►
From YouTube: Kubernetes WG IoT Edge 20191204
Description
December 4 2019 meeting of the Kubernetes IoT Edge Working Group - discussion of deployment architectural patterns, requirements for K8s implementations at edge
B
B
A
Relevant,
so
you
know
you
and
I
were
both
speakers
and
I
think
we
even
had
an
a
couple
press
interviews
and
when
you
get
that
busy,
you
can't
go
to
all
the
sessions
yourself
so
understand
that
my
report
is
pretty
much
based
on
having
a
limited
time
window
to
go
out
there
and
look
at
things
and
do
things.
Although
I
did
do
an
analysis
trying
to
get
prepare
a
cfp
for
the
nexucon
europe,
I
went
and
looked
at
some
stats
on
what
was
popular
in
the
iot
edge
space.
A
Our
own
session
was
lightly
attended
compared
to
past
ones,
but
I
think
it's
because
we
got
kind
of
the
worst
time
slot
where
they
put
us
in
literally
the
last
slot.
On
the
last
day,
in
the
farthest
room
yeah,
a
lot
of
people
have
already
left
town
or
burned
out
by
then
I'm
afraid,
but
we
still
did
get.
I
think
it
might
have
been
approaching
200
registry
registrations,
not
all,
I
think,
maybe
a
large
part
of
those
never
showed,
but
they
at
least
had
the
intent
at
one
time
yeah.
We.
C
A
B
D
Yeah
I
talked
to
sean
about
the
k3s,
but
they
said
they
are
trying
to
conquer
the
problems
from
different
angle,
so
they
it's
more
likely.
It's
not.
There
like
a
cloud
connected
edge,
is
more
like
a
independent
edge,
so
they
try
to
say,
give
a
light
platform
to
control
the
application
running
on
the
edge
without
communication
I
mean,
without
I
mean
tight
communication
with
the
cloud,
so
yeah
tend
to
run
it
independently.
D
A
Yeah,
I
think,
actually,
this
group
and
my
own
attitude
is
that
we're
actually
open
to
a
number
of
different
ways
to
approach
the
challenges
at
edge
and
not
everything
is
going
to
work
for
everybody.
So
the
k3s
falls
in
the
camp
of
running
a
full
out
kubernetes
cluster
at
the
edge
location
that
won't
work
for
everyone,
but
it
will
clearly
work
for
a
number
of
people.
I
mean
it
isn't
just
rancher
doing
it.
The
the
session
a
year
before
this
past,
one
in
seattle,
had
chick-fil-a
running
a
kubernetes
cluster
in
the
retail
location.
E
C
E
A
You
know
what,
before
you
go
on
james,
I'm
wondering
some
of
us
are
new
to
others.
Maybe
we
should
have
started
with
some
brief
intros.
Rather
yeah
I've
been
popping
in
and
everybody
talking,
because
there's
even
some
handles,
I
see
in
the
participants-
I
I
don't
recognize
myself.
So
why
don't
you
start
james
and
because
I
think
it's
been
a
while
since
you've
joined
us
on
one
of
these
meetings,
just
get
one
or
two
sentences
on
yeah.
E
F
F
F
Oh,
it's
it's
it's
a
it's
a
startup,
but
also
we
tried
we
put
together
like
a
framework
in
order
to
deploy
this.
It's
we
we
with
some
colleague
of
work.
Here
we
created
an
organization
in
gitlab,
called
iot
ops,
pretty
much
everything
we
deployed
is
there
and
we
focused
mostly
trying
to
to
develop
this
and
to
share
with
some
faults.
For
the
time
being,
we
created
the
kubernetes
with
some
applications
like
home
assistant,
in
order
to
deploy
our
at
our
houses
and
mainly
control
our
all
our
home
automation
stuff.
F
A
F
C
A
At
a
future,
talk
if
you
wouldn't
mind
giving
a
presentation
of
say,
20,
25
minutes
on
what
your
project
looks
like
and
maybe
a
picture
or
two
that'd
be
great.
You
can
just
look
at
our
public
agenda
document
and
put
yourself
in
at
some
future
date
that
you'd
like,
but
this
sounds
really
interesting
and
I
think
there's
potentially
a
lot
of
other
people
who
are
interested
in
maybe
the
problem
you're
trying
to
solve,
or
one
very
much
like
it
thanks,
let's
move
on
frederick,
do
you
want
to
introduce
yourself.
G
B
E
G
C
G
C
E
G
G
A
C
For
sure
no
problem
so
hi
kilton
hopkins
co-founder
of
a
company
called
edgeworks
and
the
creator
of
a
project
called
eclipse,
io
fog
and
which
is
now
now
going
on
its
sixth
the
year
of
development.
Although
the
last
three
have
really
been
the
good,
the
good
solid
uptake
years
and
co-heading
up
the
the
this
new
eclipse
edge
native
working
group,
that's
about
to
be
announced
next
week.
C
My
purview
is
to
make
sure
that
we
figure
out
the
right
architectures
here
in
this
working
group
and
then
find
ways
to
continue
to
implement
the
the
latest
and
greatest
ideas
and
and
requirements
from
people
in
the
the
open
source
code
base
of
biofolk.
So
you'll
hear
me,
talk
about
io
fog,
a
lot
most
of
the
time,
it's
in
context
of
what
we're
looking
to
do
or
or
you
know
how
we're
how
it's
being
leveraged
and
but
that
doesn't
mean
that
I
only
want
to
talk
about
io
fog.
A
I'll
give
you
one
that
this
came
up
in
the
demo
that
we
gave
at
kubecon
just
what
was
it
ten
days
issue
ago,
but
we
gave
a
demo
that
wasn't
kubernetes
centric,
but
perhaps
in
an
upcoming
meeting
you
could
actually
go
through
the
mechanics
of
setting
one
up,
maybe
not
production
grade,
but
just
for
people.
Who'd
want
to
do
a
quick,
eval,
yeah.
C
A
C
A
great
idea,
and
so
I'll
pass
the
baton
to
the
next
person
now,
but
just
final
comment,
I
think,
to
do
that
presentation
we
could
make
that
session
of
this
working
group,
one
that
is,
is
more
interactive
where
people
are
able
to
to
do
the
work.
Alongside
that,
that's
a
great
idea
for,
like
a
turn
of
the
year
topic,.
C
C
E
C
D
E
C
A
D
Me
yes,
yeah
yeah,
I
my
name
is
ending.
I
work
on
the
ad
about
two
years
now
so
especially
I
when
I
started
the
cooperage
about
early
2017,
however,
because
I
have
a
couple
people
left,
so
I
will
come
back
and
re
redrawn
the
group
and
participate
more.
Hopefully,
we
can
have
something
to
share
before
the
next
kubercon
europe,
so
yeah,
I'm
working
on
mainly
curb
edge
and
also
work
with
the
eclipse
I
mean
or
group
to
see
what
we
can
collaborate
together.
E
E
E
E
E
D
E
E
E
E
Yeah
yeah
kelton,
and
I
were
talking
about
that
at
kubecon-
that
you
know
you
know
he's
open
to
to
enabling
that
crd
based
on
io
fog
as
to
to
match
sort
of
what
kubej
has
done
and
he's
he's
also
done
some
stuff
and
demonstrated
integrating
octane
and
you
know
through
apis.
You
could
integrate
anything,
but
octane
was
a
good
one.
So
I.
E
E
E
Yeah,
so
it's
it
yeah,
it's
the
it's!
You
know
it
may
not
be
one
process,
it
may
be
multiple
process,
but
I
mean
if
you
go
and
look
at
k3s,
you
can
see
what
they
do
to
strip
stuff
out.
You
know
the
the
quite
the
question
is,
I
mean
there's
a
million
ways
to
solve
this
and
there's
pros
and
cons
with
everything.
E
E
E
E
A
There's
kind
of
two
levels
of
configuration
going
on
right.
The
first
level
is
configuring.
What
you've
got
out
at
the
edge
location
yep
as
a
single
instance?
But
if
you
look
ahead,
two
or
three
moves
on
the
chessboard.
Your
real
problem
isn't
that,
just
like
configuring,
an
individual
docker
container
wasn't
enough
and
you
needed
an
orchestrator
in
a
way.
This
is
the
same
sort
of
thing
where
you
need
kind
of.
C
A
E
E
E
Dionne
and
I
went
through
it,
but
so
yeah
if
you
can
proofread
it,
that
would
be
great
I'll
share
it
and
then
I'll
submit
it
today,
but
you
know,
I
think
we
need
to
have
an
opinion
just
like
we
did
with
the
security.
We
need
to
have
an
opinion
on
this
stuff
and
have
recommended
practices
and
architectures
and
people.
You
know
people
begin
to
use
other
deployment
mechanisms
and
patterns.
Then
we
can
capture
those
as
well
as
time
goes
on,
but.
E
E
The
best
thing
is
to
educate,
right
yeah
and
give
them
all
the
information
they
need,
and
then
they
can
select
the
deployment
pattern
and
the
projects
and
software,
because
my
implementation
is
going
to
be
different
than
everybody
else's
and
it's
going
to
have
its
own
requirements
and
they're
going
to
have
to
pick
based
on
that
and
the
pro
and
the
projects
are
going
to
have
to
decide
what
use
cases
that
they
want
to
meet
and
not
meet
and
what
their
priorities
are.
Yeah.
A
I'm
going
to
think
this
is
going
to
fall
out
just
like
container
orchestration
did
where.
If
you
go
back
three
years,
you
still
had
docker
swarm,
mesos
kubernetes
and
eventually,
if
you
get
large
numbers
of
people
going
down
one
path,
it
ends
up
getting
paved
to
where
it
becomes
more
attractive
than
the
dirt
footpath,
even
if
the
other
one
in
theory
was
better
suited
for
your
use
case,
it
there's
value
in
connecting
with
numbers.
I.
E
D
E
Either
kubernetes
or
kubernetes
adjacent,
and
you
know
integratable
and
compliant,
so
you
know
I
I
think
I
I
think
everybody
would
agree
that
you
know
that
that's
what
we
we
want
and
that's
kind
of
our
market
and
then
beyond
that.
It's
just
like
you
said
laying
out
the
facts,
facts
only
no
opinions,
and
then
you
know
that
way.
People
can
make
educated
choices.
E
The
the
other
thing
is,
I
think
we
need
to.
We
need
to
as
a
separate-
and
I
I
don't.
I
won't
have
time
to
submit
this,
but
you
know
maybe
this
could
be
adjacent
or
whatever,
or
maybe
it's
something
we
do
at
kubecon
europe
as
a
workshop,
or
you
know
a
hack
day,
beginning
to
talk
about
the
the
bootstrapping
and
provisioning
stuff
for
the
that
we
talked
about
that.
You
know.
A
E
C
E
B
E
C
A
Yeah,
well,
I
think
one
other
perspective
is,
I,
I
think
I'll
agree
with
these
five
categories.
I
think
in
some
cases
I've
seen
these
things,
some
of
them
coalesced,
like
your
single
binary
and
small
k8s.
The
fact
is
they're
both
small,
so
you
could
yeah
call
it
four
categories
instead
of
five
but
yeah.
I
don't
know
of
anything
you
left
out
that
I've
come
across
cool,
but
another
perspective.
Is
that
really
kubernetes
itself
the
existing
control
plane,
even
if
enhanced
with
crds.
G
A
E
You
know
it
may
be
using
ceph
as
a
shared
file
system
for
the
disks
in
there
or
a
jbod,
or
it's
going
to
be
a
wide
variety
of
stuff.
So
I
think
if
we
can
cap,
the
other
thing
we
haven't
really
gotten
to
is
capturing
the
cni
patterns
and
the
csi
patterns
and
all
all
that
sort
of
stuff.
I
think
those
are
things
that
we
need
to
document
as
well.
I
agree
because
what
people
use
you
know
whether
they
need
load
balancing
or
they
just
need
standard
networking
with
some
simple
ingress.
A
Yeah,
I
I
think,
actually,
if
service
mesh
makes
sense
in
the
traditional
public
cloud
kubernetes,
it
makes
even
more
sense,
not
now
I'm
not
saying
a
specific
instance
of
a
service
mesh
like
sdo,
because
they
were
targeted
at
it,
apps
using
https
and
they're,
not
necessarily
appropriate
for
edge
workloads,
but
that
concept
of
reducing
your
operational
cost
to
manage
this
and
relieving
application
developers
of
the
burden
of
handling
security
and
networking
yeah.
That
is
really
compelling
at
edge.
I
mean
I.
E
E
You've
got
to
provide
a
secure
it
environment
and,
if
they're
already
using
kubernetes,
and
they
know
how
to
code
against
all
that
and
operate
and
build
helm,
charts
and
write
all
the
yaml
and
everything
and
get
everything
right
from
an
operational
and
an
app.
The
app
developer
is
going
to
be
much
more
comfortable.
Writing
these
apps
and
deploying
them
out
and
providing
standard
edge
services
that
are
available
too.
So
I
you
know
he
he
asked
me
the
question
I
was
like:
well,
people
are
doing
it
and
I
had
to
step
back
and
think.
Well.
A
Well,
I
I
think
I
found
I've
had
that
same
question
but
as
as
you
give
them
a
rhetorical
question:
well,
are
they
running
containerized
software
advanced
and
is
there
a
lot
of
it?
Well
guess
what
was
engineered
to
hand
to
manage
running
a
lot
of
containerized
software,
and
I
think
then
they
they
have
the
aha,
but
I
think
the
thing
that
maybe
is
even
a
more
interesting
opportunity
that
people
don't
recognize
is
that
with
crds,
potentially
the
kubernetes
control
plane
could
manage
things
that
aren't
even
containerized
apps,
but
they're
similar.
A
E
Yeah,
you
gotta
distribute
it
and
you
could
distribute
it
in
some
form
down
and
have
it
there
and
available
and
staged
I've
seen
that
in
with
radio
firmware
updates
over
amqp,
we
staged
the
locally
staged
and
the
particular
direct
sftp
directory,
and
then
the
radios
just
reach
out
to
a
known
ip
and
ftp
into
it
and
pull
down
pull
down
their
updates.
I
agree:
it's
not
the
stuff
beneath
us,
we
may
be
staging
data
or
pulling
you
know.
The
southbound
command
and
control
stuff
is
also
going
to
have
to
go
through
this
plane.
C
This
is,
this
is
a
wonderful
pattern
by
the
way,
which
is
that,
if
we're
really
going
to
orchestrate
full
systems,
we
need
to
stop
thinking
that
a
container
is
equivalent
to
a
piece
of
application
functionality
and
ask
how
it
should
be
implemented.
Maybe
it's
a
serverless
function.
Maybe
it's
a
vm,
but
the
whole
world
should
be
orchestratable
by
kubernetes
and
that
it's
a
workflow
yeah.
A
I
think
there's
very
similar
things
too.
I
mean
the
reality.
Is
that
kubernetes
doesn't
go
drill
down
and
analyze.
What's
in
this
thing,
that
happens
called
the
container
image,
so
it
really
is
just
moving
around
blobs.
But
right
you
have
some
of
these
common
problems
that
now
some
of
them
are
challenges.
C
I
had
a
nice
talk
with
with
with
m2
michael
michael
from
the
harbor
project,
and
he
said
that
and
everyone's
welcome
to
participate
in
helping
them
build
this
out,
but
he
said
that
they
are
adding
in
the
capability
to
do
a
a
single
single
grab
of
a
new
image,
verify
it
and
then
distribute
bittorrent
style
to
right
at
node
edge
node.
I'm
super
interested
in
this,
and
I
think
it
might
be
time
to
have
you
know
some
more
harvard
talking
here.
A
A
I
don't
know
that
it's
on
the
current
roadmap
for
kind
of
next
release
or
anything,
but
I
think
that
that
harbor
project
is
aware
that
there's
an
interest
in
potentially
using
it
to
distribute
things
like
firmware
images,
and
that
could
you
combine
that
with
the
bittorrent
that
could
be
just
a
killer
building
block
for
a
lot
of
these
edge
deployments
key
piece?
It's
a
key
missing
piece
now.
The
one
thing
I
think
could
use
some
work
in
harbor
is
the
back
end
storage
for
where
those
blobs
reside,
yeah.
G
A
It's
s3
or
a
postgres
database,
but
I
think
there
might
be
some
real
benefit
to
using
a
content.
Addressable
store
that
does
d
do
because
at
these
edge
locations
you
do
have
finite
resources
and
there's
the
potential
issue
of
if
you
just
use
random
storage
like
say,
use
a
relational
database
for
that
back-end
store
and
keep
multiple
versions.
A
C
Well,
it's
there's
the
bigger
problem
of
if
you,
if
you're
talking
about
any
edge
location
too
yeah
storage
is,
is
cheap
enough,
that
maybe
you
can
throw
a
lot
at
it,
but
you
can't
go
service
it
as
easily
as
a
data
center.
So
you
really
want
to
avoid
running
out
of
storage
just
because
of
day-to-day
operations.
E
And
you've
also
got
to
be
cognizant
of
drive
life
for
mechanical
drives,
but
you've
also
if
it's
compact,
flash
or
if
it's
ssd
or
whatever
you
know,
at
least
on
the
stuff.
I've
done.
If
you're
deploying
a
10-year
solution,
you've
got
to
be
you've
got
to
buy
something,
that's
going
to
have
the
right
wear
patterns
and
algorithms
within
the
drives
to
make
sure
that
you
don't
end
up
killing
them.
Yep.
A
There's
real
advantages:
it's
small,
I
mean
you
get
lightning
strikes
people
steam,
use
of
equipment
that
you
have
to
you
know
repopulate,
and
if
you
used
a
content,
addressable
store
with
edu
that
kind
of
cloning,
repopulation
rehydrating
should
be
a
lot
easier
and
probably
the
technology
that
does
v2
could
be
relied
upon
for
your
replication.
You
know.
Hardware
harbor
also
has
north
south
and
east
west
replication
available.
G
H
E
What's
going
over
that
line
over
over
the
wire
back
back
and
forth
north
and
south,
because
otherwise
you've
got
to
do
it
at
a
network
layer
and
it's
much
harder
because
it's
going
to
be
you're
going
to
be
doing
packet,
inspection
or
you're
going
to
be
trying
to
tag
tag
priority
on
on
different
types
of
traffic,
because
traffic's
based
on
the
content
and
it's
not
readily
apparent
to
the
networking
layer.
What
the
priority
is
on
this
stuff.
C
C
We
don't
want
to
force
people
to
implement
a
network
layer
underneath
on
their
own
to
solve
the
problem.
If
we
already
know
ahead
of
time
that
there's
a
better
way
to
manage
traffic,
then
it
should
just
be
something
that
comes
as
a
package
deal
and
rides
over
whatever
connectivity
you
got
this.
Will
this
avoids
the
problem
of
saying
my
implementation
can't
do
what
this
other
one
can,
because
I.
G
F
G
Underneath
all
of
that,
before
joining
the
foundation,
I
was
at
pivotal
working
on
cloud
foundry
and
specifically
cloud
foundry
bosch
and
one
one
of
the
things
we
were
doing
well
in
that
platform
was
the
fact
that
it
was
all
based
on
on
some
some
messaging
capabilities
under
underneath,
and
I
wouldn't
I
wouldn't
say
that
bosch
is
a
model
for
for
everything.
But
one
thing
it
was
doing
well
was
the
fact
that
yes,
the
the
deployment,
what
was
always
running
on
a
synchronous
protocol
under
under
the
hood
and
that's
really
really
important.
F
A
F
A
G
C
G
Yes,
and
that's
that's
especially
important,
considering
well
how
frequent
our
air
gap
environments
are.
Certainly
you
don't
want
you
don't
want
to
have
to
pull
from
from
the
public
internet
whatever
you
are
deploying
all
the
time,
and
if
it's
not
baked
in
in
whatever
deployment
solution
you
have,
then
you
run
into
problems,
especially
considering
I
worked
a
few
years
at
nato
and
we
had
completely
distinct
networks,
completely
physically
separated
for
all
the
secret
and
higher
stuff.
So
that's
an
important
thing
to
take
into
account.
G
E
E
A
A
Okay
sounds
great,
so
it
looks
like
even
when
we
started
with
no
agenda.
This
has
been
a
nice
topic
that
is
unpeeled
and
onion,
where
we
discover
a
lot
of
things
to
do
going
forward
for
the
next
year
for
january,
I
did
stop
at
the
ubuntu
booth
and
got
a
demo
of
their
lightweight
kubernetes
for
edge
and
we
talked
to
them
and
they
were
going
to
put
themselves
on
the
agenda
for
one
of
the
january
meetings,
to
give
a
say,
a
30
minute
demo
of
the
ubuntu
edge
solution.
D
H
A
A
C
So
the
fun
the
fun
history
of
that
paper
is
the
first
draft.
I
wrote
in
six
hours
in
a
in
a
in
a
tea
fueled
evening,
then
followed
by.
It
was
about
six
months
of
edit
and
then
and
and
integration
of
of
comments
and
then
finally
publish,
although
maybe
it
was
a
bit
under
six
months
and
there,
and
I'm
saying
this,
because
I
think
the
first
drafts
of
things
can
generally
come
out
pretty
quick.
H
C
E
C
That's
great,
so
I
think
we
can
get
first
drafts
out
pretty
quick
that
have
a
lot
of
content
and
then
I'm
setting
a
deadline,
you
know,
seemed
to
help
us
get
a
bunch
of
revisions
in
and
we
tried
once
to
set
the
deadline,
for,
I
think
one
event
and
didn't
quite
make
it.
And
then
we
said
that
we
would
then
get
the
paper
ready.
For
you
know
another.
C
Event
and
that
worked-
and
we
got
it-
we
got
it
out
in
time
and
so
then,
after
we
got
it
ready,
it
took
another
month
to
get
it
through
the
cncf
process
of
getting
it
published
out.
So
it
seems
like
maybe
it
was
more
like
four
or
five
months,
but
it's
getting
close
to
that.
That
six
month
time
frame.
C
C
C
A
Yeah
the
one
issue
I
found
in
the
past,
because
I've
worked
on
kind
of
not
necessarily
white
papers
but
actually
designs
like
caps
and
there's
some
advantage
to
leaving
it
in
the
google
doc
form
as
opposed
to
github,
because
I
think
the
commenting
and
proposed
edits.
I
personally
like
working
with
a
google
doc,
far
better
than
having
people
log
issues
and
you
know,
submit
prs
against
a
markdown
document.
B
E
B
And,
and
even
after
that,
I
find
that
the
original
white
paper
that
formed
this
group-
and
you
know
just
defining
terms
and
and
everything
it
still
lives
on
and
there's
the
people
coming
up,
putting
comments
there
and
and
getting
things
from
it
so
yeah,
but
definitely
I
think
at
some
point,
you're
marking
it
as
a
you
know
at
least
putting
it.
You
know
a
version
on
it
and
publishing
it
is
is
a
good
thing
to
do,
but
you
know
writing
it
in
google.
Doc
is
much
much
faster,
so
we've
got.