►
From YouTube: SigBigData Meeting Demo (June 14, 2017)
Description
Cassandra on Kubernetes by Matthew Stump
A
B
B
Historically,
we
were
a
shop
or
the
dealt
with
big
data
systems
provide
without
a
word,
I'm
rarely
Cassandra,
but
a
lot
of
Sagara
can
cough.
Hang
things
like
that.
The
Clinton
name
was
the
perfect
and
driver
and
worked
with
custom
Iago
problem,
many
X's
such
an
excellent.
So
the
customers
that
we're
dealing
with
are
really
large
deployments
like
one
of
our
deployments,
us
Cassandra's
around
six
thousand
nodes
and
for
migrating
that's
in
GCE,
and
then
we
we
also
primarily
deal
in
regulated
environment.
Oh
thanks.
B
How
come
one
of
the
issues
that
we've
been
into
is
that
these
people
didn't
have
the
appropriate
staff
to
run
these
cultures
like
there's,
not
a
whole
lot
of
large-scale
Cassandra
people.
So
we
thought
that
instead
of
trying
to
build
these
people,
we
would
automate
it
on
top
of
kubernetes
so
that
it
would
make
it
more.
B
B
We
did
a
couple
of
things.
First,
is
that
staple
sets,
although
they're
very
good
at
getting
us
at
ease
and
reattachment
volumes
and
the
the
naming
conventions
around
cause
making
sure
that
once
husband's
other
times,
if
you
did
things
like
lose
and
EBS
volume,
the
data
would
go
away
and
like
nodes,
wouldn't
know
whether
or
not
they're,
a
new
node
or
a
replacement
zone.
B
There's
a
lot
of
things
that
you
need
to
do
when
doing
rolling,
upgrade
and
flushing
first
or
it's
like
all
the
mechanics
of
actually
running
a
cluster,
and
so
we
put
a
lot
of
effort
into
making
the
please
continue
for
Cassandra
viable
and
we've
been
pushing
that
to
open
sources
well,
and
we
can
put
your
little
utilities
can
make
this
more
viable.
In
addition
to
that
and
I'll
start
sharing,
my
screen.
B
So
we
can
maybe
up
the
Cassandra
staple
set
which
isn't
that
interesting,
but
when
you
dig
into
each
one
of
these
pods,
we
do
things
like
add
additional
containers
around
it.
So
we'll
have
the
Cassandra
container,
but
then
we'll
also
add
in
the
a
log
forwarder
and
which,
in
this
instance
we're
using
file
deep
and
we'll
use
the
JMS
exporter
from
Commedia.
B
Initially,
we
tried
to
just
have
all
the
logs
forward
to
whatever
the
default
log
sources
for
standard
out,
but
one
of
the
things
that
we
ran
into
is
that
there
are
different
teams
that
own
different
namespaces
and
the
ability
to
segregate
logs
and
make
it
visible
to
that
application
developers
that
are
difference
in
the
option.
People
was
necessary
and
then
also
some
of
these
clusters
are
are
secured
in
private
or
may
have
personally
identifiable
information.
Some
of
them
are.
B
B
Standard
out
is
just
in
their
standard
out,
but
then
we
also
will
dump
Jason
encoded
logs
to
a
shared
volume
between
the
Cassandra
container
in
the
logs
container
and
the
log
border
monitors
those
logs
in
for
them
to
a
log
session.
The
other
thing
that
we
went
into
is
that
old
Cassandra
for
jam
max
forwarding.
It
was
by
default
for
pushing
out
something
like
5,000
measures
and
when
you
pull
that
using
the.
B
Via
from
easiest
it
could,
especially
if
the
nose
was
under
load
because
we're
doing
a
lot
of
load,
testing
right
now,
testing
viability
of
media,
the
he
could
kill
the
Prometheus
scraper
I'm.
Sorry,
the
Prometheus
JMX
exporter,
like
you,
did
jetty
errors
like
not
enough
resources,
so
we
had
to
break
that
off
into
its
own
container
as
well.
B
So
it's
really
a
three
container
pod
for
Cassandra
itself
within
the
rest
of
the
ecosystem
we
went
through
and
we
made
it
Selenia
Cassandra,
yes
one,
but
we've
also
done
things
like
create
the
staple
sets
for
a
three-tier
deployment,
action,
three-tier
deployment
of
log
stash
and
I'm.
Sorry,
a
free
tier
deployment
of
a
lawful
search
as
well.
So
we
have
a
client
here,
master
tier.
Then
we
have
a
faithful
steps
of
the
actual
data
notes
as
well
and
then
and
which
makes
that
we
can
get
the
data
notice
up
as
necessary.
B
B
The
and
we've
also
done
the
same
for
from
medias
as
well.
We
are
starting
to
run
into
the
limitations
of
a
number
of
metrics
per
second
that
Prometheus
was
able
to
collect
right
now
that
the
Prometheus
server
with
a
load
of
a
standard
process
under
load
X
coefficients
LGR,
is
routinely
pushing
50
60
megabits
per
second
down
to
disk.
We
are
trying
to
work
on
reducing
the
number
of
metrics
that
are
collected
from
these
employees
to
make
it
smaller,
in
addition
to
Cassandra
we're
doing
the
same
stuff
for
zookeeper
and
Kafka
as
well.
B
C
B
Away
is
in
really
initiative
and
they're
really
really
nice.
The
other
thing
that
we've
been
doing
speaking
on
the
topic
of
is
that,
as
we
are
in
a
regulated
environment,
one
of
the
things
we've
had
to
do
is
get
through
audits,
with
all
the
AWS
roles
and
all
the
containers
and
PCs
and
all
that
sort
of
stuff.
B
So
we've
sponsored
Chris
to
make
it
so
that
we
are
able
to
use
our
own
AWS
rules
and
we
are
able
to
the
audit
the
interesting
that
it's
going
to
spin
up
and
we
do
have
self-hosted
containers
for
the
entire
community
employment.
After
that,
we
can
bring
it
through
ways,
security
skins
and
produce
that,
as
an
audit
report
to
our
customers
right
now,
all
of
our
workloads
are
pre-production,
but
we're
looking
at
moving
a
bank
in
the
asia-pacific
region
to
cops
for
a
Kassandra
workload.
B
As
you
said,
Sandra
I'm,
sorry
that
cue
grenade
uses
internally
to
the
mystical
Sutton
to
the
legacy
network.
So
all
these
people
they
have
existing
Cassandra
deployments.
Those
won't
necessarily
go
away,
at
least
not
initially
ability
and
includes
McLeod
and
Cassandra,
because
they
use
constant
protocol
is,
will
broadcast
the
IP
address
that
it's
that
urban
a
signs
to
it,
but
that's
not
how
real
network.
B
So
what
we're
having
to
do
right
now
is
we
have
to
use
the
first
Network
you
can
cause
like
really
where
we
find
sorry
that
note
that
you
were
reminded
Todd
to
a
node
as
one
of
one
Sandra.
You
know
it
allowed
convert
for
kubernetes
and
that
and
we
just
sized
the
instances
for
freely.
So
we
use
different,
auto
scaling
groups
and
different
affinity
and
anti
penny
rules
to
make
sure
that
they
know
to
get
the
points
to
the
correct,
kubernetes
node
and
that's
a
great
thing.
It
is
dedicated
to
a
production,
exchanger
engineering.
B
We
are
evaluating
whether
or
not
calico,
with
its
support
for
bgp
is
an
alternative
moving
forward,
so
that
we
can
join
that
two
networks
at
the
moment.
We're
just
really
first
lightly
think
what
else
one
of
the
other
interesting
things
is
that
we
were
responsible
for
the
original
recommended
addition
early
on,
because
we've
been
doing
this
a
number
six
years
to
not
use
EBS
and
only
use
in
store
volumes
or
rebound
and
testing
for
both
search
for
floats
and
for
Cassandra
workloads.
That
EBS
is
fast
enough,
especially
the
GPU
instances
and
the.
B
The
io1
instances
the
court
production
will
close
under
look
right
now.
We
will
typically
push
somewhere
between
10,000
and
20,000
PDF
or
cassandra
noted
with
10
gigahertz
dedicate
CPU,
64
gigabytes
and
memories
independent
good
night.
So
it
wasn't
that's
pretty
much
what
I
wanted
to
share.
Oh
one
other
thing
that
we're
doing
so
we
started
off
with
helm.
Helm
has
a
lot
of
issues.
B
The
ability
to
understand
who's
deployed
what
and
understand
what
versions
things
are
on.
Essentially,
all
the
control
that
you
would
want
coming
enterprise
standpoint
that
right
now
is
enforced
through
labels
in
helm
and
Convention,
and
there's
nothing
that
really
stops
somebody
from
modifying
those
up
from
underneath
you
are
not
doing
the
correct
thing
so
we're
switching
to
Service
Catalog
anybody's
workers
for
each
of
these
technologies,
so
that
we
can
not
only
spin
up
the
the
core
service
of
everything
around
it.
One
of
the
things
that
we're
doing
is
good,
generating
operators.
C
B
Developed
it
for
medius
operator
to
manage
all
these
instances
to
automate
the
dividend
events
so
the
chain
don't
necessarily
need
a
operations
manager.
So
that's
what
we're
doing
and
I
can
say
anything.
The
questions
are
to
show
you
code
about
what
we're
doing.
It's,
not
the
most
exciting
demo.
It's
just
watching
vehicle
cuts
going
up.
E
F
E
B
Not
at
the
moment,
I
mean
we're
still
early
on.
It's
pretty
simple,
ready
and
brokers,
pretty
simple!
So
it's
it's
right!
Now
it's
working
for
us.
We
haven't
run
into
issues.
The
main
issue
for
us
has
actually
been
stability
of
QPS
eyes
on
those
pretty
pretty
going
on.
You
know
it's
like
we
have
to
be
monitoring
app
and
you
ready
to
bounce
like
you
ready,
snow
fairly
regularly.
E
F
B
That
we
contributed
x
to
kubernetes,
it
wasn't
doing
any
wreck
wariness,
but
what
we're
doing
is
we're
swing
in
the
labels
from
the
underlying
code
on
the
weeds,
the
move,
cops
deployments
across
three
AG's
and
and
we
use
the
the
affinity
of
rules
or
anti
family
rules
to
make
sure
that
the
nodes
are
placed
evenly
across
the
three
APs.
One
thing
that
we
have
run
into
I
haven't
doubled,
listen,
1/7,
but
there's
a
lot
of
issues
with
paints
in.
B
C
B
There's
not
more
and
so
different
running
here
instance,
groups
dealing
with
the
FAA
become
a
little
bit
like
you.
We
do
plan
on
now
we're
using
just
the
DNS
name
of
the
underlying
pod
to
do
speed
discovery,
but
we're
going
to
switch
that
to
a
Cuban,
Eddy
specific
seed
provider
and
they
proven
a
it's.
A
big
snitch
they'll.
Make
it
easier
to
consume
is
amazing.
Okay,
now
we're
pushing.