►
Description
Speaker: Luke Tillman
Company: DataStax
You’ve heard the talks, followed the tutorials, and done the research. You are a font of Cassandra knowledge. Now it’s time to change the world! (Or at least build something to make your boss happy). In this talk we’ll walk through the process of building KillrVideo, an open source video sharing website where users can upload and share videos, rate them, comment on them, and more. By looking at a real application, we’ll talk about architectural decisions, how the application drives the data model, some pro tips when using the DataStax drivers, and some lessons learned from mistakes made along the way. You’ll leave this session ready to start building your next application (world-changing or otherwise) with Cassandra.
B
A
Kind
of
the
sum
up,
some
of
the
the
information
that
you've
gotten
today,
I
wanted
to
kind
of
walk
through
what
it's
like
to
build.
Your
first
application
on
top
of
cassandra,
so
you
kind
of
got
a
little
bit
about
data.
Modeling
you've
learned
a
little
bit
about
languages.
You've
learned
a
little
bit
about
how
to
monitor
things
in
production,
maybe
how
some
of
the
internals
work
and
everything
like
that.
So
what's
it
like
to
actually
build
a
build,
an
application,
and
so
I
didn't
get
a
chance
to
do
this
earlier.
A
So
I'll
probably
do
it
at
the
end
of
the
beginning
of
the
but
kind
of
introduce
myself
and
sort
of
my
background
and
what
maybe,
why
I'm
giving
this
talk?
So
I'm
a
language
evangelist
which
is
a
totally
made-up
title
language
part
of
my
title,
basically
means
that
I
focus
on
a
particular
community,
in
my
case,
the.net
community.
So
I'm
a
long
time
developer.
I've
been
writing
obligations
for
like
15
years
now,
and
I
recently
got
a
chance
to
co-present
at
at
the
cassandra
summit
in
san
francisco.
A
So
I
don't
know
if
anybody
went
to
the
casino
summit
in
san
francisco
we're
doing
it
again
this
next
year.
I
think
it's
san
jose,
but
I
got
to
co-present
microsoft,
which
was
kind
of
cool,
but
I'd
love
that
so
that
I
can
show
you
this
picture
of
what
I
think.
Most
people
thought
about:
microsoft,
being
a
presenter
at
the
cassandra
summit.
A
A
So
this
is
what
we're
going
to
talk
about.
I
mentioned
this
earlier.
In
my
my
talk.
It's
called
killer,
video.
It's
a
it's!
A
video
sharing.
My
application
so
think
about
youtube.
Basically,
it's
a
youtube
computer
or
youtube
clone.
If
you
prefer,
users
can
add
videos,
they
rate
them.
They
comment
on
them.
You
can
search
for
videos
by
tag
that
kind
of
thing.
I
think
everybody
kind
of
knows
youtube
at
this
point.
A
A
You
know
if
you're
looking
for
just
some
schema
examples
of
what
kind
of
like
cql
looks
like
it
originally
was
a
cql
schema,
so
my
boss,
patrick
used
to
take
this
ul,
schema
around
and
show
it
to
people
that
were
kind
of
learning
the
center
to
kind
of
show
off
some
different
data,
modeling
techniques
and
how
it's
different,
maybe
than
relational
database
modeling.
So.
A
An
interesting
use
case
not
just
to
show
off
data
modeling
kind
of
techniques,
but
it's
also
kind
of
cool
if
you
think
about
the
scale
that
something
like
youtube
operates
at.
So
I
put
some
stats
from
from
youtube
up
there.
More
than
a
billion
unique
users.
Each
month,
100
hours
of
video
are
uploaded
to
the
site
every
minute
to
youtube.
The
scale
must
just
be
enormous.
If
you
were,
if
you
were
really
going
to
build
something
like
this
to
compete
with
youtube,
the
scale
would
be
pretty
awesome,
so
you
probably
have
lots
of
machines.
A
This
might
be
a
great
use
case
for
something
like
cassandra,
probably
multiple
data
center.
It's
just
it's
a
good
use
case
from
that
perspective
as
well,
so
this
is
kind
of
how
we're
going
to
go
through
it.
We're
going
to
talk
a
little
bit
about
things
to
do
before
you
get
data
modeling,
then
we're
going
to
talk
about
the
actual
data
model
itself.
Then
we'll
talk
about
the
application.
A
Software
architecture
we'll
talk
a
little
bit
about
kind
of,
what's
going
on
architecturally
with
the
site
and
just
because
I
love
software
architecture
and
the
future
as
well
so
kind
of
since
we
built
this
thing
on
cassandra,
what
are
some
directions
in
different
directions?
We
can
take
this
application
and
take
this
demo
and
kind
of
run
with
some
other
interesting
things.
Maybe
we
could
we
could
do
with
it
in
the
future.
A
That
you're
doing
when
you're,
when
you're
building
an
application
on
top
of
a
relational
database,
you
should
keep
doing
it
like
a
lot
of
it
still
applies.
So
first
thing
is
getting
to
know
your
data,
so
one
of
the
things,
one
of
the
things
that
I
have
in
the
system,
what
what
are
the
relationships
between
these
things?
This
is
also
a
lot
of
times
referred
to
as
sort
of
conceptual
data
modeling.
So.
A
Entities
a
lot
of
times
this
is
entities
and
relationships
between
entities
and
in
order
to
do
this,
so
if
you're
already
doing
this
in
the
relational
database,
we'll
keep
doing
it.
There's
no
reason
why
you
shouldn't
think
through
you
know
what
are
the
things
I'm
going
to
have
in
my
system
and
how
are
they
going
to
relate
to
each
other?
So
you
might,
you
might
end
up
with
something
that
looks
a
little
bit
like
this.
Have
anybody
seen
a
chat,
notation
diagram
before,
like
anybody
glaze
over
when
they
see
a
background
like
this?
A
A
Database
stuff
for
a
living,
this
is
their
job
to
produce
diagrams
like
that,
if
you've
never
seen
them
before.
Basically,
in
generation
diagram
the
big
boxes,
there
are
some
of
the
entities,
some
of
the
things
that
we've
got
so
users
and
comments
and
videos.
Then
we've
got
kind
of
the
relationships
between
them
and
the
diamonds
here
so
user.
They
post
comments
and
they
use
your
ads
videos
to
the
site,
we've
even
labeled
them,
one
too
many
and
many
too
many.
You
know
kind
of
label
the
size
of
the
relationship
and.
A
A
Double
circle
down
in
the
bottom
bottom
corner
there,
I've
never
seen
one
of
these
before
is
basically
a
collection
of
related
things,
so
videos
when
they
get
added
to
the
site.
This
is
a
they
get
tagged,
so
we
let
we
let
users
go
ahead
and
add,
like
basically
categorize
them
kind
of
add
tags
that
are
keywords
that
are
appropriate
for
them.
So
videos
have
tags,
they
have
this
collection
of
tags
associated
with
them.
A
A
So
once
you
kind
of
figure
out
okay,
what
are
the
things
in
the
system?
You
know
how
they
relate
to
each
other,
then
you
want
to
start
thinking.
Okay,
what
are
my
workflows
and
again?
This
is
something
that
you
are
absolutely
already
doing:
you're
not
going
to
build
an
app.
You
know
a
web
app
or
a
mobile
app
or
whatever
it
is,
or
a
service
whatever
it
is.
You're
not
going
to
build
it
without
knowing.
Why
am
I
building
this?
A
A
A
This
is
an
example
of
some
of
the
workflows
we've
got
going
on.
So
if
we
start
like
in
the
top
left
corner
user
shows
up
at
the
site
they're
going
to
log
in
and
then
we
go
down
to
the
bottom
there
we
might
want
to
show
them
their
profile.
Information
so
show
them.
Some
basic
profile.
Information
maybe
show
the
videos
they've
posted
to
the
site,
show
the
comments
that
they've
added
to
the
site.
You
know
to
various
videos
on
the
site
and
then
in
the
top
right
hand,
corner
there.
A
So
you
know
here's
the
latest
comments
and
then
also
show
ratings.
So
we
kind
of
got
like
a
one
to
five
star,
like
amazon,
netflix
kind
of
style
rating
system
on
there
as
well.
So
you
might
want
to
show
like
here's,
how
many
people
upgraded
and
here's
the
average
rating
for
a
video
or
something
like
that.
A
So
again,
once
you
kind
of
have
those
workflows
a
lot
of
times,
what
you'll
find
out
is
that
there's
sort
of
a
one-to-one
correlation
between
the
workflows
on
your
site
and
the
queries
that
you
need
to
do
so
here.
I
kind
of
try
to
group
the
workflows
by
sort
of
like
what
entity
they
deal
with.
So
you
can
see.
We've
got
two
different
ways:
two
different
workflows
related
to
users
in
two
different
ways,
essentially
that
we
need
to
query
user
data
so
when
user
logs
into
the
site,
we're
probably
going
to
want
to
look
them.
A
A
B
A
So
now
that
we've
kind
of
got
now,
we've
kind
of
got
this
idea
of
conceptual
data
modeling,
and
we
got
that
out
of
the
way
we
kind
of
thought
about
our
data.
We
figured
out
our
workflows
from
our
workflows.
We
know
our
queries
now
we
can
go
about
actually
building
a
data
model
and
I
call
this
a
data
model
for
cat
video,
because
everybody
knows
that
the
internet
loves
cat
videos.
We
were
really
building
youtube.
A
That
would
be
like
90
of
the
traffic
we
served,
probably
and
if
you're
curious,
just
how
popular
or
maybe
you're
curious.
I
know
I
was
just
how
popular
cats
are
on
the
internet,
so
of
course,
this
being
the
internet.
Somebody
had
already
done
the
research
for
us,
and
so
these
are.
These
are
from
a
national
national
infographic
from
2013.
B
A
This
is
this
was
looking
at
how
many
monthly
google
searches
there
are
for
cat
videos.
So
it's
like
something
like
31
million
for
cats
and
almost
three-quarter
of
a
million
for
cat
videos.
It
was
fun
they're,
actually
looking
at
cats
and
bacon,
and
how
popular
cats
and
bacon
were-
and
I
was
like
that's
funny-
but
I
think-
grab
a
bog's
bacon
for
some
reason,
but
ask
john
later
what
he
thinks
about
bacon.
You
will
get
an
enthusiastic
response.
I
guarantee,
then
this
is
the
bottom
half
of
the
infographic
and
there's
like
two
things.
A
I
wanted
to
point
out
here
so
first,
the
keyboard
cat
statistics,
so
keyboard
cap,
video
everybody's
seen
the
keyboard
video
before
it's
like
it
says
here:
31
million
views.
These
are
these
stats
are
a
little
bit
out
of
date.
But
if
you
go
and
look
now,
keyboard
cat
actually
has
over
37
million
views,
since
it
was
uploaded
in
back
in
2007.
A
If
you
do
the
math
that
averages
out
to
about
13
000
views
a
day
or
one
view,
every
nine
seconds
which
just
kind
of
reinforcing
again
the
scale
of
something
this
is
for
a
single
video
right
so
like
the
scale
of
something
that
something
like
youtube
must
operate
at
it.
That's
pretty
impressive.
The
other
thing
that
I
thought
was
funny
was
the
grumpy
cats
that
so
they
said,
facebook,
the
facebook.
B
A
So
let's
talk
about
data
modeling
again,
so
we're
going
to
build
the
data
model,
so
casino
allows
us
to
queries
that
can
scale
across
the
cluster
right.
So
we're
always
going
to
include
a
value
for
our
partition,
key
and
optionally
clustering
columns.
If
we
can-
and
we
know
our
queries
so
we're
going
to
follow
this
sort
of
table
per
query
sort
of
pattern-
we're
going
to
build
tables
to
answer
queries
similar
to
this
idea
of
materialized
views
from
from
the
relational
database
world
and
we're
going
to
try
and
denormalize
at
right
time.
A
A
Take
care
of
our
second
query
of
being
able
to
find
users
by
id
and
then
to
do
the
second
query
where
we
need
to
find
the
first
query,
I
guess
to
find
a
user
by
email
address.
We
probably
just
use
a
secondary
index,
we
have
an
index
to
the
email
address
column
and
we
call
it
a
day,
I'll
always
well
good.
The
problem,
though,
like
once
we
you
know,
maybe
we
built
the
most
popular
application
ever
and
all
of
a
sudden.
A
A
Would
be
really
great,
like
it's
really
important
that
people
be
able
to
log
into
the
site
and
this
single
box
that
we're
keeping
all
the
user
data
on,
like
that's
a
pretty
bad
single
point
of
failure,
if
that
ever
goes
down,
like
users
won't
be
able
to
log
into
the
site
and
that's
bad
like
we're
not
making
money
when
that
happens.
So.
A
B
A
A
A
Between
the
two
tables,
so,
for
example,
email
addresses
in
both
tables
user
ids
in
both
tables
and
that's
okay,
when
users
sign
up
for
the
site,
we'll
be
doing
probably
a
batch
inside
of
cassandra
to
insert
data
into
both
of
those
tables
and
it's
okay,
it's
hard
to
get
your
mind
around
this
whole.
Third,
normal
form
is
not
the
way
to
do
it,
but
that's
the
way.
We
do
it
at
the
same
time
and
we
don't
always
duplicate
all
the
data.
So
that's
very
much
driven
by
the
query
requirements
right.
A
So
what
data
do?
I
actually
need
to
support
that
workflow
or
support
that
query
like
we
don't
need
the
user's
first
name
and
last
name
when
we're
verifying
their
password
when
they
log
into
the
site.
Here's
another
example.
So
we
got
videos
like
we
said
there
was
like
all
kinds
of
different
ways.
We
wanted
to
query
videos,
so
you
can
see,
we've
got
a
videos
table
again
sort
of
an
entity
table.
It's
got
a
primary
key
of
video
id
be
really
easy
to
look
videos
of
id.
A
Then
we
also
have
the
table
to
be
able
to
find
find
videos
by
users.
So
we
want
to
show
you
hey.
Here's
the
latest
videos,
this
user
has
added
to
the
site
and
you
can
see:
we've
got
primary
user
id
so
that
we'll
store
all
of
their
videos
together
in
a
partition,
and
then
we've
got
added
date.
Video
id
combined
as
the
as
the
cluster
of
columns,
and
you
can
see
we're
taking
advantage
of
it
with
clustering
order
by
again,
because
we
decided
hey
when
we
query
this,
we're
probably
going
to
be
doing
hey.
A
A
So
this
has
come
up
a
few
times
today.
So
here's
two
more
of
the
video
video
use
cases,
I'm
not
going
to
show
you
the
schema
form
again,
it's
open
source.
You
can
absolutely
grab
this
on
my
head
out
if
you're
interested
in
the
schema
one
you
yes,
you
know
you
guys
have
seen
enough.
I
think
everyone
here
has
seen
enough
at
this
point
to
kind
of
guess
what
the
schema
would
look
like,
but
keep.
A
A
You
know
how
frequently,
I
guess,
will
it
change
and
then
then
you
have
to
start
thinking
about.
Okay,
do
I
need
to
maintain
consistency
or
integrity
across
all
these
copies?
You
might
be
in
a
scenario
where
you
don't
need
to
right,
so
it
might
be
totally
fine
and
some
of
these
some
of
these
copies
are
out
of
sync.
But
if
you
do
need
to
maintain
integrity,
then
you
need
to
start
asking
yourself.
A
A
A
So
we
said
when
videos
get
uploaded
to
the
site,
users
get
to
tag
them
with
keywords,
so
they
get
to
add
sort
of
a
small
collection
of
tags,
and
so,
if
you
have
a
small
collection
of
related
things
and
small
being
the
keyword
here,
sequel
collections
can
be
really
a
really
nice
way
to
sort
of
model.
That
a
good
example
of
something
that
maybe
isn't
a
small
collection
of
related
things
would
be
say.
The
comments
on
a
video
like
we've,
already
showed
you
here.
A
A
The
other
thing
I
wanted
to
talk
about
is
this
idea
of
client-side
joints
so
a
lot
of
times.
This
is
actually
the
real
data
model
right
now
inside
of
killer
video
and
a
lot
of
times.
If
you're
coming
to
the
relational
database,
background
you'll
be
tempted
to
get
into
this
normalization
game,
this
third
normal
form
kind
of
game
again,
and
so
here's
an
example
of
the
view.
A
Video
page,
the
top
of
this,
the
screenshot
here
is
from
the
view,
video
page
when
somebody
actually
clicks
on
a
video,
and
we
want
to
show
them
the
details,
so
we
can
play
it
back
and
what
happened
was
when
I
originally
built
the
site,
and
I
originally
did
the
schema
for
some
reason.
I
didn't
include
the
author's
information
on
this
page.
So
for
some
reason
I
mean
it
makes
total
sense
to
put
the
author's
information
on
this
page.
A
A
That
query
comes
back.
I've
got
to
pull
the
user
id
out
of
the
data
that
comes
back
and
then
go
to
a
follow-on
query
to
the
user's
table
where
I
load
the
user
up
by
their
user
id,
and
when
you
get
in
this
situation,
where
you're,
basically
recreating
sql
joins
but
you're
doing
it
in
your
app
here,
you're
saying:
hey:
well,
cassandra
won't,
let
me
do
joins.
I
can
just
get
around
it
by
you
know,
kind
of
doing
it
manually
in
the
app
you
need
to
start
thinking.
A
Okay,
maybe
maybe
I'm
doing
something
wrong,
so
you
need
to
start
thinking
what
is
the
cost,
and
these
might
be
okay
in
small
situations.
So,
like
this
example
here,
where
we're
doing
one
query
and
then
we're
only
doing
one
follow-on
query
to
get
user
information,
it
might
be
okay,
but
these
don't
scale
over
the
long
term
so
avoid
these
absolutely
avoid
these
when
possible.
If
you
can't
so.
B
A
A
Solve
it
in
the
in
the
demo
application
at
some
point
we
will
I'm
sure,
but
it's
such
a
good
illustration
of
things
not
to
do
that.
I
kind
of
left
for
that
purpose,
so
one
possibility
might
be
to
alter
the
videos
table.
So
you
can
see
I've
got
three
alter
table
statements
up
here
where
we're
adding
the
columns
for
the
user's
first
name
last
name
and
email
address.
We
need
their
email
address
to
generate
that
pretty
little
gravatar
picture
of
them.
B
A
We
might
just
put
that
data
directly
into
the
videos
table
and
then,
when
we
do
our
query
against
the
videos
table,
we
got
all
the
data
that
we
need
right
there
in
one
table.
The
other
option
we
have
is
we
can
create
this
user's
name
table.
So
you
can
see.
We've
got
this
second
table
here
on
the
right,
where
we've
created
a
table
full
of
the
same
user
data,
so
we've
got
users
first
and
last
name
email,
but
it's
got
a
primary
key
and
video
id.
A
So
it's
you
know:
users
buy
the
videos,
they've
uploaded
user
data
by
the
videos,
they've
uploaded
and
you
might
say,
okay
well.
Why
is
that
solution
on
the
right?
Why
is
that
any
better
like
why
any
better
situation?
I
still
have
to
do
a
second
query
in
that,
in
that
situation
like
how
is
that
any
better
than
the
solution
you
have
now?
The
difference
is
that
the
table
on
the
right
is
key
by
video
id
and
the
video
is
capable
as
key
by
video
items.
A
A
So
again,
just
to
just
to
kind
of
reiterate
this
remember
the
considerations
when
you
duplicate
when
you
duplicate
data,
so
you
know,
if
I
did
do
that
scenario
either
of
those
scenarios
where
I
end
up,
duplicating
the
user
information
into
another
table.
So
what
happens
when
the
user
changes
their
name
or
their
email
address?
You
know
I'm
probably
going
to
want
to
update
that
data
on
the
videos
page,
how
how
likely
are
they
to
change
their
name
and
email
address?
How
frequently
will
that
happen?
A
How
frequently
should
we
allow
them
to
do
that
sort
of
thing
and
do
I
have
all
the
information
and
update
those
duplicate
copies
so
another
thing
that
you'll
run
into
with
cassandra?
We
didn't-
I
don't
know
at
least
actually
come
up
today,
but
another
thing
that
you'll
run
into
is
sometimes
cassandra
rules
can
sort
of
impact
the
design
choices
that
you
make
and
consider.
It
has
a
few
things
kind
of
funky
rules,
maybe
that
you're
not
used
to
from
from
relational
databases
so,
for
example,
video
ratings.
A
It
would
be
really
nice
if
we're
going
to
track
video
ratings,
that
we
can
use
counter
columns.
So
cassandra
has
a
column
type
called
counter.
That's
used
to
do
approximate
accounts
of
things,
and
so
it
would
be
really
nice
is
we're
always
going
to
show
the
ratings
data
on
the
you
know,
along
with
the
video
data,
to
just
put
those
columns
directly
into
the
videos
table
so
we'll
be
storing
two
things
here,
we'll
be
storing
a
rating
counter,
which
is
basically
how
many
people
have
rated
a
given
video.
A
You
know
so,
every
time
somebody
gives
a
rating,
we
increment
it
by
one,
and
then
we
also
have
a
readings
total
column
here,
which
is,
if
somebody
gives
it
five
stars,
we
increment
five.
Somebody
gives
it
three
stars
a
week
and
a
three,
and
since
we've
got
a
total
we've
got
a
count.
We
can
figure
out
an
average
rate
because
we
said
we
wanted
to
show
an
average
on
the
quantity.
A
So,
unfortunately,
the
rules
center
has
rules
about
counters
and
that
rule
is
and
I'll
try
and
explain
this.
Clearly,
the
rule
is,
if
you
have
columns
like
these
radiant
counter
and
rating
total
columns
that
are
of
type
counter,
then
every
other
column.
That's
in
that
table
that
has
the
counter
columns
every
other
column
has
to
be
part
of
your
primary
key
and
that's
kind
of
a
weird
rule.
A
You
know
that
cassandra
has
so
if
we
wanted
to
do
this
idea
on
the
left,
where
we
put
the
counter
columns
directly
in
the
videos
table,
we
would
have
to
add
name
and
description
and
user
id,
and
all
these
other
things
we
would
have
to
add
all
of
them
to
the
primary
key
to
be
able
to
satisfy
that
restriction
that
the
sander
has.
So
what.
A
Is
you
end
up
with
these
this
idea
a
lot
of
times
with
countings
in
particular
counter
tables,
so
you
have
the
counters
kind
of
table
by
themselves
with
the
primary
key
of
some
kind.
So
again,
we've
got
video
ratings.
You
can
see,
it's
got
two
counter
columns
and
it's
got
a
primary
key,
video
id
and
again
because
it's
key
by
video
id
when
somebody
shows
up
at
this
page,
we
can
run
this
query
in
parallel
with
our
query
to
go,
get
the
videos
data
because
it's
got
the
same
same
video
id
as
its
primary
keys.
A
A
A
And
so
the
thing
with
the
thing
with
this
is
that
all
reads
and
all
rights
as
people
are
adding
videos
to
the
site
and
people
are
reading
back
hey.
What
are
the
latest
videos
that
have
been
added
to
the
site
for
a
given
day,
all
the
reads
and
writes
are
going
to
go
to
a
single
partition,
and
that
means
that
all
the
reads
and
writes
are
going
to
go
to
a
single
node
or
a
set
of
nodes
or
replicas,
and
that
could
be
bad.
That
could
create
a
hotspot.
A
We
could
absolutely
end
up
overwhelming
that
machine
with
the
traffic
from
from
all
these
reads
and
rights
go
through
a
single
go
to
school.
Node
you'll
find
this
especially
with
anytime.
You
try
to
use
like
anything
time
based
as
part
of
your
partition.
Key
like
that's
a
lot
of
times.
That's
a
signal
that
maybe
you
should
start
thinking
about
hey.
Maybe
maybe
I'm
going
to
create
a
hotspot
here.
A
So
what
are
some
possible
solutions
to
this?
So
really,
what
we
want
to
do
is
we
want
to
spread
the
load
out
amongst
more
nodes
and
cluster
right.
So
we
need
to
do
something
to
change
the
partition
key
something
to
spread
that
load
across
other
nodes
in
the
cluster,
and
so
we
mitigate
this
by
adding
data
to
our
partition
keys.
So
this
is
the
first
example
I
think
you've
seen
in
this,
where
we
do
a
compound
partition
key.
A
So
I
mentioned
that
you
can
do
this
you'll
see,
there's
an
extra
set
of
parentheses
in
our
primary
key
definition,
where
we're
saying
use
your
year
month
month,
day
date,
plus
some
arbitrary
bucket
number
in
this
example
here.
So
there's
really
two
ways
that
we
can
do
this
when
we're
adding
data
to
the
partition
key
one
might
be
to
add
some
some
part
of
some
piece
of
data,
that's
already,
naturally,
part
of
our
domain.
So
let's
say
instead
of
tags
like
maybe
when
big
people
add
videos
to
the
site,
we
would
make
them
categorize.
A
A
In
the
example
here
we
can
just
add
some
arbitrary
piece
of
data
to
it,
so
it
could
be
a
buck
number
in
this
case,
we're
just
using
an
integer
bucket
number,
and
then
we
would
round
robin
at
the
application
level.
At
our
apps
data
access
level,
we
would
send
reads
and
writes
to
different
buckets
in
a
sort
of
round-robin
fashion
to
kind
of
balance
out
the
load
across
the
cluster.
A
A
In
my
in
my
cluster,
by
doing
this
sort
of
partitioning
so
now,
let's
talk
a
little
bit
about
building
the
actual
application
itself,
so
you've
kind
of
done
conceptual
data,
modeling
you've
got
you've,
got
physical
data,
modeling,
that's
going
on
with
the
site.
Now,
let's
talk,
what's
it
like
to
actually
build
an
application,
so
john
showed
you
python
earlier
today,
awesome
if
you're
a
python
user
or
if
you're
new
to
programming
and
you're.
Looking
to
absolutely,
I
find
this
awesome.
These
are
when
we
actually
have
drivers
available,
they're
all
open
source.
A
So
these
are
official
data
stacks
drivers
that
we
have
available
on
our
github
they're.
All
under
apache
2
license
pretty
much
all
the
all
the
languages
that
you
would
expect
on
the
right-hand
side.
There
are
some
that
will
probably
happen.
So
you
know
there
have
been
some
pretty
pretty
high
level
talks
and
maybe
some
progress
has
even
started
on
them
behind
the
scenes
and
then
we've
got
some
early
discussions
for
go
and
rust.
A
It's
very
very
similar
a
lot.
The
same
concepts
apply
across
all
of
the
across
all
the
data
stacks
drivers,
so
it's
kind
of
cool,
because,
if
you're
somebody
who
works
in
multiple
languages,
you
can
take
pick
up
and
take
a
lot
of
the
knowledge
that
you
have
in
the
language
that
you're
using
now
and
use
cassandra
in
another
language
have
a
lot
of
the
same
concepts
and
a
lot
of
the
same
knowledge
that
you
already
have
about.
The
driver
applied
other
languages
as
well.
You'll
see
just
how
similar
some
of
the
bootstrap
you
know.
A
Code
is
kind
of
a
cluster
object.
This
builder
and
contact
points
very
very
similar,
so
I'm
going
to
talk
about
three
things:
three
yeah
three
sort
of
features
of
the
of
the
driver
that
kind
of
apply
and
the
code
samples
that
I
show
you
will
be
c
sharp
because
killer
videos
and
c
sharp,
but
but
absolutely
these
these
ideas.
These
concepts
apply
to
your
language
of
choice
as
well.
So
the
first.
B
A
Is
prepared
statements-
and
I
imagine
john
probably
talked
about
these
in
his
python
presentation
this
morning,
but
this
is
a
performance
optimization
for
queries
that
you're
going
to
run
over
and
over
and
over
again.
So,
if
you
think
about
it,
if
you've
built
built,
say
web
applications
or
applications
before
or
services
a
lot
of
times
when
we're
building
against
a
relational
database,
we'll
be
running
the
same
queries
over
and
over
and
over
again
just
with
different
variable
values.
A
A
Select
from
the
user
credentials
table
like
so
when
somebody's
going
to
log
in
and
we
want
to
look
them
up
by
email
address.
This
is
probably
the
query
that
we're
going
to
run
and
if
you
haven't
seen
this
question
mark
before
that's
just
like
in
sql,
we
have,
I
think,
it's
the
ad
symbol
in
sql
server
and
it's
colon
and
oracle,
as
the
placeholder
for
variable
values.
A
Is
that
you're
going
to
prepare
once
so
I'm
going
to
pay
this
cost
to
do
the
round
trip
to
let
sandra
figure
out
how
to
execute
this
thing
once
and
then,
with
that
prepared
statement
that
comes
back,
I'm
going
to
find
and
execute
against
it.
Many
many
times,
I'm
going
to
run,
run
that
statement
over
and
over
again,
maybe
with
different
variable
values.
When
I
do
and
cassandra
then
doesn't
have
to
do
this,
this
whole
parsing,
the
cql
string,
it
doesn't
have
to
figure
out
how
to
execute
it.
A
We
don't
have
to
send
the
cql
string
over
the
wire.
All
we
have
to
do
is
send
variable
values
over
the
wire,
so
it's
a
great
optimization,
if
you're
going
to,
if
you're
going
to
be
running
the
same
statement
over
and
over
again,
just
with
different
variable
values.
So
absolutely
take
advantage
of
these.
A
A
Where
they
prepare
and
bind
and
execute
and
prepare
and
buy
and
execute
every
time
through
this
method
and
if
you're
not
doing
the
repair
once
and
then
kind
of
saving
it
off
and
reusing
it
and
then
doing
bind
and
execute
many
if
you're
preparing
every
time
through
you're,
basically
doing
two
round
trips
to
the
server,
so
you're
actually
going
to
get
worse
performance
than
you
would.
If
you
were
just
executing
using
one
of
the
other
statements
in
the
driver.
So
remember,
prepare
once
buy
mexican.
A
Batch
statements,
so
this
this
has
certainly
come
up
multiple
times
today.
So
a
couple
things
to
know
about
batch
statements,
first
of
all,
what
they
aren't.
So
these
are
not
a
optimization
to
vocal
of
data.
So
a
lot
of
times
batches
and
other
databases
are
used
to
kind
of
okay.
Let's,
let's
load
some
data,
let's
use
a
batch,
not
what
they
are
boring.
A
A
So
you
use
these
when
you
want
that
sort
of
atomic
guarantee
and
you
can
mix
and
match,
there's
multiple
standing
types
in
the
driver,
so
I
mentioned
prepared
and
bound
statements,
kind
of
showed.
That
example
there's
another
one
called
symbol
statement,
for
example
in
the
c-sharp
directory,
there's
an
analog
in
java
as
well,
where
it's
just
kind
of
like
hey.
I
just
want
to
execute
this
query:
real,
quick.
You.
A
And
attach
the
different
different
statement,
types
that
are
available,
so
this
is
going
to
be
hard,
probably
to
read
back,
but
this
is
an
example
and
there's
a
couple
things
to
call
out.
I
guess
in
this.
In
this
example,
this
is
a
sample
data
access
class
that
we've
got
that's
powering
our
video
catalog
and
a
couple
things
to
call
out
so
first
is
that
we're
mixing
and
matching
bound
statements
and
simple
statements
in
our
batch.
A
A
Let's
see
with
the
constructor,
which
is
what
that
top
method
is
it's
the
constructor
for
that,
for
that
class
constructors
are
guaranteed
to
only
run
one
time
for
a
given
instance
of
an
object,
same
thing
I
imagine
in
most
programming
languages
as
well,
and
so
that's
where
we're
actually
doing
our
prepared
we're
doing
session
dot,
preparer
and
providing
the
statement.
So
this
is
just
one
example
of
how
you
can
do
the
prepare
once
and
make
sure
that
you
only
prepare
once
and
then
mine
can
execute
many
times
so
one
one
way
to
potentially
do
it.
A
So
last
thing-
and
I
know
these
have
been
mentioned
a
couple
of
times
as
well,
so
lightweight
transactions,
so
these
are.
This
is
sometimes
referred
to
as
linearizable
consistency
or,
if
you're
from
a
relational
database
background
you
might
know
serial
isolation
level.
So
maybe
you
have
multiple
writers
writing
to
the
same
table
and
you
want
to
make
sure
that
rights
don't
step
on
each
other.
So
you
want
to
make
sure
that,
like
two
rights,
don't
try
to
write
conflicting
records.
You
know
with
the
same
sort
of
key
to
that.
So
this.
A
A
A
So
this
is
what
it
kind
of
looks
like
to
actually
do
that
scenario
in
killer
videos,
so
youtube,
never
inserts
it
we're
inserted
into
that
user
credentials
table
that
we
saw
earlier,
and
we
say,
email,
password,
user
id
and
you'll
see
at
the
end
of
the
statement.
That
says,
if
not
exists,
that,
if
not
exist,
syntax
is
one
of
the
one
of
the
ways
that
you
can
trigger
a
lightweight
transaction,
some
of
the
syntax
that
you
use
to
use
a
lightweight
transaction,
and
so
when
you
do
that,
you're
basically.
B
A
Hey
insert
this
record
if
one
with
that
same
primary
key
doesn't
already
exist.
So
if
one
with
that
that
email
address,
because
you'll
remember
this
table
was
keyed
by
email
address
answer
this.
If
somebody
with
that
email
address
doesn't
already
exist
and
what
we'll
actually
get
back
from
the
driver
is
we'll
actually
get
back,
a
special
column
called
square
brackets
applied
and
that
basically,
is
just
successive
failure.
A
successor,
failure
to
a
false
boolean
that
says
whether
or
not
your
insert
was
successful
or
not.
A
A
A
A
Of
the
logical
architecture
of
what's
going
on
with
the
demo
so
top
layer,
we've
got
a
web
ui
html5
javascript,
typical
sort
of
thing
that
you've
got
then
on
the
server
level.
We've
got
an
asp.net
snbc
app.
So
this
is
the
thing
that's
actually
serving
up
the
web
ui.
It's
answering
json
requests
and
sort
of
routing
them.
You
know
typical
sort
of
mpc
style
web
application,
then,
underneath
that
we've
got
a
sort
of
tier
of
services,
so
we've
got
you
know
video,
catalog
and
user
management.
This
is
just
four
of
the
services
involved.
A
I
think
there's
eight
or
nine
total,
where
things
are
kind
of
grouped
together
and
I'm
trying
to
put
them
together,
sort
of
by
business
capabilities,
so
less
about
that's
sort
of
about
the
entity
involved
than
it
is
sort
of
what
does
it
actually
do
for
the
site?
What
is
the
capability
that
it's
providing
on
the
site
so.
A
Keeps
track
of
the
available
videos
that
are
there
to
view,
whereas
the
upload
service
it
handles
processing
of
uploaded
videos,
because
we're
gonna
have
to
take
them
from
some
source
format
and
convert
them
into
something
that
can
actually
be
played
back
in
a
web
browser
or
a
mobile
device.
So
its
job
is
to
kind
of
do
that,
storing
and
encoding
and
re-encoding
and
generating
thumbnail
previews.
A
That
kind
of
thing
on
the
bottom
layer,
then
we've
got,
we've
got
some
big
infrastructure
pieces
involved,
so
we're
using
app
center,
which
is
a
free,
managing
and
monitoring
laundry
tool
for
data
stacks
for
for
provisioning
due
to
sander
clusters
standard
clusters,
you
get
this
nice
little
web
dashboard.
That
kind
of
shows
you
like,
what's
going
on
with
your
reading
right
throughput
and,
what's
so
very
very
nice,
to
have
very
easy
to
like
provision
clusters
as
well
provision
new
nodes
and
that
cluster
we've.
A
The
same
cluster
itself,
dse
powered,
that's
actually
storing
all
our
app
data,
so
user
accounts,
video
metadata
and
all
that
kind
of
stuff,
and
then
we've
got
three
three
azure
services
that
we're
using
as
well,
because
the
live
demo
was
built
to
run
in
azure,
so
we're
taking
advantage
of
azure
service
bus,
which
is
just
for
pub
sub
messaging.
So
if
you've
ever
used,
rabbitmq
or
zero
mq
or
any
particular
messaging
technology
of
choice,
you've
probably
used
12.
A
messaging
before
same
idea:
azure
storage,
which
is
blob
storage,
so
we
actually
need
to
store
the
video
files
and
the
thumbnails
that
are
generated.
You'd
actually
store
those
somewhere.
So
we
store
those
in
azure
storage
and
then
we
have
azure
media
services,
which
is
this
kind
of
cool
thing.
That
azure
has
that's
a
finished
service
that
will
actually
do
re-encoding
of
videos
and
actually
extract
thumbnails
from
from
videos
and
whatnot
from
finished
files,
so
kind
of
a
cool
little
piece
of
infrastructure
as
well.
A
A
You
know
things
like
the
name:
the
description
where
the
thumbnail
is
located,
where
the
video
file
itself
is
located
and
then
it
also
does
stuff,
like
it
publishes
events
about
hey,
interesting
stuff
that
happened
inside
the
system.
So
you
might
say,
hey
a
new
youtube.
Video
was
added
or
hey
you
uploaded.
We
just
accepted
a
new
uploaded
video,
something
like
that.
A
A
To
store
data
about
uploaded
videos
in
cassandra
stuff
like
where,
where
are
the
files
getting
uploaded
to
like?
What's
going
on
with
the
coding
jobs?
How
far
along?
Are
they?
What's
the
status
that
sort
of
thing
it's
going
to
use,
azure
storage
again
to
store
the
actual
raw
files,
so
video
files,
thumbnail
files,
that
sort
of
thing
use,
azure,
media
services
to
do
the
re-encoding
jobs
and
to
extract
thumbnails?
A
B
A
So
this
is
nice
because
things
are
very
decoupled
right,
everything's
kind
of
decoupled
from
from
each
other.
We
have
services
like
calling
you
know
so
they're,
not
temporarily
coupled
kind
of
a
cool
cool
way,
and
it's
also
nice,
because
the
services
now
can
be
sort
of
scaled
independently.
They
can
be
deployed
independently
of
each
other.
They
can
be
versioned
independently
of
each
other
so
long
as
they
maintain
the
message
contracts
about
what
events
they
publish.
A
A
A
Catalog
that
says:
hey,
I
added
this
new
youtube
video
to
the
catalog,
and
then
you
have
two
other
services
subscribing
to
that
event
and
they
might
say,
suggested
video
systems
like.
I
need
to
figure
out
what
videos
to
suggest
for
that
new
video
that
just
got
added
to
the
catalog
or
you
might
have
a
search
that
says
I
better
index
that
so
that
it
starts
showing
up
in
search
results.
So
it
kind
of
happens
async.
A
So
last
the
future.
What
are
we
gonna?
What
could
we
do
now
that
we've
built
this
application
on
top
of
cassandra?
You
know:
we've
got
a
pretty
solid
architecture.
Things
can
be
scaled
out
really
nicely.
You
know.
Where
could
we
potentially
take
this
application
in
the
future?
And
so
I
don't
know
about
you
but
anytime.
Somebody
says
the
future,
like
I
always
think
of
the
future.
Does
anybody
know
that
the
year
3000
sketch
from
from
governor
ryan?
A
So
so,
basically
the
premise
is
like
come
in
and
andy
consider
like
they
dress
up
in
really
ridiculous
outfits
they're
supposed
to
be
from
the
year
3000.
They
make
ridiculous
predictions
about
what
it's
going
to
be
like
in
the
year
3000.,
and
I
thought
this
one,
because
we
were
doing
because
we
were
doing
youtube.
I
thought
this
was
like
especially
appropriate,
so
I
don't
know
if
everybody
would
be
able
to
hear
this,
but.
A
Okay,
so
what
what's
some
possibilities
where
we
can
possibly
expand
so
spark
or
azure
machine
learning,
potentially
for
doing
video
suggestions?
So
right
now
it's
very
sort
of
manual
process,
video
suggestions,
but
it
would
be
kind
of
cool
to
do
like
a
smart
machine
learning
job,
maybe
based
on
other
stuff.
You
viewed
or
other
people
similar
to
you
what
they've
viewed
and
these
could
suggest
videos
for
you.
We
can
do
video
search
via
solar,
so
right
now
we're
doing
video
searching
tags.
A
A
Service
like
what's
going
on
internally
and
sort
of
all
the
different
things
that
are,
you
know,
monitoring
with
media
services
and
everything
I'm
going
to
use
to
sort
of
sort
of
model
that
instead
and
store
the
actual
state
of
the
actors
inside
of
inside
of
the
standard.
So
we
might
use
it.net
reduced.net,
which
is
a
straight
port
of
the
scala
framework
that
I'm
sure
jdm
developers
are.
A
Orleans,
which
is
the
new
actor
framework
you
know
district
virtual
distributed
freeway
from
microsoft
and
then
the
other
thing
we
could
do
that
might
be
cool.
We
could
actually
store
the
file
data
itself
inside
of
cassandra
center
if
you
guys
are
familiar
with
the
pithos
project,
but
the
pivots
project
is
basically
an
s3
compatible
wire
store.
So
now
you
can
use
any
of
the
s3
amazon
s3
drivers
that
are
available
in
talk
to
fit
those,
and
it's
a
file
store
built
entirely
on
cassandra.
A
So
you
can
actually
store
your
binary
data
and
it
takes
care
of
chunking
it
up
efficiently
and
being
able
to
retrieve
the
binary
data
efficiently,
so
kind
of
a
cool
cool
source
project
for
those
I
think
their
website
is
like
fitness.io
or
something
like
that.
So
check
that
out.
If
you're
interested
in
a
file
store
on
top.