►
From YouTube: Kubernetes WG IoT Edge 20210113
Description
January 13 meeting of the Kubernetes IoT Edge Working group. Small group open discussion of the Robot Operation System (ROS) and Kubernetes
A
A
B
So
most
likely
I
I
am
working
on
ros,
that's
a
robot
operating
system,
but
that
I'm
interested
in
the
kubernetes
to
use
that
against
the
agility
devices
to
control
distributed
system
constru
I
mean
and
inside
the
kubernetes
in
the
container
I
I
would
use
the
robot
operating
system
version
2
to
construct
the
distributed
system.
So
maybe
there
would
be.
You
know,
like
hundreds
robots
working
simultaneously.
C
B
A
B
A
Have
never
used
it
although
long
ago,
in
a
job
I
did
robots
for
industrial
automation,
but
it
was
so
ancient
yeah.
You
know
20
years
ago
that
I
don't
think
it's
remotely
comparable
and
then
right
now,
I'm
actually
coaching
a
high
school
competitive,
robotics
team
where
I
live
and
wow.
So
I
don't
know
if
you're
familiar
with
any
of
those.
There
are
programs
called
vex
and
first
that
are
made
for
high
school
kids
to
have.
A
They
essentially
play
things
like
sports
games
using
innings.
They
they
have
really
constrained
budgets,
so
it's
cheap
equipment
in
the
league,
I'm
in
which
is
first.
They
allow
them
to
have
two
android
phones
as
their
only
electronics
to
control
them
and
they
don't
and
they're,
not
even
latest
generation.
A
A
B
A
And
what
kind
of
guarantees
does
it
provide
for
latency
like
I
assume
it
connects
to
io
and
one
of
the
potential
issues?
You
have
routing
things
through
all
of
these
layers.
You
know,
let's,
let's
take
a
worse
case
of
starting
with
bare
metal,
you
go
to
a
hypervisor
you
go
to
linux,
then
you
go
to
a
container
runtime
and
then
maybe
something
in
a
container
runtime.
A
You
know
the
hardware,
acceleration
features
of
modern
chips
to
some
degree,
make
the
penalties
of
those
layers
low,
but
they're,
not
zero,
and
you
know
in
terms
of
what's
being
done
on
these
robots.
Do
you
need
some
millisecond
response
times
to
io
or
do
you
need
sub
10
millisecond
or
what?
What?
What's?
The
typical
use
case
like.
B
There
is
a
lot
of
use
cases
and
actually
there
is
the
qos.
That
is
that
it
stands
for
quality
of
service.
So
you
can
just
define
your
latency
hey.
I
want
this
latency
or
something
to
deliver
the
messages
or
something,
and
if,
if
it
is
not
achieved,
you
can
get
the
event
or
something
so
actually
you
can
define
you
and
configure
the
latency
or
a
lifetime
for
the
specific
messages
or
something?
That's
that's.
We
provide
as
a
system
so
but
but
there
is
a
use,
a
lot
of
use
cases.
B
C
C
A
Yeah
yeah,
I
think
I
would
throw
out
there.
This
is
my
personal
opinion,
so
I
don't
know
if
you
can
find
this
in
the
official
specs,
but
for
kubernetes
I
think
the
minimum
would
be
arm
or
x86
and
personal
opinion
is
4.
Gig
of
memory
is
minimum,
although
technically
there
are
plenty
of
people
who
can
demonstrate
in
a
video
that
they
got
it
to
run
on
2
gig
of
ram,
but
it
my
experience
is
that
if
you
push
it
that
low
there's,
hardly
anything
left
to
actually
run
your
application
so
yeah.
A
B
A
B
C
A
So
and
then
there's
another
potential
thing
you
can
do
here:
that's
used
for
some
control
systems
when
you're.
If
your
real
goal
is
to
have
a
control
plane
to
control
a
thousand
robots,
there
is
a
kubernetes,
has
a
really
nicely
architected
control,
plane,
that's
extensible,
and
by
extensible
it
means
you
write.
These
things
called
custom
resource
definitions
where
kubernetes
the
control
plane
is
designed
to
be
based
on
these
statements
of
intent.
A
You
describe
what
you
want
to
happen
or
the
condition
you
want
the
world
to
be
in
and
then
they
they
go
and
take
your
statement
of
intent
and
put
it
in
a
database
it.
It
goes
through
an
api,
but
it's
held
in
that
ncd
database,
which
is
a
pretty
it's
resilient.
If
you
have
a
production
grade
kubernetes-
and
you
have
these
controllers
that
subscribe
to
changes
in
the
statements
of
intent,
so
they
don't
have
to
sit
there
polling
for
changes
they.
A
You
know
it's
pretty
efficient,
even
at
scale
that
if
somebody
says
that
I
want
robot
75
to
now
be
version
15
of
the
software
or
to
go
to
this
location,
they
don't
have
to
sit
there
polling.
They
can
subscribe
and
be
notified
when
somebody
called
the
api
and
altered
something
they're
interested
in,
and
then
they
can
find
the
new
set
points
and
drive
the
system
towards
those
desired
goals
and,
interestingly
enough
in
kubernetes
in
a
data
center.
Normally
these
controllers
run
in
kubernetes
itself,
but
they
don't
have
to
be
so.
A
A
A
D
B
C
B
B
You
know
like
applications,
engineer
one
needs
to
adapt
and
adjust
the
code,
the
based
on
the
platform,
which
we
don't
really
like
it.
So
you
know
I
just
want
to.
I
just
want
the
application
engineer
to
be
happy
to
you
know
once
they
push
the
container
that
can
be
run
anywhere.
It
doesn't
matter
if
it's
cloud
or
edge,
but
you
can
just
specify
the
specific
target
or
capabilities
based
on
the
label
and
we
have
a
girl.
B
D
Yeah,
yeah
and
and
sorry
so,
I'm
completely
new
to
ross.
So
so
I
I
just
asked
him
for
beginner
questions.
So
so
is
there
anything
specific
about
applications
running
on
ros
that
that
you
know
need
to
be
concerned?
You
know
how
the
how
would
the
edge
communicate
with
the
with
with
the
cloud
or
or
or
with
the
devices
it
controls.
C
B
C
C
C
A
A
B
C
C
B
Distributed
service,
it's
a
it's
a
different
framework,
but
the
the
problem
that
we
have
in
ros
one
is
like
a
transport
framework
is
originally
made.
So
sometimes
you
know
it's
just
a
rust
one
specific
framework
and
a
message
passing
and
which
we
don't
like
it.
So,
the
in
the
raw
studio.
We
take
advantage
of
the
implementation
from
the
dds,
so
we
define
the
interfaces
and
then
any
vendor
can
plug
in
their
implementation.
B
A
A
B
A
B
B
B
A
A
B
Yeah,
I
I
mean,
like
we
have
already
tried
rust,
2,
ross,
2
and
dds
and
kubernetes
raspberry
pi
and
including
cloud
servers.
Everything
works
okay,
so
we
we.
We
are
positive
to
use
kubernetes
edge
devices.
But
what's
the
pain
here
is
like
a
cpu
consumption
by
kubelet
and
if
we
change
the
container
image
it,
it
takes
time
to
switch
over
the
old
one.
So
some
it
depends
on
the
use
case.
You
know
it
taking
one
minute.
B
A
A
C
A
He's
curious
about
getting
it
to
run
in
kubernetes
and
maybe
even
having
kubernetes
control
thousands
of
instances
of
robots
using
robot
operating
system
in
conjunction
with
the
kubernetes
control
plane.
So
I
imagine
that
I
don't
know
there
might
be
a
little
bit
of
machine
learning
going
on
with
these,
but
I
really
don't
know
the
in
terms
of
machine
learning
in
general
at
edge
at
one
time
there
was
a
kubernetes
machine
learning
working
group,
but
I
believe
it's
been
disbanded.
A
C
A
B
A
B
C
A
C
C
So
this
is
something
I'm
trying
to
address
on
with
some
exploration
at
vmware,
but
I
want
to
talk
about
the
hai
on
this
project
too,
on
some
partner
or
customer.
They
are
very
interested
if
we
can
do
something
with
kubernetes
aux
treaties,
aji
workload
deposit,
something
like
that
so
yeah,
one
of
them
I'm
trying
to
reach
out
here.
A
Well
there
there
is
somebody
who
shows
up
at
these
calls
occasionally
murritz
who's
at
a
university
in
germany.
I
believe
and
he's
working
on
a
system
that
controls
a
laser
cutting
machine
using
high-speed
cameras
yeah,
and
I
think
that
he's
got
some
machine
learning
going
on
for
image,
recognition
in
that
system
and
they're
trying
to
get
it
to
work.
It's
a
research
project
I
believe,
but
he
doesn't
seem
to
be
here
tonight,
he's
often
here,
but
not
this
evening.
A
A
Yeah
we
we
maybe
could
in
advance
at
a
future
meeting,
go
contact
the
people
interested
in
this
field
to
get
a
quorum.
I've
never
personally
encountered
someone
who
professes
to
be
an
expert
in
it.
So
I
think
this
might
be
sort
of
a
leading
edge
kind
of
thing
where
there's
a
lot
of
people
trying
to
explore
and
get
it
to
work.
A
But
you
know
maybe
finding
production
level
instances
of
kubernetes
at
low
resource
edge,
actually
doing
machine
learning
could
be
kind
of
a
rare
thing
or
something
that
maybe
the
company's
doing
it,
consider
it
early
stage
or
proprietary
enough
that
they're
not
necessarily
willing
to
share
it.
At
this
point,
okay,
I
think
for
resource
constrained.
B
A
C
A
You
can
correct
me
if
you
think
I'm
wrong,
but
I
don't
think
it's
likely
to
be
plausible
to
run
it
on
a
straight
raspberry,
pi
right,
even
a
raspberry
pi
4..
But
if
you
could
get
up
to
the
a
gateway
situation
where
you
could
put,
I
don't
know
a
custom,
nvidia
chip
or
some
other
brand
of
accelerator
chip.
C
Yeah,
I
have
some
pattern
like
edgy
link
and
even
other
hardware
vendor
they
already
produced
some
hk
tv
encrypted
with
some
hardware
accelerator
like
I
mentioned
that
intel
mobids
vpu
and
google
http,
or
this
kind
of
stuff,
so
for
wherever
I'm
I'm
trying
to
bring
out
this
field.
Furion,
I'm
not
sure
you,
you
know
that
be
the
filling
vmware
visually
yeah
like
dynamically,
attach
that
gpu
anywhere
in
the
data
center
by
intercepting
the
code
called.
C
C
A
A
We
did
have
a
speaker
who
was
talking
about
using
bitfusion
for
you
other
two
who
aren't
familiar
with
what
bit
fusion
is
this.
It
is
unique
to
the
vsphere
hypervisor,
but
the
idea
is
that
you
can
have
hypervisor
nodes
that
have
fairly
powerful
gpus
installed
and
the
gpus
then
are
carved
up
to
form
virtual
gpus,
so
they
can
be
sub
allocated
out
to
smaller
workloads
and
they
were
running
in
kubernetes.
A
B
C
B
B
Yeah,
actually
I
I
am
the
one
to
have
had
to
have
the
keynote
speech
for
roscon
20.
I
remember
it's
that's
18.,
so
the
eyeball
based
on
the
ros1
and,
of
course,
machine
learning
and
speech.
Recognition
is
processed
in
the
edge,
but
the
training
is
different.
So
in
the
cloud.
So
that's
that's
what
we
do.
So
we
are
interested
in
the
machine
learning
use
cases
also
yeah,
okay
and.
B
C
A
C
C
C
C
A
A
C
A
A
C
A
C
A
I
ask
you
something:
I
don't
know
when
you
know
bit
fusion
technology
and
by
the
way
on,
if
you
edit
your
slide,
I
think
you
bit
fusion
has
a
typo
where
it's
missing
an
s,
but
what
kind
of
bandwidth
is
consumed
with
that
link
to
the
remote
gpu,
I'm?
What
I'm
getting
at
is
I'm
wondering
if
it
would
tolerate
going
over
bluetooth
or
something
like
that
where
you
had
robots
connected
wirelessly.
A
A
A
A
A
A
Well,
it's
been
an
interesting
conversation
as
a
warning.
We've
got
about
15
minutes
left,
but
I
I've
learned
some
things
on
in
this
meeting
and
I'd
like
to
follow
up
on
it.
If
people
are
interested
to
maybe
try
to
declare
a
meeting
on
these
subjects
of
both
machine
learning
and
robotics
and
trying
to
build.
A
A
I
don't
know
I'll
volunteer
to
do
a
little
bit
of
playing
around
prototyping
with
ross
on
kubernetes,
just
because
I'm
curious
so
long
as
I
can
manage
to
get
a
little
bit
of
a
hardware
in
a
home
lab
scenario
where
I
can
try
and
play
with
this
and
I'm
even
willing
to
go
out
and
buy
a
little
bit
within
reason
to
go.
Educate
myself,
like
I
said,
I'm
coaching,
a
high
school
robotics
team
and
I
think
they
might
like
learning
this
as
an
experience
anyway.
So
maybe
I
could
justify
it
on
that
basis.
B
Yeah
something
I
can
do
is
I
can
as
a
first
step.
Maybe
I
can
share
more
details.
What
we
have
done
using
you
know
like
kubernetes,
with
rust2
or
as
construct
distributed
system
in
edge
and
what's
the
what's,
the
problem
that
we
have
and
what
we
are
trying
to
address,
that
the
issues
include
the
contribution
to
the
open
source,
that's
something
I
could
do.
Yeah,
okay,.
B
B
A
D
B
A
A
D
D
A
C
A
A
To
get
this
done
in
time,
but
just
you
know
I'll,
let
both
of
you
know
that
the
cncf
is
contemplating
having
a
kubernetes
at
edge
pre-event
for
kubecon,
the
the
online
kubecon
europe,
and
I
suspect
that
they
may
put
out
calls
for
presentation
sometime
in
the
next
month.
If
you
know,
if
that
would
be
enough,
lead
time
to
try
to
get
a
speaking
proposal.
Put.