related
open
source
projects,
and
they
have
been
hosting
an
annual
survey
of
iot
and
Edge
commercial
adoption
and
developer
usage,
so
that
survey
recently
opened
and
there's
a
link
in
the
agenda
node
stock.
But
let
me
put
it
in
the
chat
as
well,
so
they
welcome
your
participation.
That's
been
running,
I!
A
Think
just
short
of
300
people
participating
in
that
survey
last
year,
if
I'm
not
mistaken,
so
in
terms
of
how
these
things
go,
I
think
it's
a
reasonably
decent
Source
of
information
on
Trends
in
the
iot
edge
space.
You
can
go
back
and
find
last
year's
survey,
but.
A
Know
they're
getting
this
year
started
and,
as
you
answer
the
questions
at
the
very
end,
you
have
the
option
of
providing
an
email
to
get
notified
of
the
results
when
they
tally
them.
Second
item
kubecon
China
is
going
to
be
held
in
Shanghai
in
September.
It's
interesting
that
they
just
announced
this
recently
and
slipped
it
in
before
kubecon
North
America,
the
cfps
are
open
right
now,
I
believe
they're
open
until
July
18,
but
you
better
check
on
that.
A
When
they've
done
China
in
the
past
also
for
the
rest
of
world,
they
often
close
the
nominations
at
midnight
Pacific
time
in
the
U.S,
but
in
China
they
base
it
on
whatever
the
calendar
runs
there.
So,
depending
on
where
you
live,
it
could
be
even
a
full
day
off
from
where
you
are
so.
Keep
that
in
mind.
A
If
you're
interested
in
doing
a
cfp
there
you
check
what
time
zone
and
reference
points
they're
using
for
the
cfp,
if
you're
the
typical
procrastinator
and
are
going
to
push
the
envelope
on
how
late
you
submit
there
is
a
networking,
Edge
and
Telco
track
available.
There
then,
in
the
agenda,
notes
I
also
put
a
link
to
the
kubecon
North
America's
event:
that's
going
to
be
November
6th
through
9
in
Chicago.
A
That's
also
currently
open
for
cfps
I
think
that
one
is
open
until
June
18th,
so
it's
open
for
a
little
more
than
a
month,
and
also
they
have
the
same.
Networking
Edge
and
Telco
track
available
I've
been
informed
that
there
will
be
another
kubernetes
on
the
edge
Day
event,
or
at
least
they
are
recruiting
volunteers
to
help
organize
it.
So
I
think
that's
going
to
take
place,
but
for
now
cfps
aren't
open
and
you.
B
A
They
haven't
announced
a
day
or
any
details,
but
I
believe
it's
likely
to
happen
and
I
hope
so
because
the
that
event
in
Europe
I.
A
A
This
satellite
feature
is
made
for
Edge
locations
that
are
either
air,
gapped
or
intermittently
connected,
and
they
posted
a
YouTube
video
that
covers
this.
The
link
I
just
dropped
in
the
chat
has
a
Time
index
to
skip
ahead
to
the
part
on
the
satellite
feature,
which
was
only
you
know.
It
was
a
hour-long
presentation
on
Harbor
in
general,
but
this
jumps
to
the
spot
where
they
talk
about
the
satellite
feature,
so
I
have
a
feeling
that
people
operating
in
the
edge
Community
might
be
interested
in
that.
A
Finally,
an
interesting
item-
I
just
came
across
in
the
past
day
is
a
have
become
interested
in
the
opportunity
for
AI
and
machine
learning
at
Edge,
and
there
is
an
org
called
tiny
ml
that
is
pursuing
machine
learning
at
Edge
and
I
found
that
they're
hosting
a
hackathon
that
is
I,
think
partially
based
on
the
city
of
San
Jose,
smart
cities.
But
it's
a
hackathon
that
challenges
you
to
come
up
with
AI
apps
that
use
image,
recognition
to
count
pedestrians
in
crosswalks
so
and
they
have
cash
prizes,
sounded
kind
of
interesting.
A
It
looks
like
the
nomination
started
a
few
weeks
ago,
but
it
isn't
clear
what
the
deadline
was
when
I
looked
I
just
glanced
over
it,
but
looked
like
kind
of
an
interesting
project
and
I'm
convinced
that
this
whole
field
of
AI
at
Edge
is
just
wide
open
at
this
stage.
But
it's
going
to
be
huge.
You
know
the
world
got
an
Awakening
by
the
popularity
of
chat
GPT,
which
is
text
in
Cloud,
but
the
fact
is
that
most
of
the
world's
data
originates
at
Edge
and
I.
A
Think
the
opportunity
for
using
Ai
and
ml
is
actually
bigger
at
Edge
than
it
is
with
text
processing
based
on
chat,
we'll
see
what
happens
but
like
I
say
this.
This
thing
is
something
I
found
interesting
with
that
said.
A
That
runs
out
of
the
things
that
were
submitted
formally
as
agenda
items.
If
anybody
else
wants
to
bring
up
something
go
for
it
anything
you've
come
across
interesting
or
questions.
Observations
suggested
discussion,
topics.
D
Hey
Steve,
just
as
I
know,
I
meant
to
watch
that
Harbor
video
and
did
not
get
a
chance
thanks
for
putting
it
back
up
and
secondly,
we
just
we
put
in
an
RFP
for
kcp
Edge
for
for
kukan
North,
America
I,
just
put
it
in
yesterday
and
kubecon
China
as
well.
A
That
sounds
good
if
you
want
to
give
us
an
update
on
any
I've
missed
the
last
community
meeting
for
kcp
Edge.
But
if
you
want
to
give
us
a
brief
update
on
anything,
there
that's
been
happening
there.
Yeah.
D
I
can
keep
it
brief,
actually,
for
you
and.
A
D
Yeah,
so
we've
built
kcp
edge
with
our
sister
technology
kcp
community
and
the
the
real
gist
of
it
is,
is
to
try
and
give
people
the
ability
to
have
a
one-to-many
relationship
with
Edge
locations
so
to
be
able
to
deploy
workloads
to
many
different
devices
without
having
to
being
encumbered
with
having
to
hand
pick
or
manually
recraft
each
time
you
wanted
to
put
a
yaml
and
select
and
put
it
out
to
some
match
label
Etc.
So
that's
the
main
gist
of
it.
D
We
were
originally
achieving
that
using
kcp,
but
we've
now
moved
past
kcp.
The
kcp
is
now
being
they're.
Looking
for
new
suitors
there,
red
hat
is
no
longer
going
to
be
backing
that
project
as
the
maintainer,
so
I
don't
have
any
update
there
as
to
how
that
project's
coming
along,
but
we've
moved
our
attention
to
other
kubernetes
distributions
and
we
feel
that
we
can
do
the
same
type
of
work
with
any
kubernetes
distribution.
So
that's
going
to
be
pretty
cool
if
you're
interested
in
learning
more
about
it.
D
A
Now
you
tickled
my
interest
with
that
comment
that
you
know
about
Red
Hat,
maybe
reducing
the
contribution
level
to
the
parent
kcp.
How
how
strongly
is
kcb,
Edge
tied
to
kcp
itself?
Do
you
share
components,
or
is
it
just
kind
of
a
fork
with
an
Associated
mission
that
is
largely
independent.
D
Yeah
so
at
this
point
we
are
using
it
for
the
next
month
or
two
we'll
be
using
kcb
componentry
underneath
the
covers,
but
we
are
moving
away
from
it
to
something
more
generic.
So
for
the
next
two
months
we
have
an
exposure,
but
we're
moving
past
it
because
we
see
value
in
other
kubernetes
distributions.
A
Now,
in
terms
of
architecture,
when
you
attach
to
kubernetes
distributions
based
on
the
way
KCT
Edge
works,
could
you
actually
even
go
at
sort
of
I
I
call
it
like
I,
don't
know
if
it
were
all
public
cloud
and
I
realize
Edge?
Isn't
they
call
it
multi-cloud?
When
you
deal
with
kubernetes
on,
say
AWS
versus
Azure
versus
gcp?
A
D
So
there's
two
there's
two
areas
where
the
different
kubernetes
distros
come
into
play
so
first
is
delivering
or
deploying
a
workload
to
a
given
distro
that
we
can
do
today.
So
anything
that
speaks
kubernetes
we
can
deploy
to
at
this
point,
then
there's
the
underlying
API
machinery,
and
today
we
use
for
the
next
few
weeks
we're
going
to
be
using
kcp,
but
we're
looking
to
make
connectors
that
will
work
across
any
kubernetes
distribution
using
something
we
call
a
denatured
View.
D
So
that
allows
us
to
deploy
a
workload
to
a
kubernetes
distribution
and
not
have
it
necessarily
be
applied
there.
So
what
does
that
mean?
So
we,
if
you
take
a
deployment,
a
pod,
you
take
a
config
map,
a
secret
and
you
apply
it
to
any
given
namespace
in
a
kubernetes
distro
it
activates
it
deploys.
It
creates
a
pod.
It
can
use
those
resources.
What
we're
creating
is
what
we
call
denatured
views
whereby
you
can
deploy
or
apply
a
kubernetes
resource
to
a
namespace
and
nothing
happens,
and
why
is
that
important?
D
So
this
is
the
this
is
where
we
think
the
power
of
what
we're
experimenting
with
gets
us
into
hierarchical
distribution
of
workloads
at
a
much
faster
pace
and
I.
Think
it's
going
to
become
I.
Think
people
will
catch
on
to
how
this
works
quite
quickly
and
start
to
leverage
it
in
other
projects
fairly
soon.
A
Talk
to
I'll
call
a
mezzanine
kubernetes
control
planes
in
Metro
areas
like,
let's
just
say
you
put
one
in
America
East
another
in
America
West,
one
in
Europe,
and
then
they
in
turn
go
down
another
tier
to
talk
to
Edge
locations
running.
You
know
really
kind
of
tiny
reduced
resource,
kubernetes
clusters
and.
A
You
could
take
this.
You
could
bubble
this
down
from
the
top
tier
to
the
regional
level.
They
don't
really
act
on
it
other
than
sort
of
pass-through,
maybe
tolerating
intermittent
connectivity
even
to
the
lower
level
and
acting
as
a
caching
or
CDN
like
layer
right
that
multiplexes
out
to
large
numbers
of
edge
nodes.
Then
they
get
to
the
edge
node
and
kind
of
this
distilled
dehydrated,
spec
yeah
ads
you
pour
water
on
it,
it
grows
and
expands
and
actually
takes
root.
So
do
is
that
do
I
have
that
right,
yeah.
D
So
we
called
you
know
the
mint
that
mezzanine
level
we
call
an
intermediary
at
the
moment.
We
we
use
the
term
denature
for
the
most
part,
but
inert
is
another
one
that
comes
up.
Dehydrated
is
another
term
that
you
can
be
used
to
characterize
it
and
then,
when
it
finally
gets
to
its
final
resting
place
where
it's
intended
to
be
deployed
and
manifests
itself
as
a
pod
or
a
config
map
or
a
secret.
D
That
is
the
the
spot
where
we
rehydrate,
if
you
will
right
or
create
or
allow
the
workload
to
run,
and
so
it's
really.
You
know
it's
great,
because
now
you
have
two
plus
tiering,
three
plus
tiering
effect,
not
just
Hub
and
spoke,
and
it
allows
you
to
move
workloads
around.
As
you
mentioned,
you
know
the
regional
areas,
or
even
places
where
you
might
have
you
know
not
Regional
clouds.
What
do
they
call
those
things?
D
Sovereign,
Sovereign
clouds
as
well
right,
so
you
might
actually
look
for
places
where
you
can
do
that,
so
the
workload
you
know
will
bounce
through
there
will
be
applied
through
there,
but
will
not
necessarily
take
effect,
and
you
can
think
of
this.
As
if
you
change
the
API
or
the
type
of
object,
let's
say
from
deployment
to
deployment
dot
denatured,
you
could
get
that
effect
right
in
that
intermediary
level,
because
there
is
no
controller.
D
That
knows
how
to
to
react
or
respond
to
or
apply
something
that
has
that
extension
and
then,
once
you
get
it
to
its
final
spot,
where
it
needs
to
be
rehydrated.
You
remove
the
let's
say
the
dot
denatured
off
the
end
of
the
object
type
and
then
it
will
apply
correctly
because
the
controller
could
absorb
and
apply
it
and.
A
You're
saying
this
just
uses
kind
of
generic
kubernetes
primitive,
so
it's
not
tied
to
a
particular
disc
row.
D
Right
so
kcp
created
something
that
did
something
similar
had
some
similar
effect,
and
then
we
took
it
and
made
it
into
something:
kcp
Edge.
We
took
it
and
made
it
into
something:
that's
a
bit
more
generic
and
that's
where
we're
in
the
process
of
building
now.
A
Does
your
project
Implement
and
prescribe
a
particular
transport,
or
is
this
a
plug-in
mechanism
where
you
could
write
your
own?
The
reason
I
ask
is
there's
particularly
when
you
get
to
these
Edge
Leaf
nodes,
there's,
obviously
the
common
scenario
of
intermittent
connectivity,
but
there's
even
some
like
air
gapped,
where
right
I've
heard
of
agricultural
use
cases,
for
example,
where
somebody
there
is
literally
no
network
connectivity,
Wireless
or
otherwise.
D
D
We
have
a
we're
doing
a
joint
studies,
agreement
with
the
folks
in
Cornell
right
now
on
agriculture
and
it's
something
that
we're
going
to
experiment
with,
because
you're
right,
you
know,
when
you
get
to
these
remote
locations,
you
may
only
have
passing
connectivity
right
or
or
maybe
it's
I
think
somebody
used
the
analogy
if
it
was
you
last
week
about
somebody
putting
a
thumb,
drive
or
a
phone
in
their
backpack
yeah
that
was
driving
through
town
and
you
know
unbeknownst
to
them
or
no
action
required
by
them.
D
They
went
ahead
and
transmitted
data
Upstream.
So
that's
very
interesting
for
us
I
think
an
equitable
distribution
I
think
we
have
to
I'd
like
to
explore
that
use
case,
some
more
so
yeah.
We
are
going
to
to
look
closer
at
that
today.
I,
don't
have
that
specific
I
don't
have
that
on
the
truck,
but
the
Community
is
looking
at
that,
and
the
facility
for
synchronizing
the
workloads
to
the
edge
location
is
something
we
call
the
kcph
Sinker
and
it's
detailed
in
that
Medium
blog
post
that
I
just
put
in
the
chat.
Okay,.
A
E
So
I
was
also
curious
with
this,
so
I
just
had
other
questions
and
I'm.
You
mentioned
how
this
is
a
way
of
kind
of
it
was
denatured.
The
word
you
used
for
okay
of
getting
that
deployment
to
the
farthest
spoke
or
the
desired
spoke
I'm
curious.
How
strong
is
the
Hub
in
this
scenario,
because
I
think
the
Counterpoint
to
this
would
be
that
the
control
plane
knows
how
to
directly
deploy
to
that
farthest
spoke
that
you're
trying
to
Target.
E
So
is
the
Hub
applying
this
deployment
to
everything,
because
it
knows
all
these
nodes
and
clusters,
or
is
it
only
aware
of
that
first
layer
and
then
the
second
layer
is
doing
the
the
applying.
D
D
So
in
the
intermediary
layer,
you
could
add
more
smarts
there
to
put
another
invent
a
set
of
inventory
there
as
well,
but
I
see
that
last,
probably
less
so
a
popular
option,
I
see
it
more
so
being
more
of
a
you
know,
a
main
Hub
like
a
Smart
Hub,
so
to
speak
right.
That
has
all
those
smarts
good
question.
E
F
Yeah,
so
this
is
interesting:
India
I
think
it
sort
of
brings
to
mind
what
we
do
with
long-range
networks
like
low
raw
and
and
pre
like
istio
mesh
networks.
Really,
you
know
pushing
data
across
nodes
there.
There
is
no,
it's
not
a
multi-tiered
model
so
much,
but
it
might
be
interesting
for
your
team
or
the
community
to
look
into.
F
Essentially,
every
node
connects
with
other
nodes
with
it
creates
a
peer
group
and
through
that
peer
group,
you
can
you
can
then
forward
data
or
messages
to
other
nodes.
Eventually,
some
node
knows
what
to
do
with
the
data.
So
so,
in
this
case
a
node
would
need
to
know
hey,
it's
I
need
to
rehydrate
this
thing
and
actually
run
it,
but
you
in
in
that
mesh
Network.
You
can
actually
basically
set
it
up
so
that
you,
you
could
query
nodes
for
like
who,
who
are
your
peers,
like?
F
How
close
can
you
get
me
to
my
final
destination
and
find
and
you'd
need
almost
like
a
control,
plane
element
that
then
says
you
know,
send
it
down
this
path,
but
the
nice
thing
about
the
mesh
Network
in
this
scenario
also.
F
Is
that
because
of
the
multi-node
connectivity,
because
each
node
connects
to
multiple
peers?
If
a
node
goes
out,
you
have
multiple
paths
to
that
final
endpoint
that
you're
trying
to
get
to
but
yeah
it's
interesting
to
to
see
this
pulled
into.
You
know
the
kubernetes
and,
of
course,
mesh
networking's
really
been
overtaken
by
other
things
in
this
space,
but
but
if
we
can
set
that
aside
for
a
minute
and
look
at
how
it
was
done,
you
know
on
microcontrollers
in
the
past
it's
interesting
yeah.
A
I
didn't
think
of
them.
That
would
be
a
really
cool
concept.
You
know
Laura
might
be
a
little
low
bandwidth,
given
the
payloads
of
those
packets
for
doing
very
much
in
terms
of
a
spec
but
mesh
in
general.
A
Maybe
you
could
even
use
like
you
know,
Wi-Fi
or
something
to
move
through
a
metro
area,
and
there
are
techniques.
There
was
an
old
thing
used
for
control,
Telemetry
called
the
Ricochet
Network,
where
every
node
was
aware
of
its
latitude
and
longitude,
and
you
would
tag
a
destination
with
that
and
somebody
that
that
heard
a
packet
with
a
Target
destination.
That
was
where
they
knew
they
were
closer
to
that
destination
than
the
sender
would
take
it
and
try
to
forward
it
and
it
was
pretty
resilient.
A
G
Actually
mentioned,
the
a
lot
of
those
concepts
are
in
open
ZT,
which
I
was
presenting
last
week,
operating
as
a
as
a
smart
route
in
mesh
Network.
So
you
can
effectively
because
we,
in
fact
we
don't
have
root
traffic
or
an
IP
and
DNS.
You
root
it.
According
to
the
identity,
you
just
say:
hey
I
want
to
send
this
to
to
the
the
to
the
node.
That's
at
that
that
far
end
and
for
it
to
come
back
to
life
or
shrink
back
down
and
the
overlay
will
just
handle.
All
of
that.
G
It
does
so
all
of
the
nodes
are
communicating
to
one
another
and
the
control
plane
is
facilitating
that
information
as
to
what
is
the
the
latency
between
each
hop.
So
when
it
says
I
want
to
send
something
from
a
source
to
destination,
it's
effectively
running
a
diapers
algorithm
to
go
well,
the
lowest
latency
way
is
to
go
across
all
of
these
hops,
which
are
available.
A
G
A
G
C
A
Fact
is
that
if
one
were
to
do
a
plug-in
architecture,
the
cool
thing
to
do
would
be
not
just
to
use
it
for
control
plane,
although
I'm
a
firm
believer
in
isolating
control
plane
from
data
plane.
But
when
you,
at
the
end
of
the
day,
a
lot
of
these
apps
that
you're
running
these,
could
you
know
containerized
apps
that
are
going
to
run
at
the
edge
Leaf
nodes
are
going
to
have
demands
to
talk
to
one
another
as
well.
A
So
you
know
having
in
place
some
sort
of
a
mesh
architecture,
for
that
would
be
highly
desirable
and
maybe
also
calling
for
a
plug-in.
But
if
you
really
wanted
to
amp
this
up
and
do
the
world
of
service
having
this
plug-in,
be
versatile
enough
that
you
could
earmark
yeah
one
level
of
this
for
control,
plane,
meaning
higher
priority
than
the
data
plane,
that
that
would
be
really
taking
it.
A
The
dial
up
to
11,
because
you
know
Things,
Fall
Apart,
when
the
data
plane
can
hog
all
the
bandwidths
of
the
control,
plane,
no
longer
works
and.
G
I'm
not
sure
we've
implemented
that,
but
we
have
had
that
discussion
before
when
talking
to
Automotive
companies,
so
you
could
effectively
say
well,
it's
yeah
for
a
Telco,
it's
like
if
it's
VoIP
or
or
entertainment,
don't
care.
Critical
systems
need
to
make
sure
they
have
the
priority
and
the
overlay
can
force
that
to.
D
G
H
Yeah
I
was
just
curious
about
the
whole
thing,
especially
when
it
comes
to
like
air
gap.
Environments.
I
was
quite
curious
about
what's
the
opinion
of
of
the
people
around
here
in
this
meeting,
because
I
get
to
work
with
air
gap
environments
almost
on
regular
basis,
because,
let's
be
honest,
smes
and
big
infrastructure,
big
big
companies
like
oil
and
gas,
they
don't
really
like.
H
Some
of
them
are
very
skeptical
about
connecting
their
Edge
devices
direct
to
directly
to
a
public
cloud
or
something
from,
and
what
my
understanding
is
it
like
at
least
the
containerization
part
of
the
solution?
Sometimes
is
not
the
best
thing
when
it
comes
to
air
gap.
Environments
right,
like
I,
have
I
had
to
work
with
someone
where
they
have
a
a
backlink
channel
to
some
certain
control
entities
where
it's
just
5
Mbps
per
second,
and
they
wanted
to.
H
You
know
it's
everybody,
since
Stills,
jumping
on
the
whole
containerization
aspect
of
software
development
right,
so
they
have
containers
which
are,
which
are,
let's
say,
a
couple
of
hundred
megabytes
here
and
there,
and
sometimes
when
you
want
to
let's
say,
send
a
update
the
container
on
certain
Edge
nodes,
which
are
not
very
like
very
healthily
collect
connected
to
the
internet.
It
becomes
much
much
more
of
a
tricky
situation
to
get
updated
containers
on
these
kind
of
edge
networks.
These
are
and
when
I
say
Edge
Network.
These
are
extremely
far
rich
like
take
some.
H
Some
of
them
are
just
absurdly
out
there
on
their
own,
but
somehow
they
have
a
small
backlink
that
gets
connected
to
a
central
entity
and
then
that
somehow
is
meant
to
be
used
not
just
for
updates,
but
also
for
data
that
is
being
sent
Upstream.
So
it's
it's
quite
an
interesting
challenge.
Actually,
when
it
comes
to
working
with
these
kind
of
air-gapped
environments,.
D
Yeah
so,
like
we've
heard
this
before
with,
like
you
know,
cruise
ships,
oil
rigs,
Etc,
energy,
entertainment,
travel
yeah,
so
these
are
all
important
areas
to
include
that
we
we
have
not
yet
delved
into.
H
Yeah,
we're
also
adding
to
Marx's
Point
sorry
Steve,
just
a
second,
but
is
the
fact
that
I
think
the
the
discussion
that
that
goes
on
when
it
comes
to
the
Mesh
networking
stuff.
Have
you
guys
heard
of
this
IEEE
standard
for
low
power
networks?
So
it's
an
IEEE
standard
called
802.15.4
and
these
guys
Focus
solely
on
extremely
low
powered
networks,
but
the
plus
side
of
it
is
they
have
some
really
nice
protocols.
So
I
wrote
a
thesis.
H
I
wrote
my
master
thesis
by
picking
one
of
these
routing
protocols
or,
let's
call
it
back
off
protocols,
and
it
actually
works
wonders
when
you
think
about
it.
It's
so
simply
designed
because
these
kind
of
networks
anyways
have
to
sleep
more
and
do
less
and
when
they
have
to
do
something,
they
have
to
do
a
lot,
but
some
of
the
protocols
are
insanely.
Good
and
I
was
quite
curious.
H
If
someone
in
the
kubernetes
community
stumbles
on
such
protocol
and
starts
implementing
some
some
completely
cut
off
protocol,
which
is
like
a
backing
back
off
mechanism,
they
have
very
nice
features
with
IPv6.
So
you
can
do
a
lot
of
interconnecting
nodes.
You
can
design
shortage
routes
when,
when
a
certain
set
of
subnet
sub
network
devices
go
down,
and
also
it's
quite
nice,
if
you
guys
want
to
have
a
look
into
it,.
A
Yeah
I
kind
of
think
that
the
kubernetes
project
would
steer
away
from
dealing
with
protocols
directly
because
right
now,
I
believe
they
kick
that
down
to
the
Linux
kernel
host.
You
know
going
through
the
container
runtime.
Even
they
did
have
a
major
Reformation
of
kubernetes
to
deal
with
ipv4
and
IPv6,
but
even
there
I
don't
think
anything
going
on
in
the
kubernetes
project
tries
to
go
below
that
water
line.
They
try
to
keep
the
abstraction
layers
a
little
simpler
than
that,
but,
that's
not
to
say
it
wouldn't
work
with
these
protocols.
F
Yeah
so
I
I
guess
one
of
the
the
challenges
that
that
I
see
is
that
I
use
retail
as
an
example
is
that
e,
so
retailers
have
thousands
of
sites
that
they
deploy
to
and
if
they
have
kubernetes
deployed
into
those
sites.
F
You
end
up
with
this
Hub
and
spoke
model,
where
not
only
do
the
retail
sites
not
talk
to
each
other,
that
they
really.
They
have
intermittent
connectivity
to
some
centralized
Hub
and
they
sort
of
allow
it,
because,
if
they're
using
kubernetes
that
that's
the
way
you
know
it
needs
to
work
that
way.
F
The
challenge
that
that
we
run
into
and
that
you
know
it
kind
of
looked
to
this
group
for
for
ideas
about
is
that
with
the
Hub
and
spoke
model
you
can
really
overwhelm
the
API
server
when,
when
thousands
of
sites
check
in
all
just
trying
to
maybe
get
an
update
or
something
but
but
they're
all
trying
to
communicate
with
the
API
server
at
the
same
time,
and
it
just
overloads
the
system
and-
and
so
you
know,
I've
been
toying
with
some
ideas
about
buffering,
communication
and
sort
of
you
know
in
the
in
the
vein
of
it
eventually
consistent
data
sort
of
eventually
consistent
communication.
F
Where
you
know,
if
they're
just
checking
in
to
get
an
update,
do
they
really
need
it
immediately?
Maybe
they
do.
Is
there
a
way
to
to
build
sort
of
cues
that
are
you
know
where
you
could?
You
could
say:
I
really
need
this
now
versus
like
I.
Just
I
need
an
update
in
the
next
three
hours,
or
you
know
so
so
trying
to
I
think
this
is
a
problem
that
we're
all
going
to
be
facing
for
Edge
and
and
that's
why
I
suspect
Andy?
F
Why
you
guys
are
looking
at
a
not
Hub
and
spoke
model,
but
that
particular
solution
doesn't
work
for
the
retailers
that
we've
got
because
they
won't
let
each
retailer
talk
to
each
other.
So
we
can't
we
can't
pass
data
between
stores
we
have
to.
We
have
to
go
through
some
central
location,
so
you
know
just
really
interesting.
Challenges
am.
A
D
Yeah
so
yeah
I
was
just
about
to
jump
in
here.
I
wanted
to
wait
for
an
opportunity
right
so
Mark,
exactly
where
you're
double
clicking.
There
is
exactly
where
we're
going
next
and
we've
got
a
working
on
our
next
proof
of
concept
for
September,
where
we'll
be
doing
work
on
scalability
and
one
of
the
things
you
know.
Two
of
the
things
that
come
up
consistently
for
us
right
is
that
these
SED
clusters
get
overwhelmed
with
you
know
a
few
more
than
a
few
thousand.
You
know
objects
right.
D
So
that's
first,
so
one
of
the
just
to
go
a
little
bit
further
with
that.
Is
that
that's
one
of
the
reasons
why
kcp
Edge
we've
chosen
not
to
go
with
any
kind
of
bundling
it's
the
raw
ref.
It's
a
raw
data
types
themselves
right,
the
real
API
types,
so
deployments,
config
specs
Etc
we're
not
using
bundles
of
any
sort
right,
no
manifest.
The
second
thing
that
you
hit,
though,
right
after
you
get
past
that
hurdle
and
we've
done
tests
to
prove
this
is
the
API
Machinery.
D
So
we've
actually
got
work
that
we
did
and
we
submitted
to
Upstream
where
we
do
API
priority
and
fairness.
So
we've
we've
made
corrections
to
that
algorithm
to
help
in
this
case,
so
we're
ready
for
that.
But
we
have
to
get
past
the
object,
store
limitations
and
we're
working
on
that
next.
A
Yeah
I
I
was
going
to
say
when
you
mentioned
eight
the
API
server
Mark,
whether
you
really
met
the
API
server
or
NCD,
because
etcd
is
kind
of
notorious
they.
They
knew
from
the
get-go
that
it
was
architected,
I,
think
four
or
five
thousand
cluster
notes.
D
A
Even
then,
when
you
say
5
000
cluster
nodes
to
even
get
to
that
level,
it
requires
heroic
efforts.
I
mean
that
isn't
like
a
a
tiny
three
node
SCD
cluster
you're
gonna
have
to
invest
a
lot
of
resources
in
to
feeding
a
monster,
basically
to
to
keep
up
with
a
five
thousand
note
cluster,
and
it
it
wasn't
that
there's
something
magical
against
about
clusters
either
I
suspect.
A
It's
objects
in
general
that
if
you
have
large
numbers
of
objects
of
any
type,
whether
clusters
or
not,
the
FCD
back
end
is
going
to
be
challenged.
If
you
could
tier
this,
that
might
be
a
solution,
but
it's
interesting
to
hear
that
maybe
Andy
you've
done
research
and
figured
out
that,
even
if
you
magically
made
the
NCD
issue,
go
to
nothing
that
the
API
server
would
just
be
a
bottleneck
that
wouldn't
be
too
much
further
behind
it.
Yeah.
D
D
An
interesting
one
I
mean
eventual
consistency
of
any
implementation,
would
be
I've
got
I've
got
to
get
you
I'll.
Look
you
up
and
figure
out
how
I
can
get
you
a
a
link
to
you
know
the
cap
that
was
for
associated
with
the
API
priority.
In
fairness.
Maybe
you
can
take
a
look
at
that.
D
So
if
you
have
a
three
note,
SCD
cluster
right
on
a
raft
protocol
and
then
you
do
that
100
times
Well,
we
blew
up
our
budget.
We
we
came
up
to
the
conclusion
that
if
we
wanted
to
get
to
a
million
Edge
locations,
we'd
be
spending
on
the
order
for
each
instance
of
of
putting
out
a
million
Edge
locations
roughly
about
360
000
a
year
and
just
the
and
just
the
compute
cost
to,
and
most
of
that
was
to
host
the
object
store.
D
It
was
a
a
20x
increase
over
what
we
could
have
done
with
you
know,
tin
cans
and
and
baling
wire
and.
D
Either
yeah
this
wasn't
even
approaching
multi-cloud.
So
if
you
got
to
multi-cloud
the
egress
Ingress
would
probably
you
know
be
one
or
two
times
more
than
that
alone.
Yeah.
A
So
Shan
back
back
to
your
original
comment
about
how
you
deal
with
air
gapped
and
nervousness
rightfully
so
about
connectivity
to
the
internet.
I
think
maybe
something
I.
My
personal
opinion
is
underutilized.
Is
this
concept
of
using
you
know?
Non-Network
transport?
Like
you
know,
there
was
kind
of
a
half
ingest
RFP.
A
Somebody
wrote
about
the
bandwidth
of
strapping
sdram
cards
to
the
legs
of
carrier
pigeons
being
higher
in
bandwidth
and
fiber
connections,
and
you
know
the
fact
is
that
some
of
what
goes
on
to
these
Edge
locations
like
delivering
updated
container
images
doesn't
really
have
to
happen.
You
know
on
real-time
latencies
yeah.
You
would
like
to
update
them
right
away,
but
the
realistic
Attitudes
by
people
who
are
running
at
those
Edge
nodes
is
that
they
would
like
to
do
those
updates
when
production
is
shut
down.
A
You
know
they're
not
going
to
go,
disrupt
a
working
Refinery
at
1
pm
in
the
afternoon
during
Peak
productivity
to
update
all
their
containers.
That
would
probably
be
lunacy,
so
the
bottom
line
is
that
there
is
a
lot
of
this
traffic
that,
if
you
could
get
it
there
by
you,
know
flash
drive
sent
by
FedEx
or
in
rural
AG
areas,
just
somebody
who
takes
it
out
there
on
a
motorcycle.
That
would
be
perfectly
workable
for
a
lot
of
these
use
cases
even
going
back
the
other
direction,
I
I
think
a
common
thing.
A
That's
going
to
happen.
If
ml
takes
takes
root
at
Edge,
like
I,
think
it
will,
there
will
be
demand
to
run
inference
engines
on
live
data,
but,
looking
ahead
a
couple
moves
on
the
chessboard.
You
have
to
keep
those
trained
and
optimized
so
every
month
or
two,
maybe
you
want
to
retrain
it,
take
advantage
of
new
data.
Well,
that
requires
collecting
training
data
from
potentially
a
hundred
thousand
nodes
out
at
Leaf
nodes.
A
A
So
you
know
it's
authentic
and
then
somebody
collects
these
flash
drives
once
every
week
or
once
a
month
and
sends
it
physically
to
be
ingested
with
by
a
public
Cloud
where
you
would
run
your
retraining
exercises
that
should
be
perfectly
viable
and
it
probably
would
save
a
whole
lot
of
money.
I
mean
even
delivering
from
Cloud
to
Edge.
It
might
be
a
way
to
get
around
egress
costs.
Sending
it
back
up
to
the
cloud
might
might
be
a
way
to
deal
with
no
or
intermittent
connectivity.
A
So
I
think
that
you
want
to
keep
your
an
open
mind
about
ways
to
solve,
to
deal
with
reasonable
desires,
to
not
connect
to
the
internet.
Maybe
ever
and
buffering
things
to
physical
storage
that
are
physically
transported
would
seem
to
be
a
way
to
to
do
this.
That
there's
no
reason
it
shouldn't
work.
You
do
want
to
have
Security
on
this,
but
I
think
that's
a
solved
problem
I
mean.
Obviously,
if
you're
going
to
FedEx
this
data
and
it's
sensitive,
you
need
to
worry
about
that
being
lost
or
stolen.
D
H
Yep
I
think
that
Steve,
you
just
summarized
my
the
complete
business
strategy
of
the
product
that
I'm
working
on
so
it
we
took
the
other
way
around
when
I
started
with
the
company,
where
I
am
right
now
they
actually
had
to
start.
They
actually
were
shipping
products
based
on
where
you
would
put
in
a
USB
stick
and
things
would
get
updated
this
and
that
because
that
generally
was
the
criteria
or
the
requirement
for
for
a
lot
of
customers
who
generally
are
very
air-gapped
or
decide
to
have
very
air
gap
networks.
H
So,
and
it's
only
because
of
recent
trends
that
that
some
of
them
are
now
slowly
and
steadily
coming
out
of
this
Rabbit
Hole
of
oh
guess,
what?
Maybe
there
might
be
a
benefit
of
having
certain
updates
over
the
air
in
the
sense
that
it
might
some
part
of
the
network
would
be
connected
to
the
internet
to
keep
to
keep
certain
things
up
to
date
or
be
compliant
with
certain
security
requirements
and
this
and
that
so
I
I
mean
we
might
be
laughing,
but
guys.
A
A
Where
you
know
in
the
old
days,
Mainframe
OS
vendors
would
literally
send
mag
tapes
that
had
the
updates
and
they'd
come
with
packaging
that
had
the
vendors
labels,
and
there
are
rumors
that
somebody
stole
the
boxes,
put
a
fake
tape
in
the
usual
Behavior.
The
operator
was
they'd
Mount
the
tape
and
run
it
without
yeah.
The
authentication
would
just
be
that
it
arrived
by
UPS
in
the
Box
looked
authentic
and
that's
a
recipe
for
a
disaster.
A
So
I
think
that
when,
when
you
send
it
out
to
the
edge
nodes,
it
needs
to
be
signed,
it
needs
to
be
encrypted
and
authenticated
more
than
than
just
the
package
looks
legit
and
even
with
USB
keys.
I
would
not
recommend
that
people
would
mount
a
random
USB
key,
because
there
are
all
kinds
of
hacks
with
those
where
they
could
have
a
key
logger
in
in
this
key,
even
if
the
actual
key
was
at
authentic
sign
data.
A
It's
almost
like
you
need
kind
of
a
mezzanine
device
that
is
a
DMZ
to
ingest
this
inbound
data
and
not
route
through
USB
signaling,
which
is
potentially
far
more
than
just
data
Maybe.
Maybe
another
form
of
storage
other
than
USB
would
be
more
desirable,
like
an
sdram
card,
because
you
wouldn't
expect
the
system
to
tolerate
an
SD
ram
card
is
incapable
of
reporting
that
it's
a
keyboard.
For
example,
Andy.
D
You've
got
your
hand
yeah
Steve.
Just
on
that
note.
I'm
heading
up
to
Vancouver
for
github's
con
I'm
presenting
there
next
week
and
one
of
the
things
that
I
thought
was
interesting.
Is
my
company
required
me
to
get
a
loaner
laptop
this
time
around
to
travel
internationally?
Apparently
there
are
people
at
confiscating
laptops.
You
have
to
either
log
into
them.
Let
them
see
what's
on
it
and
then
they'll
install
spyware
or
keyloggers
of
et
cetera,
et
cetera,
so
they
don't
want
to
take
that
chance
anymore.
A
C
The
comment
there's
a
lot
of
good
discussion
here,
actually
has
been
reviewing
that
imaginative
white
paper,
so
I
think
it'd
be
great
to
capture
the
discussion
here
and
a
kind
of
a
continuation
of
that
what's
covered
in
that
white
paper
by,
for
example,
how
to
classify
Edge
right.
You
know
far
Edge
remote
Edge,
you
know
what
is
and
then
what
is
the
business
value.
A
Yeah
I
lately
I,
haven't
been
the
one
of
the
more
active
people
on
that
white
paper.
I'm,
not
sure
if
we
have
any
of
the
white
paper
people
on
in
this
meeting
today,
but
I
agree
with
you
that
that
would
be.
You
know
either
putting
it
in
the
existing
one
of
the
existing
white
paper
efforts
or
starting
a
new
one
would
be
worthwhile.
C
D
Victor,
are
you
part
of
the
edge
application?
Are
you
contributing
to
the
edge
application
document
that
we're
working
on
at
the
moment
with.
C
The
not
at
this
moment
but
yeah
it
would
be
interested.
A
D
Okay,
here
I'm
putting
the
link
in
here
now
for
everyone.
This
is
the
link
to
the
edge
application,
Edge
native.
D
Design
behaviors
that
we're
working
on
together,
Brandon
Wick
myself,
who's
the
gentleman
from
Cisco
who's,
also
working
with.
F
So
I
thought
this
meeting
was
every
other
week.
Is
it
it
sounds
like
there
was
a
meeting
last
week
that
I
missed
did
we
are
we
actually
meeting
every
week
now?
No.
A
The
reason
was
we
canceled
the
meeting
that
conflicted
with
kubecon
so
rather
than
really
cancel
it.
We
just
kicked
it
down
the
road
a
week.
If
you
will,
we
debated
just
canceling
it
all
together,
but
I
think
we
wanted
to
have
that
open,
ZD
presentation,
so
we
held
one
that
was
sort
of
out
of
band
I.
See
no
problem.
I
did
notice
that,
even
though
we
canceled
it
I
get
noticed
notified
when
people
attempt
to
join
the
meeting
and
about
three
people
missed
the
notification
of
cancellation
and
I.
Don't
know,
I
was
in.
A
F
Yeah
I've
I've,
like
basically
blocked
out
my
calendar,
with
a
link
to
the
Google
meeting
every
two
weeks
and
just
that's
how
I
kind
of
keep
track
of
it,
so
less
so
keeping
my
eyes
on
the
Google,
Doc
and
mm-hmm
I
I
would
like
to
maybe
for
for
the
agenda
for
next
week.
If
we
don't
already
have
something
so
I
have
some
some
curiosity
questions
for
this
group
and
I.
F
What's
going
on,
I've
been
watching
some
videos
and
stuff,
and
then
the
other
one
that
that
just
is
constantly
tugging
at
me
is
whether
or
not
kubernetes
is
is
the
right
mechanism
for
the
edge
I
think
it's
got
some
some
of
the
right
aspects,
but
but
I'd
like
to
have
more
deeper
conversation
with
people
who
are
in
this
space
about
kubernetes
and
some
of
the
challenges
in
general
about
using
container
orchestration
at
the
edge.
A
I
would
really
welcome
that,
for
you
know,
I
I
usually
put
the
shell
of
the
agenda
up
slightly
after
a
meeting
ends,
so
it
you
can't
go,
throw
it
in
there
just
literally
today,
but
you
you
can
nominate
those
on
the
slack
Channel
and
then
as
a
reminder
to
me
to
put
them
in
the
agenda.
How
about
that
and
I'll
put
them
in
the
agenda
for
next
time.
So.
F
B
A
You
know
Andy
I
hate
to
task
you
with
kcp,
but
as
long
as
you're
doing
pass-through
that
kubernetes
doesn't
act
on
how
about
this
for
I'm
just
brainstorming
here,
but
a
crazy
feature
where
we
put
it
in
there
and
it
gets
passed
through
all
these
layers.
But
it's
earmarked
is
this
isn't
kubernetes
at
all,
but
this
this
Leaf
Edge
node,
let's
just
throw
it
out.
There
is
running
Docker,
compose
or
podman
in
compose
mode,
and
it's
really
just
a
Docker
compose
spec.
A
That's
a
single
node,
kubernetes
and
I
think
there's
been
plenty
of
debate
on
whether
single
node
kubernetes
even
is
some
kind
of
anachronism
like
what
is
it
if
it's
single
node,
maybe
half
of
what's
in
kubernetes,
is
senseless
because
it
doesn't
do
anything
useful
and
just
brings
along
a
big
code
footprint,
that's
subject
to
cves
and
things,
and
what,
if
this
pass-through
could
even
have
generic
blocks,
I
mean
to
make
it
really
versatile
and
future
proof
just
earmark
this?
Is
we
don't
really
know
what
this
is?
A
This
is
just
a
blob
like
it
just
like
the
oci
container
concept
was
broad
enough
that
it
didn't
have
to
be
a
Docker
image.
What
if
what
you're
forwarding
to
the
leaf
node
doesn't
have
to
be
anything
you
even
have
declared
you're
aware
of,
but
we'll
just
get
it
there
for
you,
and
that
would
be
a
really
cool
concept
that
wouldn't
lock
you
into
these
Leaf
nodes
being
kubernetes,
but
maybe
you
could
still
manage
them
with
the
kubernetes
control
plane
at
the
top.
That
was
scalable.
F
A
F
D
That
it
shouldn't
matter,
we
believe
that
kubernetes
is
the
is
the
control
plane
or
the
platform
you
should
use
for
Edge
right,
obviously,
I'm
a
bit
biased,
but
the
reason
being
is
is
that
it
does
allow
afford
you
the
anonymity
of
the
workload
of
the
payload
within
it
right.
So
you
could.
We
are
looking
at
things
like
oci
we're
looking
at
wasm
right,
so
this
gets
two
birds
of
one
stone
there
mark
so
wasm.
The
same
sort
of
effect
right
is
that
we
know
that
this
is
lurking.
D
You
know
sitting
in
the
background,
ultimately
will
be
used
by
certain
providers
or
vendors
to
do
certain
things.
I
see
like
the
world
of
printers
gets
opened
up
as
soon
as
you
include
something
like
wasm
right.
I
can
see
that
being
a
attractable
technology
for
them,
so
just
to
keep
it
short.
Yes,
on
kubernetes,
yes,
on
ubiquity.
E
E
Is
my
mic
working?
Oh
okay,
yeah
I
wanted
to
go
ahead
and
jump
in
on
the
web
assembly
side.
I'm
happy
to
kind
of
feel
liaison
for
a
lot
of
those
conversations
and
also
wanted
to
point
out,
were
there's
work
under
a
way
to
create
a
webassembly
working
group
under
the
runtime
tag
as
well.
So
that's
happening
right
now,
they're
building
out
the
charter,
and
so
that's
going
to
be
a
great
place
to
also
ask
those
questions
around.
How
is
this
evolving
and
forming
and
I'll
probably
attend
those
meetings
too?
E
So
I
can
kind
of
bring
our
questions
there.
That
might
be
specific
to
what
does
webassembly
look
like
on
the
edge
and
I
think
the
question
of
also
is
kubernetes
made
for
the
edge
is
also
a
question.
People
have
around
webassemblies
kubernetes
made
for
webassembly
like.
Is
that
the
way
we
want
to
orchestrate
webassembly
or
there
are
other
orchestrators
we
want
to
use
in
other
application
paradigms
so
I
think
those
those
two
discussions
will
pair
well,
just
like
Andy
just
mentioned.
C
B
Just
to
be
succinct,
I'd
also
be
interested
in
just
defining
what
that
leaf.
Node
would
be
and
what
its
capabilities
would
be,
because
the
worker
node
brings
with
it
some
requirements
for
having
an
overlay,
Network
and
talking
to
other
nodes
and
if
you're,
in
the
medical
space
and
you're
making
a
defibrillator.
Your
defibrillator
probably
doesn't
need
to
talk
to
other
defibrillators.
It's
probably
quite
happy
just
talking
to
the
hub,
so
there
might
be
a
need
for
defining
a
leaf
node
with
fewer
requirements
for
conformance
and
introducing
that
into
the
kubernetes
spec.
H
Just
a
quick,
quick
update
on
on
what
Rob
was
mentioning
I
think
there's
a
very
nice
framework
called
Web
of
things.
Maybe
that
might
be
interesting
when
it
comes
to
describing
what
a
leaf
node
is.
Maybe
might
try
to
solve
a
lot
of
problems,
at
least
for
the
far
Edge
logics
I.
Don't
know
that's,
maybe
it
might
be
worth
discussing
it
next
time.
A
Okay,
we're
two
minutes
after
the
hour,
so
I'm
going
to
call
this
to
a
close,
but
thanks
everybody
for
coming
and
it
sounds
like
we've
got
a
topic
nominated
for
next
time,
which
is
great.
We'll
cover
a
couple
of
things,
both
web
assembly
and
kubernetes
suitability
for
Edge
anything
else.
People
want
to
bring
up,
throw
it
into
the
chat,
throw
it
into
the
slack
Channel.
A
Okay,
bye,
everybody
thanks
for
coming,
cheers.