►
From YouTube: Movile: Thoughts About Growing and Scalability — Bumping up Startup Business with Apache Cassandra
Description
Speaker: Eiti Kimura, Software Engineer at Movile International
Apache Cassandra was adopted by Movile in 2009, and became a fundamental piece within the robust and scalable architecture to support more than 50 products, impacted by over 200MM users in Latin America. In this case we present the architecture of our ring, configuration details, detailed tuning, hardware used to be able to achieve our performance requirements (order of a few milliseconds), information storage strategies for network and disk space optimization, and best practices, in addition to showing the evolution of the architecture of simple systems to become scalable and distributed platforms. We introduced our cluster with a relatively low number of nodes (6) using commodity hardware to support critical high-performance applications. After this talk, you'll understand how Apache Cassandra was essential to evolve our systems and leverage the growth of our business. Movile is the leading mobile content company in Latin America. Movile’s products include mobile content, mobile TV, mobile learning, mobile games, mobile payment, mobile marketing and mobile commerce. Every month, it publishes content and services to more than 20 million mobile costumers. It has grown substantially over the last few years (with a more than 25-fold increase in its revenue over the last five years) both organically and through an aggressive M&A strategy, including five acquisitions in the last five years. Movile is positioning itself as a kind of Silicon Valley company based in Brazil. For the last two years, Movile has been named in the “Great Place to Work” list for technology companies in Brazil. The company shareholders include the founders of the company plus Naspers, a South-African media conglomerate.
A
A
A
A
What
was
that
asking
my
boss?
Maybe
a
manga
monkey,
an
electrical
current
from
a
dynamo.
He
questioned
it
or
a
hornet
bus
or
a
rabbit.
A
raptor,
nope
I
replied
it's
something
really
fast
that
can
see
beyond
what
our
eyes
can
see.
Unbelievable
he
yelled
excitedly,
yes!
Well,
as
I
said
it's
Apache
Cassandra
so
good
morning,
everyone,
my
name,
is
HD
camera
from
Mali,
Brazil
and
I'm,
going
to
show
you
a
little
bit
about
our
use
case
in
how
we
use
Cassandra
and
morally
to
start
up
to
bump
up
our
startup
business.
A
Here
are
some
of
our
products:
we
have
the
play,
kids
offer
hours
of
Ages,
appropriate,
fun
and
educational
shows
and
have
books
when
entertainment
for
kids.
It's
a
our
app
and
I
consider
as
a
top
kids
app
in
Latin
America
Brazil
in
some
countries
over
world,
and
so
we
have
the
free
zone.
They
can
scan
and
connect
you
to
open,
Wi-Fi
hotspots.
A
So
we
have
the
I
food.
You
can
easily
ask
food
using
your
smartphone,
which
doesn't
click,
it's
pretty
easy
and
fun,
and
we
have
ze
to
find
coolest
videos
on
the
web
and
see
what
our
friends
are
seeing
you
can
share,
and
so
it's
very
nice
and
so
for
companies.
We
have
solutions
for
learning
through
news
games
and
TV
with
cover
business
right,
so
I'm
going
to
show
a
little
bit
about
our
platform.
Them
are
gonna
present
you
the
architecture
here,
so
we
have
a
subscription
and
billing
platform.
So
it's
a
rest-based
service
API.
A
It
handles
subscriptions
in
charge
and
carers
and
it's
responsible
to
renew
the
expired
subscription
as
well.
The
system
must
be
at
fault,
tolerant
and
platform
is
basis
for
the
camera
business
to
build
service
products
in
application.
It's
responsible
to
keep
running
our
car
business
with
carers.
It's
a
pretty
critical
platform.
A
Now
some
numbers
we
have
about
50
million
subscriptions
right
now
and
we
have
our
API
receiving
about
50
requests
per
second,
where
the
main
operations
are
the
SUBSCRIBE
cancer
and
profile,
and
our
renew
engine
has
a
it
does
about.
52
million
of
billion
tries
a
day
it's
about
60,
26,
603
requests
per
second
to
the
carers,
and
this
is
how
our
how
our
data
are
distributed.
A
Most
subscriptions
are
concentrated
in
Brazil,
which
representing
green
in
Mexico
in
purple
and
on
the
rights
on
the
right
side,
so
the
spots
are
in
the
countries
where
we
have
subscribers
in
South,
Central,
America
and
a
few
in
Norfolk
Erica.
So
now
that
we
have
an
idea
about
the
nature
of
data,
the
volume
and
how
we
use
it,
let's
see
how
simple
architectures
have
to
evolve
to
accommodate
and
process.
A
large
volume
of
data
working
and
in
a
distributed
fashion
is
a
quote
for
from
Martin
father's
book.
A
There
isn't
just
one
way
to
state
a
system
architecture.
Rather
there
are
multiple
architectures
in
a
system.
Interview
of
what
is
architectural
significance
is
one
that
can
change
over
the
system's
lifetime,
so
the
systems
are
evolving.
All
time
the
requirements
change
your
business
change,
we
need
to
be
prepared
to
handle
this
change.
A
Well,
here
I
have
a
very
basic
architecture:
it's
simple:
it's
common:
it
works
for
the
most
low
traffic
and
low
processing
applications.
That's
the
way
we
started
in
about
I.
Think
in
2007.
We
have
this
kind
of
architecture.
Yes,
it's
working,
then
we
have
to
to
increase
the
load,
increase
the
people
accessing
data
concurrently,
and
we
have
our
renew
system
attached
it
to
the
relational
database.
This
model
was
was
working,
but
what
our
user
database
was
growing
and
growing,
but
something
changed.
A
You
know
our
requirements
change,
so
we
had
to
to
be
scalable
high
availability
in
high
performance,
so
the
first
step
we
did
was
just
to
scale
and
do
a
scale,
our
application
layer
and
we
had
a
cluster
of
service.
Ap
is
connected
to
our
relational
database.
Our
database
starts
to
suffer
in
this
times.
A
Then,
with
a
very
high
usage,
the
system
goes
very
slow.
Okay,
very
slow
system
responds
so
overall
throughput
decreased,
low
availability,
single
point
of
failure
and,
even
worse,
then,
stopping
us
to
only
work.
Sometimes
you
can
see
it
so
in
the
center
of
the
image,
the
database
is
burning.
It's
our
bottleneck,
okay,
so
one
lessons
learned
here
is
that
you
must
you
have
to
to
scale
your
your
application
layer,
but
you
have
to
scale
your
database
storage,
your
cash,
your
cues.
You
have
to
scale
as
well
to
follow
your
system
and
business
growing.
A
A
A
Okay,
so
now
you
can
see
how
the
Cassandra
fits
to
our
solution
to
make
our
applications
be
more
kind
of
a
consistent
increase,
availability
and
I'll.
Show
you
a
little
bit
how
we
did
is
so
how
we
change
our
data
model
to
try
to
achieve
the
this
kind
of
setup,
it's
kind
of
architecture
and
how
we
did
our
hardware
configuration
and
a
little
bit
about
our
cluster.
A
A
There
are
three
different
ways
to
modern
solution
that
there
are
different
ways
to
model
solution.
There
is
no
right
or
wrong
way.
There
is
the
one
that
best
works
for
nodes
in
requirements.
So
it's
very
important.
You
need
to
model
according
to
your
requirements.
According
to
what
do
you
need
how
you
are
planning
to
retrieve
data
from
the
database
so
try
to
model
thinking
about
and
how
you
can
retrieve
the
arekhaka
can
write
to
do
to
match
to
our
performance
requirements.
A
Plenty
of
queries
and
how
you
need
to
get
the
information
before
modelling
use
it's
a
driver
for
modeling
decisions,
so
I'm
sure
you
so
two
versions
of
data
models,
the
version
one
that
we
use
at
the
beginning,
using
Cassandra
and
the
version
to
an
improved
version
that
we
have
to
move
to
to
get
more
performance
that
we
need.
So
our
version,
one.
A
So
here
the
CQ
oil
syntax,
all
right
I
have
so
this
model
relates
to
the
users
and
their
subscriptions.
It's
consistent.
It's
consistent.
It
looks
like
a
traditional
relational
database
model,
and
so
we
have
subscribers
the
people
that
has
subscriptions
to
our
products
to
our
service
and
so
on,
and
that
was
the
first
representation
we
used
and
here
the
tables
with
some
data.
A
So
it's
a
relationship
of
one
too
many.
The
users
are
identified
by
their
by
the
non-id,
it's
the
the
phone
number.
Then
each
user
has
maybe
one
or
more
subscriptions,
and
this
is
our
first
model
representation.
So
it
works
it's
okay,
but
so
we
map
that
the
the
most
used
function
of
our
API
or
was
the
the
profile
one
to
get
a
user
subscriptions
and
a
status
update
each
subscriptions.
So
I'm
going
to
show
you
here
to
simulate
how
to
get
data
to
to
mount
a
profile
response
in
our
api.
A
A
A
Noticed
I
had
to
do
a
query
per
record
that
has
returned
it
from
the
previous
query.
I
have
the
index
I
have
the
subscription
IDs
from
a
given
user
and
I
have
to
query
the
subscription
taped
able
to
get
the
subscription
for
this
user.
So
I
have
three
queries
here,
but
if
in
case
that
this
user
has
more
subscriptions
like
I,
don't
tan
or
something
like
I
have
to
do
a
lot
of
queries
to
my
table
and
it's
it's
very
expensive.
A
It's
a
kayak
snow
is
known
as
client
signed
joints,
and
you
have
to
do
so
lots
of
a
lot
of
remote
calls
and
it
costs
in
terms
of
performance.
It's
it's
a
bad
pratt
see.
So,
yes,
we
changed
a
bit
about
our
data
model
and
are
in
our
data
model
version
2.
Now
that's
what
we
have.
We
have
just
only
one
table
for
subscriptions
that
represents
relate
subscriptions
and
its
subscribers.
Okay.
Now
I
still
have
my
phone
number
that
identifies
a
user.
They
subscription
ID
and
here
the
serialized
blob.
A
It's
a
new
design
decision
that
we
take.
So
it's
we
decide
to
to
serialize
all
subscriptions
object
into
a
blood
feud.
We
hadn't
about
50
about
50
fields,
attributes
and
also
description,
but
we
decided
to
serialize
it
to
keep
the
data
it
smaller
to
traffic
network
traffic
and
so
on.
So
in
this
model,
so
I
have
this:
is
our
the
data
in
our
new
model
so
notice
that
now
I
have
the
serial
is
serialized
data?
A
It's
a
array
of
bytes,
okay
in
no
more
diffuse
of
the
values
so
to
query
now
the
upper
5
/
user,
it's
become
so
pretty
easier.
Now
I
can
do
just
a
select
star
from
subscription
subscriptions
where
phone
it's
like
that
phone
number
I
have
my
matches
and
I
have
all
of
data
that
I
need
in
my
hands.
I
don't
need
to
do
so
a
lot
of
other
requests
or
queries,
but
it's
a
very
pretty
better
model
than
the
previous
one.
A
So
that's
a
question
of:
why
did
we
decide
to
change
our
stores
or
storage
strategy
to
and
start
to,
serialize
those
fields?
It's
it's
simple.
It's
a
part
of
our
storage
strategy,
so
represent
information.
So
it's
complex.
It
depends
on
your
proposed
and
how
we
want
to
use
it
okay,
so
we
tried
some
various
ways
to
store
information.
It
optimizes
network
traffic
as
well
it
with
this
solution
that
we
we
have
adopted.
A
So
here
have
the
types
of
object,
representation
and
in
the
object
size
and
bites
the
same
subscription
representation
in
xml
java
by
Ray,
J's
on
and
Google
protocol
buffer.
So
now
we
have
this
as
our
solutions
to
use
the
Google
protocol
buffer,
it's
about
5,
500,
500
bite
and
it's
as
very
small
objects
to
go
to
the
database
and
the
result
to
use
in
the
sky
dive
approach.
You
can
see
right
here
where
we
have
our
database
size
and
our
storage
strategy
sizing
gigabytes.
A
A
Well,
data
model,
so
performance
increases
significantly,
so
reduce
it
complex
from
two
tables
to
one
table
it's
simpler.
It's
lighter
so
reduce
the
number
of
remote
cause
in
our
data
model
version.
One
I
had
one
query
to
index
table
and
X
queries.
21
/
index
return
it.
So
in
our
version
2,
we
just
have
one
one
query
that
brings
all
all
of
the
data
and
the
data
one
has
reduced
it.
So
that
shows
how
how
important
it
is
to
how
do
you
do?
Did
you
plan
your
day
tomorrow
base
it
on
your
needs,
all
right.
A
A
A
So
we
had
now
have
a
very
good
hardware
infrastructure,
but
so
I'm
sure
you
I'm
going
to
show
you
a
version
one
of
our
hardware
infrastructure
and
then
improve
it,
one.
It's
a
call
diversion
tube
in
our
first
experience
with
Cassandra.
We
just
had
so
far
servers
running
with
once
one
SATA
disk
m1
SSD
with
Apache
Cassandra
version,
one
on
106.
A
Ok
and
that's
our
first
configuration
it
works
very,
very
nice
and
it
was
able
to
handle
about
40
million
subscriptions
and
subscribers
with
no
problem.
So
we
decided
to
to
upgrade
our
cluster
to
scale
it
up
and
scale
it
up
scale
it
out
as
well.
So
the
new
version
of
infrastructure
configuration
is
this
one.
Now
we
have
six
servers
and
six
virtual
machines.
A
A
Alright,
our
key
space,
our
key
space,
is
configured
so
with
a
replication
factor
of
three
per
data
center.
Here
are
calm
families
we
have
subscription
organized
by
subscribers,
delivered
tickets,
use
for
a
synchronous
billing
with
carriers,
so
Harley
control
and
rules
are
used
for
safety
mechanism
to
prevent
over
charts.
We
have
an
index
config
too.
It's
a
lookup
table
to
find
subscriptions
for
a
given
configuration,
so
user
directives
used
restore
our
blacklist
and
white
list
information.
A
A
So
our
overall
cluster
response
time
we
have
I
just
separated
to
the
data
center
one
and
data
center
too
so
pay
attention
to
the
right
and
read
latency
per
node
that
I
have
in
our
total
number
of
operations
per.
Second,
it's
about
thirteen
thirteen
thousand
reads
and
nine
thousand
writes
per
second
or.
A
Item
of
operation
operating
system
and
software
stomach
customizations.
According
to
the
Cassandra
docks
the
recommended
recommended
settings
for
production.
We
are
using
Oracle,
Java,
1.7
plus
je
na
configured
with
Zabel,
the
swap
and
the
install
NTP
servers
in
all
of
servers
to
synchronize
the
server's
clock.