►
Description
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects - Learn more at https://kubecon.io
Kubernetes Federation for Edge - Vishvesh Vijaywargiy & Madhusudan N, NVIDIA
What does it take to run Kubernetes at the edge? Running applications on edge devices is a fundamental shift from the cloud native world and this talk explores if Kubernetes is ready for it. We will start by exploring the different shades of the edge through the lens of several use cases: retail store, edge antenas, highways, cars and cargo ships. We will also highlight the challenges that applications and infrastructure faces when running at the edge (NAT boundaries, intermittent connectivity, limited compute resources etc). Finally we will present a kubernetes federation solution that can handle the scale of managing thousands of edge devices. An example deep learning application will be used to demonstrate the solution.
https://sched.co/YWIw
A
A
B
For
the
agenda
first,
we
need
to
understand
what
it
scenarios
are
we
looking
at
and
what
are
the
type
of
issues
that
we
would
encounter
and
what
we
are
actually
trying
to
solve.
So
first,
we
will
start
off
with
that,
and
we
will
also
explore
what
are
the
different
kinds
of
workloads
that
is
different
from
what
runs
at
the
edge
and
what
runs
in
a
regular
Kuban
at
a
server
which
is
probably
be
hosting
micro
services.
B
A
B
A
A
You
see
that
it's
a
right
candidate
for
the
real-time
analysis.
Let
me
give
you
an
example.
Let's
say
an
ambulance
come
traffic
signals
are
meant
for
the
traffic
flow
and
if
the
signal
is
red
and
ambulance
is
there,
will
this
traffic
signal
be
able
to
take
the
intelligent
decision?
I
stopped
the
other
direction
traffic
flow
and
let
the
ambulance
go.
B
Well,
let's
look
at
a
different
scenario.
Also,
you
know
you
have
a
ship
over
there
and
especially
when
it's
in
the
sea,
it
doesn't
have
any
network
connectivity
and
it
doesn't
really
host
a
web
application.
Also,
but
what's
really
happening
is
it's
running
AI
workloads,
which
is
looking
at
the
real-time
information
of
whether
the
tides
or
you
know
the
crews
and
giving
real-time
information
to
the
captain
of
the
ship,
and
we
have
to
really
focus
on
this
network
isolation
aspect
as
well,
very.
A
B
Well,
we
had
really
three
problems
or
three
scenarios
that
we
were
discussing
right.
The
real
time
processing
became
a
real
constraint
and
we
had
the
isolated
Network
and
the
security
which
is
really
important,
and
we
also
mentioned
that
the
inference
the
I
inference,
which
is
really
happening.
It's
important
that
it
happens
at
the
edge
as
well
in
the
scenarios
that
we
showed
you.
It
doesn't
happen
somewhere
in
the
cloud.
You
know
you
can't
really
wait
for
that
long.
You
cannot
wait
for
the
network
connectivity
exactly.
A
A
B
We
have
also
seen
use
cases
where
it
had
to
be
single,
node
plus
I,
know
using
the
word
single,
not
in
class
or
don't
go
well
together,
but
let's
assume
that
it's
a
single
node
device
that
you
really
have
and
defining
high
availability
for
this
use
case.
What
does
high
availability
for
this
use
case?
Yeah.
A
You
must
have
seen,
as
the
first
use
case,
that
it's
a
traffic
signal
you
will
not
have
the
luxury
of
setting
up
the
high
available
kubernetes
cluster,
most
of
the
H
compute
for
the
a
use
cases
that
the
data
sources
are
minimalistic
set
up
like
a
single
node
master,
which
is
working
as
a
worker.
Also,
and
then
it
is
doing
the
data
processing,
inferencing,
yeah.
A
Federation
is
not
new,
which
we
bring
here.
It's
it's.
It's
a
well
solve
problem
that
you
have
multiple
kubernetes
clusters
behind
the
federated
cluster.
You
deploy
your
CR
DS
or
CRS
to
the
federated
cluster,
which
takes
care
of
pushing
the
data
to
all
the
kubernetes
cluster
which
are
behind
the
Federation.
Unfortunately,
all
the
use
cases
which
talk
about
does
not
have
luxury
of
all
the
features
like
network
connectivity.
It
has
to
be
secured.
It
has
to
be
always
processing
the
real
at
the
real
time.
A
In
that
scenario,
how
we
will
update
the
date
update
the
workloads
or
raid
our
models
running
on
the
edge
control?
So
here
we
talk
about
this
architecture,
which
is
not
push
based.
It
is
more
of
a
pull
based
architecture.
In
this
we
have
a
federated
cluster.
This
federated
cluster
has
a
namespace
and
workloads
are
running
on
that,
and
this
is
a
namespace
ABC.
A
Then
you
have
the
H
controllers,
these
H
controller
H
clustered,
which
has
multiple
controllers
running
on
that
one
of
the
controller
is
the
Federation
controller,
whose
main
responsibility
is
to
watch
all
the
CRS
or
C
already
submitted
to
the
Federation
controller
as
when,
as
and
when
the
network
reestablishes
it
downloads,
those
c
ARS
and
CRS,
and
create
a
local
representation
of
those
years.
There
is
another
controller
called
H
controller,
which
does
it
performs
the
action
on
the
new
submitted
request
and
create
the
final
objects
for
that.
Well,.
B
To
help
you
understand
and
simpler
terms,
assume
when
I'm
a
teacher
and
you're
all
my
students
instead
of
writing
the
homework
in
each
of
your
individual
notebooks
every
day
and
asking
you
to
finish
it
I
would
just
write
it
on
a
notebook
and
ask
you
to
copy
it.
So
in
a
real
world
analogy,
I
believe
that
it's
better
right
is.
A
B
Would
like
to
bring
up
a
very
interesting
question
here:
how
do
you
expose
services
from
the
edge
to
the
outside
world?
So
let's
say
you
know,
I
resumed.
My
network
connectivity
but
and
I
want
to
really
export
some
of
the
really
important
information
like
the
logs
that
help
metrics
the
monitoring
or
what
really
happened
in
the
cluster.
How
do
I
really
expose
these
services
yeah.
A
That
is
another
challenge
which
you
are
going
to
talk
about
later.
So,
as
I
mentioned,
that
you
may
have
to
expose
your
services
outside
the
cluster
for
n
number
of
reasons,
altmetrics
log
collection
and
all
in
this
example,
we
are
talking
about
a
middleware
where
most
of
you
and
your
day-to-day
work
must
have
used
the
reverse
tunnel.
When
you
connect
to
a
private
cloud
set
up
on
the
AWS,
you
have
to
use
the
bastion
connectivity
or
your
own
data
center.
You
use
the
jump
boxes
there.
We
use
the
same
concept
in
in
this
design.
A
Your
H
cluster,
which
is
running
the
H
controller,
is
responsible
to
create
the
reverse
tunnel
and
open
all
the
ports
for
services
which
are
running
inside
that
cluster.
These
ports
are
opened
up
and
a
reverse
tunnel
is
set
up
for
the
federated
or
cluster
Federation
cluster,
where
your
SSH
server
pod
is
running,
and
this
SSH
server
pod
is
responsible
to
expose
these
services
behind
load
balancer.
A
So
when
outside
user
connects
to
the
federated
cluster,
this
federated
cluster
can
expose
the
services
and
the
data
which
it
is
requesting
for
the
important
thing
to
note
here
is
the
network
boundaries.
I
would
like
to
remind
that.
We
don't
have
a
luxury
of
push
or
getting
query
query
to
the
H
clusters.
We
have
to
pull
the
data
right.
B
So
you
know
we
have
been
talking
about
edge
deployments.
So
let's
also
look
at
a
use
case
where
a
I
federated
learning
use
keys
could
also
be
solved
so
for
people
who
do
not
know
what
federated
learning
is
I'll.
Give
you
a
simple
example:
let's
say:
I
have
finals
and
I.
Ask
these
fine
notes
to
train
the
data
continuously
and
you
know
update
the
models,
but
it's
very
local
to
them.
B
The
models
are
pretty
local
and
now
I
want
to
build
a
combined
model
so
from
these
individual
models
which
have
been
generated-
and
this
is
the
concept
behind
the
federated
learning
and
how
it
essentially
works
as
pools
of
the
nodes,
combines
the
models
and
written
new
model
and
then
pushes
back
the
updates.
That's
the
important
process.
How
do
we
really
solve
this
with
you
know
using
the
reverse,
SSH
proxy
architecture?
It's
pretty
much
the
same,
so
what
we
also
did
was
with
the
same
intent.
We
introduced
the
service
mesh
and
the
service
mesh.
B
The
init
containers
update
the
IP
tables,
so
essentially
an
application
developer
doesn't
really
have
to
focus
on.
You
know
routing
his
data
to
certain
Federation
clusters.
He
doesn't
have
to
even
have
the
configuration
to
do
it.
All
he
would
do
is
forward
his
data
to
a
local
host
local
port,
and
you
know
at
the
infrastructure
layer
based
on
the
business
rules.
We
would
be
configuring
these
endpoints
using
the
service
mesh,
and
there
are
a
lot
of
available
solutions
in
the
market
on
the
service
mesh.
You
could
explore
and
that's.
A
Really
interesting
as
a
application
developer,
I
don't
have
to
worry
about
infrastructure.
I
will
just
send
the
information
to
the
local
host
at
some
port
and
it
always
be
like
that.
So
yeah
we
spoke
about
a
is
workload
deployments
on
the
edge
clusters.
We
spoke
about
how
we
can
expose
the
services
outside
the
edge
clusters,
how
we
can
use
the
service
mesh
to
actually
send
the
application
data
and
forwarded
to
the
federated
cluster
and
finally
consume
it,
but
that's
it
edge.
Computer
is
still
primitive.
Mother,
yeah.
B
That's
a
very
difficult
question
to
answer,
because
we
have
to
define
a
lot
of
things
about
what
really
high
availability
means.
You
know
I
can
have
refrigerator
at
my
home,
which
is
really
a
smart
refrigerator,
and
do
I
mean
that
I
have
two
refrigerators
in
case
one
goes
down,
and
so
we
have
to
really
define
what
high
availability
means
and
also
about
backup
and
restore.
How
do
we
really
do
it
at
the
edge
when
there
is
no
person
who
is
really
specially
trained
to
perform
this
activity
just
sitting
there
doing
it?
A
True
and
as
clusters
are
very
minimalistic
single
node
clusters,
what
if
a
lot
of
data
is
coming
there,
how
will
be
load
balanced,
doing
a
luxury
of
adding
more
nodes?
In
fact,
if
you
want
to
reconfigure,
the
workloads
which
are
deployed
at
the
edge
edge
from
the
security
perspective
is
very
sensitive.
You
have
to
even
restrict
the
cluster
administrator
access,
you're
not
allowed
to
give
them
SSH
access,
so
mostly
the
practicer
we
have
noticed
is
when
the
H
node
comes
up,
it
bootstraps
itself.
A
B
Also,
a
few
things
about
edges
that
you
know
you
may
have
real
data
sources
could
be,
the
sensors
could
be
the
surveillance
cameras.
But
how
do
you
discover
them?
And
how
do
you
you
know,
configure
them
and
also
about
the
sources
of
information?
It
could
be
the
container
itself
which
is
running
and
it
could
be
about
the
models.
You
know
where
do
I
fetch
this
data
from
trusted
sources,
and
this
has
to
be
pretty
much
Auto
configured
so
that
human
intervention
does
not
cause.
You
know
anomalies,
because
we
tend
to
do
errors
right
so.
A
So
this
is
where
we
are
looking
for.
The
community
help
I'm
sure
that,
as
I
saw,
a
lot
of
people
working
under
h,
SE
and
nvidia
is
working
on
AI
workloads
at
the
edge
for
the
h
compute
we
are.
This
is
exploring
these
solutions,
it's
an
ongoing
process
and
we
expect-
or
we
at
least
looking
for
a
similar
kind
of
contribution.
If
community
also
provides-
and
we
we
are
available
after
this
talk,
you
can
have
a
chit
chat
around
it
right.
So.
B
The
takeaways
for
you
are
how
to
use
kubernetes
for
the
edge
compute,
and
you
also
learned
about
the
federated
architecture.
You
know
which
we
are
proposing
based
on
the
pool
models.
You
know
instead
of
a
push
model,
and
so
please
feel
free
to
write
back
so
his
wish
wish
and
I
mother
Susan,
our
email
ids,
are
on
the
screen
so
have
a
nice
day.
Thank.