►
From YouTube: Tech Talk: Cassandra Data Modeling
Description
Check out the full post here: http://www.hakkalabs.co/articles/cassandra-data-modeling-talk
In this talk, Patrick McFadin (Chief Evangelist for Apache Cassandra, DataStax) breaks down topics like storing objects, indexing for fast retrieval, and the application life cycle.
This talk was given at Cassandra Day Silicon Valley 2014.
ABOUT DATA COUNCIL:
Data Council (https://www.datacouncil.ai/) is a community and conference series that provides data professionals with the learning and networking opportunities they need to grow their careers. Make sure to subscribe to our channel for more videos, including DC_THURS, our series of live online interviews with leading data professionals from top open source projects and startups.
FOLLOW DATA COUNCIL:
Twitter: https://twitter.com/DataCouncilAI
LinkedIn: https://www.linkedin.com/company/datacouncil-ai
Facebook: https://www.facebook.com/datacouncilai
Eventbrite: https://www.eventbrite.com/o/data-council-30357384520
A
So
why
is
it
the
data
modeling
in
time
series
fit
together,
I'm
going
to
try
to
explain
a
little
bit?
Why
so
wake
up
here
are
some
interesting
facts
right
well,
I'm,
just
going
to
establish
my
use.
Why
I
want
to
talk
about
this,
so
the
Internet
of
Things
is
very
interesting
new
development.
It's
kinda
like
it's
almost
getting
to
be
like
a
cloud
buzzword
now,
item
T
was
that
I
owe
to
you.
Yes,
it
is,
and
but
I
think
it's
pretty
realized,
whatever
we
define
as
regardless.
A
Analysts
as
your
gazillion
dollar
bargained
for
it
still
comes
down
to
one
thing:
we're
catching
keys
to
the
internet
really
quickly.
You
guys
know
what
that
thing
is
up
there.
It's
right
here!
That's
the
net!
You
just
when
I
said
that
in
Europe
we
were
like
there
was
done.
That's
a
net
right,
Google
just
bought
nest
for
three
billion
dollars.
Now.
Do
you
think
Google
is
in
there
a
board
meeting
in
there
McMahon?
You
know
we
got
this
cool
driving
car,
self-driving
car,
great
search
engine,
Google
Apps,
that's
cool!
A
We
need
a
thermostat
now
that
is
it
what
they
said
it
and
I
have
like
one
in
my
house
to
you,
yeah
dude.
Do
a
good
knows
we're
here.
We
need
that
data
they
bought.
They
bought
three
billion
dollars
of
what
the
f
you're
doing
yours
creepy
eyes.
It
knows
when
I'm
there
it
knows
what
I
need.
He
knows.
A
temperature,
no
Google
wants
that
information.
Air
say.
C
A
A
Right
and
your
car
now
Ellen
talks
with
sink,
so
all
these
devices
are
going
to
be
talking
so
15
billion
devices
estimate
about
forty
billion
devices
by
2020.
That's
a
lot
of
mouths
to
feed
and
that
all
that
means
pretty
much.
One
thing
this
all
has
to
go
in
a
database
somewhere.
So
let's
talk
about
what
that
is,
and
it's
almost
always
going
to
be
time:
series
theater
temporal
data.
Here's.
A
At
this
time
the
sander
is
a
great
home
for
that
and
it
wasn't
really
designed
for
that
initially.
But
why
can
standard
for
time
series
and
then
I'm
going
to
use
this
as
a
data?
Modeling
launching
point
so,
as
you
saw
my
last
talk,
one
slide
proved
it
all
and
if
you're,
not
a
Cassandra
user,
you
can
prove
it
to
yourself.
But
Cassandra
scales.
That's
good
thing
for
you
IOT,
because
there's
going
to
be
a.
A
Resilient,
which
means
it
is
going
to
be
online
all
the
time
and
it's
going
to
deal
with
all
the
crap
that
I
throw
at
it
it's
great
day
tomorrow.
Aha,
let's
talk
about
that.
That's
going
to
be
a
big
part
of
this,
but
then
beyond
that-
and
this
is
the
kind
of
a
secret
sauce,
a
commander
of
wide
time
series
data
is
so
good.
Is
this
efficient
storage
model
now
the
efficient
storage
model
as
a
programmer
I?
Really
don't
care
too
much
about
right?
A
That's
not
something
I'm
going
to
concern
myself
with
as
I'm
writing
code,
but
I
want
you
to
understand.
What's
going
on
so
when
you're
writing
code
and
you're,
looking
at
the
data
model
to
understand
what's
going
on
under
the
scene
because,
let's
face
it,
what
about
that
I
mean
that
this
is
the
power
with
what
Cassandra
could
do.
So,
let's
get,
let's
start
out
with
the
data
model
example,
but
I
like
examples,
good
old
weather
station
and
classic
right.
This
is
a
classic
time
series
temperature
over
time.
C
A
So
that's
a
that's
an
easy
one.
So
how
do
you
do?
What
do
you
start
with
Cassandra
date
alone?
Data
modeling
is
a
little
different
in
Cassandra,
because
what
we're
doing,
instead
of
taking
the
data
and
then
building
models,
we're
taking
our
application
and
building
the
model
so
I'm
going
to
start
out
at
the
top
with
what,
with
my
application,
going
to
need
for
a
query
how
much
the
application
is
going
to
ask
for
this
data
I'm
going
to
graph
this
data
on
that
on
my
website.
A
What
are
those
queries
look
like,
so
the
first
thing
is
I
want
to
get
your
my
needed
queries.
I
want
to
get
all
my
data
for
one
weather
station.
Okay,
I
want
to
get
all
the
data
for
a
single
date
in
time
and
I
want
to
get
data
for
a
range
of
date
and
times
so
all
very
similar
around
the
single
weather
station
ID
weather
station
itself,
but
a
little
different
I
want
a
single
date
time.
I
want
to
get
a
range.
So
how
do
I
model
that
so
by
data
model
support
these?
A
It's
pretty
simple
I'm
going
to
store
data
per
weather
station.
That's
my
first
requirement.
My
second
requirement
are
going
to
store
in
time
series
order
first
event,
the
last
of
it.
So
what
does
that?
Look
like
so
I'm
going
to
create
a
table?
My
create
tables,
pretty
simple:
I
have
a
weather
station
ID,
which
is
just
a
text
I
use
text,
because
it's
here's
actual
value
down
here.
It
has
some
letters
in
it
event,
time
which
is
a
timestamp.
A
A
This
is
what
makes
the
data
model
rich
and
I'm
going
to
show
you
why
it
makes
a
difference
in
a
minute,
but
this
primary
key
is
an
important
piece
of
your
cable
creation,
because
what
you're
specifying
here
is
the
uniqueness
that
partition
partition
is
here's
all
my
data
that
I
would
order
in
a
certain
way
and
then
our
codomain
together
so
much
like
how
a
relational
database
has
a
primary
key.
This
is
d2
combined.
These
two
values
are
unique
same
thing,
but
what?
A
A
A
That's
what
I
want
now,
if
you
look
at
the
order
of
that
data,
again,
I
specified
in
the
primary
key
before
that,
so
the
order
was
I
want.
The
first
thing
that
happened
to
the
last
going
back
to
this.
The
event
time
is
that
second
part
of
the
primary
key,
the
first
part
sets
up
the
partition.
The
partition
is
this
weather
station
ID
every
element
in
that
partition
will
be
ordered
by
event
time.
That's
how
I
would
describe
that
to
you.
So
yes,
these
two
are
unique
weather
station.
A
A
So
what
happens
on
the
storage?
I?
Remember,
I
told
you
I
teased
you
on
that
a
little
bit
the
storage
engine
is
the
cool
part
of
the
story.
Now
this
is
the
part
where,
as
a
programmer,
you
may
depart
and
say
long
I'm
good,
with
this
I'm
moving
on
I'm,
going
to
write
my
code
and
ship
it
because
it's
an
agile
thing
and
I
got
to
do
a
demo
day
tomorrow.
But
let's
look
and
see
what
really
happens.
I
think
the
most
interesting
part
what's
really
happening
on
the
storage
engine.
A
A
A
This
is
great
for
storing
data.
But
what's
the
point,
why
are
we
doing?
The
point
is:
whenever
I
do
something
like
this
now
this
is
my
this
is
my
query,
where
I'm
going
to
ask
for
that
range
of
time
again,
keep
in
mind
that
that
was
one
of
my
application
requirements.
Is
that
what
I
worry
my
dear
I
want
to
be
able
to
grab
a
range
of
time
in
this
case,
I
wanted
to
grab
it
from
one
minute
to
the
four
minute,
not
very
exciting
time
period,
but
that's
what
I
want.
A
A
Unfortunately,
Al's
right
next
door
telling
you
why
this
suck,
if
you
were
to
be
over
there
in
here
at
the
same
time,
you're
freaked
but
that's
cool,
but
you
would
hear
him
say
well:
here's
the
problem,
you're
going
to
be
running
into
disk
problems,
I'm
going
to
tell
you
yes,
you're,
going
to
run
into
disk
presence.
Trust
me,
but
the
way
to
get
around
some
of
those
issues
is
to
minimize
the
amount
of
seeks.
A
You
do
on
disk
a
seek
on
a
disk
Hertz,
because,
let's
look
at
everything
in
your
computer,
you
know
the
CPU
is
in
nanoseconds.
Memory
is
in
micro
or
even
the
a
no
seconds
now.
But
what's
the
disk
do
spinning
is
just
seconds.
That's
like
orders
of
magnitude
slower.
Then
we
haven't
really
fixed.
That
problem
sort
I
mean
we
got.
We
knew
how
new
disk
technologies
that
are
now
available,
but
we
still
look
at
most
most
discouraged.
A
Turned
into
a
10
millisecond
query,
just
based
on
that
one
thing,
and
if
I'm
chasing
milliseconds
that's
going
to
hurt
okay,
so
getting
more
less
than
less
than
that
one
seek.
Are
you
getting
I'm?
Sorry
less
than
one
seek?
You
can
do
that,
but
getting
a
less
seeks
is
a
good
thing
so
that
single
seek
on
disk.
A
That's
what
I
want
I
want
to
put
the
disc
ad
right
at
the
beginning
of
my
data
and
because
what
I'm
trying
to
do
here
is
get
this
range
I'm
going
to
grab
all
this
data
right
here
and
I,
whatever
I
ask
Cassandra
to
get
it.
It'll
move
the
disk
there
and
sweep
across
the
head
and
puts
it
right
back
into
a
CTL
table
for
me
programmer.
A
A
C
B
C
C
A
B
A
A
Done
this
data
model,
similar
ones
out
of
the
Oracle,
and
you
will
limit
out
real
fast
and
mainly
because
you
run
out
items
sequential
operations
are
going
to
be
faster,
a
larger
dataset,
a
larger
bigger
the
dataset
you
pop
I
have
one
other
reason.
This
is
going
to
be
faster
as
well,
and
that's
because
you're
going
to
distribute
that
many
queries,
so
let's
hold
off
on
the
questions
for
a
minute.
Let
me
move
on
to
the
next
topic
and
maybe
that'll
answer
some
of
these
questions.
A
So
what
about
some
additional
help
on
the
storage
engine?
So
let
me
show
you
what
a
weed
with
the
read
path
looks
like
and
Cassandra.
In
a
nutshell,
when
a
client
asks
for
digging
it,
it
goes
to
the
key
cache.
The
key
cache
is
like
hey
I.
Have
this
that
first
part
of
my
primary
key.
Where
is
that
on
the
dips
I
want
to
know
where
that
is
on
the
disk?
A
A
A
A
Time,
500,
microseconds,
95th
percentile,
is
very
doable.
If
you
do
this
right,
so
the
next
place
is
a
bloom
filter.
I
talked
about
that.
My
last
talk.
Bloom
filters
are
just
a
probabilistic
function
and
all
they
do
is
they
say
where
the
data
is
not
so
again
we're
trying
to
find
out
where
the
data
isn't.
A
So
the
key
to
the
speed
in
Kassandra,
based
on
that
on
that
particular
read
model
now
like
what
how
the
reeds
work
in
Sandra
here's
the
keys
to
speed-
and
this
is
actually
this
is
important-
point
number
one.
The
first
part
of
the
primary
key
that
first
part
the
weather
station
ID,
gives
you
location,
that's
going
to
put
you
on
a
single
server
and
I'm
going
to
explain
how
that
works
in
a
second.
But
that's
going
to
put
you
on
a
single
server.
A
So
if
you
have
a
thousand
node
cluster,
that
first
part
would
put
you
on
one
of
those
servers
in
that
Buster.
It
will
not
be
a
random,
seek
the
second
part
that
partition
key
will
get
you
closer
to
the
SSD
and
you
get
right.
You
want
to
minimize
that
seek
you
want
to
get
into
that
file
as
soon
as
you
can.
You
want
to
minimize
and
then
third
thing
which
I'm
putting
up
here
just
for
brevity,
but
we
don't.
A
We
won't
cover
as
much
I
do
a
lot
of
performance
tuning
type
talks,
but
the
other
thing
is
how
fast
you
can
find
that
data
in
the
SS
table.
If
you
have
a
three
gig
file,
you
can
use
indexes
to
find
that
data
immediately.
So
once
you
do
is
seek
you
go
to
where
the
data
is
instead
of
having
a
move
that
diskette
across
the
disk.
So
there
is
a
new
feature
as
a
2.0
that
will
help
you
get
to
that
data
faster.
A
This
is
actually
there
for
time
series
data
and
in
those
cases
where
I
might
have
three
files
on
the
file
system
that
could
have
potentially
have
my
data
I
want
to
minimize
how
much
seats
I
have
on
the
disk.
So
those
same
query:
I.
Have
this
I
want
I'm
looking
for
data
in
between
1
and
4
I
know,
there's
a
bug
in
my
code.
I
forgot
to
put
an
equal
filter,
I
get
pointed
out
off
the
shark
route,
but
here's
three
files
that
it
could
potentially
be
in
now
this
min/max
hint.
A
What
we
do
is
when
we
create
those
files.
Now
we
create
an
index
an
index
built
on
top
of
that
file.
That
says:
hey
here's,
the
minimum
maximum
values
for
this,
for
this
primary
key
that
your
app
that
you've
set
up
in
your
data
model
and
you
can
use
that
later.
So,
how
does
that
use
so?
The
minimum
event
time
here
is
the
index
shows
that
it
was
a
march
27th
to
31st.
This
one
was
April
1st
to
4th,
and
this
was
April
5th
to
the
ninth.
That's
the
event
time.
A
A
A
This
is
how
I
do
time
series
data
models
in
less
than
a
millisecond
net
mean
a
peyote.
You
know
when
I'm
messing
around
with
my
nest
by
the
way
nest
uses
Cassandra
I
want
to
make
sure
I'm
changing
the
temperature
of
my
house
and
pissing
off
my
wife
really
fast
I.
Don't
wanna
wait
so
I'm
gonna
change,
a
temperature
really
quick,
funny
story
about
this
little
video,
our
co-founder
Matt
filed.
A
He
has
a
nest
and
he
had
a
roommate
Mike
Blauman,
who
runs
apps
in
her
team
and
when
Matt
was
on
the
road
at
night,
he'd
jack
up
the
temperature
Mike
could
wake
up
sweating
us
I'm
here.
What's
going
on
the
thermometer
or
the
thermostats
broken?
So
what
did
he
do?
He
called
support:
hey
this
thing's
broken
it's
like.
Oh
man,
that's
terrible!
So
they're
all
trying
to
diagnose
as
they're
running
all
these
things
on
they're
trying
to
figure
it
out.
They
they
got
back
to
him
on
email,
I.
A
A
Busted,
eventually
fun
so
now,
I
can't
do
it
anymore.
So
whenever
I
go
to
nest
and
now
they're,
like
oh,
say,
hi
to
match
for
us,
you
me
just
do
all
that
extra
work,
so
I,
don't
real
quick
I
want
to
go
through
some
ingestion
models
right
now
that
about
getting
that
data
there,
because
this
is
the
next
question
I
would
get
if
you're,
if
you're,
if
you're
at
the
point
where
you
have
a
database,
that
you
do
millions
rights
per
seconds,
you
gotta
have
an
application,
they
could
do
millions
all
right.
A
A
Is
a
new
game
right
you're
talking
about
internet
things,
you
to
get
a
lot
of
data
together,
so
when
I
hear
so
much
about
right
now
in
mana,
since
I
talked
to
this
after
that
cop-killers
project
at
Lincoln
graduated,
it
grew
up
in
a
big
world
of
interest.
Apache,
as
talking
on
the
project,
is
based
on
Scala
cool
queueing
system.
A
It
is
big
based
of
basically
the
same
type
of
paradigm,
that
Cassandra's
horizontal
scale,
yeah
more
delicate,
more
scale,
lots
of
people
using
this
now
and
it's
a
really
cool
project,
but
I
see
it
in
conjunction
with
the
food
pools.
Storm
is
the
other
one.
Storm
is
a
processing
framework.
So
if
you're
taking
it
a
lot
of
stuff.
C
A
Know
refrigerators
all
over
the
world
tell
you
what
temperature's
is
decided
with
refrigerant
going
through
Kafka,
making
sure
that
that
gets
once
and
only
ones
delivering
storm
will
then
roll
it
up
and
do
different
things.
As
idea,
she's
got
a
lot
of
data
coming.
What
are
you
gonna
do
with
it?
You
need
a
database
that
can
scale
up.
Cassandra
seemed
to
be
the
place.
People
have
another
one
is
one
of
my
projects
of
flu.
A
I
was
recently
at
the
Hadoop
survey,
which
is
kind
of
funny,
for
example,
personal,
and
one
of
the
guys
in
my
talk,
was
Jonathan
Shay,
who
was
the
original
author
flew
and
as
I'm
talking
about
flume
I,
see
this
guy
like
smiling
in
the
back
I
used
to
work
on
this
project,
so
it
was
kind
of
cool
seeing
that
back
back
in
old
school.
This
is
an
old
project
that
it's
fused
about
four
or
five
years
old
and
it
was
originally
built
for
congestive
logs
I
use
it
for
bringing
in
syslog.
C
A
Syslog
HTTP
there's
a
channel
which
is
like
a
bus
and
then
the
sync
which
picks
up
those
events
off
the
channel,
and
then
you
can
do
transforms
what
I
did
is
I
built
a
system
that
basically
took
in
HTTP
events
from
our
f5
load
balancers
and
it
parsed
out
the
entire
log
string
and
then
created
90
different
inserts
based
on
that
one
web
hit.
So
I
created
this
massive
like
web
amplifier
right
and
it
was
the
only
thing
I
could
I
was
trying
to
put
it
into
an
Oracle
database.
A
He
failed
and
then
I
just
ran
out
of
money.
Really
I
tried
to
do
it
and
the
only
thing
that
really
could
scale
up
was
exadata
and
I
didn't
like
spending
two
million
dollars
of
this
one.
Little
blog
processing
thing.
So
I
started
using
Cassandra.
It
was
when
I
first
started,
and
that
was
on
point
seven
and
it
just
worked.
So
I
noticed
that
this
thing
is
producing
tons
of
data.
I
couldn't
do
anything
with
it.
The
only
other
choices
I
could
dump
it
into
like
s3.
A
You
know
it
doesn't
mean
anything
I
needed
that
data
in
milliseconds,
so
Cassandra
countries.
So
let's
talk
about
that
data
with
speed,
and
this
is
going
to
move
us
into
a
different
dated
modeling
topic.
So
if
you're
going
to
do
a
million,
writes
per
second,
do
the
math
it
and
insert
every
millisecond
or
microsecond
microsecond.
That's
fast,
I,
don't
know
about
you,
but
I,
don't
know!
A
lot
of
applications
could
deal
with
that
too
much
you
got
to
work
as
a
team.
If
you're
going
to
do
is
microseconds
right,
everyone's
got
a
kitchen.
A
You
put
that
on
one
single
server
that
is
going
to
be
the
best
microsecond
reverse
of
that
thing's
gonna
fit
so
there's
other
promise
collisions.
These
are
real
problems.
If
you're
in
time,
series
data
collisions
or
an
issue,
I
did
a
talk
at
ING.
This
is
the
Big
Bang
and
they're,
pretty
public
they're
doing
financial
transactions
using
Cassander
they're.
A
B
A
A
You
how
that
works,
because
I
think
you
need
to
know
that
primaries
our
case.
This
is
how
David
it
gets
distributed,
but
it's
also
how
you
can
do
a
lot
of
things
and
how
your
data
model
really
depends
on
making
that
right,
also
the
rate
of
partitioning
how
your
data
can
spread
around
that's
a
part
of
the
standard
and
then.
A
Tool
for
avoiding
collisions
with
time
since
data,
let's
talk
about
what
happens
with
our
weather
stations,
I
got
a
lot
of
weather
stations
to
feed,
so
let's
quickly
go
into
the
replication
story
of
Cassandra
and
how
that
this
is
your
data
model.
So
the
primary
key
that
first
part
of
the
primary
key
that
I
told
you
about
that
is
the
important
piece
of
figuring
out
what
server
belongs
to.
So
how
do
we
do
that?
A
So
we
have
this
really
basic
data
month
on
our
primary
key
is
the
first
name
of
people
which
is
really
a
poor
primary
key,
but
for
our
limited
data
model
it'll
work,
so
we
have
Jim
Carroll,
Johnny
and
Susie.
That's
our
data
points.
What
we
do
in
Cassandra
with
that
primary
key?
Is
we
hash
it
md5
or
murder?
It's
a
writ!
That
is
a
consistent
hashing
operation,
so
that,
if
you
put
in
the
same
string,
you'll
get
the
same
under
twenty
bit.
Number
up,
awesome.
C
A
This
is
how
many
technologies
work.
If
you
look
at
memcache
back
that
far
back,
that's
how
I
got
data
distributed
in
a
reliable
way,
it
was
using
md5,
it
would
hatch
it
and
then
throw
it
out
into
a
reliable
way.
So
all
these
keys
are
hashed
now.
How
does
that
deal?
How
does
that
do
with
Cassandra
in
this
regard-
and
this
is
what
I'm
talking
about
the
data
locality
like?
Where
is
that
data
in
relation
to
the
physical
layout
of
your
server
how's
it
finding?
A
So
we
have
a
four
node
cluster
here:
node,
a
b
c
and
d,
and
each
node
will
store
a
range
of
data,
a
range
of
tokens.
So
we
take
that
big
128
bit
number,
which
has
a
name
I,
don't
remember
like
a
septa
bike
or
something
and
we're
going
to
break
it
into
quarters.
So
this
one
this
is
one
quarter
of
the
range
on
a
1/4,
a
range
of
be
one
quarter
range
on
C
and
one
quarter
on
D
each
one
of
those
nodes.
A
Analysis,
I'm
primarily
responsible
for
a
quarter
of
the
range
of
that
120
bit
number,
meaning
that
when
a
one
of
these
is
hashed
I
will
see
if
it
fits
inside
my
range
and
then
I'll
hope
and
no
one
else
will
take
it.
So
I
take
Jim's
number,
hey
it
magically
fits
that
5e
goes
in
between
4
and
8
awesome.
So
now
C
is
responsible
for
Jim's
data
Carol.
She
goes
to
node,
D
Johnny
goes
to
know
de
now,
because
md5
is
somewhat
random
and
of
your
mathematician.
A
Random,
okay,
but
it's
pretty
random
and
it's
going
to
randomize
the
placement
of
that
data
around
so
as
I
did,
is
getting
set
in
there.
It's
randomizing,
so
you're
getting
good
distribution,
but
it's
also
going
to
different
nodes.
Now
what
about
the
replication?
Each
one
of
those
nodes
can
also
say
you
know
what,
because
you're
my
neighbor
I'm
going
to
hold
your
range
too.
A
So
if
I
change
my
replication
factor,
that's
all
what
that
means
is
the
nodes
become
responsible
for
its
own
data
and
then
its
neighbors
data
may
be
two
neighbors
data,
which
is
a
lot
better
than
my
neighbor.
My
neighbor
would
never
put
anything
in
mine
in
his
garage,
but
cassandra
is
more
like
working
like
a
tea.
So
whenever
Carol,
which
is
the
primary
responsibility,
is
Dee
well,
a
and
B
also
say
you
know,
we'll
hold
that
data
for
you
as
well,
so
asynchronously
replicates
that
David's
not
synchronous,
easy.
A
A
Pick
as
a
programmer,
you
can
make
a
determination
on
consistency,
and
we
just
have
to
mr.
consistency
here.
Are
you
going
to
do
the
next
doc
you're
just
dream
cases?
Oh
yeah,
eventually,
you'll
eventually
be
consistent.
Yes,
Chris
Chris
house
is
out
for
Netflix
he's
a
going
to
do
a
talk
and
eventual
consistency
which
I
think
that
will.
Thank
you.
We're
going
to
painful
painful
consistency,
man.
A
A
Now,
how
does
that
fit
in
with
the
rest
of
it?
Well,
whatever
I'm
writing
that
million
writes
per
second,
if
I
have
a
very
small
cluster
for
nose.
That
means
that
there's
a
potential
later.
That
means
it's
going
to
be
a
lot
of
writes
per
second
per
node.
Now,
that's
because
each
one
of
them
holds
a
quarter.
So
how
does
that
linear
scaling
story?
Work
if
I
double
the
amount
of
nodes?
Now
each
node
is
only
responsible
for
1/4,
1/4
or
1/2
of
the
quarters.
1/8
I
do
fracture.
A
A
A
A
Now
that
UID
is
added
at
the
end
in
case
you
have
the
exact
same
time
Santa,
if
you're
putting
data
in
your
database
like
that,
then
you
want
to
make
sure
these
valleys
are
unique.
That's
a
time
UID!
So
what's
cool
about
these
are
also
known
as
a
version
one
UUID,
which
means
it's
somewhat
of
a
standard.
You
can
get
driver
support
for
it
in
Python
and
Java.
That's
what
I
think
they're
sorta
bull,
meaning
that
if
you
present
a
bunch
of
tiny
new
IDs
they'll,
be
in
sorted
order
and
you
could
reverse
it.
A
B
A
Example,
where
we
use
that
time
UID
so
again,
our
data
Molins,
how
we
date
a
month
if
I'm
going
to
my
needed
queries,
I
want
to
get
all
trades
and
symbol
for
it
and
add
a
crystals
you're.
Gonna
love
this
example
by
the
way
get
a
trade
for
a
single
date
time,
and
then
you
get
the
last
10
trades.
The
last
10
trades
that's
kind
of
a
tricky
one
right.
A
How
do
I
get
the
last
tip,
so
my
data
model
support
that
is
I'm,
going
to
store
my
data
per
symbol
and
date
I'm
going
to
store
it
time.
Series
in
reverse
order,
last
two
first
and
then
I'm
going
to
make
sure
that
every
single
transaction
is
one
percent
unique.
So
even
if,
through
the
same
microsecond,
so
I'm
going
to
start
buying,
some
Netflix
talked
a
lot
of
actually
sold
a
lot
to
so
here's
an
interesting
thing
with
I'm
going
to
point
out
a
couple
of
deals
here
on
my
day
tomorrow.
A
So
these
are
just
stock
ticker
data,
typical
type
of
stock
ticker
data,
where
it's
coming
in
a
rapid
pace,
so
there
could
be
potentially
some
collision,
but
what
I've
done
here
are
my
primary
key
is
a
couple
of
things:
I've
put
this
symbol
and
date
together
as
that's
the
location.
That's
my
private
part.
First
part
of
my
primary
key.
That's
going
to
give
me
my
locality.
That's
going
to
mean
that
for
every
date,
which
is
just
like
for
one
day
and
one
symbol,
it'll
always
go
to
the
same
group
of
servers
throughout
the
day.
A
So
the
next
day,
it'll
move
to
a
different
group
of
servers,
and
then
the
trade
is
a
time
UID.
So
if
I
have
the
exact
same
at
the
exact
same
second
microsecond
I'll
still
have
uniqueness
to
them
and
it'll
be
ordered
the
next.
The
last
thing
here
is
this
clustering
order
by
this
is
a
very
fancy
way
of
saying
I'd
like
to
reverse
the
order
store
it
on
disk
by
the
trade,
which
is
that
tiny
UID.
So
that's
kind
of
an
interesting
plan.
A
A
The
first
thing
that
happened
is
at
the
bottom
of
the
list,
so
now
I'm
a
reverse
order
based
on
the
trade
details
or
that
trade
time
that
UID
gets
parsed
and
reversed,
but
right
now
I
can't
even
tell
you
what
times
I'm
sure
it
could.
So
what
does
this
look
like
on
the
disk,
though,
now
that
I've
told
you
to
reverse
sort?
This
is
when
it
gets
really
interested,
because
now
what
I
did
is
because
I
told
it
to
reverse
or
and
merge
now
it
actually
does
that
on
the
disk
too.
A
So
whenever
I
go
look
for
that
data,
I'm
going
to
say,
hey
I
want
a
limit
of
three
limit:
three
cool
coolness.
In
times
what
I'm
saying
when
I
say
limit
3
is
I
only
want
three
things
returned
but
because
I
know
I
reversed
it
I'm
going
to
get
just
three
items.
The
last
three
items,
the
last
three
trades
that
happen
boy.
This
data
model
is
used
so
many
places
for
fraud,
detection
for
user
sentiment
for
identity
tracking
on
websites.
A
A
Now,
I
just
have
the
last
three
trades
that
have
happened.
I
think
that
is
a
cool
way
to
do
things
like
I,
said
I've
seen
I've
done
personally
done
a
lot
of
these
data
models
in
production,
using
a
reverse
that
clustering
order,
by
with
a
limit
to
create
really
interesting
interactions
with
time
series,
because
all
those
things
flying
in
as
fast
as
they
are.
You
need
to
be
able
to
do
that.
Yeah
is
descending
7
house
going.
A
And
then
that
file
that
the
way
that
Cassandra
writes
out
is
a
sequential
right,
so
it
will
write
that
file
out
there's
a
weekend
we're
getting
into
likes
of
operations
things
but
later.
If
there's
two
roads
of
data,
it
will
do
a
compaction
operation
as
a
normal
background
operation
and
then
burn
short
those
in
memory
again,
but
that
merge
short
memory.
It's
a
hashmap
will
sort
that
properly.
So
it's
not
going
to
slow
it
down
order
is
independent
of
how
fast
it
means
our
data.
A
B
A
Is
like,
if
you're
looking
at
URLs
all
right,
can
you
say
here
give
me
here's
the
row
keys
URL
a
day
well,
all
day
long,
that
row
is
going
to
be
the
hot
spot
there.
In
that
case,
what
what
I've
done
is
just
put
like
modulus
in
there
or
something
else
to
make
it
more
unique
and
then
how
to
consistent
way
of
managing
that?
That's.
B
A
Will
just
be
writing
to
the
same
note
of
the
same
replica
set
over
and
over
now.
What
we
hope
for
in
that,
like
with
the
stock
trade
example,
is
there's
a
company
called
circa
that
collects
all
stock
trades
around
world
2
million
a
second
I
doubt
you're,
going
to
find
a
hot
spot
in
that
yeah.
You
know,
Netflix
stock
may
be
hot
at
one
moment,
but
there's
still
zillion
other
things
going
on,
so
it
gets
washed
out
the
noise.
A
It
may
cause
a
little
bit
of
a
blip,
but
with
that
many
rights
per
second
going
on
in
your
cluster,
that's
not
really
a
good
consideration.
Now,
if
I
was
only
looking
at
a
select
group
of
stocks
and
I
knew
that
one
was
going
to
weigh
more
than
the
other
okay,
that's
a
different
game,
so
something
you
have
to
consider
for
sure.
But
if
you
look
at
the
big
picture,
distribution
of
that
won't
be
a
problem.
I
think.
C
A
A
Really
and
the
wide
row
concept
all
right
so
that
partition,
how
big
can
a
partition
be
two
billion
cells?
Now,
that's
that's
great.
It
sounds
like
a
lot,
but
if
I'm
collecting
like
weather
station
data
every
second,
it's
not
a
last,
you
can
do
that's
when
you
get
into
to
what
or
if
I
could
just
collect
all
trades
for
one
stock,
one
symbol
forever,
then
yeah
two
billion
is
going
to
get
eaten
up,
pretty
quick.
A
So
that's
when
you
probably
want
to
think
about
what
am
I
doing
in
my
data
model,
the
reason
I
added
date
that
date
in
there
let's
go
back
to
this.
The
reason
I
add
a
date
in
there
is
to
eliminate
that
problem.
Now
that
I
know
every
single
day
will
be
a
new
day
for
a
new
partition
that
will
help,
as
many
of
that
now
do
I
think
I'm
going
to
get
two
billion
in
a
day.
I'm
saying,
though,
I
won't.
C
A
I
know
and
that's
the
right-
maybe
that's
a
different
consideration
in
here
so
and
that
that
day
tomorrow,
I
showed
you
with
the
WebP.
Is
that
actually
had
a
how'd
it
feel
to
sustain,
but
what
I
did
was
you
know?
We
just
had
a
little
bit
of
logic
in
our
system.
Organize
grab
particles
at
those
off
moments,
the
chances
that
somebody
we
did,
everything
in
ddd
the
chances
and
everything
everybody
being
at
work
at
midnight.
A
Vt
we
pretty
high,
so
there
wasn't
even
an
entrance
for
us
to
and
absolutely
but
we
just
had
enough
code
in
there
to
deal
with
that
situation
and
our
it
was
pretty
efficient
out.
We
didn't
really
have
any
problems
with
that,
but
something
to
consider
right,
you're
going
to
you're
going
to
be
transitioning
to
in
our
new
partition.
A
Of
times
yes
and
get
to
know,
data
modeling
I
put
up
here
the
planet
Cassandra.
Oh
there's
my
animation
again,
because
I'm
so
proud
of
it.
Lana
Cassandra
has
a
lot
of
cool
interviews.
Those
five-minute
interviews
like
use
cases
but
also
white
papers.
Things
like
that,
so
I
would
hope
you
wouldn't
have
to
reinvent
a
nice
round
wheel
again,
you
know
people
like
myself,
I've
done
thousands
of
data
models,
I
just
and
now
I'm
kind
of
like
the
horse.
Whisperer
I
just
walk
up.
A
You
know
because
I
were
to
get
all
the
reasons
we
do
things
and
you'll
get
to
that
point
too.
I've
worked
with
teams
now
for
a
long
time,
I
work
with
Christmas's
team.
Quite
a
bit
nobody's
you
know.
No
one
is
really
like
digging
too
deep
anymore
because
they
get
all
the
points
and
then
they're
just
building
the
data
models,
and
sometimes
you
just
it's
like
relational
when
I
first
started
doing
SQL
I'm
like
what
is
this.
This
is
crazy
talk,
but
then,
after
a
while
is,
is
doing
in
it.