►
From YouTube: C* Summit EU 2013: Real-Time BI with Cassandra
Description
Speaker: Tim Moreton, Founder & CTO at Acunu
Cassandra is a superb platform for building scalable real-time analytics applications. And the highest value applications will be those that put simple visualizations of critical KPIs in the hands of business users.
But there's a tension between your development team crafting complex, denormalized data models in CQL and your business users constantly evolving their understanding of the data and their analytic needs.
Acunu Analytics helps to resolve that gap by using Cassandra to continuously maintain OLAP cubes that support instant, SQL-like queries and offering those up through a BI dashboarding tool.
In this talk, I'm going to explore these concepts, and dig into how we've automated the process of going from a stream of JSON events to a live dashboard of the indicators business users need.
B
I
think
it
was
April
the
year
before
last
that
we
first
announced
the
beta
of
hakuna
analytics
witches
our
product
to
to
enable
this
sort
of
real-time
bi
that
runs
over
Cassandra
and
is
powered
by
Cassandra
and
I'm
sort
of
going
to
tell
you
a
little
bit
about
what
we've
learned
about
real-time
use
cases
the
best
way
of
using
cassandra
for
this
sort
of
for
this
sort
of
purpose.
So
I
guess
the
first
question
that
we
often
hear
is
really.
B
When
should
you
be
deploying
a
no
sequel
solution,
and
when
should
you
be
looking
towards
a
Hadoop
solution
for
solving
analytical
problems?
You
know
you
have
some
big
data
workload
in
some
dimensions,
either
velocity
variety
or
volume
and
the
the
first
question
that
people
often
ask
is
well.
You
know
what
is
the
sort
of
set
of
tools
that
I
should
be
using
to
try
and
tackle
those
those
problems,
and
we
like
to
categorize
these
things
into
two
broad
sets
of
you
cases,
so
I
talked
about
over
on
the
left
here,
discovery
analytics.
B
B
So
it's
the
use
of
this
sort
of
unstructured
data
warehouse
where
your
bi,
guys
or
perhaps
they're
called
data
scientist
now,
because
you
need
them
to
be
able
to
program
come
at
a
large
set
of
data,
perhaps
put
together
disparate
sources
and
types
of
data
and
look
for
I
guess
mainly
unknown
unknowns.
You
know
the
sort
of
so-called
needles
in
a
haystack
to
try
and
inform
business
processes
that
perhaps
you
need
to
change
months
down
the
line.
B
So
how
should
I
change
my
supply
chain
based
on
the
analysis
of
marketing
results
and
meteorological
data,
that
sort
of
use
case?
It's
where
the
data
sort
of
supposed
to
sing
to
you
I'm,
not
really
that
much
of
a
believer
in
data
necessarily
singing
to
you
wiih,
weird,
a
we
focus
a
lot
more
on
what
I
would
describe
as
operational
intelligence.
So
that
is
where
you
have
some
idea
in
advance
about
the
sort
of
metrics
that
matter
to
your
business
or
to
your
operation,
your
your
systems
and
we're
five
minutes
matters.
B
So,
where
there's
a
sort
of
better
decision,
but
a
business
decision
you
could
make
on
the
back
of
having
access
to
timely
information
or,
where
there's
corrective
action
that
you
can
take
if
something's
going
wrong.
If
you
could
predict
it
either,
just
in
time
or
potentially
proactively,
using
predictive,
analytics
and
spot
something
ahead
of
time
going
wonky
so
that
you
can,
you
can
fix
that.
So
that's
really
the
area
that
we're
we're
really
interested
in
and
I.
Think
that
you
can.
You
can
characterize
these
things
in
two
ways.
B
Really
you,
over
on
the
left,
you've
got
complex,
long
running
analytics
a
total
lack
of
structure
on
the
right.
You
know
the
structures
of
des
trous,
typically
well-known
you're
talking
about
click
streams,
internet
of
things.
You
know
telemetry
data
from
telcos
financial
market
data
and
by
the
way
those
are
all
applications
that
hakuna
analytics
has
been
deployed
against.
B
Since
we
launched
the
beter
18
months
ago
and
across
a
wide
range
of
verticals
and
there
you
know
you
don't
need
a
Hadoop
cluster
if
you're
an
advertising
to
tell
you
that
click
through
rate
is
an
important
metric
for
your
business.
If
you're
doing
infrastructure
and
operations
stuff,
you
know,
you
know
that
the
latency
of
API
calls
is
something
that
you
should
be
watching
and
if
you're
doing
manufacturing
and
you're
doing
production
line
analytics,
you
know
that
the
pass/fail
rate
is
obviously
something
that
you
should
have
a
good
handle
on.
B
You
don't
need
a
Hadoop
cluster
to
tell
you
that.
So
when
you
have
some
insight
into
the
metrics
matter,
this
is
where
I
think
Cassandra
comes
into
its
own
as
a
platform,
and
you
know
we,
we
love
Cassandra
de
kunu,
we
organized
the
last
Cassandra
you
and
we've
been
working
closely
with
datastax
on
on
this
one.
We've
been
working
with
Cassandra
and
contributing
to
it
for
a
number
of
years
now
and
I
think
some
of
the
key
things
that
we
like
about.
B
It
are
common
to
all
use
cases
for
Cassandra,
but
in
particular,
resonate
around
analytics.
So
you
know
Cassandra's
very
scalable.
It
has
no
no
single
point
of
failure.
I
there
in
terms
of
actual
availability
or
in
terms
of
or
in
terms
of
performance,
either
so
degradation
in
performance
in
10
can
often
be
worked
around
pretty
pretty
easily.
It's
very
fast
and
in
particular,
is
very
fast
for
rights
and
that's
great
for
real-time
analytical
workloads.
If
you
think
about
what's
happening
there,
you
have
this
big
stream
of
data
coming
at.
You
is
typically
immutable.
B
So
the
fact
that
Cassandra
doesn't
really
have
you
know
very
strong
transactional
semantics
can
can
be
sort
of
like
set
to
one
side.
It
doesn't
really
matter
as
a
good
design
trade-off
for
this
particular
application,
and
you
know
very
often
we
see
a
rate
of
rights
of
about
a
thousand
to
one
rights
to
reads
so.
People
are
often
collecting
this
huge
stream
of
data,
not
often
updating
the
various
elements
or
events
that
are
coming
in
and
trying
to
summarize
or
analyze
the
sort
of
the
output
of
that
stream.
B
So
you
know
in
a
basic
sense:
they're
counting
it
that's
the
sort
of
fundamental
building
block
their
of
analytics
and
Cassandra,
of
course,
has
great
counter
support.
That
was
a
feature
that
was
originally
contributed
by
Twitter
for
one
of
the
for
basically
an
analytical
application,
real-time
analytics
application
that
they
were
building
on
on
top
of
Cassandra.
B
So
you
cassandra,
cassandra
is
a
great
fit
for
those
reasons
and
what's
become
really
clear
to
us
in
the
last
12
months
of
we've
been
working
with
cassandra.
Customers
is
that
it
is
really
arrived
in
terms
of
being
deployed
for
mission-critical
use
cases
in
mainstream.
What
I
would
characterize
that
you
know
probably
traditionally
quite
risk
averse
organizations
and-
and
that's
that's
very
interesting-
cassandra
is
really
here
right
now
for
those
use
cases,
so
the
some
of
you
may
be
familiar
with
this.
B
This
cartoon-
and
you
know
no
sequel,
giveth
and
no
sequel,
taketh
away,
and
one
of
the
things
that
you
do
encounter
when
you
move
to
a
no
sequel
platform
is
quite
often
you.
If
you're
coming
from
a
relational
background,
you
lose
the
you,
lose
the
ability
to
you
lose
the
sort
of
live,
the
availability
of
a
very
rich
query
framework.
So
this
is
a.
This
is
a
great
cartoon
which
sort
of
wood
has
been
around
since
the
early
days
of
no
sequel
and,
of
course,
it's
poking
fun
at
our
friends
over
react
and
bash
a.
B
But
you
know
the
point
still
really
does
stand
for
the
Cassandra
and
by
the
way,
I've
not
ever
seen
a
Cassandra
or
react
developer
in
a
suit,
but
there
you
go
so
the
point
still
stands
with
Cassandra
as
well
right.
You
know
so
a
key
value
interface,
regardless
of
the
sort
of
additions
and
great
enhancements
that
cql
3
has
added.
B
So
you
there
are
still,
there
are
still
barriers
there
to
use
Cassandra
just
on
its
own
for
building
these
analytical
applications.
So
this
is
what
we
originally
built
hakuna
analytics
to
tackle.
We
said
okay,
so
the
database
Cassandra
xert,
superb
database,
but
the
notion
of
what
is
the
database
perhaps
is
going
to
layer
down
so
rather
than
talking
about
a
system
where
we
can
ask
rich
queries
of
it.
A
database
is
something
that
manages
your
data
in
terms
of
getting
you
great
performance,
great
availability
and
durability
across
across
data
centers
and
the
query
layer.
B
I
guess,
is
something
that
you
want
to
build
on
top
of
it.
So
we
built
hakuna
analytics
to
keep
to
work
with
Cassandra's
best
characteristics,
I
guess
to
sort
of
keep
the
spirit
of
denormalization,
but
to
bring
you
back
the
ability
to
ask
complex
questions
of
data
and
to
do
so
in
real
time
so
I'm
going
to
talk
very
quickly
about
some
of
the
technology
that
underpins
underpins
this
so
hakuna
analytics
is
basically
a
system
which
takes
in
events.
We
have
a
range
of
different
API
for
doing
that.
B
We
have
a
HTTP
API,
flume
integration,
mq
integration,
we've
just
added
kefka
integration
as
well,
so
you
can
just
throw
day
to
the
you
in
JSON
format
and
by
the
way,
one
of
the
things
that
used
to
be
a
bit
of
a
pain
with
cassandra
is
getting
data
and
it's
less
of
an
issue
now.
But
it's
still
nice
to
be
able
to
connect
many
different
data
sources
easily
and
so
koonu.
Really,
you
can
think
of
from
Cassandra
point
of
view,
at
least
as
a
sort
of
automatic
data
modeling
tool
for
Cassandra.
So
we
sit
Cassandra.
B
We
have
this
analytics
engine,
and
but
we
also
have
a
bunch
of
built-in
visualization
tools,
so
we
have
a
browser-based
UI
for
building
rich
visualizations,
widgets
and
dashboards
and
being
able
to
see
those
change
and
update
in
real
time
and
a
framework
for
raising
alerts
out
the
other
side
as
well.
So
it's
really
sort
of
for
analytical
applications.
It's
really
about
adding
functionality.
On
top
of
the
great
foundations
that
Cassandra
already
provides.
I
can
tell
you
a
little
bit
about
how
it
works.
B
So
the
first
thing
that
the
way
that
we
allow
you
to
express
the
metrics
that
matter
to
you
and
the
way
that
we
can
give
you
instant
answers
without
having
to
trawl
across
billions
of
roar
events
is
to
ask
you
to
define
I
guess
what
are
analogous
to
OLAP
style
aggregate
cubes.
So
I
a
cube
is
a
definition
of
an
aggregate,
so
you
a
proxy
or
a
top
K
or
count
or
standard
deviation,
or
something
like
that
and
a
set
of
dimensions.
B
So
I
want
to
be
able
to
take
this
click
stream
and
I
want
to
be
able
to
see
the
unique
number
of
visitors.
I
want
to
be
able
to
break
that
down
by
browser
and
geography
and
by
customer
segments,
for
example,
and
so
armed
with
that
information.
We
can
then
allow
you
to
collect
events,
and
those
events
will
potentially
result
in
multiple
updates
into
into
Cassandra.
B
There
is
many
of
the
sort
of
concepts
that
are
familiar
to
you
from
a
sequel
world,
so
select
where
group,
by
limit
having
join
and
all
of
those
sorts
of
keywords,
but
we
decompose
those
queries
on
two
slices
of
the
aggregate
data.
So
you
never
have
to
go
back
to
the
roar
events
to
be
able
to
answer
queries
that
that
you
prepare
for
an
advance
by
setting
up
these
cubes.
We
also
maintain
a
mapping
to
the
raw
event
so
that
you
can
drill
down
to
see
the
underlying
result.
B
To
give
you
historical
context
on
on
that
newly
defined
cube
and
obviously
that's
a
process
which
does
need
to
touch
the
real
events
and
does
need
to
get
paralyzed
across
all
the
nodes
in
the
cluster.
But
it
means
that
when
your
boss
asks
you
actually
it'll
be
really
nice
if
you
could
visualize
these
results
or
get
me
an
answer
to
this
other
question,
which
you
didn't
know
about
previously,
it's
much
much
easier
to
do
that.
You
don't
need
to
write
any
coding,
you
don't
need
to
do
any
gluing.
B
You
don't
need
to
be
changing
your
your
your
Cassandra
data
model
to
do
that.
It's
just
a
click
of
button.
Are
you
in
a
UI,
and
so
really?
What
we're
doing
here
is
is
providing
some
tools
that
if
you're
a
developer,
you
make
your
life
easier
for
building
analytical
applications.
So
we
simplify
the
ingest
process.
We
give
you
rich
queries.
We
make
denormalization
nor
the
process
which
is
going
to
hurt
your
agility
and
we
give
you
a
bunch
of
built-in
visualizations,
which
I'll
talk
a
little
bit
about
later.
B
So
that's
that's
great,
but
so
it's
all
really
for
a
developer,
and
so,
if
you
look
at
the
workflow
around
how
you
actually
set
these
things
up
in
our
current
version
or
until
yesterday,
our
current
version
of
a
coup
new
analytics
there's
a
lot
of
there's
a
lot
of
the
developer
in
there
right.
So
you
know,
you've
got
it.
You've
got
to
install
cassandra
and
you've
got
to
install
a
cocoon
of
analytics.
B
That's
that's
that's
great,
but
then
the
whole
process
of
setting
up
new
cubes
to
refine,
that
is
the
business
guy,
is
basically
going
to
have
to
go
and
pester
the
developer
or
the
technical
guy
again
to
be
able
to
allow
that.
So
that
all
changes
from
our
perspective
today
and
what
we're
doing
with
the
Kuno
analytics
v5,
which
we're
announcing
today
and
which
you'll
be
able
to
download
I
think
this
afternoon
from
our
website,
is
to
make
this
process
something
that
really
empowers
the
business
user.
B
So
I
sort
of
like
to
describe
it
as
the
tableau
ization
of
streaming
data
I,
don't
know
whether
that's
a
message
that
resonates
with
you
but
I
love
tableau.
It
makes
it
really
really
easy
to
get
value
out
of
CSV
files
and
databases
that
are,
you
know
you
have
access
to
on
your
machine.
It
empowers
people
who
can
use
pivot
tables
in
excel
right,
that's
an
incredibly
powerful
tool,
but
for
streaming
data
and
big
data.
B
That
tooling,
is
just
not
been
there
so
far,
and
I
think
that's
that's
something
that
we're
really
really
looking
to
tackle
and
we're
really
excited
about.
So,
of
course,
you're
still
going
to
need
your
dev
to
install
it.
But
we've
done
a
number
of
things
that
mean
the
business
user
can
get
stuck
in
two
doing
real-time
analytics
on
Cassandra
and
the
first
one
is
to
basically
automatically
infer
the
structure
of
tables
and
and
these
dimensions
that
you
need
to
set
up
in
a
Cooney
from
a
sample
event
file.
B
So,
in
the
same
way,
the
you
know
in
Excel
you
go
into
the
import
wizard
and
you
see
a
sort
of
set
of
data
and
a
table,
and
it
guesses
what
the
format's
that
data
is,
and
then
you
just
click
import
and
ninety
percent
the
time
it
gets
it
right
and
occasionally
you
have
to
change
them
stuff.
We
do
exactly
the
same
thing,
so
you
just
upload
a
file
of
a
set
of
JSON
objects
or
instead
of
CSV
events
and
well
I.
B
Okay,
so
you're
collecting
clickstream
data,
and
these
are
the
fields,
and
these
are
the
ways
that
you'd
like
to
set
that
up
and
then
we'll
pull
those
events
in
and
the
next
step
is:
we've
added
a
drag-and-drop
Explorer
view,
so
you
can
basically
build
queries
in
the
UI
just
by
saying.
I
want
this
on
this
axis.
I
want
this
on
this
axis,
and
then
you
know
show
me
that
graph
and
the
point
about
this
is
that
you
know
previously.
You
have
this
slightly
awkward.
B
You
have
this
slightly
awkward
dependency
because
you
needed
the
cubes
to
be
able
to
ask
the
queries.
What
we've
done
now
is
enable
ad-hoc
queries
which
to
use
a
sample
of
the
data,
that's
already
available
in
the
system,
to
give
you
sort
of
a
rough
sketch
of
what
the
results
are
going
to
look
like
and
when
you
click
Add
to
dashboard.
We
automatically
infer
the
cubes
that
you're
going
to
need
to
get
real-time
answers
there
and
we
we
build
those
cubes.
We
populate
them
with
historic
data.
If
you
want
and
then
you
get
real-time
results.
B
So
turning
on
the
data
feed
to
actually
keep
this
process
happening
in
real
time
is
something
that
we
still.
We
still
need
a
developer
for,
but
you
know
we're
working
on
that
in
a
sense,
because
we've
added
a
mechanism
for
plugging
in
new
connectors
through
a
very
simple
public
API.
We
push
that
API
to
github
and
we
provided
some
examples.
So
new
careful
integration
is
just
one
of
those
examples.
B
We
have
our
own
business
guy
so
who
could?
Who
could
screw
this
up
worse
in
your
company
than
the
vp
of
marketing?
No
offense
die
better
now,
so
what
I'd
like
to
do
is
invite
our
vp
marketing
guy
Craig
onto
the
stage
and
just
to
demonstrate
on
a
live
instance
that
even
the
marketing
guy
can
could
can
do
real
time.
B
So
so,
745
this
morning,
guy
had
a
quick
play
just
to
check
that
he
just
to
check
that
he
sort
of
knew
where
things
were
so,
hopefully
this
this
real
life,
real-time
experiment
will
give
you
some
insight
into
into
just
how
easy
it
is
so
I
mean
you
know
to
go
back
to
the
cartoon
and
diet.
Could
you
come
on
the
stage
in
build
some
dashboards,
I?
Think.
C
B
C
Whoa
whoa
no
worries
okay,
so
first
thing
I
want
to
do
is
I
want
to
play
with
some
tweets
because
I'm
going
to
build
a
Twitter
analytics
application,
so
I've
got
a
file
which
has
got
some
of
the
kind
of
some
sample
tweets
in
that
already
n
emoni
kind
of
load
that
up
there's
all
my
sample
tweets.
We
see
those.
C
C
Series
of
strings
and
then
all
things
twigs
and
half
that's
a
string,
so
I've
got
an
inferred
structure
that
I
get
the
opportunity
to
kind
of
play
with
that
and
modify
it.
I've
also
got
automatically
ticked.
For
me
is,
do
you
want
me
to
insert
the
events
from
that
file?
Obviously,
it's
kind
of
parsing
those
out
so
I'll
create
my
table
uploaded
those
it's
okay.
What
do
you
want
to
do
with
them?
Well,
like
all
good
tweet
analytics,
apps
will
start
with
count.
Shall
we
and
we'll
group
it
by
all?
C
Let
me
see-
let's
group
it
by
the
the
key
word
point
outside
before
I-
do
that
you
can
see
I've
only
got
40
it
him
first.
He
says
what
you
want.
You
asked
me
to
count,
so
it's
just
as
well
I'll
give
you
a
metric
and
it
counts
them.
There's
40
of
them
in
this
file.
So
now
I
group
them
by
keyword
and
by
default.
It
can't
show
a
group
by
on
a
metric,
so
it
works
that
well
pie
chart
any
good
to
you.
C
I,
could
let
me
just
so
I
can
see
it
better
I'm
going
to
limit
it
to
50
the
top
15
of
that,
but
I
could
also
go
off-piste.
I
don't
have
to
use
the
visualizations
that
it
wants
to
use.
I
can
maybe
show
that
as
a
bar,
that's
cool
and
I've
got
my
my
tweets
by
the
different
keywords,
but
there
may
be
actually
enough
gone
off
that
I'm
exploring
my
date
after
that's
the
interesting,
but
it's
not
grabbing
me
too
much.
Why
don't?
I.
C
C
And
then
in
fact
leads
color-coding
the
map.
For
me,
if
I
look
at
my
widget
properties,
there
are
those
it's
actually
a
calorie
coding,
those
by
lowest
it's
a
higher
values
of
those
by
default
automatically.
If
I
use
the
maps-
and
I
can
adjust
those
of
course
I
want
to-
and
then
next
thing
I
can
do
is
you're.
Looking
very
nervous,
Tim.
C
C
C
B
C
It's
counter
to
2000
and
I've
got
it
here
on
the
speed
and
the
precision
is
about
in
the
middle.
Let's
move
it
down
to
that
yep!
So
that's.
If
I
took
the
high
speed,
I
am
I'm
now
sampling
the
the
live
Twitter
feed,
but
at
very
high
speed,
high,
very
low
sample
set
in
order
to
keep
this
up
to
it.
Why
I'm
going
and
now,
if
I
group,
that.
C
C
B
B
So
that's
really
all
I
wanted
to
cover
just
to
give
you
a
quick
overview
of
the
key
things
that
are
that
are
coming
up
in
a
communal,
analytics
v5,
and
we
have
a
couple
of
things
you
can
do
if
you
want
to
download
it
it'll
be
available
this
afternoon
from
a
it'll,
calm,
/
download,
and
if
you
want
a
exclusive
limited
edition,
a
coup
new
t-shirt
with
that
very
message
on
come
to
the
coup
new
boots
very
quickly
does
going
to
model
it
here.
For
you,
oh.
C
B
B
B
400,000
events,
a
second
which
equation
about
40
billion
events
a
day.
Can
somebody
check
my
math
designs,
its
answer,
your
question
so
I
mean
typically
we
see
and
we
I
mean
the.
We
see
typical
setups,
where
we
have
a
sort
of
right
behind
buffering
in
analytics
to
a
sort
of
absorb
counter
increments.
So,
basically
we're
sort
of
where
we
build
a
number
of
column
families
behind
the
scenes
we
mapped.
B
These
sort
of
high-level
group
buys
and
wears
and
cubes
at
your
building
onto
those
column,
families
and
every
we'll
increment,
a
set
of
counters,
and
so
we
absorb
the
plus
ones
in
analytics.
You
can
choose
to
basically
absorb
the
minute
in
a
sort
of
right
behind
buffer,
which
is
also
a
cache,
and
we,
you
know
we
we
limit
the
load
on
your
Cassandra
cluster
by
light
thrown
plus
35
s
into
Cassandra,
rather
than
plus
1.
So
that
helps
a
lot,
but
yeah
I
mean
you
can
do
an
awful
lot
on
very
small
clusters.
Actually,
yes,.
B
Right
so
that
this
is,
there
is
a.
There
is
obviously
a
trade-off
in
terms
of
the
amount
of
aggregation
you
do
ahead
of
time,
so
we
have
in
fact
it's
just
a
factor
right.
So
the
more
events,
the
fewer
fewer
cubes
but
I
mean.
Typically,
we
see
we
see
setups
that
run
very
very
nicely
where
the
cubes
that
you
set
up
can
contribute
to
say
a
hundred
counter
increments
for
every
event
that
comes
in
and
that
they
work
absolutely
fine,
especially
if
there's
locality
in
your
data
and
using
the
the
buffer.
B
B
Okay,
let
me
let
me
just
answer
that
verbally
yes,
so
there
is
a
chance
of
the
dashboard
use,
HTTP
API,
where
you
can
write
a
ql2
and
that
HTTP
API
is
is
available,
of
course,
to
any
application
you
want
to
build.
On
top
of
it
and
the
the
same
API
is
used
by
the
dashboards
and
yeah.
We
see
very
often
people
writing
both
applications
and
dashboards
through
these
charts.
In
fact,
if
I.
B
If
I
do
this
and
show
you
one
of
these
things,
I
can
actually
take
one
of
these
dashboards
and
embed
it
as
well.
So
we
have
a
JavaScript
library
which
allows
you
to
embed
this
directly
into
your
own
site
and
because
it's
backed
by
Cassandra,
which
has
you
know
pretty
low
great
high
availability,
characteristics,
low
latency
reads,
and
we
do
a
lot
of
the
work
ahead
of
time.
It's
very
easy
to
embed
these
widgets
in
your
own
site
and
we
have
store
procedures.
B
B
B
Good
question:
so
how
do
we
do
joins
so
the
joins
in
aql
are
designed
to
be
low.
Cardinality
joins
that
you
do
across
so
there's
there's
no
there's
no
silver
bullet
here
right.
So,
if
you
need
to
do
joins
across
high
cardinality
dimensions,
data
warehouses
have
been
optimizing.
Those
sorts
of
techniques
for
many
many
years
and
the
whole
the
whole
idea
about
a
coup.
New
really
is
to
try
and
do
as
little
work
after
the
fact
as
possible.
So
we
to
pull
a
form
of
in
jes
time
join,
whereas
an
event
comes
in.
B
You
can
enrich
that
event
with
other
information.
So,
for
example,
if
you're
collecting
tweets-
and
they
didn't
have
that
long
associated
with
them,
you
could
basically
do
a
look
up
on
that
long
from
somewhere
and
without
affecting
a
throughput
too
much.
We
could
then
enrich
that
event
and
then
basically
use
those
those
additional
fields
in
the
cubes.
As
you
go.
B
B
B
B
Good
good
question,
so
we
work
in
our
own
key
space
on
a
Cassandra
cluster,
so
you
just
say
in
the
configuration
what
key
space
you
want
us
to
create
or
create
yourself
and
will
will
operate
in
that
some
focus
on
an
existing
cluster.
Some
folk
create
their
own
cluster.
We
also
have
an
ami
image
which
comes
with
Cassandra
in
it
as
well,
which
avoids
the
hassle
of
setting
up
separately,
not
that
it's
very
hard.
These
days,
I'm
question,
though,.
B
Do
we
duplicate
data?
So
I
guess
we
do
we
duplicate
a
little
bit
of
data.
I
guess
because
the
same
event
will
feature
in
different
different
points
in
your
analytic
results
if
you
like,
but
that
is
because
what
we
do
is
mostly
quantitive
analytics.
That's
that's
just
count
typically,
and
we
keep
a
standard.
We
keep
a
single
instance
of
the
roar
event
and
Cassandra's
built-in
compression
is
actually
very,
very
good.
You
we
really
don't
see
space
or
ever
being
ever
been
a
real
issue
in
production
clusters.
B
B
Good
question:
so
we
and
we
implement
a
distributed,
work
you
so
we'd.
If
the
schemas
we
distribute
the
we
differ
schemas,
we
work
out
exactly
what
what
bits
of
what
column
families
we
need
to
update.
We
push
that
work
on
to
a
queue
which
is
pulled
by
every
node
in
the
analytics
clusters
or
analytics
runs
as
a
scale
out
here
above
Cassandra
or
typically
on
the
same
boxes.
It's
a
common
deployment
pattern
and
each
of
those
will
do
a
particular
slice
of
time
and
then
those
those
jobs
will
be
pushed
back
in
a
bit
I.