►
From YouTube: Kubernetes WG IoT Edge 20230531
Description
May 31, 2023 meeting of the CNCF IoT Edge Working Group. Open discussion: Identifying ways to extend Kubernetes and orchestration to lower resource devices. Edge Native Principals Paper Review.
A
Hi
welcome
to
the
May
31st
2023
meeting
of
the
cncf
iot
edge
working
group
on
today's
agenda.
We
got
a
nomination
to
discuss
a
white
paper
draft
of
behavior
for
Edge
native
applications.
A
The
people
who
put
it
on
the
draft
I
think
would
like
to
do
that
in
the
second
half
of
this
meeting,
because
they
said
some
people
would
be
arriving
a
little
late.
So
until
that
happens,
we're
open
to
member
nominations
for
other
things
to
discuss
can
be
com.
Anything
from
a
question
to
just
some
topic.
You
became
aware
of
something
interesting
on
to
a
question
and
you'd
like
to
have
the
group
resolve.
A
So
the
floor
is
open
for
free
form,
discussions
for
anybody
who
wants
to
put
something
on
the
table
and
for
the
people
who
wanted
to
have
that
white
paper.
Discussion
just
go
in
there,
I'm,
not
even
sure,
if
who's
here
and
who's,
not
other
than
I
heard
that
they'd
be
arriving
late.
So,
let's
plan
on
that,
starting
about
9
30ish.
Unless
somebody
in
that
group
who
put
that
on
the
agenda
wants
to
suggest
an
alternate
time.
B
That
sounds
good
Stephen
yeah.
This
is
Brandon
and
it's
mainly
me
Frank
Joel
and
Andy
who's
who's
going
to
be
the
one
joining
at
the
bottom
half
of
the
hour.
Okay,.
A
D
I
missed
our
last
meeting,
but
two
meetings
ago,
at
the
end,
we
were
talking
a
little
bit
about
identifying
ways
to
extend
kubernetes
and
orchestration
to
lower
footprint
devices
topic
interest,
people.
A
I
think
it
does
I
think
we
may
even
have
a
few
people
on
the
call
who
are
maybe
Authorities
on
that
subject.
It
depends
on
whether
you
know
there's
two
flavors
of
that
either
reducing
kubernetes
itself
or
not
using
kubernetes
and
they're
both
valid
depending
on
kind
of
the
the
right
thing
to
for
you
is.
It
depends
I.
Think
Frederick
here
has
given
a
talk
on
that
subject
at
least
a
couple
times
that
I'm
aware
of
yes.
C
Yes,
I
did
and
yeah.
The
ultimate
answer
to
the
question
was:
it
depends
because,
of
course,
anything
that
has
to
do
with
actual
Mission
critical,
real-time
requirements.
You
know
even
even
the
Linux
kernel,
just
the
Linux
kernel
doesn't
cut
it
for
that.
If,
in
the
eyes
of
some
people,
you
know
even
with
the
preempty
patches
and
all
of
that,
so
of
course
I
think
I
got
any
orchestrator
on
the
top
of
that
yeah
ruins
the
whole
story.
C
From
from
a
real-time
perspective,
that
said,
you
know
you
will
you
will
likely
find
it
somewhere
in
the
infrastructure
or
in
the
in
the
overall
architecture
of
solution
anyway,
not
necessarily
in
the
in
the
robots
doing
actual
Wells,
or
you
know
all
of
those
well
machines
with
Computing
capabilities,
let's
say
or
connected
machines
that
you
could
you
could
find
in
a
factory
or
industrial
infrastructure,
but
certainly
the
how
do
I
say
that
the
the
Border
or
the
the
threshold
for
kubernetes
issues
versus
something
else
is
malleable
depending
on
the
use
case,
I
guess,
yeah.
D
So
some
of
the
work
that
I've
done
has
been
interacting
with
real-time
operating
systems
with
kubernetes.
So
one
of
the
things
we
explored
was
creating
a
custom
resource
definition
so
that
we
could
use
MCU
boot
to
do
an
OTA
update
of
a
microcontroller
running
Zephyr.
So
it
would
trigger
the
second
partition
to
flip
over
and
do
a
reboot
so
that
you
could
be
into
another
workload
that
second
workload
happened
to
be
tensorflow
Lite
being
downloaded
to
the
microcontroller.
So
it
could
do
some
very
primitive
gesture
recognition.
D
Another
approach
we
tried
was
using
kubernetes
to
talk
to
vxworks
to
do
application
over
the
air
updates,
so
you
could
download
new
real-time
applications
to
vxworks
and
we've
also
sort
of
explored
with
the
idea
of
eliminating
some
of
the
apis
that
the
Cuba
has
access
to,
so
that
it
only
focuses
on
the
the
core
kubernetes
API
set,
which
could
be
interesting
for
a
lot
of
use.
D
We're
not
we're
not
putting
together
like
a
separate
framework,
we're
trying
to
stick
to
stock
kubernetes
wherever
possible,
so
I
mean
the
crd
was
custom.
Of
course
we
used
the
the
generator
the
open
API
generator
to
create
a
different
set
of
bindings
in
some
cases,
because
we
obviously
can't
use
golang
on
some
of
these
very
small
devices
that
have
deterministic
requirements,
but
so,
as
far
as
I
can
think,
it
was
like
a
hundred
percent
vanilla,
kubernetes
and
kubernetes
apis.
A
Okay,
so
just
to
make
sure
I
understand
this:
it
it
it's.
What
I
would
call
in
the
traditional
control
application,
what
they
used
to
call
or
still
do
supervisory
control,
where
you
have
a
high
level
control
Loop
that
gives
objectives
to
the
lower
level
devices
that
are
actually
where
they
execute
them,
with
real-time
reaction
parameters
right
so
you're,
using
kubernetes
kind
of
as
a
API
framework
to
effectively
build
your
own
API
to
control
the
actual
devices
in
a
supervisory
mode.
It's
that
about.
D
We've
experimented
with
both
like
with
the
small
microcontrollers.
Your
comment
is
bang
on,
because
we
would
have
like
a
Linux
gateway
to
run
the
cubelet,
and
then
the
crd
would
use
the
cubelet
running
on
the
gateway
to
interact
with
the
microcontroller
over
Bluetooth.
Because
is
it
there's
that
air
gap
literal
air
gap
between
the
two?
D
But
we
also
have
done
experiments
where
it
goes
beyond
just
supervisory
control?
And
we
have
done
a
number
of
different
things
to
play
out
different
scenarios
where
the
SLA
for
one
device
might
fail,
and
we
actually
have
orchestration
going
on
in
the
background
to
fall
back
to
another
device
that
is
able
to
meet
the
the
objective.
A
Well,
I
have
to
say
that
being
in
this
group
for
years
now
that
going
down
that
path
by
my
observation
is
really
common,
that
you're,
you
know
you
don't
have
to
be
nervous
that
you're
the
first
one
to
ever.
Do
it.
It's
almost
maybe
a
little
shocking
that
there
are
more
standardized
ways
of
doing
that,
like
off
the
shelf,
where
somebody
has
turned
that
into
a
platform
so
that
users,
one
at
a
time,
don't
effectively
go
do
the
research.
A
D
And
and
that
hits
on
the
reason
I
wanted
to
bring
it
up,
I
mean
making
sure
that
we
adhere
to
the
standard
and
create
some
sort
of
Open
Source
standard
for
people
that
want
to
do.
This
is
something
that
we'd
be
interested
in
interested
in.
In
particular,
when
I
started.
Looking
at
the
worker
note
requirements,
there
were
some
things
that
make
perfect
sense
for
a
standard
worker
node.
That
may
not
be
required
for
some
of
these
forage
devices
like,
for
example.
D
Perhaps
you
don't
actually
need
an
overlay
Network,
because
your
small
little
microcontrollers
may
not
be
talking
to
each
other
directly.
They
may
just
be
talking
to
the
supervisor.
To
use
the
terminology
you
use
Steve,
so
I'd
be
interested
in
sort
of
figuring
out
whether
we
could
Define
some
sort
of
leaf,
node
or
device
node
in
the
specification,
so
that
we
could
say
that
we
were
adhering
to
The
Standard.
E
Yeah
Liam
Randall
I
work
on
a
webassembly
in
plasma
Cloud
long
time,
kubernetes
person
as
well.
You
Rob,
if
you
explored,
you,
know
using
webassembly
or
runwasi
or
any
of
the.
You
know,
webassembly
tools
on
the
real-time
OS
stuff
in
your
work
here
and
I
apologize.
I
was
a
couple
minutes
late
onto
the
call
I'm,
not
sure.
If
you
already
covered
that.
D
That
is
something
I'm
interested
in
both
web
assembly
and
things
like
ebpf
or
ubpf.
Depending
on
what
the
scenario
is.
One
of
the
restrictions
that
we
face
is
we
often
work
on
safety,
critical
devices,
so
the
languages
we
use
need
to
be
certifiable
so
that
restricts
Us
in
some
regards
yeah.
C
D
The
main
advantage
that
we're
perceiving
is
the
ability
to
have
a
single
pane
of
glass
for
all
of
your
devices
so
that,
when
you're,
looking
at
your
Enterprise
I.T
systems
and
your
OT
systems,
you're
able
to
have
those
currently
separate
teams
come
together
and
do
common
planning
for
how
you
coordinate
software
updates
across
all
those
systems.
How
you
roll
out
security
fixes
and
how
you
make
sure
that
you're
able
to
keep
all
of
your
systems
up
to
date.
Using
the
same
sets
of
policies
and
procedures.
C
Yeah,
that
makes
sense.
I
was
expecting
that,
let's
say,
but
in
any
case
yeah
I
was
I
was
wondering
how
you
you
viewed
it
yourself.
Thank
you.
F
That
would
be
kubernetes
because
to
me
that
sound
a
little
bit
like
what
they
say,
cross-plane
folks
or
the
likes,
are
doing
so
that
you
can
go
and
represent
everything
as
as
a
resource
in
kubernetes,
including
small
sensors
and
I,
saw
a
demo
a
year
ago,
where
they
had
actually
real
container
ships
represented
as
resources
in
kubernetes,
and
they
were
able
to
go
and
orchestrate
them
and
manage
them
with
Shivers
in
which
Landing
Dock
and
whatever
anything
they
can
be
a
resource
right.
F
So
I
think
that's
more
like
your
use
case
rock,
where
you
use
one
one.
Generic
control
plane
for
well
virtually
any
endpoint,
rather
than
you're,
trying
to
go
and
run
kubernetes
on
this
endpoint
right,
yeah.
A
I
think
Frank
you're
exactly
right,
but
it
once
again
it's
a
Continuum
kind
of
based
on
how
much
resource
you
have.
So
some
of
these
things
that
people
call
Edge
have
enough
to
effectively
run
three
nodes
of
compute.
You
know
three
Intel
Nooks
or
something
in
which
case
they
can
be
a
kubernetes
cluster
and
it
can
make
sense,
but
others
I
think
it
becomes
kind
of
a
gray
area.
When
you
get
to
one
compute
node,
you
know
you
could
run
that
as
a
so-called
single
node
kubernetes
cluster.
A
A
Kubernetes
cluster
actually
makes
sense,
because
if
you
look
at
the
components,
how
much
of
kubernetes
is
actually
delivering
value
just
versus
just
overhead-
and
you
know,
points
of
failure,
points
of
security
risk
like
I,
say
it's
a
gray
area
and
some
people
do
it
anyway
and
it
works
for
them,
and
then
you
get
out
to
low
resource
where
it
can't
even
run
one
node
of
a
kubernetes
cluster,
simply
not
enough
CPU
in
memory
and
really
low.
A
You
might
have
some
Edge
nodes
that
have
big
resource,
but
others
that
are
just
very
tiny
outposts
and
a
solution
that
could
embrace
them
all
is
attractive
because
you
know
there
there
are
going
to
be
situations
likely
that
at
least
your
I
o
devices
probably
need
to
be
managed
if
they're
smart
at
all,
they
probably
have
security
risks
and
updates
firmware,
and
if
your
one
system
could
do
all
of
that,
that
would
be
really
attractive,
so
that
I
think
even
some
of
these
people
with
three
tiny
kubernetes
nodes
still
have
additional
smart.
A
I
o
devices
that
they
would
like
to
also
manage
in
I.
Don't
know
I
hate
the
term
one
pane
of
glass
because
it
really
isn't
a
world
where
operators
are
involved
at
the
scales
we're
talking
about,
but
one
overall
supervisory
control
plane
that
can
put
all
of
these
things
together,
like
I,
say:
I,
don't
think
that
there's
a
solution,
maybe
that's
a
good
thing,
because
the
technology
at
Edge
is
certainly
evolving
rapidly,
so
locking
in
a
design
early
might
be
actually
undesirable.
A
Things
like
advances
in
microcontrollers,
getting
smaller,
cheaper,
lower
power,
web
assembly
being
Universal
across
CPU
boundaries,
but
they're
they're,
sort
of
all
like
bit
spot
solutions
to
one
aspect
of
the
many
problems
you
have
and
yeah.
No
one
has
There's
an
opportunity
for
somebody
to
to
do
an
over
embracing
Suite,
I
think
so.
D
So
I
I
would
suggest
perhaps
writing
out
the
requirements
or
standards
for
that
leaf,
node
or
device
node
might
be
a
good
place
to
start.
Is
there
anyone
that
would
be
interested
in
doing
that
with
me.
E
I'd
work
on
that
I
mean
we've
been
working
through
this
problem.
You
know
through
webassembly
and
I,
think
we
kind
of
hit
across
the
Continuum.
You
know
we
obviously
support
kubernetes
out
of
the
box
with
cncf
Blossom
Cloud.
You
know
you
can
deploy
natively
or
via
operators
right
there,
but
where
we
definitely
see
the
biggest
opportunity
is
way
down,
Market
or
in
complex
devices.
Think
like
the
AMD,
zolynix,
fpgs
or
fpgas.
E
You
know
they've
got
multiple
processors
on
them
and
you
look
at
the
different
lines
of
processors
that
arm
builds.
For
example,
they've
got
the
real-time
CPUs,
as
well
as
like
the
server
side
or
mobile
CPUs
as
well,
and
you
often
find
these
things
integrated
together
in
complex
packages,
so
I
do
think
that
starting
with
defining
the
domain
and
what
are
the
characteristics
that
sort
of
would
be
a
far
Edge
versus
a
middle
I
think
would
be
helpful.
There
is
an
awesome.
E
You
know
Linux
Foundation
sub
project,
that
I'm
sure
you
guys
have
seen
that
infographic
of
you
know
that
they
did
for
the
edge
Edge
across
and
pulled
up
real
fast
here.
You
know
we,
you
know
kind
of
made
a
custom
version
of
this
I
think
that
includes
browsers.
E
So
it's
a
little
bit
more
web
assembly
focused,
but
I
know
that
we're
not
the
first
people
to
have
this
discussion,
so
it's
probably
worth
looking
at
some
of
the
prior
art
in
and
around
the
ecosystem
here
in
order
to
in
order
to
see
in
order
to
see
what
people
have
done,
if
I
can
get
your
screen
share,
I'd
show
this
image
real
fast
and
folks
could
take
a
look
at
it.
Okay,.
E
D
Well,
he's
doing
that
I
think
that
sounds
great,
like
identifying
all
the
different
permutations
and
combinations
I
liked
what
you
said
about
the
the
arm,
cortex
R
series,
because
in.
D
You
might
have
the
cubelet
running
on
the
a
core
yeah
and
you
might
be
using
something
like
RP
message
to
send
messages
back
and
forth
between
the
two
and
Trigger,
like
a
update
of
a
workload
on
the
r
series.
So
those
would
be
things
that
I
would
love
to
explore.
You.
E
And
I
and
I
are
thinking
exactly
the
same
here
on
you
know,
I
mean
these
are
the
these:
are
the
brains
behind
like
5G.
You
know
things
like
that
when
we
think
about
the
last
hop
before
you
get
to
a
device
or
on
a
device
itself.
You
know
that's
essentially
what
you're
looking
at,
and
this
is
from
the
from
the
Linux
Foundation.
They
have
a
whole
separate
working
group
that
where
this
image
is
sourced
from
and
we
made
a
custom
version
of
it,
I'd
be
happy
to
open
source.
E
If
we
wanted
to
tweak
one,
because
we
also
threw
in
our
version,
we
threw
browsers
in
there
because
we
feel,
like
you
know,
things
like
webassembly,
really
just
give
you
the
world's
tiniest
virtual
machine
so
that
you
can
think
of
a
browser
or
other
non-traditional
execution
targets,
as
as
a
domain
where
you
can
be
operating
devices
but
I
think
that's
certainly
is
a
different
definition
than
what
you
would
use.
If
we
were
thinking
about.
Where
does
kubernetes
work
and
in
a
kubernetes
in
a
world
dominant?
You
know
where
just
kubernetes
is
an
option.
A
I,
ask
you
a
follow-up
question
on
that
to
make
sure
I
get
this
proposal
straight,
but
using
a
browser
to
host
webassembly
are.
Is
this
something
you
would
think
for?
Edge
could
be
something
where
the
browsers
effectively
the
equivalent
of
a
Docker
runtime
for
a
real
low
resource
thing
that
is
already
there
guaranteed
to
be
installed
in
the
OS
and
you're,
not
even
using
as
a
browser.
Just
something
you'd
invoke
to
run
a
piece
of
code.
Yeah.
E
Absolutely
Let
me
give
two
examples
of
that.
So
Adobe
did
a
case
study
with
our
open
source,
wasn't
cloud
and
one
of
the
big
takeaways.
This
was
in
background
removal
as
a
microservice.
One
of
the
big
takeaways
was
that
in
over
80
percent
of
all
cases,
they
were
actually
able
to
push
the
workload
into
the
customer's
browser,
and
that
meant
that
it
was
faster
because
you
weren't
uploading
images
and
back
that's
400,
milliseconds
versus
just
pure
browser
was
40
milliseconds.
E
It
was
cheaper
because
they
weren't
paying
for
the
compute
and
it
was
the
same
quality.
So
I
think
that's
where
webassembly
and
kubernetes
really
start
to
tell
the
Better
Together
story,
because
on
their
back
end,
Adobe
is
all
kubernetes,
so
they're
able
to
seamlessly
you
know,
run
workloads
that
extend
in
both
ways.
E
This
is
a
your
first
part
of
the
question
in
the
second
part,
I
would
point
out
when
you're
looking
at
you
know
some
of
the
Modern
Edge
type
people
like
fastly
fastly
uses
a
runtime
to
run
their
compute
products
for
webassembly.
That's
called
wasn't
on
and
that's
from
the
BCA,
but
cloud
flare
actually
uses
headless.
They
use
V8
in
in
running
their
webassembly
processes,
so
they're
actually
using
you,
know
the
core
engine
out
of
the
browser.
The
other
big
engine
is
spider.
E
D
E
You
know,
there's
there's
a
couple
projects
that
integrate
this
directly
into
kubernetes,
there's
a
if
you
wanted
to
run
webassembly
natively
I'm
going
to
do
some
big
air
quotes
there,
because
the
deeper
you
get
into
kubernetes
the
more
you
discover
that
kubernetes
is
really
tightly
coupled
to
the
idea
of
a
container,
but
there
is
a
shim.
That's
called
runwazi.
E
E
Actually,
we
have
kubernetes
operators
and
things
like
that
for
pushing
and
deploying
as
well
so
I
think
there's
multiple
strategies
depending
on
the
architecture-
and
you
know
what
you're
looking
at
Nats
does
speak
in
qtt
right
out
of
the
box,
which
is
one
of
the
reasons
that
we
selected
it
for
our
sort
of
Transport
layer
between
house.
A
So
Liam
to
summarize
I
see
that
there's
the
transport
layer
kind
of
the
orchestrator
that
would
tell
a
a
leaf
or
an
edge
node
What
It
Wants
running
in
what
version
and
the
observation
that
perhaps
you
can
use
a
web
browser
engine
to
be
your
runtime
for
the
Wazi.
But
what
about
actual
containerization
of
it
and
by
container
I,
don't
necessarily
mean
Docker
container,
but
maybe
I
do
mean
oci,
because
it
strikes
me
that
you
still
need
packaging
for
these
web
assemblies.
Where
you
have
provenance.
E
E
That's
a
great
question,
so
this
is
the
image
that
I'm
happy
to
donate.
A
version
of
this
that,
where
we
just
put
the
browsers
in
if
we
wanted
to
you
know
like
add
this
and
I'll
have
our
Graphics
guy
is
awesome.
E
I'll
have
him
just
you
know,
release
a
version
of
this
that
we
can
just
pick
up
and
use
and
we
can
customize
our
definitions
and
use
whatever
we
want
here
and
you
know,
lay
kubernetes
on
it
as
a
layer
more
clearly
or
something
I've
tried
to
do
that
here
lightly
and-
and
you
know,
that's
our
corporate
cut
sheet
kind
of
stuff,
but
when
you're
thinking
about
how
do
you
move
these
webassembly
modules
around?
They
are
actually
moved
around
as
oci
artifacts,
and
there
were
huge
reasons
for
this.
You
know
we
worked.
E
If
you
look
at
who's
in
the
by
code
Alliance,
you
know
it's
Google,
Microsoft
Amazon
and
all
of
these
companies
have
huge
ecosystems
around
oci
scanning
of
artifacts
itemizing,
artifacts,
Distributing.
Caching,
you
know
all
of
those
kind
of
things
now
underneath
the
hood,
though
there
is
a
little
bit
more
introspection
that
goes
into
webassembly.
E
Webassembly
is
really
about
moving
around
a
unit
called
component
and
what's
powerful
about
components
is,
is
that
they
can
be
combined
into
new
components
or
they
can
be
satisfied
at
runtime.
So
one
of
the
big
ideas
that's
happening
behind
webassembly
is
just
right
here,
your
business
logic
and
then
deploy
or
attach
your
components
for
web
servers,
key
value
and
other
non-function
requirements
at
runtime,
as
opposed
to
compile
time.
E
The
advantage
there
is
is
that
you're,
obviously
getting
the
latest
version,
and
when
you
look
at
the
devex
around
like
a
fastly
or
a
cloud
flare
or
my
company
cosmotic,
that's
really
what
we're
driving
for
as
well,
which
is
you
know,
containers
are
great,
but
we're
really
driving
for
that
next,
higher
level.
Abstraction
where
you
look
at
the
common
complexity,
that
is,
you
know
the
things
that
the
95
of
the
source
code-
that's
all
open
source
that
goes
into
each
and
every
app.
E
You
know
that's
the
really
kind
of
big
idea
there,
but
underneath
the
hood
it
is
tightly
coupled
and
aligned
to
a
Better
Together
story
with
kubernetes
I'm
sure
CNC
Blossom
day
as
well,
and
we
actually
just
announced
wasm
con
recently
I'm
sorry
I'm,
trying
to
not
turn
into
trying
to
turn
this
into
a
web
assembly
call
just
just
on
on
things
that
I'm
super
excited
about.
So.
E
A
This
sounds
great,
injecting
this
into
the
birds
of
a
feather
and
I.
My
personal
latitude
is
maybe
I'd
like
to
turn
this
into
a
longer
discussion.
If
you
want
to
put
yourself
on
the
agenda
for
a
future
meeting,
you
know
to
go
into
this
in
more
depth.
Our
agenda
is
open
to
members
to
just
nominate
self-nominate
things
just
go
put
it
in
a
future
date.
A
F
G
A
Wanted
to
start
an
organized
thing
you
know
under
the
auspices
of
the
cncf
or
Frederick
here
represents
the
eclipse
foundation
too,
and
I've
got
nothing
against
having
this
be
a
joint
effort
of
multiple
groups
as
long
as
they're,
open
source
and
yeah
or
multiple
groups
in
the
cncf,
because
I
think
there's
a
web
assembly
thing
coming
together
under
tag,
runtime
I
believe
yeah.
G
E
A
Could
live
with
efforts
that
cross
domains.
E
Yeah
yeah
we're
we're
plugged
in
there
Kevin.
The
creator
of
Blossom
cloud
is
the
tech
is
the
technical
advisor
to
that
new
subgroup.
That's
spinning
up,
so
we
have
a
couple
members
that
attend
there,
but
it's
just
pure
coincidence
that
I
decided
to
start
attending
this
and
we
needed
to
and
I'm
actually
trying
to
work
on
a
Wind,
River
Project
right
now
so
for
Rob
on
my
ptme
after
this
call,
I've
already
got
questions
about
a
whammer
support
on
on
the
on
Wind
River.
Real-Time
OS
sounds.
A
Yeah
I've
done
seen
you
at
any
other
meetings
Leanne,
but
just
another
thing
that
might
be
of
interest
to
you.
That
came
up
a
few
meetings
ago
is
that
you
know
one
of
the
hottest
topics
going
on
here
and
Edge
isn't
going
to
be
left
out,
but
it
it's
Ai
and
machine
learning
and
you
know
to
reinforce
that
it.
A
For
this
bof
I
was
saying
what
have
you
done?
Cool
and
I
kept
my
mouth
shut,
but
the
coolest
thing
I've
done
in
the
last
week
is
I,
went
to
my
first
physical
Meetup
of
What's
called
the
tiny
ml
foundation
and
they're
having
physical
meetups
around
the
world,
and
it's
kind
of
astounding
you
know
chat.
Gpt
gets
all
the
popular
press,
but
man
the
amount
of
r
d
being
Unleashed
for
Edge
machine
learning,
accelerators
inference.
Engine
accelerators
is
just
incredible.
A
A
few
weeks
back,
I
put
references
to
ICS
with
accelerators
in
our
slack
Channel
and
I
came
up
with
16
just
easily
I'm
sure,
there's
many
more
than
that
went
to
this
meeting
and
it
was
local,
Southern
California,
but
there
were
about
30
people
there
and
man.
Everybody
was
excited.
They
were
generally
representing
Hardware
vendors
in
the
form
of
Chip
vendors,
but
the
world
world
is
going
to
see
some
big
advances
there.
Now.
What
goes
along
with
this
I
think
is
a
need
to
have
a
control,
plane
or
and
orchestration.
A
If
you
will,
although
you
know,
there's
yet
another
thing
to
be
managed
that
I,
don't
think
Docker
containers
or
web
assemblies
the
solution
and
that
is
getting
updated.
Training
model
training
data
to
these
Edge
nodes
that
are
actually
going
to
run
it,
and
maybe
bi-directional
Communications
with
all
the
typical
challenges
of
low
and
intermittent
connectivity
with
these
Edge
nodes.
A
But
I
think
that
this
is
going
to
be
a
huge
opportunity
in
the
world
is
just
really
taking
on
some
of
these
opportunities,
at
least
in
proofs
of
Concepts
and
early
arrivals,
are
having
success
from
what
I.
What
what
I've
been
seeing
at
this.
E
Yeah
yeah,
that's
that's
where,
when
you
look
at
the
you
know
the
design
of
awesome
Cloud,
it's
really
positioned
around
some
of
the
Native
orchestration
challenges
and
with
webassembly
specifically
and
we've
been
focused.
There.
We've
been
a
cnci
project
for
a
couple
years
now
and
we've
exceeded
the
effort
to
go
incubating
so
we're
in
the
process
of
applying
for
incubation
now,
which
we
hope
will
raise
the
kind
of
profile
in
and
around
the
ecosystem.
E
We've
done.
That
was.
We
did
a
talk
at
arm
Summit
in
the
fall
where
folks,
from
Intel
and
BMW
had
a
extended
the
webassembly
standard
to
support
interactions
with
machine
learning
models
that
were
running
in
tensorflow
underneath
webassembly,
so
it's
sort
of
as
a
sidecar
to
webassembly,
not
in
webassembly
itself,
but
the
logic
was
in
webassemble.
In
order
to
do
you
know
to
Route
data,
to
a
local
machine
learning
model
or
to
a
cloud-based
model
depending
on
security,
privacy
or
availability.
E
So
I
think
that
we're
all
circling
around
the
same
elephant
and
looking
for
all
these
amazing
pieces
of
it
that
we
can
pick
off
and
I,
definitely
think.
There's
a
lot
of
opportunity
and
greater
collaboration
Steve
to
come
back
to
your
point
in
around
open
communities
where
we
can
build
on
each
other
and.
A
We're
going
to
call
this
to
a
close,
because
I
think
the
people
are
here
for
our
second
topic,
but
Liam
just
go
back
in
the
agenda
notes
and
look
at
it
might
be
of
interest
to
you
of
something
from
Sony
media
quora.
That's
an
attempt
to
use
webassembly
to
create
an
abstraction
layer.
Virtualization
around
machine
learning,
inference
engines.
It's
go.
A
So
for
the
next
topic,
the
discussion
of
the
edge
native
application
draft
white
paper
who,
who
wants
to
run
that,
if
you're,
not
Coastal,
co-host
already
I'll,
give
you
that
permission.
B
Well,
I
can
start
off
introducing
the
topic,
but
I
think
Frank
or
Andy
should
should
probably
share
their
screen
and
do
the
running
through.
G
B
So,
to
remind
folks
of
what
this
is,
a
group
of
us
have
been
meeting
now
periodically
over
the
last
few
months
to
further
develop
the
idea
we
had
of
a
supplementary
paper
to
go
as
an
accompaniment
with
the
original
Edge
native
applications
white
paper
that
was
published
a
while
back
and
in
talking
about
using
that
paper
as
a
foundation.
B
You
know
we
thought
about
what
would
be
the
next
logical
extension
of
the
topic
in
order
to
produce
an
asset
that
would
be
valuable
for
the
ecosystem,
so
rather
than
modify
that
original
paper,
which
I
we
all
think
is
still
a
great
Foundation,
we
decided
to
create
a
supplementary
piece
with
the
focus
on
design,
behaviors
or
developers
to
put
those
Edge
native
application
principles
into
practice
and
make
them
more
actionable
and
to
be
used
in
accompaniment
with
the
original
white
paper
as
a
guide
for
developers
designing
and
building
applications
for
the
edge.
B
So
we
had
a
discussion
here
with
this
forum.
I
think
it
was
about
five
six
weeks
ago,
where
we
talked
through
the
concept
and
an
early
draft.
We
then
collected
feedback
live
from
that
group
and
then
shared
the
this
draft
of
the
paper
on
the
slack
Channel
and
with
the
Google
group.
Since
that
time,
we've
seen
comments
coming
in
and
we've
now
reconciled.
The
majority
of
those
I
think
there's
still
a
few
comments
down
below
that.
B
We
can
maybe
talk
through
here
today,
but
then
just
following
sort
of
the
same
process
as
last
time
with
the
creation
and
finalization
and
Publishing
of
the
first
white
paper.
I
think
we're
at
the
stage
now
where
we're
presenting
this.
As
our
proposed
final
draft
back
to
this
group,
we
can
take
this
opportunity
to
incorporate
any
final
feedback
from
the
group
and,
if
desired,
you
know
we
could
leave
this
stock
open
for
a
little
bit
longer
for
any
additional
comments
or
questions
from.
A
B
But
for
my
role
here,
I
just
took
the
opportunity
last
night
to
go
through
the
paper
again
provide
and
edit
a
very
light
edit
for
clarity,
30
or
flow
and
a
proofread,
and
in
order
to
sort
of
match
the
format
of
the
last
paper.
I
added
an
author
and
reviewer
section
to
page
one
and
included
some
names
of
folks.
That
I
know
are
working
on
the
project.
There
might
be
some
names
missing.
F
B
So
that's
a
that's
a
request
to
the
group
and
then
once
we
figure
out
a
publish
date,
that's
something
that's
finalized
there
at
the
end
and
then,
as
we
wrap
up
with
this
group,
I'll
then
Channel
it
the
work
over
to
the
cncf
marketing
committee
and
they
have
a
designer
team
there
who
can
put
this
into
a
PDF
format
and
then
work
into
their
upcoming
promotion
schedules
like
we
did
for
the
last
white
paper.
B
So
that's
what
I
had
and
I
know
that
you
know
Frank
and
Andy
and
Joel
have
been
pretty
instrumental
in
getting
this
paper
to
where
it
is
now
so
I,
don't
know
if
one
of
you,
gentlemen,
want
to
pick
it
up
from
here
and
talk
through
where
we
are
and
maybe
talk
through
any
final
questions
or
comments
that
the
group
might
have.
H
Frank
you
wanted
to
do
something
like
basic
introduction
and
I'll.
Add
color
along
and
Joel
and
I
can
add.
Color
yeah.
F
F
F
This
high
level
flow
already
so
as
Brandon
was
saying
like
we
are
trying
to
go
and
have
those
white
paper
complement
the
original
white
paper
with
the
principles
to
go
and
help
people
design
an
application
to
be
quote
unquote,
Edge
native
and
when
we
did
the
review
here
in
in
this
group
last
time
we
said
well,
it
would
be
really
helpful
to
have
a
an
example
like
an
anchor
or
a
beach
at
that
we
can
go
and
use
an
example
application
that
we
would
then
say:
okay,
this
is
how
we
can
go
and
architect
and
use
these
both
the
design
principles,
as
well
as
the
software
application,
design,
behaviors
that
were
distilling
in
this
white
paper
and
we'll
make
that
a
reality
and
if
you
scroll
down
to
the
very
bottom,
that's
the
main
addition
that
we've
done
over
the
past
couple
of
weeks.
F
If
you
go
to
the
very
bottom
Brendan,
then
you'll
you'll
see
that
section
that
says
I,
don't
even
recall
the
name.
So
we
we
still
have
these
individual
behaviors
that
we're
calling
out
and
yeah.
F
We
have
this
imaginative
application
sample
scenario
and
the
scenario
that
we're
picking
is
one
that
I
think
many
applications
use
today,
where
you
have
the
the
situation
that
you're
acquiring
video
at
the
edge
and
that
you
need
to
go
and
turn
that
video
signal
into
information,
and
that
includes
something
that
we
discussed
earlier
on
typically
running
something
AI
or
another
way
to
go
and
interpret
the
video
and
turning
that
into
a
signal
and
uploading
that
that
signal
to
the
clouds
where
it's
going
to
go
and
get
further
processed,
and
that
might
be
for
bandwidth
reasons
that
you
you
do
this
or
it
might
be
for
reasons
of
privacy,
regulatory
policy
that
you
want
to
go
into
that
processing
at
the
edge.
F
So
we
basically
said:
okay.
Well,
let's
take
this
very
simple
application
where
we
have
an
edge
portion
and
that
would
live
like
a
smart
device.
Hatch,
so,
relatively
far
out,
where
you
know
that
this
site
is,
is
going
to
be
detached,
it's
every
now
and
then,
where
it
might
not
have
a
load
of
Upstream
connectivity
where
the
Upstream
connectivity
might
be
constrained
over
multiple
layers
of
an
ad.
And
what
have
you
where
we
would
acquire
the
image
then
do
some
processing
in
it
on
it.
That
might
be
simple
resizing.
F
It
might
also
be
more
sophisticated
AI
models
than
you
upload
the
results
and
then
on
the
on
the
cloud
side.
You
you
take
the
input
and
then
you
you
do
further
stuff
with
it
right.
So.
G
F
To
go
and
come
up
with
a
very
generic
architecture
for
that,
and
then,
if
you
scroll
further
down,
then
say
well
the
original
design,
principles
or
behaviors
that
we
had.
How
do
they
apply
to
this
particular
use
guys
so
that.
I
F
Become
really
like
more
that
you
can
have
a
way
to
mentally
attach
to
these
individual
design
principles
and
design
behaviors
in
a
more
natural
way,
using
this
example
application.
Yes,
it's
a
very
simple
one,
but
simple
I
think
it's
not
necessarily
bad
here,
because
it
probably
resonates
with
every
everybody.
So
we're
saying
like
what
do
I
need
to
go
and
do
for
concurrency
and
scale.
What
do
I
do
need
to
go
on
for
articulating
dependencies?
How
do
I
deal
with
failures
at
the
edge
for
that
particular
application?
F
How
do
I
deal
with
the
fact
that
edges
need
to
go
and
be
able
to
discover
their
environment
and
be
aware
of
their
environment?
Like
yeah,
well
discover
the
endpoints
that
you're
dealing
with
from
a
sensor
perspective
in
this
case?
That's
the
camera.
How
do
I
deal
with
storage
I.E?
You
don't
want
to
go
and
persist
data
at
the
edge
you
want
to
persist
data
in
the
cloud.
You
only
want
to
go
on
Cache
so
that
the
disposability
of
of
the
edge
location
is
still
a
given.
F
How
do
you
deal
with
metrics
and
logs
I.E
were
only
expose
actionable
logs
because
we're
assuming
a
non-expert
operator
at
this
very
far
Edge
and
that
dovetails
for
the
with
the
operational
piece
so
I
think
that's
the
main
addition
of
what
we
had
so
making
the
the
additional
or
the
behaviors
that
we
distilled
from
from
a
white
paper
perspective,
making
them
say
more
visual
or
more
more
articulated
by
by
using
this
simple
example,.
H
Yeah
so
Frank
just
to
back
up
on
that
I
think
we
were
also
talking
about
if
you
scroll
up
just
a
bit
Brandon
on
to
the
the
tiering
discussion
here.
So
this
is
one
of.
H
No
problem,
I'm
gonna,
add
I'm
gonna
dovetail
that
in
because
it
adds
a
layer
of
complexity
here
so
yeah.
So
one
of
the
other
things
that
we
talked
about
you
know
within
the
working
group
is
that
there
was
this
notion
of
you
know.
Well,
today,
people
are
very
much
used
to
HUB
and
spoke,
but
when
you
get
into
spaces,
where
disconnected
operations
is
more,
you
know
a
familiar
concept
like
oil
rigs,
offshore
cruise
ships,
Interstellar
Communications,
etc,
etc.
There's
becomes
this
notion
of
well.
H
H
This
notion
of
intermediaries
or
esps
Edge
service
providers-
and
there
may
be
multiple
of
these
that
are
in
the
middle
that
provide
either
some
kind
of
DMZ,
where
there's
a
division
of
responsibility,
Matrix
between
what
is
the
Hub
responsible
for
versus
the
spoke
or
it
could
be
just
merely
a
transport
layer.
That's
ignored
for
the
for
the
sake
of
targeting
deployment
of
workloads
and
it's
just
used
as
a
you
know,
a
means
to
get
to
the
outer
edge,
and
this
could
be
something
I
mean
Steve.
You
brought
this
up
in
the
past.
H
This
could
be
just
somebody
like
that
works
on
a
on
an
agricultural
setting
in
an
agricultural
setting.
That
is,
has
a
cell
phone
that
collects
all
the
data
and
then
at
some
point
in
the
future,
they're
going
into
town
to
get
supplies,
and
they
happen
to
brush
up
against
the
Wi-Fi
or
or
some
kind
of
cell
spot
and
then
are
able
to
transmit
data.
H
That's
not
something
you
really,
you
know,
could
connect
with
a
hub
and
spoke
model,
but
it's
something
that
we're
thinking
about
in
more
of
you
know
at
these
intermediary
layers
in
between,
and
it
gives
a
lot
more
flexibility
to
those
that
are
building
Edge
native
applications
now
and
in
the
future.
H
H
So
you
can
imagine
putting
an
intermediary
tier
in
the
middle
of
these
two
gray
boxes
here,
something
that's
sat
in
the
middle
and
provided
some
layer
of
abstraction
or
layer
or
means
of
transportation
between
the
The
Edge
and
the
cloud
provider
or
or
the
cloud
space.
F
G
If
you,
if
you
take
the
this
here,
it
has
an
example
of
a
single
edge
where
the
tax
stocks
at
the
bottom
is
just
putting
sample
building
on
that,
to
give
example,
deployment
models
of
here's,
one
Edge,
here's
what
it'll
look
like
this
has
to
get
added,
then
I
have
to
add
here's.
G
What
one
Edge
looks
like,
and
now
here's
what
you
know
this
goes
now:
200,
maybe
500
and
building
that
from
one
Edge
to
a
region
to
Global,
if
that
made
sense,
the
gist
of
from
one
location
and
how
it
scales
into
Andy's.
Point
of
here
are
the
different
tiers
and
what
that
might
look
like
and
again,
the
intent
is
to
provide
an
example
of
people
can
then
go
and
utilize,
for
whatever
their
specific
project
is.
I
H
Past
was
there
a
question.
Sample
scenario,
looks
really
interesting.
I'll
definitely
take
a
look.
Okay,.
A
Hey
I,
just
added
a
comment:
you
can
look,
I
got
long-winded
like
I
do
verbally
too,
but
my
comment
was
that
there
is
a
section
there
that
appeared
to
me
to
ban
persistent
storage,
declaring
that
all
data
is
persisted
in
the
cloud
and
it
struck
me
as
maybe
to
a
fair
authoritarian
that
there
are
use
cases
where,
for
legal
privacy
reasons
you
explicitly
or
maybe
even
doing
your
compute
at
the
edge
specifically
because
you
cannot
have
data
persisted
to
cloud
and
that
this
should
be
open-ended
to
allow
for
those
scenarios
and
as
I
was
writing
my
comment.
A
A
Maybe
it's
the
standard
data
is
cloud
back
and
the
edge
only
holds
is
viewed
as
a
cache,
but
perhaps
allow
for
another
category
that
isn't
allowed
to
leave
the
building,
or
you
know
the
edge
node
and
clearly
it
does
exist,
and
there
are
use
cases
for
that.
A
F
Now
I
think
that's
a
good
point
Steve
that
you
bring
up
and
I've
seen
these
use
cases
as
well,
like
a
lot
largely
around
like
anything
regulatory
privacy.
F
Whenever
those
concerns
are
risen
or
even
security,
then
data
needs
to
go
and
stay
local
and
be
be
persisted
local,
but
maybe
we
can
go
and
frame
it
as
another
category,
as
you
say,
because
what
we
want
to
go
and
continue
to
have
is
this
separation
of
of
compute
and
data
so
that
we
can
scale
compute
in
two
Dimensions
now
one
is
like
within
an
edge
location
and
even
across
the
many
Edge
locations,
and
we
would
lose
the
ability
of
disposability
if
we
make
the
application
itself
persist
data.
F
But
if
we
we
separate
the
code
from
the
data
and
say
well,
we
would
assume
that
there
is
indeed
an
instance
of
something
that
can
persist.
Storage
at
the
edge
which
is
separated
from
the
application
that
is,
is
and
and
we're
basically
talking
to
application,
design
principles
here
and
the
the
design
behaviors
we're.
A
So
on
this
screen
share
I'm.
Currently,
looking
at
the
the
section
is
entitled
data
storage,
you
maybe
could
entitle
that
cloud
backed
data
storage
and
it's
an
optional
feature
that
you
need.
Maybe
you
usually
use
it,
but
don't
declaratively
state
that
that's
all
there
is
and
that's
all
you're
allowed
to
use.
F
Now
we
need
that
second
category
and
we
we
should
go
and
reflect
that
I.
Think
that's
a
really
good
point.
Yeah.
F
I
think
the
key
thing
is
the
separation
piece
so
that
it
can
go
and
continue
to
fail
processes
that
I
can
even
fail
entire
clusters
whatever
and
that
I
persist
the
data
somewhere
else
right
and
if
I
persist
the
data
in
the
cloud
or
if
I
persist
the
data
in
another
instance
somewhere
at
the
edge.
F
That's
perfectly
fine
right,
I
think
the
key
thing
is
the
separation
and
that
we
can
keep
that
separation,
even
if
we
go
all
the
way
to
the
edge
rather
than
break
the
data
into
the
into
the
application
again,
which
is
like
not
a
good
idea.
I
think.
A
Be
another,
and
that
could
either
be
two
well
probably
more
likely
three,
but
it
could
be
two
just
a
mirror
at
Edge,
where
two
edges
in
the
same
call
it
political
domain
or
something
for
data
privacy
restrictions
could
back
one
another
up
and
it's
better
than
no
backup
right
and
okay,
I
I
like
that.
But
I
think
maybe
this.
This
has
room
for
improvements.
A
And
then
the
other
thing
that
maybe
you
call
out
is
as
I
read
this
this
strike.
It
struck
me
that
this
was
more
going
down
the
path
of
it
being
kind
of
App
application,
layer,
time,
series,
Centric
kind
of
data,
but
there's
another
category
of
data
which
I
would
call
kind
of
the
control
plane,
the
desired
State
configuration
and
that
kind
of
thing
those
are
a
form
of
data
storage.
A
You
know
that
are
aligned
with
the
control
plane
more
than
the
application
plane,
but
they
still
have
the
same
thing
where,
presumably,
even
if
you're,
using
something
like
a
get
Ops.
The
get
repository
in
the
cloud
is
the
authoritative
single
point
of
truth
and
the
stuff.
You
might
still
even
have
a
small
git
repository
down
at
the
edge
node.
A
That
is
effectively
a
cache
that
is
getting
populated
from
the
cloud,
but
would
calling
out
data
storage
use
cases
for
the
control
plane
itself,
as
opposed
to
the
applications,
be
a
useful
concept
because
I
think
they
might
use
you
know.
They'll
they'll
have
different
governance
models
that
would
be
desirable
and
maybe
even
different
places
and
ways
they're
stored.
You
know
so
the
control
plane
might
store
and
get,
but
for
application.
Level
data
like
frame
by
frame
video
or
something
that's
unlikely-
to
be
appropriate
for
storage
in
a
tool
like
that.
A
H
Right
I
mean
a
system
of
record
or
source
of
Truth,
as
you
put
it
like,
GitHub
or
NCD
depends
on
which
you
know
you're
choosing
or
maybe
you
use
both
is
different
than
is.
You
know
categorically
different
than
application
storage
right.
So
it
may
be
useful
to
think
about
the
not
only
the
mode
right,
the
sovereignty
versus
non-sovereign
data,
but
it
also
is
useful
to
I
think
think
about
the
categorization
of
the
data
type
as
well,
and
so
I
think
it's
all
useful
here.
It's
good
color.
I
Up
yeah
I
I
was
wondering-
maybe
all
this
paper
is
this
about
far
Edge
or
because
there,
when
we
talk
about
Edge,
they're,
my
understanding
their
different
categories
of
the
edge
like,
for
example,
if
you
have
a
training
facility
in
the
football
field,
you
have
an
application,
that's
provided
by
a
remote
cloud
provider
right
that
still
Edge.
I
However,
there's
a
local
database,
that's
persistent,
so
and
yeah,
and
similarly
there
there's
the
the
the
the
the
original
Edge
paper
about
Edge
native
versus
Cloud
native,
saying
that
the
the
ad
native
is
primarily
monolith
rather
than
microservices.
I
That
does
it
apply
to
all
edge
cases,
or
is
that
just
for
specific
I
I
know,
for
example,
for
webassembly
applications.
That
might
be
true
at
this
moment,
but
it
is
true
in
all
cases
that
in
edge
cases
is
not
micro
services.
H
I,
don't
think
we
double
click
on
them
on
the
monolith
as
as
an
edge
here
in
this
particular
paper
but
I.
My
answer
to
you
is
what
you
know
the
old
famous
saying
it
depends
right.
I
mean
I,
don't
know
that
we
necessarily
prescribe
any
any
particular
architecture
type
for
that
for
what
should
be
on
the
edge
and
I
think
you're
right,
I
think
it
could
be
in
many
cases.
H
You
know
you
see
like
k3s
or
you
see
microchip
right
down
there
they're
very
happy
to
run
in
microservices
rather
than
a
monolith
or
or
or
run
a
monolith.
So
I
don't
think
that
we're
in
any
position
here
in
this
paper
and
I,
don't
know
if
we've
already,
if
we've
done
it
at
all
in
the
paper,
I
can't
recall,
maybe
Victor.
You
know
that
we've
prescribed
it
being
monolith
versus
microservices.
If
that's
your
question.
F
So
how
do
we
want
to
go
and
continue
because
I
think
many
of
us
probably
have
another
call,
so
we
go
and
continue
the
discussion
Steve
next
time.
A
You're
welcome
to
do
it.
I,
don't
think,
there's
anything
on
the
agenda
for
a
meeting
in
two
weeks
and
I.
Think
you
don't
want
to
necessarily
you
know.
Kick
it
too
far,
Downstream
you,
we
said
we'd
make
a
declaration
of
how
long
you're
open
to
comments
too.
So
we
haven't
heard
that
yet
either,
but.
F
F
Time
once
everybody's
been
through
the
text
and
then
we
understand
all
the
comments
and
then
we
we
can
go
and
update
from
there
I
think
it
would
be
good
so
that
that
everybody
had
a
real
scan
through
and
right
and
then
we
can
go
and.
J
F
A
J
H
F
Key
thing
is
like
whether
you
can
treat
the
edge
as
an
extension
of
the
cloud
where
you
don't
need
to
go
and
change
anything
from
a
from
a
principal's
perspective,
then
well.
Well,
that
could
be
like
it's
not
really.
A
verification.
I
think
it's
the
the
qualities
of
that
edge.
If
that
edge
is
like
suffering
from
limited
low.
F
Whatever
connectivity
from
specific
policy
constraints
from
specific
security
constraints,
then
I
think
we
need
to
go
on
do
something
different
than
just
treat
it
as
an
extension
of
the
cloud
and
I
think
this
is
where
this
applies
so
it
it
doesn't
necessarily
say.
Well,
it's
only
far
Edge.
It's
only
near
Edge,
it's
only
smart
device
Edge.
It's
only.
We
can
go
and
well
expand
that
further
in
next
next
Wednesday
no
week
after
next
Wednesday
in.
F
A
Well,
thanks
I
think
this
was
valuable
and
we'll
call
this
to
it
close,
bye,
everybody
and
see
you
in
a
couple
weeks.
Thanks.