►
From YouTube: K8s IoT Edge working group meeting - 03/25/2020
Description
09:00 - EdgeNet project presentation - https://github.com/EdgeNet-Project
39:00 - OAM/Dapr projects presentation - http://oam.dev, https://dapr.io/
A
B
Sure
I'll
start,
my
name
is
Timur
Friedman
I'm,
connecting
from
Paris
I'm
on
the
faculty
at
Sorbonne
University
and
with
me
on
this
call,
as
well
as
Barrett's
NL.
Who
is
an
engineer
on
my
team
and
we're
going
to
be
presenting
today
edge
net,
a
platform
that
we've
deployed
and
we
look
forward
to
working
with
you.
Thank
you
for
having
us
I.
C
E
This
is
Brian
Gartner
and
I
attend
these
every
once
in
a
while,
not
very
regularly,
but
I
do
follow
the
mailing
list
and
everything,
but
I
did
want
to
learn
about
it.
Today.
That's
two
son
stuff
I
do
work
on
edge
kinds
of
things
I
play
with
kubernetes
all
the
time
and
in
fact
that's
why
there's
no
webcam
my
laptop's
running
in
kubernetes
cluster
I,
don't
want
to
mess
with
that
right
now
and
I
load,
my
other
webcam
to
my
son
for
his
or
her
own
access.
So.
F
G
H
Hi,
my
name
is
Jill
Tompkins
I'm,
the
creator
of
an
open-source
project
called
eclipse,
io
fog
Sunday,
the
Eclipse
Foundation,
has
the
name
on
now.
A
bunch
of
people
work
on
that
most
of
the
co-founder
of
a
company
called
edge
works
I'm
a
regular
on
this
group,
although
the
last
probably
four
or
five
meetings
I've
been,
these
are
barely
on
or
hasn't
been
able
to
make
it
due
to
all
this
global
turmoil,
but
on
yep
I
think
that's
about
it.
I'm
interested
in
all
things
edge
and
we
can
go
very
technical.
D
I
Hi,
my
name
is
Gary.
I
was
also
following
this
receptivity,
like
I
know,
one
year
ago
and
now
I'm
checking
back
how
things
are
working
out.
I'm,
mostly
involved
in
the
edge
computing
group
of
of
the
open
tech
foundation
and
at
this
time,
I
think
we
should
start
to
do
some
collaboration
in
terms
of
figuring
out.
I
What
are
the
possible
test
scan
areas
for
each
cloud
infrastructures,
and
maybe
we
could
somehow
standardize
this,
so
we
can
have
like
a
set
of
properties
what
you
would
like
to
measure
or
infrastructure
against,
and
we
have
a
standards
that
standardized
way
to
do
that.
Let's
see
how
far
we
can
get.
Of
course,
this
is
not
my
target
for
today.
Don't
be
scared.
A
D
L
D
Eduardo
just
so,
you
know
from
time
to
time
people
have
actually
voiced
and
asked
for
a
presentation
on
time.
Series
data
for
people
who
don't
know
Eduardo
is
behind
fluent
D
for
a
long
time
and
a
frequent
speaker
at
conferences.
So
if
you'd
like
to
put
yourself
in
for
an
agenda
slot
in
this
working
group
at
a
future
meeting,
please
go
for
it
because
I
think
there's
a
lot
of
interest
in
it.
Maybe
you
could
give
a
presentation
on
how
the
work
you
do
maps
into
kubernetes
and
IOT
edge.
Oh.
L
Thank
you
yeah.
That
would
be
great
yeah.
Definitely
I
will
do
that,
and
also
we'll
be
releasing
guys
if
you
can
join
the
cloud
native
summit
that
we're
going
to
happen
next.
Monday
I'm,
not
sure
if
you
are
aware
about
that,
like
independent
group
of
different
companies,
we
are
in
right
now
online
sunny
where
we're
going
to
provide
some
quick
updates.
Oh
that.
D
L
D
Okay,
let's
get
started,
I
think
the
first
thing
on
the
agenda
was
edge
net,
so
I'll.
Let
you
guys
go
ahead
and
kick
off.
I
think
the
on
you're,
probably
the
moderator.
So
maybe
if
they
have
a
presentation,
you
can
right-click
on
their
name
and
give
them
authority
to
present
okay
we're
good.
We
see
that.
B
So,
thank
you
so
much
for
having
us,
as
I
mentioned,
I'm,
Timur,
Friedman
and
also
in
the
skull
is
Barrett
Sanel
we're
at
Sorbonne
University.
In
Paris
we
represent
a
consortium
of
people
from
a
large
number
of
institutions
who
created
edge
net
as
it
is
today.
It's
an
effort.
That's
been
ongoing
for
a
couple
of
years,
so
we're
the
latest
in
a
long
line
of
global
networking
research
infrastructures.
We
are
maybe
a
little
bit
oddballs
in
this
community
because
we
do
come
from
academia
and
from
the
research
world,
rather
than
mostly
from
industry.
B
There
have
been
a
series
of
research
infrastructure,
some
of
quite
considerable
size
that
have
been
built
over
the
years.
Many
of
you
may
have
heard
of
planetlab,
which
is
one
of
the
best-known
the
Seattle
testbed
run
by
Justin
Kappas,
and
why
you
tend
in
school.
He
works
with
us.
Janey
was
a
very
big
NSF
funded
infrastructure.
It's
still
operating
fed
for
fire
here
in
Europe,
also
very
big,
multi-million
dollar
infrastructure
for
research,
networking
networking
savvy.
B
So
these
infrastructures
have
in
common
that
they
are
highly
distributed
around
the
world
and
they
are
open
to
public
bona
fide
researchers
who
want
to
do
research
on
the
public
Internet.
So
it
might
be
Network
measurements.
They
might
be.
People
experimenting
with
new
content
distribution
networks.
They
might
be
people
looking
at
how
web
content
looks
different
from
different
vantage
points.
B
B
We
have
dozens
and
planetlab.
We
had
hundreds,
we
run
planetlab
Europe
at
Sorbonne
University.
We
have
over
200
institutions
signed
up
each
of
them
providing
rack
servers,
so
each
node
is
under
the
administrative
control
of
a
different
individual
or
institution,
and
then
we
have
a
hundreds
of
researchers.
If
you
count
students
who
access
the
system,
thousands
of
people
and
again
these
are
from
many
different
institutions
and
all
of
them
accessing
separately
under
the
authority
of
their
individual
institution
and
for
extranet
right.
B
Forgive
me
for
the
use
of
the
term.
We
hope
to
have
a
viral
spread
of
our
infrastructure
with
previous
infrastructures.
You
were
talking
about
people
installing
physical
boxes,
rack
servers
for
planetlab
people
had
to
invest
about
four
thousand
dollars
u.s.
dollars
to
buy
servers
now
with
darker
and
kubernetes
I.
They
just
create
a
virtual
machine
and
in
just
a
matter
of
minutes
we
can
install
an
adjutant
node
on
a
virtual
machine.
B
So
it's
very
simple,
very
quick
and
then
the
tremendous
potential
that
we
see
for
kubernetes
is
that
it's
so
widely
used
so
well-known
doctor
as
well,
and
so
people
who
know
how
to
use
those
technologies
know
how
to
use
our
tests
that
they
just
use,
cube,
control
or
the
kubernetes
dashboard.
They
use
docker
containers
that
they
may
have
already
created
in
their
lab
for
for
testing
purposes.
B
B
B
B
Show
you
what
we're
doing,
how
we're
doing
it
Barrett
will
take
it
from
here
and
talking
about
that
and
we're
eager
to
enter
into
contact
with
others
in
the
community
who
have
similar
issues
and
we'd
like
to
combine
our
efforts,
share,
return
on
experience
code
and
so
on.
We
already
have
monthly
meetings
with
people
at
Google
in
New
York,
the
measurement
lab
group
there.
It
has
a
similar
architecture,
and
so
we
have
monthly
return
on
experience
meetings
and
we're
hoping
to
broaden
that
circle.
So
I'll
leave
it
to
Barrett.
F
This
is
the
king
I'm
going
to
talk
about
what
and
how
we
have
what
we
want
to
have
in
the
next
step,
so
high-level
overview.
We
want
to
contribute
back
to
the
cavernous
community
and
to
do
so
our
code
could
be
compatible
with
kubernetes.
Therefore,
the
language
that
we
used
in
the
first
version
of
H
not
to
go
language
so
far.
We
use
custom
resource
definitions
that
extent
the
commenter's
api
and
we
focused
on
two
principal
terms:
bringing
kubernetes
into
the
edge
contending
and
enabling
multi-tenancy
kubernetes.
Let's
have
a
look
at
how
CR
divorce?
F
What
is
our
architecture
for
now
in
Asianet,
we
use
the
custom
source
definition
because
it
uses
the
main
cover
to
safety,
so
hungry
customer
sources
in
the
same
way
with
Cory
sources.
Like
thought
and
be
myself,
as
you
see
in
Figure
two
new
features
we
have
plant
of
CR
bees
in
use,
such
as
selective
deployments,
authorities,
users
and
so
on.
I
wanted
to
talk
about
the
CRVs,
because
I
want
to
discuss
the
problems
we
encountered
after
the
first
presentation
to
principal
dance
in
the
first
place
as
dementia
mentioned.
F
In
our
scenario,
therefore,
we
just
mean
club-like
capabilities
located
at
the
infrastructure
edge.
So
we
we
have
established
a
campus
each
cluster
as
nodes
are
located
in
distributed
locations
by
getting
the
benefit
of
a
feature
that
makes
it
easy
for
third
parties
to
add
notes
in
clustering
to
burners.
F
This
feature
allows
users
to
do
so,
based
on
continents,
countries,
states
and
regions
and
cities,
as
well
as
polygons
described.
The
using
latitudes
and
longitudes
I
also
mentioned
about
our
tennis
aspect
quickly.
So
we
develop
a
new
tennis
model
lineage
that,
by
inheriting
the
concepts
mainly
from
plant
lab
site
is
a
physical
location,
but
lab
notes
are
located,
various
and
ultimately
image.
That
is
not
required
to
provide
note,
but
they
can
take
both
own
users
and
slices.
F
Additionally,
a
tortoise
village
that
have
an
added
feature
called
team
to
group
users
to
keep
the
motorisation
great
slices
independently
done.
Authority
admins
I
want
to
make
sure
that
authorizations
to
usage
that
are
handed
out
theoretically
so
that
authority
admins
approve
local
users
who
they
know
the
central
admins
of
as
not
approve
the
establishment
of
authorities
and
each
authority
having
its
own
local
admin
and
admin
welcome
the
users
to
their
authorities
in
age
nine.
We
also
use
CRD
for
e-mail
verification
during
the
authority,
registration
or
user
registration.
F
Let's
focus
on
current
is
at
the
edge
I,
even
explain
how
much
networks
withdraw
beings
at
following
slides
today,
kubernetes
cluster
consists
of
nodes
that
are
all
grouped
together
in
centralized
data,
centers
and
managed
by
one
user
group
in
general.
Therefore,
it's
not
so
difficult
to
add
notes
into
the
cluster,
so
users
deploy
their
applications
across
the
cluster
by
using
keep
city
I'll
count
line,
interface
to
or
dashboard
in
Asianet.
We
want
users
to
add
their
notes
into
the
edge
net
cluster
because
of
peers.
We
made
it
easy
to
add
nodes
into
the
cluster.
F
Additionally,
although
we
just
used
geni
nodes
are
sort
of
ble
knots.
In
this
way,
each
contribution
of
users,
as
you
see
in
the
slice,
help
us
to
build
a
much
more
widely
distributed
edge
cloud
by
the
day,
thanks
to
Jeannie
and
ple,
we
is
established
at
each
cluster
so
far
with
the
notes
that
provide
vantage
points
to
users
to
deploy
their
applications.
For
now,
we
aim
to
reach
the
limit
of
5,000,
not
just
to
say
it
in
our
cluster
and
when
we
get
there,
we
might
consider
participating
in
a
federation
studies
as
well.
F
F
The
here
we
go
in
this
case
beside
the
class
redmond,
our
authority
admins,
who
have
particular
namespaces
and
users
in
the
authorities
technically
in
the
authority
namespaces,
but
I
mentioned
before
daughter.
The
namespaces
are
another
broad
space
which
users
collaborate
with.
It
means
it
is
not
a
lot
to
deploy
applications
in
an
opportunity
space
thanks
to
the
resource
cutter.
F
There
are
two
solutions
that
auto
tiredness
can
create
teams
by
choosing
users
made
from
different
authorities.
This
creates
a
team
in
space,
but
still
it's
not
able
to
deploy
applications
to
and
the
part
particles
of
the
team
can
create
near
slices
fight
using
users
to
deploy
their
applications
across
the
cluster.
So
this
great
says
slice
namespace
its
controller.
They
can't
matter.
This
is
just
a
shortcut
which
is
creating
a
slice
directly
in
dr.
Kim's
face.
F
F
Arnaud
is
ready
to
use
in
this
outer
version.
We
also
have
another
feature
if
the
knot
gets
crushed.
It
also
starts
a
recovery
procedure
that
contains
clean
installation
as
well
to
recover
not
from
that
crash
situation
occur.
Tears
also
can
enable
or
disable
scheduling
or
not
that
they
contribute
to
the
asian
cluster.
I
am
fading,
not
contribution
object.
They
tried
it
before.
F
So
after
this
point,
I
talk
about
location
based,
not
selection,
they
will
be
of
distributed
systems.
Is
it
offers
us
vantage
points
and
allows
us
to
deploy
application
selectively
according
to
locations
edge
net?
Has
that
feature
hereby
I
will
try
to
explain
how
that
works.
There
is
a
custom
controller
without
CRT
that
just
watch
the
notes.
F
So
this
custom
controller
includes
an
even
handler
for
the
even
types
of
add,
update
and
plate
as
usual,
to
get
inform
about
them
and
then
ground
related
functions,
so
that
generates
label
contents
and
could
on
if
even
there's,
a
big
editing
or
adding
during
the
creation
of
peers
labels.
The
controller
looks
for
the
IP
address
of
the
node.
Even
if
it
exists,
queries
makes
minds
Jo
like
to
database
to
get
geolocation
information
according
to
that
IP
address
and
then
assigns
the
data
to
the
labels.
F
As
you
seen,
the
slides,
not
labeler
label
the
nodes
in
various
levels
corresponding
to
their
city,
state
region,
country
continent
and
a
little
bit
and
longer
that
to
be
consumed
by
selective
deployment
service.
But
since
Qantas
has
alphanumeric
character,
restrictions
whitespace
in
the
builders
are
replaced
by
underscores
and
the
directional
aberrations
are
added
to
the
values
of
negative.
F
F
Paris
at
this
point,
selective
deployment
CRV
joins
the
game,
one
because
we
just
want
to
deploy
our
applications
in
Berlin.
San,
Francisco
and
Paris
mean
an
edge
cloud
scenario.
Users
want
to
select
notes
at
particular
locations
for
the
airports
by
defining
its
a
cluster,
for
instance,
a
be
Merced
defined
as
having
five
notes
in
Germany
and
poor
notes
in
France
will
ensure
that
their
ports
are
deployed
to
nine
notes
that
meet
those
criteria.
F
F
Select
deployment
of
object,
configures
team
assets
deployments
are
stateful
sets
that
are
the
types
of
controllers
to
deploy
a
brioches
across
the
cluster
in
Caracas
Noah
days.
Instead
of
creating
a
replica
set
priority,
users
are
more
likely
to
create
deployment
that
creates
and
manipulates.
Replicas
starts
selecting
deployment.
Cr-V
is
inspired
by
this
method
tests
and
it
employee
must
can
control
demon
sets
deployments
as
stateful
cells
to
deploy
an
application
based
on
users
desire
geolocations.
F
But
how
does
selecting
point
configure
dos
the
controller
picks
up
the
nodes
according
to
the
selector
defined
by
the
user
and
configures,
the
not
affinities
of
controllers,
not
affinity
allows
us
to
constrain
each.
Not
the
port
is
eligible
to
scheduled
on
base
on
Labour's
on
the
North.
Here
it
is.
We
succeeded
in
a
deployment
based
on
geolocations
and
the
next
versions
of
selective
deployments,
so
it
is
not
working
like
a
demon
set,
but
in
the
next
versions
of
selective
deployment
will
also
include
a
custom
schedule
as
well
to
provide
a
feature
like
deployment.
F
Software
kubernetes
I
want
to
point
out
certain
improvements
that
we
want
to
have.
We
want
to
allow
our
providers
to
open
their
not
suit
or
notes
to
authorities
team
and
flies
as
they
desire
in
this
way
as
an
individual
user.
For
example,
me
if
I
contribute
to
the
regional
cluster,
I
can
open
my
not
only
sorbonne
universe,
so
the
Meo
turkeys
accept
so
I
can
feel
myself
to
say
if
I
want,
they
don't
really
plan
to
get
the
benefits
of
the
admission
controller.
D
Thank
you.
We
do
have
a
second
presentation,
but
maybe
we
have
room
for
just
one
or
two
questions
before
we
turn
it
over
to
them.
Microsoft
people
or
look.
We
can
put
questions
on
the
edge
IOT
slack
on
kubernetes
and
if
you
have
any
questions
for
us,
maybe
you
can
put
them
or
other
group
members
I.
Think
the
slack
channel
for
the
IOT
edge
group
would
be
a
good
place
to
put
those.
D
C
C
C
It
has
great
value
when
you
have
a
lower
granularity,
probably
down
to
the
city
level,
but
when
you
work
with
like
cell
towers
directly,
is
that
a
the
the
cell
tower
coverage
is
not
like
subdivided
perfectly
a
geolocation,
but
in
the
same
latitude
and
longitude
there
could
be
multiple
cell
terrorists.
The
truth
from
so
in
that
case,
selecting
knows
based
on
geolocation
is
not
necessarily
reliable,
and
some
thinking
is
that
maybe
we
have
to
her
practically
collect
Network,
telemetry
and
based
on
historical
predictive
latency
to
choose
different
nodes.
C
Understand
clearly,
but
the
first
question
is
that
I
think
the
assumption
here
is
that
the
newly
joined
worker
knows
will
talk
to
the
master
knows
over
in
tonight,
right
yeah,
but
already
in
tonight
we
have
kind
of
unpredictable
agencies
and
between
the
nose
I
think's,
when
communities
must
not
check
the
loudness
of
the
worker.
Note,
if
you
have
variant
latencies,
there
could
be
false
negatives,
about
nose
becoming
available.
F
Okay,
I
see
your
point,
so
we
are,
we
are
in
the
very
beginning
of
the
development,
so
we
now
start
creating
procedures.
Any
kind
of
experience
like
this
you
pointed
out
would
be
welcome
because
to
create
a
better
procedures
to
recover
the
notes
and
automatically
and
also
to
create
a
service,
migration
and
location-based,
not
selection,
so
it
is
important.
I
think
yeah.
D
Some
projects
have
gone
with
solutions
or
workarounds,
like
virtual
cubelets,
to
convince
the
control
plane
that
there
still
is
a
presence
related
to
that
node
as
a
facade
but
yeah
I.
Think
a
lot
of
people
have
encountered
problems
with
latency
connections,
dropping
off
and
coming
back
as
well
as
some
people
concerned
with
security
issues.
With
that
long
communication
chain
going
from
a
cloud
hosted,
control
plane
out
to
edge
nodes
may
be
over
transports
that
aren't
fully
trustworthy.
But
that's
a
big
topic
that
I
don't
think
we're
going
to
cover
for
this.
D
M
M
F
Hi
Kareem,
so
we
established
as
a
cluster
we're
nearly
and
we
started
the
trade,
our
own
organized
sale
and
monitoring
tools
from
the
greeting
from
the
plan
table
and
I
think
we
well.
We
will
have
this
kind
of
measurement
data
three
or
four
months
later,
when
we
have
this
kind
of
data,
I
would
be,
cannot
share
it
with
the
community.
So
if
we
can
keep
in
touch
in
the
slack
Channel,
when
we
have
this
kind
of
data,
I
would
be
more
than
happy
to
share
it
with
you.
A
K
C
Understand
I
go
fast.
Can
you
see
my
screen?
Yes,
we
can
see
it
yeah.
My
network
naturally
is
kind
of
spotty.
My
wife
had
this
great
idea
to
buy
a
house
in
the
forests.
That's
the
instructor
here
is
really
in
stable,
okay.
So
today
my
topic
is
to
introduce
ohm
and
pepper.
Om
is
the
open
application
model
and
there
is
the
distributed
application
run
time.
So
we
designed
this
project
for
cloud
native
applications
and
our
first
implementations
are
highly
tailored.
C
It
was
communities
and
when
we
create
designs
these
models,
we
had
both
cloud
and
edge
in
mind
and
criminals
being
pervasive
across
cloud
and
edge
actually
helps
the
project
to
be
applicable
to
both
cloud
and
edge
as
well.
So
I
thought
I
was
knew.
I
was
going
to
introduce
myself
I'll
just
skip
all
that,
and
I
will
skips
from
the
mental
discussions
about
why
we
made
the
project.
D
D
C
I
would
just
go
right
into
the
opening
of
an
application
model.
So
basically
the
idea
here
is
a
simple
we've
seen
a
lot
of
developers
working
on
kubernetes,
but
in
our
point
of
view,
chrominance
is
really
very
instructure
heavy
platform
and
we
really
need
a
good
modeling
language
for
developers
to
model
their
applications
instead
of
using
profanities
primitives.
C
But
the
idea
of
open
application
model
is
to
create
this
platform:
agnostic,
modeling
language,
for
a
macro
service
application.
Essentially,
an
application
is
made
up
by
a
bunch
of
components
and
you
can
define
scopes
or
boundaries
across
arbitrary
components.
For
instance,
you
have
a
network
segment
and
if
you
have
a
security
boundary
you
can
define
different
scopes
and
you
can
apply
different
trees
to
different
components.
For
instance,
if
you
have
a
component
that
is
auto
scalable,
you
can
attach
an
auto
scaling
trait
to
to
say:
hey.
C
This
component
can
be
auto
scaled
between,
let's
say
ten
instance,
instances
and
twenty
instances.
So
the
idea
is
to
provide
the
developer
this
modern
language
so
that
he
can
describe
his
application
in
entirety
and
then
hand
it
over
to
operations.
Then
the
operations
people
take
the
manifest
can
make
different
decisions.
For
instance,
they
will
see
the
application
requires
a
ingress
or
load
balancer.
Then
operation
and
people
can
decide
what
kind
of
load
balancer
and
what
kind
of
ingress
she
will
put
in
place
for
that
particular
deployment
and
same
like
auto
scaling.
C
What
kind
of
auto
scaling
mechanism
the
operational
month
we
use.
So
those
are
separation
of
concerns
that
developers
describe
the
intention
and
operation
people
make
that
a
reality
on
different
platforms,
and
there
are.
There
were
actually
a
numerous
attempts
of
creating
the
application
model
before
and
why
we
think
we
we
could
do
it,
because
there
were
several
reasons.
The
first
is
that
with
the
recent
not
recent,
you
know
last
few
years,
the
development
of
containers
and
the
service
measures.
We
have
the
necessary
abstraction
language
to
create
such
a
model.
C
Now,
before
we
didn't
have
that
now
we
can
actually
model
each
component
as
a
black
box,
which
is
a
container
and
they
are
somehow
interconnected
by
a
service
mesh.
So
we
feel
we
have
the
necessary
abstractions
nowadays
to
actually
come
up
with
this
common
model
and
with
the
popularity
of
kubernetes
growing,
and
we
also
hear
the
increasing
voice
and
complaints
about
crew
Benitez
being
too
complex,
not
necessarily
developer
friendly.
C
So
there's
a
need
to
separate
the
preventive
complicity
away
from
the
developers
and
that's
another
motivation
for
us
to
do
that,
and
also
the
third
one
is
that
when
we
see
a
lot
of
addcom
computing
scenario
arise,
people
are
thinking
wow.
If
my
application,
if
I
can't
anomaly
deploy
to
different
environments,
maybe
I
want
to
have
a
cloud
ear.
A
deployment
maybe
I
want
to
deploy
to
any
seas
when
it's
available
or
in
the
disconnector
mode.
How
I
can
make
some
of
the
components
available
on
edge
to
provide
base
discontinuity
without
internet
connection?
C
An
idea
of
home
is
to
allow
people
to
describe
an
application
regardless
how
the
applications
hosted.
Then
it's
up
to
the
operation,
people
how
to
adjust
the
deployments
based
on
the
particular
environments
and
so
I
will
skip
the
architecture
details
for
now.
So
Casey
the
the
own
project
is
in
development
of
Microsoft
and
Alibaba,
and
also
some
other
partners
is
up
there
in
open.
C
Everybody
can
contribute
we're
currently
an
alpha.
It's
still
early
stage
and
we
are
making
alpha
3
as
we
speak.
But
idea
is
to
have
this
back
that
got
accepted
by
different
people
and
then
there
could
be
different
government
implements
on
different
structures,
for
instance,
where
we
are
building
our
members
of
a
dreary.
Implementation
Alibaba
has
their
Alibaba
specific
implementations
and
our
partners
are
treating
their
implementations,
and
you
can
imagine
in
the
in
the
future,
people
can
create,
like
a
genetic
implementations,
just
to
manage
low-power
devices.
C
So
that's
the
idea.
We
are
still
pretty
far
from
that,
but
I
think
that
that's
some
direction
we'd
like
to
try
on
so
then
there
are
some
details
about
how
the
each
individual
component
works
in
the
model.
I
think
I
will
skip
those
for
now.
They
again
they're
very
straightforward.
There's
nothing
fancy
about
here.
They
should
be
very
intuitive
concepts,
so
I
will
just
skip
those
two
now
and
I'd
like
to
maybe
not
mention
the
future
either,
but
I
will
switch
the
depper,
which
is
actually
more
interesting
in
the
application
model
itself.
C
So
the
deborah
project
was
actually
raised
by
young
Snyder,
one
of
my
colleagues
one
day
when,
when
we
were
discussing
home,
you
know
in
the
phone
room-
and
you
just
mentioned.
Maybe
we
can
go
a
step
further
because
from
the
perspective
of
home
each
of
the
component
innovative
services
is
a
black
box
right
it
it's
encapsulated
by
a
container.
We
don't
really
care
how
that
service
is
written.
He
was
saying.
C
Maybe
we
could
go
step
further
to
her
to
provide
a
programming
model
for
useful
to
write
those
services
at
that
time,
I
was
I
was
I
had
a
lot
of
thoughts
about
that,
because
there
was
again
numerous
attempts
of
creating
a
common
programming
model
for
distributed
applications,
but
the
success
was
quite
limited
and
but
then
he
explained
his
idea
was
to
have
this
sacral
architecture.
Basically,
we
chose
a
very
non
intrusive
way.
C
We
will
provide
common
building
blocks
for
common
distributed
application
like
servicing
location
and
state
management,
pop
zombie
cetera
as
a
same
car,
and
the
application
code
can
just
consume
those
capabilities
through
a
standard
interface
like
HTTP
or
gr
PC.
There
are
several
benefits
of
doing
this
way.
The
first
is
that
the
application,
the
application
code,
doesn't
need
to
include
any
libraries.
It
doesn't
need
to
include
any
type
or
libraries,
because
everything
is
a
HTTP
or
a
PC
endpoint.
It's
almost
like
a
service-oriented
architecture
on
the
hatch.
C
The
second
benefit
is
that
this
allows
us
to
support
any
programming.
Language
I
mean
not
necessarily
any
programming
language.
We
try
to
get
it
work
for
kobu.
Where
do
we
got
it
work,
but
we
will
feel
kind
of
cheating
a
little
bit.
We
use
a
HTTP
library
for
Cobo,
it's
not
native,
but
the
concept
regardless
is
that
if
a
modern
language
can
support
HTTP
or
G
RPC,
then
it
can
consume
pepper.
C
You
can
just
use
pepper,
HTTP,
endpoint
to
say:
hey,
save
a
state,
missus
key
will
read
the
state
with
this
key
is
as
simple
as
that
and
then
I
start
to
believe.
Maybe
the
set
car
architecture
of
you
really
take
off.
So
we
started
the
beaut
under
codebase
and
they
called
released
a
couple
months
back.
It
was
really
well
received
by
the
community.
C
We
reached
five
thousand
stars
under
like
two
months
or
something.
If
you
look
at
the
the
github
star
chart,
our
project
is
kind
of
a
unicorn.
If
this
shoot
up
straight,
it's
different
from
any
other
open
source
projects
out
there.
So
I
think
we
got
something
there
because
something
meaningful
that
we
should
keep
invest
in
investigating.
C
So
that's
the
the
sacral
architecture
and
pepper
from
day
one
is
designed
to
to
work
as
a
separate
process
instead
of
container
so
the
Debra's
off.
At
the
moment.
The
binary
is
about
fifty
eight
megabytes
and
the
wrong
time.
Memory
is
about
four
megabytes,
so
it's
really
designed
to
run
on,
like
a
lower
power,
add
devices
without
containers,
whatever
works
great
on
kubernetes,
it
works
great
in
containers,
but
it
can
also
work
as
a
separate
process
that
you
can
attach
to
a
main
process
as
a
sidecar
and.
C
The
slice
Kara
is
just
showing
different
deployment
options
of
typer,
as
I
said,
it
runs
as
a
loose
process
or
you
can
use
a
pepper
pot
injector
to
inject
dye
per
set
car
into
any
pause
and
provide
the
pepper
API,
and
we
also
in
the
community,
is
making
great
contributions.
That
was
we
work
with
SP
tonight's
team
and
Microsoft
and
there's
also
a
lot
of
great
with
Chavez
spring
boot
framework
and
there's
a
Visual
Studio
code
extension.
C
Someone
is
creating
the
Python
SDK,
so
there's
a
lot
of
community
contributions,
and
now
the
temporary
is
able
to
bind
to
most
popular
cloud
services.
You
can
think
of
you
combine
two
ayahs
on
to
Alibaba
to
Google
and
go
to
add/remove
course
and
many
third-party
services
alike
at
Twilio
and
different
different
service
providers
and
the
building
blocks,
as
I
mentioned
just
common
things,
we've
observed
that
distributed
application
developer
would
need
such
as
state
and
pops
up,
but
this
is
a
quite
straightforward.
I'll
just
take
one
as
an
example,
let's
say
the
state
of
management.
C
Basically,
we
use
diaper,
we
define
a
very
simple
API,
so
you
read
and
the
right
key
value
pairs.
So
you
can
post
a
list
of
key
value
pairs
and
you
can
read
a
key,
so
the
test
is
essentially
how
the
typer
API
is
designed.
So
basically
we
look
at
each
individual
problems
that
developers
may
face
and
we
define
very
simple
API
for
developers
to
use
and
the
P
point
of
this
I
want
precise
is
that
there
is
incrementally
adaptive,
meaning
when
you
use
pepper,
you
don't
have
to
use
everything
from
pepper.
C
You
can
you
say
you
only
want
to
use
state
management.
Sure
you
can
just
use
a
state.
Api
temperate
doesn't
force
you
to
take
everything
from
pepper.
This
is
a
rather
different
from
other
distributed
frameworks
in
which
you
have
to
adapt
everything
right.
You
have
to
adapt
the
program
model,
the
framework,
the
way
you
describe
your
code
and
language,
or
rather
Cody
in
so
you
have
to
buy
into
everything,
but
there
is
very
mallow
and
says:
hey
use
whatever
you
want,
and
also
we
don't
use
any
proprietary
structures
or
data
formats
in
the
backend.
C
So
if
you
want
to
access
your
data
store
directly
you're
welcome
to
do
so,
but
we
don't
have
well.
If
you,
you
have
to
access
the
data
stores
directly,
we
rely
on
the
date,
they'd
store
itself
to
provide
transactions,
coping
and
concurrency,
etc.
Of
course,
in
that
case,
you
don't
get
whatever
gives
you
in
terms
of
concurrency
assurance,
but
you
can
still
do
so.
C
That's
a
different
philosophy
from
other
frameworks
is
that
we
give
you
the
tools
you
can
use,
but
if
you
want
to
go
beyond
our
tooling
support,
if
you
want
to
do
more
advanced
cooking,
you
can
also
do
that.
We
don't
want
to
try
to
constrain
you
into
a
small
parts
about
what
the
Deborah
can
does.
I
can
do,
and
similarly
we
have
the
same.
Api
concept
for
servicing
cook,
invocation
and
binding
binding
is
very
powerful.
C
Concept.
Allows
you
to
even
driven
programming.
You
can
bind
to
all
kind
of
human
forces
to
be
triggered
by
different
events,
for
instance
the
asp.net
poor
guys.
What
they
did
is
to
integrate
pepper
into
SP
ionic
core,
then,
all
of
a
sudden,
your
SP
Banette
website
is
triggered
not
only
by
a
browser
traffic.
C
He
was
new
to
go
language
and
he
was
able
to
do
it
on
a
few
hours
and,
of
course,
the
pops
up
that
messaging
based
integration.
We
support
that
and
also
we
have
auto
box
distributed
tracing.
So
Debra
has
this
concept
of
a
middleware,
because
us
is
the
set
car
sitting
your
traffic
flow.
We
can
inject
all
kinds
of
middlewares
into
the
flow
into
the
flow.
Without
you
modifying
your
code,
for
instance,
we
integrated
with
open
sensors
and
in
the
future,
open
telemetry
to
inject
distributed
tracing
automatically
to
our
application.
C
Your
application
doesn't
need
to
do
anything.
Your
application
doesn't
need
to
do
any
annotations
or
incrementing
libraries,
it's
all
configured
and
enabled
by
Tepper,
and
you
can
export
to
all
kinds
of
tracing
systems
to
collect
and
analyze.
Your
your
application-
this
is
the
trellis
version.
I
have
forgot
to
dress
like
that.
I
should
have
deleted
that
and
stepper
also
some
support
actors
and
functions
which
are
two
very
popular
programming
model
in
distributed
systems.
But
the
problem
is
current
functions
or
actor
problem,
although
is
that
they
don't
have
interoperability
among
them.
C
Let's
say
if
you
use
a
SS
lambda
or
you
use
address
functions.
Basically,
you
need
to
buy
into
a
lambda
of
functions
right.
You
can't
say:
I
have
a
function
using
lambda
code
or
West
version,
but
that
protects
our
approaches
that
we
we
want
to
make
actors
and
functions,
probably
model
choices
under
tapper,
so
you
can
use
tapper
to
write
stateless
services,
stateful
services
address
functions
in
any
language
and
they
all
work
together.
You
can
change
your
code
to
say:
hey
I
want
to
configure
my
code
to
act
as
actor
today.
C
Tomorrow,
I
may
want
to
configure
my
code
to
act
as
a
function
or
if
I
have
a
c-sharp
function,
I
want
to
call
a
java-based
function.
You
can
totally
do
that
with
diaper
because
of
this
car
architecture,
and
this
is
the
virtual
actor
pattern.
I,
don't
think
I
need
to
go
into
more
details
about
this.
They
say
we
provide
the
basic,
not
basic.
We
provided
the
virtual,
a
complete
water
actor,
a
pattern.
C
This
is
a
pretty
much
comparable
to
a
service
fabric,
actress
or
a
collector
framework,
but
we
have
a
key
difference
is
that
we
don't
actually
allocate
memories
forward,
I,
actually
access
every
instance
for
us
each
actor
instance,
it's
just
a
routing
rule
in
diaper.
There's
a
lot
last
to
create
a
really
high
density
actress
on
a
single
machine,
but
the
benchmark
weight
data
was
to
launch
a
million
actors
on
a
single
node
cluster.
This
is
really
heightens
day
because
literally
we
don't
have
like
a
memory
footprint
for
actor
right.
C
Only
when
a
dream
is
activated,
we
update
the
rod
to
to
evoke
the
code
after
code.
Then
that
state
is
saved
to
the
external
state
store.
Then
there
is
nothing
capped
in
the
memory
for
the
actor,
so
this
will
also
to
ask
to
reach
really
high
density
of
water
actors
with
diaper,
and
the
other
thing
we
are
supporting
is
the.
Of
course
they
do
functions.
C
C
Nothing
about
mention
is
that,
as
I
said,
for
us
actors
are
not
instances
so
when
they
fail
over,
we
are
not
migrating
any
actor
incidences
across
the
note.
We
are
just
updating
our
routing
rules
and
we
use
an
algorithm
from
Google,
the
the
the
it's
a
flavor
of
consistent
hashing,
so
that,
if
you
add
mal
note
or
some
note
becomes
a
habitable,
the
rod
rules
are
updated
to
to
load
balance
across
the
nose.
But
again,
there's
no
reallocation
happening
behind
the
scenes.
C
D
C
So
the
they
own
itself
right,
it's
a
spec,
it's
abstraction
language
to
describe
the
application.
It
has
no
connection
to
communities
whatsoever,
but
the
control,
plane,
implementation
and
the
control
plane
implantation.
What
we
have
is
based
on
Cremona
fees,
so
in
that
case
we
do
assume.
There's
a
criminal
class
were
running
somewhere
and
the
alibaba
they
did
their
implementation
based
on
their
idioms
culture,
which
is
not
true,
manatees,
okay,.
D
C
D
C
Yeah,
there
are
a
lot
of
well,
they
do
overlap
cases
because,
as
you
said,
they
kind
of
trying
to
address
the
similar
problem
domain.
I.
Think
our
approach,
it's
a
bit
different,
like
I,
said
that
/
aims
to
add
incremental
adoption,
meaning
that's
you
can
just
choose
to
use
some
type
of
functionality
without
Hydrick
buying
the
whole
thing
and
the
by.
D
D
C
D
C
Has
two
modes:
why
is
the
continuum
node
which
we
expect
to
work
under
criminalities?
It
also
runs
as
a
separate
process
intermission.
It
takes
about
four
max
of
memory.
Money
in
runs
as
a
native
process,
so
we're
doing
vision.
You
can
run
diaper
around
lower
quality
vices,
but
in
some
extreme
cases
like
really
low
power
devices,
maybe
the
sidecar
architecture
itself
is
too
heavy.
Probably
you
don't
want
to
run
a
separate
HTTP
server
on
your
edge
devices.
So
for
that
we
are
thinking
in
two
directions.