►
From YouTube: Autoscaling a Multi-Platform Kubernetes Cluster Built with kubeadm [I] - Lucas Käldström
Description
Autoscaling a Multi-Platform Kubernetes Cluster Built with kubeadm [I] - Lucas Käldström, Upper Secondary School Student - Individual Maintainer
Does Kubernetes run on any architecture other than Intel/AMD64? Yes it does, since v1.3.0.
Lucas Käldström (@luxas), a general upper secondary school student, added support for ARM and ARM 64-bit to the core codebase earlier this year. Kubernetes is all about letting you choose the container runtime, the operating system, the environment (cloud vs bare metal) etc., so why wouldn’t you like to choose CPU architecture as well?
The talk demonstrates Kubernetes’ multi-architecture features in a very practical manner: by having a 10-node cluster of credit-card-sized devices with three architectures (amd64, arm and arm64) in front of the audience. The cluster is deployed with kubeadm. For the demo, a cross-platform nginx workload will be created from the dashboard, deployed to all nodes regardless of architecture, monitored via an auto-updating grafana dashboard and autoscaled by a Horizontal Pod Autoscaler while traffic is generated against the cluster’s Ingress controller. And just to make things more exciting, a node’s power cable may be unplugged by a volunteer from the audience, at any time during the heavy traffic to demonstrate Kubernetes’ auto-healing functions.
About Lucas Käldström
Hi there and thanks for finding my talk!
I'm a Swedish-speaking Upper Secondary School Student from Finland.
I'm really excited about Kubernetes and new emerging technologies.
I have been engaged in Kubernetes for about two years now and ported Kubernetes to ARM, both 32 and 64-bit as my first larger contribution. I've written an accepted proposal for how Kubernetes should work on multiple platforms. I've been a maintainer for Kubernetes on my spare time since April 2016. I was in the group that started Minikube as well, but then in August 2016 I moved on to maintaining kubeadm, which I have focused on since then.
Join us for KubeCon + CloudNativeCon in Barcelona May 20 - 23, Shanghai June 24 - 26, and San Diego November 18 - 21! Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy and all of the other CNCF-hosted projects.
A
You
may
also
know
me,
as
the
kubernetes
unarmed
guy,
which
was
my
first
project
that
I
launched
one
and
a
half
year
ago
about
yeah
from
Finland
and
I'm
speaking
Swedish
as
my
mother
tongue.
Some
some
people
in
Finland
do
that
and
I'm
a
maintainer
of
kubernetes
since
a
year
back,
which
I'm
really
excited
about
and
I'm
excited
to
be
here
today.
A
A
funny
thing
to
know
about
me
is
that
I've
never
have
taken
a
computing
class
I,
just
like
tried
to
learn
from
the
internet
and
from
the
community
from
others
from
you
and
what
I
have
been
doing.
Then.
My
first
open-source
projects,
as
I
said,
was
Cuban
in
Islam.
It
was
a
really
easy
solution
to
run
kubernetes
on
only
on
whatever
price.
Initially,
but
I
didn't
like
it
to
be
a
side
project
in
my
own
github
domain,
so
forth,
so
I
simply
merge
it
merged
it
into
core,
and
that
was
pretty
exciting.
A
A
A
So
a
quick
recap.
How
can
you
do
this,
for
example,
with
the
go
and
you
can
cross-compiled
pretty
easily
with
go?
That's
one
of
the
good
things
with
the
language.
For
example,
you
may
just
make
a
non
binary.
That's
with
gold
as
static
statically
compiled
you
just
set,
go
art
form
or
you
set
go
watched
on
64.
If
you
want
to
do
to
make
things
simple
main
goal,
or
whatever
binary
for
Windows,
you
can
set
go
OS
to
Windows
and
it
will
make
a
such
binder
for
you.
A
A
Let
us
be
really
great
to
see
they
interest
them.
Also,
it's
it's
possible
to
to
do
emulation
with
Linux,
so
I
can
run
on
binaries
on
an
AMD.
64
host
and
I
can
run
build
sucker
images
as
well
and
when,
when
a
non
binary
runs
in
a
container,
for
example,
with
docker
build
its
it.
Just
if
I
put
an
info
binary
inside
of
the
container,
which
is
schema,
I
can
register
a
module
in
the
kernel,
so
the
arm
binary
will
be
translated.
A
The
arm
cisco's
will
be
translated
into
amd64
ones
when
using
chemo
as
the
layer
in
between
and
the
this
approach
is
used
now,
with
all
the
builds
or
old,
akka
builds
for
form
on
64
and
so
forth,
and
this
is
pretty
convenient
because
well,
we
can
have
official
builds
without
having
to
do
anything
else.
Just
straight
from
from
google
CI
builds.
A
So
what
has
the
journey
being
like
being
like?
Well,
1.2
was
the
first
release,
I
participated
in
and
there
I
made
initially
on
binaries
that
were
released,
and
it
was
a
huge,
huge
step
for
me
personally,
at
least
to
get
that
into
core.
Then
I
continued
in
in
1:3
and
made
the
darker
images
with
this
key
new
approach
as
well.
So
after
that
kubernetes
could
release
both
on
32
and
64
bits
and
also
the
docker
images
like
hypercube,
for
example.
So
you
could
run
a
cluster
from
official
binaries
official
everything
with
1/3
from
another
machine.
A
How
many
of
you
have
have
used
kubernetes
unarmed,
by
the
way,
really
cool
that
that's
awesome,
and
also
when
talking
about
scheduling
it's
a
little
bit
difficult.
If
we
don't
know
if
the
cubelets
doesn't
say,
I'm
an
armed,
node
or
I'm,
an
AMD
64
node,
we
can't
schedule
on
it,
but
in
1/3
allayed
the
cubelets
registers
itself
with
a
specific
label,
and
we
can
then
schedule
based
on
this
label.
This
is
this:
isn't
like
the
recommended
approach,
long
term,
but
it's
possible
to
build
on
top
of
this
for
more
advanced
mechanism.
A
Then
in
1/4,
cubed
n
was
released
and
I'm
very
exciting.
To
see
this.
This
simple
and
and
straightforward
flow
with
cube,
ATM
and,
as
I
said,
I've
worked
worked
with
the
thick
glass,
the
lifecycle
team,
from
the
video
from
the
beginning
and
this
cluster
that
I
have
here
as
well.
It's
running
cube
ATM
and
it
was
just
a
matter
of
some
some
straightforward
commands
with
innocent
join
in
one.
For,
unfortunately,
we
had
to
patch
the
goal
line
because
there
was
some
Sun
bug
in
the
in
the
runtime.
So.
A
A
As
I
said,
the
community
has
been
really
great
for
primarily
here
from
IBM,
with
PowerPC
64
little-endian
and
as
390x
and
enzyme
for
1/6.
They
contributed
builds
for
these
architectures
as
well,
and
now
it
now
it's
really
cool,
because
it's
it's
mostly
a
matter
of
like
adding
to
the
list
of
supported
architectures
in
all
make
files
and
that's
how
it
should
be
and
yeah.
If
you
have
any
questions,
if
it's
just
just
to
ask
so
why
should
companies
work
on
multiple
platforms?
A
Isn't
it
isn't
that,
like
a
two-week
raspberry
pie,
for
example,
that
doesn't
make
sense,
or
how
should
we
look
at
this?
Well,
we
have
the
CNCs
charter,
which
states
this
pretty
pretty
well
and
yeah.
I
totally
agree
with
this
and
think
it's
for
the
diversity
of
machines
and
for
the
future
ecosystem,
with
cloud
native,
it's
important
to
include
all
architectures
and
all
platforms,
and
by
that
we'll
see
a
stronger
competition
and
so
forth.
A
A
In
the
use,
cases
and
and
yeah
well
feel
strong
the
ecosystem,
and
we
don't
know
in
20
years
if
until
will
be
the
dominating
platform
it
could
be
or
it
may
not.
We
don't
know,
and
if
we
get
four
new
architectures
platforms
to
the
kubernetes
ecosystem,
we'll
get
some
more
people
with
it
with
it
and
that's
great
and
the
risk
of
when
the
lock-in
will
be
significantly
reduced.
A
What
would
kubernetes
an
arm
and
just
on
32
and
64-bit
be
useful
in
the
meantime.
Well,
we
have
a
master
thesis
from
these
guys's
combination
and
Martin
Jensen.
They
wrote
about
a
year
ago
from
my
project
banana
Sonam.
They
wrote
663
page
master
thesis
about
this,
how
cubanelle
how
they
could,
how
anyone
could
teach
kubernetes
the
cloud
native
concepts
and
and
so
forth.
So
the
students-
and
this
is
a
thing
I'm
really
excited
about.
A
We
want
more
people
to
be
able
to
adopt
this
these
technologies
and
when
having
a
Raspberry,
Pi
cluster,
for
example,
it
gets
really
practical,
I
think
I've,
seen
two
or
three
here
downstairs
and
from
from
the
vendors
they've
built
their
own
clusters
and
are
having
fun
playing
with
the
with
their
small.
Our
machines
and
this
project
went
well
as
far
as
I
could
could
reach
and
I
hope
will.
We
can
continue.
A
A
A
Well,
you
see
you
buy
them,
it's
really
easy
and
straightforward.
How
many
of
you
have
heard
of
Q
Batum
cool
Bailey,
everyone?
Basically,
your
first
trusting
the
the
Google
app
key.
This
will
be
changed
in
the
future
to
the
CNCs
and
we
have
or
when
we
all
host
packages
on
CNCs
in
front,
but
for
the
meantime,
it's
Google
that
is
kind
enough
to
respond
for
this.
Then
you
can
just
up
get
installed,
cube,
ATM
and
it
will
install
it
and
if
you
want
to
bring
bring
up
your
master
with
all
the
default
set.
A
Just
if
you
cube
a
them
in
its
in
the
command
line
and
boom,
you
have
a
master,
then
you
should.
You
should
apply
a
port
network.
Kubernetes
is
extensible,
so
it
doesn't
come
with
a
default
port
network.
Instead,
third
body
providers,
like
these
one
downstairs,
may
come
up
with
their
own
solution
that
that
our
conformance
with
the
CNI
spec
and,
in
this
case
I'm
using
weave,
which
has
built-in
support
for
arm
and
arm
32
and
64
bits
out
of
the
box
and
Cabana
editions,
have
different
install
path.
A
So,
if
we're
having
five
different
install
path
for
five
different
architectures
it,
it
won't
just
work
and
we'll
need
this
troll
in
the
ecosystem.
Overall
and
now
you
might
think
how
does
this
really
work
under
the
hood?
And
since
this
is
a
deep
dive,
I'm
going
to
start
going
to
a
little
bit
at
least
a
kubernetes
releases,
as
I
said,
server
binary
for
supported,
architectures
and
also
windows.
A
Now
that
big
windows
have
has
done
a
terrific
work
with
the
cubelets
and
cube
proxy,
so
it
it's
vulnerable
on
windows
as
well,
and
these
are
released
and
uploaded
the
GCS
bucket
and
easily
downloadable.
Then
all
dark
images
in
the
core
kubernetes
reporting
dr.
repository
our
bills
and
pushed
using
this
chemo
emulation
method,
and
we
there
is
a
semi
standardized
make
file.
A
So
basically,
when
I
ported
kubernetes
to
tom
the
first
time,
I
rewrote
to
make
false
by
using
and
now
it
has
been
copied
and
being
used
used
up
as
the
base
for
new
projects
and
new
images
as
well
and
Debian
packages
are
built
for
all
architectures,
and
this
is
pretty
straightforward.
It
just
downloads
the
binary
from
GCS
bucket
and
create
some
Deb's
and
then
finally,
cube
am
is
aware
of
the
which
architecture
is
running
on
and
generates
manifests
for
the
right
architecture.
But
this
is
I.
A
Think
I
don't
want
to
do
in
the
future
because
it
gets
very
cumbersome
if
you
have
three
different
platforms
and
you
want
to
run
a
workload.
My
cool
app
for
example,
and
I
first
building
images,
I'm
building
gold,
binary,
Lagaan
binary
and
pushing
the
sadaqa
hub,
but
then
I
want
to
run
it
on
this
cluster
and
whatever
I
choose
schedule
Axford.
If
I
choose
to
run,
do
a
cube,
GTL
van
and
my
image,
my
cool
app
amd64.
A
If
it
then
lands
on
a
our
machine,
it
won't
be
available
and
will
fail.
So,
ideally
I
would
just
say
my
cool
app
and
whatever
architecture.
It's
scheduled
to
it
will
choose
the
right
one,
and
this
is
possible
and
Phil
has
done
some
great
work
on
this
topic
and
the
other
docker
registry
guys
and
dr.
ray,
says:
registry
version.
2
schema
su
supports
this
and
it's
called
manifest
list.
A
So
a
manifest
list
is
a
well
list
that
specifies
that
for
this
platform
with
these
properties,
you,
the
layer
that
should
be
used-
is
this
one
at
basically
a
reference
to
another
image.
So
if
we're
pulling
a
manifest
list
from
an
amd64
host,
it
will
choose
the
amd64
image
and
so
forth
for
almond
on
64
and
unfortunately,
this
well
gee
cRIO.
The
google
container
read
registry.
A
Doesn't
support
version,
2
schema
to
at
least
hopefully
so
it
isn't
possible
right
now
to
use
it
for
kubernetes
it
bills,
but
where
I
think
the
movement
here
is
towards
CN
CF
repositories,
also
on
the
container
and
docker
image
side.
So
maybe,
and
most
eventually
will
have
cncs
repositories,
which
I
hope
will
support,
schema
su
as
well,
and
also
we
don't
have
any
docker
built
in
tooling.
A
Yes,
Phil
has
a
repo
with
a
tool,
a
manifest
tool
that
can
push
images,
and
you
can
basically
you're
basically
saying
which
platforms
you
want
to
push
for
then
specifying
an
a
template
and
that
template
will
will
well
arch
and
always
will
be
replaced
with
with
the
real
content.
And
then,
in
this
case
my
looks
as
my
app
version.
One
will
be
pushed
as
the
manifest
list
and
for
every
architectural
specified
here
it
will
redirect
for
those
supported
platforms,
and
then
we
have
a
demo
here.
I
have
two
up
boards
in
this
cluster
up.
A
If
intel-based
board
costs
about
$100
I,
think
then
made
in
netherland,
then
we
have
all
droids
see
two
two
of
them,
which
is
on
64,
and
we
have
three
raspberry
PI's
or
the
version
3,
and
in
this
cluster
I
run
cube
area
minutes
and
joined
the
rest
of
the
node
and
all
these
instructions
you
can,
you
can
do
you
do
this
as
well.
The
instructions
are
available
in
this
repo.
A
Basically,
I'm
doing
this:
if
you
buy
10
minutes,
you
better
enjoy
very
straightforward,
then
deploy
just
with
cubes
ETL
fi
and
the
rest
of
the
one.
The
rest
of
the
images
so
I've
built
on
my
looks
as
domain
a
lot
of
manifest
lists
for
this
with
manifest
tool
and
then
I'm
using
those
images
in
in
the
kubernetes
llamó
files,
and
then
it's
just
a
matter
of
cubes
ezel,
applying
them
on
a
on
a
fresh
cluster.
A
A
So
I
have
a
public
IP
address
not
like
the
FS,
but
I
have
a
public
domain
so
to
the
Internet
to
showcase
this
on
and
it's
great
for
hybrid
clusters
especial
for
demos,
then
all
the
metrics
data
in
the
cluster
is
stored
in
in
flux,
DB
and
can
be
visualized
with
Pravana
and
it
seeps.
So
that's
pushing
metrics,
CPU
and
memory
metrics
to
in
flux.
A
Also
I'm,
using
the
primitive
operator
for
call
from
Corollas
and
from
easy
earth
itself
is
multi-platform
as
well,
which
is
really
great
to
see
and
I'm,
creating
our
Prometheus,
so
I
bought
the
resource
in
this
demo
and
then
in
one
combinators
one
six.
We
have
a
new
feature
where
you
can
order
scale
on
custom
metrics
this
offer.
So
it's
it's
not
stable
yet,
but
I
think
the
plan
is
to
make
the
custom
metrics
API
beer
in
time
for
1/7,
so
I'm
bringing
this
up
now,
because
I
hope
you
are
like.
A
You
are
interested
in
hacking
on
kubernetes,
and
this
is
a
great
place
to
contribute
as
well
or
Vicey.
Eyes
are
the
best
one
to
test
out
and
try
to
see
if
you
find
some
some
limitations.
So
you
want
to
change
something
and
be
involved
in
process
and
then
we'll
have
a
better
Benares.
When
is
API
or
function
functionality
when
it
reaches
via
and
a
sample
app
that
serves
a
matrix,
endpoint
and
auto
scaling
with
capital.
Metrics
is
possible
and
it's
not
possible
really
with
one
six
so
or
it
at
least,
is
very
complicated.
A
A
So
it's
it's
for
hackers
and
makers
of
kubernetes,
basically,
the
horizontal
port
or
the
scaler
controller
of
the
ATI
server
for
the
custom,
metrics
API,
which
in
indeed
is
running
on
cluster,
and
this
is
a
service
just
a
normal
service
that
speaks
the
the
custom,
metrics
API
language
and
my
custom.
Metrics
API
pod
is
asking
from
ETS
for
information
and
for
me
to
create
the
nemetrix.
A
A
Dashboard
is
running
on
a
PI
and
so
forth
it
just
whatever
the
scheduler
chooses
at
one
time,
then
we
have
cube,
see
CL,
get
horizontal
pod,
auto
scalars,
and
we
have
one
now
there's
coming
a
true
quests
a
minute
to
it,
but
if
we
do
some
load
make
some
load
to
the
metrics
and
the
the
sample
application.
I
hope
we'll
see
an
increase
in
some
minutes.
A
A
A
A
A
So,
based
on
some
custom
metrics
that
that
were
exposed
by
the
sample
application,
I
could
order
scale
a
cluster
here
in
real
time.
Just
and
and
now
again
it
doesn't
matter
where
they
land,
the
scheduler
just
increase
the
controller
manager
just
increase
the
replicas
and
they
were
scheduled,
maybe
on
some
pies
or
whatever,
and
this
makes
it
possible
to
to
really
really
do
mixed
workloads
on
in
the
same
cluster
and
yes,
I,
encourage
everyone
here
to
try
also
offer
features
out.
A
Lastly,
the
road
map-
the
I'm
working
with
the
CNCs,
seems
to
make
sure
that
we
have
CI
for
arm
and
on
64
in
the
future,
and
maybe
for
OPC
and
s
390x
as
well,
and
to
extend
the
CN
CF
cluster
and
then
make
sure
are
the
CRI
complaint.
Compliant
vendors
are
supporting
manifest
lists
and
it
would
be
cool
to
create
an
admission
controller.
That
automatically
does
the
scheduling
for
you
that
most
was
my
fault.
Do
we
have
any
questions.
A
A
B
A
At
least
some
something
of
the
conformance
test
run
on
unarmed
hardware,
but
still
we
someone
gotta
pay
for
the
hard
way.
If
we
do
automated
CI,
but
yeah
I
hope
hope
we
could
do
automated
CI
for
all
platforms,
but
still
the
coverage
should
be
increased
and
that's
something
I
would
be
really
grateful
if
some
of
you
could
could
be
good
contribute
to
today's
it's
a
lot
of
tests.