►
Description
Although some of today’s cloud properties, like elasticity, scalability, self-healing, durability, DR, etc. seemed revolutionary during the pre-cloud days, Cassandra had many of these properties already. With the ever growing popularity of Kubernetes, the k8ssandra open source project is intended to bring Cassandra's advantages to the cloud and help simplify operations. The biggest challenge in the Kubernetes world today is multi-cloud. Althoug easy to conceptualize, it is hard to implement. After a quick overview of Cassandra, k8ssandra, you will see how the worlds of Cassandra and Kubernetes on the cloud collide in a remarkably cohesive way to incorporate the best of both. In this primarily demo driven session, demoed at Kubecon, attendees will walk away with a good understanding of the k8ssandra project, support multi-region and multi-cloud. Along this journey we will look at one-off deployments on GKE, EKS including multi-cloud (GKE and EKS). On EKS, we leverage EKS Kubefed.
A
A
About
the
in-person
events
was
because
basically
this
is
something
that
has
evolved.
You
know
something,
that's
you
know
I
kind
of
built
the
demo
and
I'm
coming
improving
the
demo
as
we
go
along
right.
So
initially
I
set
it
up
for
multi-cluster
as
you
as
you
all
know,
you
know.
One
of
the
biggest
challenges
in
in
communities
today
is
being
able
to
do
multi-cluster.
It's
not
easy
to
do
that
right.
A
You
know
it
a
little
bit
of
it
depends
on
the
application
that
you're
building
you
know
it
depends
on
on
the
platform
that
you
build
on
top
of
communities
and
all
that.
But
but
you
know,
you'll
see
how
actually
cassandra
and
kubernetes
come
together
and
the
cloud
come
together.
You
know
in
kind
of
a
remarkably
cohesive
manner
and
that's
kind
of
what
I'll
show
in
the
demo.
A
So
the
agenda
for
today
is
basically
I'm
going
to
do
a
quick
intro.
You
know
talk
about
nosql
for
those
of
you
who
have
never
heard
of
nosql,
I'm
sure
you
have
but
but
you
know,
I
think
I
think
it
has
a
lot
to
do
with
the
cloud
as
well.
It's
really
about
horizontal
scaling,
you
know
and
being
able
to
provide
the
durability,
resilience
that
you
know
our
typical
cloud
properties.
A
So
you
know
there
was
the
cloud
before
there
was
the
cloud,
that's
kind
of
how
I
put
cassandra,
then
kate,
sandra
or
sometimes
called
cat
sandra.
You
know
it's
kubernetes
version
of
cassandra
or
cassandra
on
top
of
kubernetes.
Now
again,
the
biggest
challenge
that
we
face
today,
at
least
in
the
communities
world,
is
the
multi-cluster
and
multi-cloud
and
multi-region
and
multi-everything
right,
and
actually
we
have
introduced
a
new
operator
called
gate
center
operator
earlier.
B
A
Had
something
called
as
a
cache
operator
and
I'm
going
to
briefly
touch
up
on
all
of
those,
but
what
the
kate
sandra
operator
does
is
kind
of
sits
on
top
of
cache
operator
and
what
it
does
is
it
deploys
to
multiple
clouds
or
multiple
regions
or
multiple
clusters,
and
so
on.
Okay,
finally,
I'll
do
a
demo
and,
like
I
said
you
know
it's
an
evolving
demo.
A
I
started
it
at
kubecon
in
a
los
angeles
last
year,
where
I
did
a
demo
on
gke,
then
I
had
to
take
the
same
thing
and
do
it
in
aws
re
invent,
but
a
demo
on
gke
is
probably
not
gonna
cut
it
in
aws,
three
invent
right.
So
what
I
did
was
I
adapted
it
for
eks
and
I
used
a
project
called
eks
cube
fed.
You
know
which
takes
care
of
the
networking
aspects
and-
and
the
theme
is,
is
common
across
all
these
different
clouds.
A
A
I
even
teach
in
the
boston
area,
which
is
where
I've
lived,
pretty
much
half
my
life
and
then
one
of
my
passions
is
about
inner
loop
because
I
think
it
makes
a
whole
lot
of
sense.
I
know
in
the
earlier
talk
was
about
flux,
argo,
cd
and
so
on.
That
is
really
more
of
a
robust,
ci
cd
kind
of
loops.
But
but
here
what
we
are
talking
about
is
something
that
is
a
lightweight
inner
loop
and
I'll
talk
about
that.
A
B
A
I
work,
for
you,
know,
data
stacks
as
a
developer
advocate
and
we
do
run
workshops
every
week
pretty
much.
Every
week
we
used
to
run
more
than
one
before
all
of
them
are
free,
feel
free
to
attend
them.
They're,
all
virtual.
We
get
a
lot
of
people
from
india.
We
get
a
lot
of
people
from
u.s.
We
get
a
few
people
from
africa
every
other
continent,
really
not
many
from
australia,
because
the
time
doesn't
work
out
really
well.
But
but
you
know,
if
you
can
attend,
that'll
be
great
as
well.
What
is
nosql?
A
You
know,
it
was
just
something
that
somebody
came
up.
Actually.
I
was
in
san
francisco
in
june,
11
20
2009.
You
know
roughly
around
that
time
where
we
had
to
come
up
with
some
kind
of
catchy
hashtag
for
something
that
is
not
sql
right.
You
know,
sql
is
the
traditional
relational
data
model,
which,
which
kind
of
the
only
way
to
scale
is
to
by
going
up
and
up
and
up
right,
and
at
some
point
it's
going
to
go
so
big
that
it's
going
to
fall
off
right.
A
A
These
days
right,
you
know
I'm
going
to
give
away
my
age
if
I
say
that
my
first
you
know
pc
had
20
megabytes
of
disk
space,
okay,
think
about
that
20
megabytes
of
disk
space,
not
even
giga,
right
and-
and
you
know
I
was
so
happy
when
I
moved
from
the
xt
to
the
80,
which
went
from
20
megabytes
to
40
megabytes.
So
I
doubled
my
disk
disk
space,
which
is
like
just
superb
right.
A
But
but
you
know,
disk
is
very
cheap
these
days
and
that's
why
you
know
it's
better
to
spread
your
data
across
and
that's
exactly
what
nosql
does
in
general,
I
mean
you
know
there
are
other
things
as
well,
but
but
in
general
it's
about
horizontal
scaling,
it's
relational
and
nosql
on
two
ends.
One
is
about.
A
A
Basically,
what
this
states
is
that
you
know,
if
you
think
about
consistency,
availability
and
partition,
tolerance
and
distributed
system
when
there
is
a
failure,
you
know
and
and
failure
is
inherent
in
a
distributed
system
right
it.
It
does
happen.
You
hack
you,
you
cannot
have
all
these
three
properties.
At
the
same
time,
it
cannot
be
consistent
available
and
partition
tolerant.
A
You
have
to
sacrifice
at
least
one
of
the
three
right,
and
this
is
very
critical
and-
and
if
you
think
about
this,
you
know-
and
we
can
go
at
length
about
this,
but
partition
tolerance
actually
is
worse
than
any
of
the
other,
or
rather
sacrificing
partition
tolerances,
worse
than
sacrificing
consistency
in
sacrificing
availability,
if
it's
not
available
for
a
few
seconds,
maybe
even
a
few
minutes,
it's
probably
okay.
If
it's
not
consistent
for
a
few
seconds
or
a
few
minutes,
it's
probably
okay,
because
eventually
it'll
become
consistent.
A
If
and
not
on
the
leaderboard
immediately,
it's
not
a
big
deal
right.
You
know
it
will
eventually
get
there,
but
if
you
lose
partition
tolerance,
then
you
know,
even
my
grandmother
can
make
out
that
you
know
the
data
is
not
consistent,
a
little
bit
more
about
cassandra.
What
it
does
is.
You
know
it
uses
a
a
ring-like
architecture
as
it's
referred
to
and
and
if
you,
the
oversimplified
way
of
looking
at
this,
is
that
what.
B
A
Is
it
takes
data
right?
You
know
it
uses
something
called
as
a
partition
key
and
it
hashes
the
partition
key
into
a
particular
node
into
this
node.
You
know
ring
of
notes
now
what
happens
when
you
add?
You
know
nodes
and
subtract
nodes
and
all
that
right.
The
system
takes
care
of
all
of
that.
So
so
again,
the
oversimplified
view
is
kind
of
look
at
it
as
like
a
a
distributed
hash.
You
know
table
or
whatever
you
want
to
call
it
right,
except
that
the
platform
takes
care
of
all
these
failures.
A
B
A
Care
of
the
sharding
itself
sharding,
it
turns
out
to
be
very
hard,
so
you
know
leave
it
up
to
the
platform.
Now,
a
lot
of
lot
of
you
know:
big
companies
use
cassandra.
In
fact,
cassandra
originated
in
facebook,
and
facebook
is
still
one
of
the
contributors
to
cassandra
and
and
again
you
know
what
is
the
big
deal
about
talking
about
cassandra,
which
is
already
maybe
about
13,
14
years
old
right?
The
point.
B
A
B
A
And
thousand
nodes,
you
know
your
your
basically
your
response
or
latency
throughput
or
whatever
you
know,
should
scale
linearly
and
in
many
systems
they
do
until
a
point
and
after
that
it
kind
of
levels
off
with
cassandra
it
keeps
on
taking.
You
know
it
keeps
on
going
and
it's
very
linearly
scalable.
A
You
know
without
without
too
much
effort
now.
Having
said
all
that,
what
is
kate
sandra,
which
is
you
know,
because
we're
talking
about
kubernetes
here
right?
Can
you
really
run
a
database
on
communities?
This
was
one
of
the
articles
that
one
of
our
colleagues
right
now
christopher
bradford,
who
was
on
the
other
side.
You
know,
did
not
work
for
datastax.
At
that
time,
basically
wrote:
can
you
really
run
a
database
and
communities
now?
A
This
was
about
four
or
five
years
ago,
and
and
obviously
a
lot
of
things
have
evolved
and
you
will
see
you
know
today
that
there
is
something
called
as
the
cass
operator
that
we
rely
on.
So
a
lot
of
things
have
changed,
but
today
I
you
know
definitely
data
and
communities
kind
of
go
together,
and
you
can
talk
about
that.
It's
not
just
about
stateless
apps,
it's
not
just
about
12
factor
apps.
You
know,
if
you
remember
all
that,
but
but
you
know
it
is
really.
A
You
know,
prime,
for
you
know
putting
data
on
kubernetes.
One
of
the
cool
things
about
kubernetes.
Again,
you
know
is
whether
you
want
to
do
it
on
on-prem
or
whether
you
want
to
do
it
on
the
you
know,
different
clouds.
I
I
think
I
saw
siebel
in
the
previous
session.
So
really
it
doesn't
matter
right.
You
know
very,
very
straightforward.
A
You
want
to
install
it
on
your
laptop.
You
want
to
install
it
on-prem
everything
you
know
just
pretty
much
works
the
same
right
to
be
able
to
install
cassandra.
We
install
a
number
of
different
components,
and-
and
you
can
see
here
that
we,
you
know
have
all
of
these
installed
with
a
you
know
something
called
as
a
cache
operator.
You
can
see
here
in
the
right
the
cache
operator,
installs
cassandra.
A
A
A
Very
simple
right,
so
this
is
it
for
a
single
cluster.
Now
what
happens
in
the
case
of
a
multi-cluster
right,
so
we
introduced
a
operator
called
cate
center
operator
and
we'll
see
that
today
we
basically
what
happened
was
we
pushed
him
to
the
limit
and
then
built
a
new
operator
called
the
kit
center
operator,
okay
and
and
if
you
want
to
understand
a
little
bit
of
the
rationale
of
why
we
did
this,
you
can
kind
of
look
at
this
particular
article
on
newstack.
A
Why
multi-y
multicluster
cassandra
has
always
been
designed
for
multi-region.
You
know
it's
partition.
Tolerance
is
extremely
important
for
us.
So
even
if
one
of
the
you
know
the
nodes
goes
down
or
maybe
a
set
of
nodes
goes
down,
you
still
should
be
able
to
service
the
request
from
you
know
from
the
customer
right
nodes
automatically
route
traffic
to
the
nearby
neighbors.
You
know
they
use
something
called
as
a
gossip
protocol
and
really
the
theme
here
is
going
to
be
once
you
set
up
the
networking
everything
just
kind
of
falls
into
place.
A
Okay
data
is
automatically
and
asynchronously
replicated
and
really
from
a
cassandra
perspective.
The
cluster
is,
you
know,
very
homogeneous
humanities,
on
the
other
hand,
was
not
designed
for
multi-region.
There
were
some
design
decisions
that
were
consciously
taken.
You
know,
I
recently
did
a
panel
on
networking
for
kubecon
and
essentially
even
there
right.
You
know
there
were
certain
design
decisions
that
you
know.
Sometimes
the
designers
probably
thought
you
know
are
right
now,
considering
maybe
they
probably
should
have
had
it
back
right.
A
There
was
it's
really
about
you,
know
kind
of
taking
this
apps
today
and
and
adapting
it
to
the
cloud
and
making
it
in
such
a
way
that
it
works
on.
You
know
all
the
clouds
and
so
on.
That's
really
the
goal
for
communities
it
was
really
built.
You
know
it
was
a
platform
for
building
platforms
and
was
not
you
know
designed
for
multi-region.
So
how
do
you
adapt?
You
know
humanities
to
be
able
to
do
this,
for
multi-cluster,
multi-region
multi-cloud
and
so
on
is
using
the
what
we
call
as
the
kate
sander
operator.
A
A
So
how
does
this
work
in
terms
of
you
know
the
kubernetes
pures
right?
Essentially,
what
you
do
is
you
inject
the
configs
of
these
two
clusters,
the
data
plane
clusters
into
the
control
plane
so
that
the
control
plane
is
able
to
manipulate
the
data
plane
and
that's
really
what
you
know
how
it
works.
So
what
I'll
do
is
I'll
in
my
demo
I
will.
A
Just
a
few
weeks
back,
I
used
the
kate
sander
operator
to
do
the
demo.
I
will
show
you
the
evolution
during
the
demo.
Okay,
but
essentially
there
is
a
concept
of
what
is
referred
as
a
kit
sandra
cluster,
and
here
you
specify
what
is
the
context
you
know
remember
the
humanities
context
I
was
talking
about
right.
One
is
the
east
context
here
and
the
other
is
the
west
context,
and
basically
you
can.
You
can
set
up
each
of
these
data
centers.
A
You
know
I
you
will
see
in
my
example,
I
set
up
these
different
contexts
and
each
of
those
will
have
their
own
kind
of
disk
provider,
because
you
know
each
of
these
clouds
have
different
disk
providers
and
all
that
and
you
can
tweak
all
that
and
then
finally
install
the
kit
sandra
cluster
via
the
case
center
operator.
Okay-
and
this
is
kind
of
it
talks
about
how
you
can
do
this
inject
the
client
configs
into
the
you
know
into
the
data
I
mean
into
the
control
plane.
A
A
The
critical
part
of
this
multi-cluster
with
cassandra
is
to
be
able
to
set
up
the
networking
in
a
way
that
all
of
this
happens
pretty
much
automatically
right
for
gke,
which
is
where
I
started.
It's
pretty
straightforward
to
do
that
on
on
eks,
not
so
much
right.
Luckily,
for
me,
what
I
did
was
I
I
stumbled
upon
upon
a
project
called
eks
cube,
fed,
which
is
based
on
cube,
fed,
but
is
really
for
eks
and
what
it
does
is
it
sets
up
all
of
this.
For
you,
it's
a
bastion
host.
A
A
And
then
I
had
a
cluster
two,
which
is
rags
fed22
at
eu
central
one.
You
know
this
is,
I
believe
in
dublin,
and
this
is
in
frankfurt
and
you'll
see
both
of
these
in
action.
The
nice
thing
about
ek's,
cube
fed
is
that
it
sets
up
all
this
networking.
I'm
not
a
networking
guru
myself,
but
once
you
set
this
all
up,
it's
very
easy
to
install.
You
know
cassandra
because,
like
I
said
what
it
does
is
it
looks
for
the
neighboring,
node
and
kind
of
builds
the
cluster
and
forms
a
big
cluster.
A
So
all
of
this
happens
automatically.
Okay.
So
with
that
said,
let's
see
if
I
can
jump
to
the
demo.
This
is
the
multi-cluster
with
the
gate.
Center
operator
goes
into
a
lot
more
details
of
exactly
what
it
does
like.
You
know
it
has
the
different
contexts
that
are
injected
here
and,
and
you
know,
the
different
components
that
can
be
adjusted
and
so
on.
But
let's
not
worry
about
too
much
about
that.
Let's
go
to
the
demo
and
see
how
it
goes
all
right.
A
A
A
So
you
can
see
here
that
I'm
I'm
gonna
show
the
context.
Okay,
so
you
can
see
here.
There
are
two
clusters:
one
is
the
fed
two
one
and
the
other
is
the
fed
two
two
two
right.
So
all
that
I
need
to
do
is
kind
of
look
at
this.
We
just
describe
nodes
and
get
the
topology
right
and
you'll
see
here
that
this
is
in
eu
central
one
1b1a1c.
A
You
will
see
all
these
different.
You
know
like
the
reaper,
the
different
components,
the
dc
twos,
which
are
in
different
tracks,
and
then
you
will
see
somewhere
here
that
it's
running
the
cache
operator
underneath
okay.
So
so
so
that's
exactly
what
is
what
is
going
on
in
in
this
particular
node?
Now,
if
I
use
context-
and
if
I
go
to
the
other
cluster
right,
you
will
see
if
I
do
the
same
thing.
If
I
look
at
the
topology.
A
There's
a
white
okay,
and
this
might
give
us
a
little
bit
more
insight
into
this.
So
you
can
see
here
it's
all
172.21,
right,
172.21
and
so
on.
So
if
I
go
back
to,
for
example,
you
know
two
two
you'll
see
that
those
are
probably
172
or
22..
So
all
of
this
networking
aspects
and
all
that
are
set
is
set
up
by
eks
and
I
really
didn't
have
to
do
anything
okay,
so
this
is,
as
far
as
you
know,
just
eks
is
concerned.
A
Called
beatrix,
and
essentially
what
happens
here
is
it
it
enables
the
networking
very
similar
to
what
happened
in
eks
right.
What
this
does
is
it.
You
know
it
sets
up
the
multi-cloud
and
you
can
see
here
for
aks.
It
sets
up
10.2
for
eks,
it
sets
up
10.1
and
for
gk
it
sets
up
10.3.
It
does
all
the
you
know,
the
routing,
the
peering
and
all
that
so
to
look
at
this
multi-cloud.
A
A
You
know
if
I
do
the
same
thing
again,
you
know
if
I
look
at
the
parts
and
if
I
look
at
the
operator,
okay
you'll
see
that
you
know
again.
This
is
10.1
which
is
for
eks,
okay,
and
you
can
see
here
that
the
you
know,
cassandra
control
plane
is
said
to
be
false,
so
this
is
one
of
the
data
planes
right.
So
likewise,
if
I
go
to
the
gke
cluster
right
same
thing,
you
will
see
the
same
thing
as
well.
A
Hopefully,
right
oh
yeah,
dude,
look
at
the
parts
I
mean
look
at
the
cage
center
operator
and
you'll
see
here
that
you
know
the
control
plane
is
false
and
essentially
you
know
all
of
this
is
being
managed
by
the
by
the
kid
center
operator.
Okay.
So
now,
if
I
want
to
take
a
look
at
the
unified
cluster
right,
what
I
can
do
is
I
can
do
a.