►
From YouTube: 34 Truths We Learned About Kubernetes and Edge - Saiyam Pathak, Walmart Labs & Karthik Gaekwad
Description
Don’t miss out! Join us at our upcoming events: EnvoyCon Virtual on October 15 and KubeCon + CloudNativeCon North America 2020 Virtual from November 17-20. Learn more at https://kubecon.io. The conferences feature presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects.
34 Truths We Learned About Kubernetes and Edge - Saiyam Pathak, Walmart Labs & Karthik Gaekwad, Oracle
Running Kubernetes on the Edge has gained a lot of popularity this year, but it is still complicated to develop, deploy, and maintain at scale. In this talk, the speakers will demo how to run Kubernetes on Raspberry Pi clusters at scale from their experiences at Walmart Labs and Oracle Cloud. They will demo machine learning use cases on Raspberry Pi's using the K3s distribution. Further, they will go into details of why K3s might be a good fit for k8s running on edge scenarios, how it is different from the regular kubernetes distribution, and how it works under the hood. Talk outline: Introduction to kubernetes and Edge computing Problems faced with Edge devices involving kubernetes K3s solutions for the Edge Demo: Running kubernetes on Raspberry Pi's using k3s Challenges using k3s on edge devices Speaker Recommendations for kubernetes on edge
https://sched.co/ZeoY
A
My
name
is
sayam
parthak
and
I
am
a
software
engineer
working
at
walmart
labs.
My
twitter
handler
is
saying
patrick
I'm
cka
ckad
certified
I'm
a
cncf
ambassador,
I'm
a
community
leader.
I
run
a
lot
of
meetup
groups,
including
rancher
influx
docker
at
bangalore,
I'm
an
influx
ace
and
I
also
run
a
youtube
channel
where
I
do
lot
of
live
streams
with
other
community
members.
Today,
I'm
joined
with
karthik.
B
Thanks
siam,
my
name
is
karthi
gayquad
and
I'm
the
head
of
cloud
native
engineering
at
verica,
I'm
the
author
of
learning,
kubernetes
and
a
bunch
of
other
cloud
native
courses
on
linkedin
learning
prior
to
verica.
I
built
the
oracle,
managed
kubernetes
engine
and
was
a
principal
developer
at
oracle
cloud.
I
also
do
a
bunch
of
community
stuff,
like
devops
js,
austin
cloud
austin
all
day,
devops
and
container
days.
I've
been
in
industry
for
a
while,
and
you
can
find
me
more
on
iteration
one
on
twitter.
B
Today,
we'll
go
into
more
about
edge
computing
and
talk
about
the
intersection
of
edge
and
kubernetes,
we'll
discuss
some
of
the
complexities
that
we
see
from
an
edge
perspective
and
how
kubernetes
and
k3s
helps
us
from
that
perspective.
We'll
talk
more
about
cloud
native
at
edge.
Do
a
really
sweet
demo
about
how
to
run
k3s
on
a
raspberry
pi
cluster,
talk
more
about
complexities
at
edge
and
then
give
you
some
recommendations.
B
So
what
is
edge
computing?
If
you
look
at
wikipedia
it
describes
edge
computing
as
a
distributed
computing
paradigm
that
brings
computation
and
data
storage
closer
to
the
location
where
it's
needed
to
improve
response
time
and
save
bandwidth,
and
also,
if
you
look
at
edge
computing,
there's
a
lot
of
growth
from
the
gartner
and
forster
reviews.
Gartner
said
that
by
2023
there
could
be
more
than
20
times
as
many
smart
devices
on
the
edge
of
the
network.
As
in
conventional
I.t
roles,
forester
actually
said
that
the
edge
cloud
market
will
grow,
at
least
by
50
percent.
B
But
what
does
this
all
really
mean?
Being
an
engineer,
I
like
to
break
this
problem
down
by
looking
at
specific
use
cases.
In
my
time,
at
national
instruments
and
mentor
embedded,
I
spent
a
lot
of
time
looking
at
the
intersection
between
the
cloud
and
edge
devices
or
embedded
devices
in
general.
One
of
the
my
favorite
things
that
used
to
come
up
a
lot
of
times
was:
how
do
we
get
all
this
stuff
to
work
in
the
case
of
wind
turbines?
B
You
might
want
to
understand
what
the
generator
speed
is:
try
and
increase
it
or
decrease
it.
Look
at
the
blade
adjustment
angle,
so
the
blade
adjustment
angle.
The
turbines-
can
actually
move
in
a
horizontal
or
vertical
way,
and
there
are
times
when
you
actually
want
to
change
that
angle,
to
be
able
to
generate
more
electricity
and
also
the
wind
turbine
rotation.
B
So
you
can
actually
rotate
the
turbine
left
to
right,
and
you
know
there
are
times
when
you
might
want
to
do
that
too,
because
the
wind
direction
might
be
from
a
different
kind
of
perspective.
So
those
are
the
different
controls
that
you
might
have
from
a
turbine
standpoint
from
a
wind
farm
control
center.
So
the
control
center
is
more
of
a
place
where
you
might
have
a
lot
of
folks
kind
of
looking
at
your
whole
turbine
farm
from
a
holistic
perspective.
B
So
I
threw
in
a
picture
of
your
hybrid
rolla,
renewables
control,
control,
room
and
from
from
that
perspective,
you're
kind
of
looking
at
the
whole
farm.
So
you
know
you
might
be
collecting
your
average
turbine
speeds
in
there.
Looking
at
the
health
status
of
each
of
your
individual
turbines
monitoring
performance
of
the
turbines,
you
know,
if
there's,
if
there
are
things
that
might
not
be
working
with
an
individual
turbine,
you
might
need
to
send
a
technician
out
and
go
get
that
fixed,
etc.
B
Also
monitor
the
general
efficiency
of
the
farm.
Is
it
actually
generating
a
lot
of
electricity
or
you
know?
Is
there
something
wrong
from
like
a
farm
perspective
and
then
also
one
of
the
big
goals
is
to
be
able
to
generate
electricity,
so
you
know
be
able
to
calculate
that
from
a
whole
farm
perspective.
B
You
can
think
of
this
more
of
a
you
know
from
the
high
level
from
the
control
center
perspective,
so
it's
responsible
for
the
overall
operations
of
the
farm.
You
know
big
data
processing,
so
you
have
a
bunch
of
data
coming
in
to
your
to
a
wind
farm.
You
might
want
to
throw
that
into
the
cloud
to
do
additional
processing
aggregate
different
data
et
cetera
across
you
know
a
lot
of
times
that
you
might
want
to
use
the
cloud.
B
B
So
if
you
have
to
send
data
up
to
the
cloud
you're
able
to
do
that
and
also
connected
to
all
the
turbines
in
the
farm
as
well
typical
use
case
that
I've
seen
over
here
is
the
turbines
are
connected
via
ethernet
or
you
know
some
other
some
some
other
kind
of
way.
But
ethernet
is
pretty
popular
these
days
and
then
going
down
to
the
sensor
level
or
to
the
turbine
level
the
turbine
itself.
B
B
B
The
code
deployments
need
to
be
tiny
because
there's
not
a
lot
of
ram
and
memory
on
the
actual
board,
and
then
you
also
have
you
know,
embedded
device
concerns
so
like
you
can
actually
put
something
too
big
on
there
or
you
know
other
concerns
that
we
might
not
think
about
from.
You
know
from
a
cloud
perspective
at
the
edge
level.
It's
all
about
connectivity,
so
you
know
everything
connected
up.
Is
it
easy
to
connect
multiple
devices
to
the
edge?
B
B
Typically,
this
is
not
a
problem,
but
this
used
to
be
a
problem
back
in
the
day
when
you
had
to
build
out
like
your
own
kind
of
data
center
inside
of
you
know,
inside
of
your
control,
plane
or
your
control
tower
and
then
as
maintenance
of
your
sensors
and
farm.
This
doesn't
happen
super
often,
but
it
does
like
your
sensors
might
go
offline.
There
might
be
issues
with
individual
turbines
etc,
and
then
you
need
to
go
and
address
maintenance
issues
off
your
sensors.
B
One
of
the
things
about
this
whole
ecosystem
is
that
there's
actually
a
lot
of
vendors
but
they're
very
like
vendor
specific,
and
then
you
have
walk
into
your.
You
know
to
your
iot
devices
or
to
your
edge
devices.
So
there's
a
ton
of
custom
solution
providers,
so
you
know
you
might
approach
a
specific
provider
about
hey.
I
need
like
a
solution
for
my
wind
farm
like
what
do
you
have
to
provide
and
you
can
go
talk
to
individual
vendors
that
will
help.
B
So
you
can
have
open
source
kind
of
in
the
cloud
you
can
also
have
open
source.
You
know
on
the
ed
side
of
things,
and
it's
the
greatest
thing
about
kubernetes
is
that
it's
built
for
managing
many
devices,
so
things
like
scaling
and
deployment
et
cetera.
All
of
those
things
are,
you
know,
are
ingrained
in
the
in
the
kubernetes
mindset
and
in
the
kubernetes
ecosystem.
B
A
A
Let's
now
understand
what
actually
the
problem
is
that
we,
as
engineers
are
trying
to
solve
so
we
have
edge
and
we
want
to
leverage
the
highly
complex
ecosystem
of
kubernetes
at
h
now,
in
order
to
do
that,
we
have
k3s,
which
we
are
going
to
discuss
today.
So
k3s
is
another
open
source,
cncf
certified
kubernetes
distribution,
which
has
all
the
powers
of
kubernetes,
and
it
is
extremely
suitable
for
edge.
A
So,
as
you
can
see,
it's
extremely
lightweight,
and
by
lightweight
when
I
say
it
means
it
is
less
than
a
50mb
binary,
so
which
is
huge,
so
it
removes
all
the
entry
cloud
providers.
It
removes
all
the
entry
storage
drivers
and
it
removes
all
the
alpha
features.
So
it's
basically
a
very
trimmed
down
code
based
version
of
the
native
kubernetes
distribution-
it's
specifically
built
for
arm.
So
when
the
developers
were
building
kubernetes,
so
there
were
two
things
first
like
we
have
to
make
something
for
the
edge
for
the
arm
devices.
A
A
So
when
we
say
that
the
code
base
is
trimmed,
so
the
power
of
kubernetes
reduces
no,
it's
not
the
case,
so
the
code
base
is
reduced,
but
still
you
get
the
full
power
of
kubernetes
and,
most
importantly,
everything
is
running
as
a
single
process,
so
which
is
what
makes
it
so
simple,
efficient
and
fast,
let's
quickly
dive
into
the
architecture
of
k3s.
So
on
the
left
hand
side,
you
have
the
k3
server
and
you
on
the
right.
You
have
the
k3s
agent.
A
Now
everything
is
running
as
a
single
process,
so
it's
typically
the
same
architecture
as
you
have
of
a
kubernetes
one,
but
here
what
additional
thing
is
the
tunnel
proxy,
so
the
agent
registers
itself
to
the
server
via
a
tunnel
proxy
and
once
the
connection
is
established
after
that
the
communication
happens
via
this
particular
tunnel
between
the
agent
and
the
server
rest.
It
has
the
escalate
database,
which
is
inside
the
process
itself
and
its
cd
is
not
used
in
the
back
end.
They
are
using
kind
and
it's
a
plugable
kind
of
architecture.
A
A
A
A
A
A
So
now
the
raspberry
pi
will
be
working
as
the
k3
server
and
the
raspberry
pi
agent
will
be
the
agent
which
will
connect
to
the
server
now
k3s
installation,
as
I
have
mentioned
before,
as
well.
It's
pretty
simple!
So
it's
as
simple
as
one
command.
That's
it!
So
it's
finding
the
stable
release,
it's
using
b1.18
and
it's
downloading
the
binaries
and
within
a
couple
of
seconds
your
k3s
server
will
be
up
and.
A
A
A
A
A
A
A
A
A
A
A
So
we
are
taking
the
influx
db
client,
we
are
establishing
the
client,
we
are
creating
the
database.
We
are
switching
to
that
database.
We
are
having
a
timestamp
now.
This
particular
file
gives
you
the
temperature,
the
cpu
temperature
updated
all
times.
We
create
a
json
body
where
we
send
all
the
tags
in
the
fields.
A
A
A
A
A
A
So
this
is
just
one
use
case
where
what
we
have
seen
is
we
have
our
temperature
sensing
from
a
raspberry
pi.
Now
this
particular
use
case
can
be
extended
to
n
number
of
sensors,
like
karthik
has
stock
for
windmill,
where
temperature
is
an
essential
part
and
can
be
extended
to
other
sensors
and
have
all
things
in
place.
So
you
have
one
stable
thing
ready
and
now
you
can
add
top
of
on
top
of
it.
A
B
B
B
B
The
biggest
thing
I'll
hear
also
is
cultural
challenges,
and
you
know
we've
kind
of
seen
this
as
well,
with
just
back-end
developers
and
front-end
developers,
but
also
the
folks
actually
working
on
the
iot
side.
Those
are
embedded
developers
and
their
mindset
is
very
different
from
kind
of
what
we
think
about
from
a
cloud
native
perspective
or
from
a
cloud
native
developer
perspective.
B
So
it's
a
different
tech
stack
c
is
very
common
from
that
tech
stack
over
there
so
and
then
also
for
for
us
in
the
cloud
native
world
things
like
ci
and
cd,
continuous
integration,
continuous
deployment
delivery.
Those
things
are
are
very
normal,
and
you
know
those
are
things
that
we
think
are
about
upfront,
but
from
an
embedded
developer
standpoint.
They
might
not
even
know
what
ci
or
cd
might
be
so
there
are.
B
There
are
challenges
that
you'll
end
up
facing
there
when
you
kind
of
bring
two
disparate
sets
of
developers
together,
also
understanding
your
architectural
needs.
So
you
know
we
kind
of
talked
about
raspberry
pi
and
how
to
run
k3s
and
raspberry
pi.
We
also
talked
about
the
wind
farm
example,
so
you
know
from
if
you
take
like
a
30
000
foot
view
on
this.
Kubernetes
is
really
good
for
cloud
for
cloud
and
data
center
compute
instances,
so
you're
managing
your
whole
data
center
with
kubernetes.
B
That's
why
it's
really
popular
cube
edge
from
our
reading
is
really
good.
For
you
know
when
you
wanna
have
a
cloud
control
plane
and
you
wanna
have
disparate
edge
devices
or
instances
it's
great
for
controlling
that.
So
think
of
the
you
know,
if
you're
using
nest,
kind
of
devices
etc,
where
they're
all
able
to
connect
to
the
cloud
or
to
a
cloud
control,
plane,
cube
edge,
would
work
really
well
over
there.
K3S
is
also
really
good
for
the
cloud
use
cases
and
edge
as
well
as
well
for
air
gap
environments.
B
A
Well,
thank
you
kante.
Now,
let's
dive
into
the
future
discussion
on
the
state
of
kubernetes
and
edge.
So
let's
make
it
a
informal
discussion
and
not
a
very
formal,
powerpoint
kind
of
discussion
yep.
So
I
think
that
kubernetes
will
play
a
very
important
role
for
edge.
Why?
Because
people
and
the
companies
are
transitioning
into
the
virtualization
world
and
building
apps
for
cloud
native
ecosystem
and
they
want
to
change
the
apps
change.
The
versions
all
that
stuff
and
communities
will
definitely
play
a
very
important
role
in
handling
these
operations.
A
A
Now,
if
you
go
by
the
stats,
I
see
that
it
has
more
than
13
000
github
stars,
which
is
pretty
amazing.
It
has
a
wide
community
adoption
and
people
actually
love
it.
So
that's
why
they
have
built
products
on
top
of
k3s.
So
if
you
see
a
open
source
enthusiast
like
alex,
he
built
k3sf,
which
is
pronounced
as
ketchup
for
so
it's
basically
bootstrapping
the
k3s
cluster
in
less
than
a
minute.
So
so
that's
that's
the
adoption
that
we
have.
A
Then
there
are
a
few
companies
like
sivo
who
is
who
built
up
the
first
ever
managed
k3s
service.
So
so
people
have
already
started
looking
k3s
as
a
stable
product,
and
it
is
always
in
line
with
the
upstream
kubernetes
updates.
So
whenever
the
version,
a
new
version
comes
okay,
3s
team
and
the
open
source
folks
they
work
and
make
it.
A
You
know
matching
the
upstream
kubernetes
version,
so
I
think
yes,
k3s
itself
has
a
very
bright
future
and
people
are
making
they
have
their
road
maps
very
clear
and
the
issues
the
pull
requests
say
it
all.
So
k3s
definitely
has
a
bright
future,
and
with
this
definitely
it
will
be
used
for
edge,
because
that
was
a
primary
use
case
for
g3s,
so
edge
computing
will
definitely
be
very
powerful,
become
very
powerful
with
the
usage
of
k3s.
A
B
I
totally
agree
with
you.
I
was
kind
of
trying
to
think
about
whether
you
know
to
kind
of
go
deeper
into
the
technology
or
think
of
it
from
a
high
level
perspective,
and
I
think,
from
you
know,
from
a
high
level
perspective
from
a
tech
standpoint.
I
think
we're
there
today
what
you
know
whether
it's
k3s,
whether
it's
cube
edge.
I
think
the
idea
of
using
something
like
kubernetes
that
helps
you
that
basically
helps
you
take.
B
You
know
a
single
device
put
basically
an
operating
system
on
there
and
then
scale
that,
for
you
know,
hundreds
like
nodes
of
many
clusters.
I
think
that
it's
pretty
proven
technology
today
from
a
k2
standpoint
from
a
cube
edge
tent.
However,
so
I
think
we
have
solutions,
which
is
great.
I
mentioned
before
that,
the
the
challenge.
I
think
that
we
will
face
at
least
in
this
next
year
next
couple
years,
is
to
understand
what
the
problems
are
understand,
what
the
real
problems
are
from
an
embedded
developer
standpoint.
B
So
a
lot
of
it,
you
know,
comes
from
like
bridging
communities
right
because
I
think
the
communities
that
we
have
from
a
tech
standpoint
like
you
and
I
we're
squarely
in
like
the
cloud
native
space
and
we're
like
yeah
kubernetes.
We
understand
it.
We
might
not
know
everything
about
it,
but
I
think
we
know
like
from
like
a
pretty
good
perspective
of
like
how
the
how
everything
works-
architectural
textural
implementation
whatever.
B
But
I
don't
think
anybody
in
the
embedded
space
really
knows
about
a
lot
of
the
stuff
that
you
know
we're
we're
kind
of
coming
out
with
whether
it's
k3s,
whether
it's
cube
edge,
whether
it's
just
even
kubernetes
or
prometheus
or
like
any
of
the
things
in
our
ecosystem.
So
the
the
the
challenge
will
be.
You
know
advocating
a
lot
of
these
platforms
to
folks
that
are
actually
writing,
embedded
code
and
being
like
hey.
I
know
you're
trying
to
solve
these
mqtt
problems
and
you're
trying
to
write
this.
B
This
mqtt
code,
specifically
for
your
cis,
but
did
you
know,
there's
like
a
tool
that
you
can
use
to
or
not
not
to
about
a
platform
that
you
can
use
to
kind
of
manage
all
of
this
stuff
by
default,
and
you
know
kind
of
get
get
some
kind
of
reception
from
what
you
know.
What
what
folks,
who
are
actually
in
that
ecosystem
have
so
understand.
You
know
kind
of
understanding,
their
pain
points
and
then
bringing
it
back.
B
B
But
really
we
need
all
the
communication
to
work
with.
You
know
mqtt
or
you
know,
we
need
something
even
more
low
power
like
bluetooth
or
something
like
that,
and
we
have
to
write.
You
know
more
more
apis
that
help
communicate.
You
know
through
a
cluster
with
bluetooth
or
something
something
something
similar
to
that.
B
So
you
know
it
might
sound
kind
of
lofty,
but
it's
trying
to
figure
out
how
we
can
bring
embedded
engineers
that
are
kind
of
in
their
own
ecosystem
might
be
in
their
own
silo,
but
try
and
bring
them
into
kind
of
our
cloud
native
tribe,
and
I
think
that's
where,
like
the
next
generation
of
problems
for
us,
will
be
it's
it's
not
necessarily
like
technology
but
it'll.
You
know
it'll
end
up
being
more
of
like
hey
we're.
We
built.
A
Yeah,
I
completely
agree
to
all
those
points.
Karthik
and
even
even
I
feel,
like
you
know,
the
use
case
is
like
people
want
to
do
machine
learning
at
edge
like
capturing
billions
of
images
and
not
sending
them
to
cloud
due
to
latency
issues
and
all
that
so
so
they
need
solutions
which
can
be
fit
into
their
specific
small
data
centers,
and
they
can
do
you
know
a
machine
learning
on
top
of
that
anomaly:
detection.
A
On
top
of
that,
whether
you
take
example
of
a
few
generators
and
finding
the
defects
from
that
or
you
take
example
of
the
cctv
footages
and
you
know
doing
the
capturing
the
per
second
images
and
doing
ml
ai.
On
top
of
that.
So
yes,
the
there's
a
huge
need
for
the
cloud
native
ecosystem
over
there
and
the
engineers
who
are
building
these
devices.
B
A
B
B
Yep
totally
agree,
and
once
again
you
can
read
siam
on
twitter
at
scienpatech
and
I'm
iteration1
on
twitter.
I
think
we're
probably
at
the
end
of
our
session,
so
you
know
we'll
be
around.
If
you
have
questions
for
us
use,
you
know,
use
use
this
use
the
slack
or
we're
actually
both
on
the
kubernetes
slack
as
well.
So
you
can,
you
know
you
can
finger
there
as
well,
but
we'll
hang
out
after
after
our
session.
So
you
know
we're
not
hard
to
find
so
hope
hope.