►
Description
Speaker: Edward Capriolo
SlideShare: http://www.slideshare.net/slideshow/embed_code/18573117
The ColumnFamily data model and wide-row support provides the ability to store and access data efficiently in a de-normalized state. Recent enhancements for CQL's spare tables and built-in indexing provide the capability to store data in a manner similar to that of relational databases. For many use cases hybrid approaches are needed, because complete de-normalization is appropriate for some access patterns whereas more structured data is appropriate for others. At times a single logical event becomes multiple insertions across multiple column families. Likewise a user request might require a several reads across different column families. This talk describes some of these scenarios and demonstrates how advanced operations such multiple step procedures, filtering, intersection, and paging can be implemented client side or server side with the help of the IntraVert plugin.
A
A
little
bit,
but
this
table
is
more
so
they
used
to
be
called
like
a
static
column
family
where
there's
one
row
essentially
per
user
and
all
the
columns
are
pretty
much
known
at
runtime.
There's
not
going
to
be
an
extra
column
in
this
table
called
wombats,
or
this
is
a
pretty
much
known
entity
and
primary
key
is
user
name,
meaning
if
we're
ever
looking
for
a
particular
user.
We
could
find
it
fast
because
Cassandra
is
primary
key
is
spread.
We
could
define
right
lots
of
queries.
A
You
could
say
well
give
me
the
first
name
and
last
name
where
people
are
named.
T.T
cod,
that's
a
very
fast
party,
because
the
primary
key
is
on
the
user
name,
and
so
we
could
get
it.
This
column,
information
first
name
last
name
very
quickly,
because
we're
really
most
of
the
work
is
just
finding
T
Cod's
location.
It's
not
parsing,
the
first
name
and
last
name.
Those
are
just
columns
in
a
row
we've
essentially
found
and
because
that
table
had
a
composite
primary
key.
A
There
is
some
things
I
could
do
like
I
could
look
for
a
video
ID.
This
is
almost
a
bad
example
I'm
set.
I
did
this
way,
but
because
the
video
name
is
part
of
the
primary
key,
the
column
is
sorted
by
the
primary
key.
So
if
I
happen
to
have
lots
of
things
with
the
same
uuid,
which
doesn't
make
sense,
I,
don't
know
why
I
did
that
example.
A
Cool
but
like
if
you're
coming
from,
like
a
person
who
use
the
database
before
you're
saying
so,
I
can
slice
or
I
could
query
what
can
I
sub-query.
You
know
what
about
where
clause
is
like,
where
first
name
is
bob,
and
last
name
is
last
than
the
concatenation
of
fish
sticks
or
something
that's
not
really
in
there
yet
and
by
the
way.
This
this
slide
is
not
like
a
damnation
of
Cassandra
all
the
stuff.
A
That
can't
do
there's
a
good
reason
why
you
can't
do
some
of
these
things
and
I'm
going
to
kind
of
mention
at
the
end,
here's
one
like
say
I
want
to
Union
two
slices
like
does
the
does
the
query
language
provide
me
a
way
to
say
slice,
this
data
slice,
this
data
and
send
me
back
the
intersection
of
the
Union
and
then
round
trips.
You
know
it's
if
my
dad
it
does
happen
to
naturally
spread
across
two
tables
or
column
families.
How
many
round
trips
are
gonna
have
to
do
like.
A
Given
my
examples,
if
you
saw
the
movie
table,
if
someone
only
wants
to
know
the
movie
and
the
tags,
that's
great
and
the
user,
that's
fine,
but
if
someone
says
I
want
the
movie,
the
tags,
the
user
and
the
users
first
name
now
that
might
have
to
be
two
different
operations,
because
first
I
have
to
look
up
the
movie
that
comes
down
then
I
say:
okay,
the
movie
was
written
by
user.
The
youth
did,
the
movie
was
uploaded
by
user.
Todd
now
get
me
Todd.
A
So
those
are
things
that
end
up
taking
a
lot
of
time,
the
more
round
trips,
the
more
lanes-
and
you
can
expect
you
know,
there's
procedure
issues
like
triggers,
as
you
mentioned,
Jonathan
mentioned,
he's
going
to
add
it
it's
coming
in
Cassandra,
20
and
then
also
Hagrid.
It's
like,
for
example,
say,
say:
I
want
to.
A
A
particular
slice
or
a
query:
does
that
aggregate
there
yet
and
what
I
mentioned
before
is
the
reason
that
some
of
them
aren't.
There
is
because,
as
you
guys
may
know,
Cassandra
could
store
two
billion
columns
in
a
single
row.
So
a
lot
of
these
operations
wouldn't
make
sense
for
Cassandra
like
if
I
have
two
slices
of
two
billion
rows
and
I
want
at
Union
those
that's,
obviously
a
tricky
problem
and
I
shouldn't
expect
Cassandra
to
do
that.
A
Real
time
is
when
people
kind
of
pick
up
Cassandra
because
they
want
like
the
latency
reads
and
writes,
and
databases
are
roads
ever
because
you
send
a
car
to
a
database
you're
not
always
expecting
it
to
come
back
right
away.
You
know
this
quarry
has
to
read
through
this
table,
maybe
its
genetic
data
to
a
temporary
table,
same
story
here.
A
So
in
a
lot
of
cases,
there's
not
a
good
reason
for
Cassandra
to
do
the
things
that
I've
mentioned,
but
some
of
them
theoretically
could
be
done
if
you
have
the
certain
types
of
data
conditions
for
them.
So
it's
kind
of
what
I'm
going
to
talk
about
today.
So
the
api's
you
have
are
there
600,
Eternals
and
there's
this
storage
proxy.
Sometimes
it's
called
the
fat
client
which.
A
Looked
the
thrift
API
than
you
drill
your
way
through
the
code
and
find
the
storage
proxy
they're,
actually
almost
identical,
but
cql
took
in
another
level
to
be
able
to
do
these.
These
queries
that
they
use
the
indexes
automatically
and
stuff
there's
more
application
logic
in
c
ql
right,
like
thrift,
just
assets
and
gets,
but
in
c
ql
you
can
say
where
the
key
is
in
this
or
that
or
that
and
CQ
on
those
at
the
column.
If
you're
using
an
in
flaws,
that
really
means
I
need
to
do
multiple
gets
or
the
multi.
A
A
Like
it
seems
like
I
should
be
able
to
do
that
right.
Just
two
slices
and
union
them
doesn't
seem
that
hard,
so
I
was
like
heck.
Let's
just
it's
not
like.
This
has
never
been
done
before,
like
Soul
Andrew
originally
had
its
own
fat
client
and
a
lot
of
people
have
written
their
applications
on
the
same
jvm
as
Cassandra.
For
these
reasons
and
Brian's
in
the
room
he's
one
of
the
guys
I
think
you
guys
ever
use
Virgil
anybody
all
right.
Oh
well,.
A
A
About
Virgil
right
you're
trying
to
make
me
make
hypermedia
so
people
who,
who
are
don't
have
the
cassandra
client
could
use
virgil
to
get
information.
So
if
you
imagine
that
request
flow
someone's
sending
a
request
to
virgil,
which
will
take
it
and
be
like
great
now,
I'm
going
to
send
that
request
to
cassandra
and
that's
not
super
efficient,
because
that's
two
sockets
when
it
could
have
been
one.
So
that's
kind
of
like
what
I
was
drilling
at
with
what
we
call
introverted
like.
Can
we
remove
those
scenarios
like
who
who
should
be
applying?
A
Who
should
that's
a
good
question
like
people
used
to
say
this
before?
If
you
look
at
they
way
back
the
old
trigger
to
trigger
tickets,
they're
like
when
I
do
a
client,
you
can
do
triggers
client-side.
You
know
you're
inserting
into
that
column,
family
right.
So
before
you
do
insert
into
the
ten
other
common
families
you
need
to
write.
That
was
the
logic,
but
there's
kind
of
something
about
that
like
moving
process
and
closing
the
move.
Processing.
Excuse
me
close
to
the
data
and
this
something
that
came
from
Hadoop
I,
don't
know
close
familiar.
A
A
Of
benefits,
you
know
less
network
I,
oh,
for
example,
if
I'm
doing
a
slice
that
returns
a
thousand
rows
and
I
really
only
need
it,
one
of
them.
Well,
that's
a
thousand
rows
that
had
to
come
over
the
network.
So
there's
you
know
more
network
data
and
there's
memory.
You
know,
like
we've,
heard
people
to
mention
Java
and
young
Gen.
So
when
you're
returning
more
something
you
have
to
remember,
you
have
a
client
and
server.
So
this
server
has
to
buffer
up
all
that
stuff.
A
I,
don't
know
how
it
works
and
send
it
across
the
client
and
they're
each
using
bigger
buffers
and
more
stuff,
and
you
can
do
new
stuff
like
the
Union
stuff.
We
talked
about
it.
So
the
way
we
did
it
was
we
I
was
just
going
to
be
like
hey,
let's
just
throw
a
tomcat
on
topic
Cassandra,
but
I
Nate,
who
was
one
of
the
neighbor
call.
Who
was
one
of
the
hector
committers
was
eventually
go.
There's
this
really
cool
thing
called
vertex
and
I
was
like.
A
A
Job
and
that
could
be
wrong-
maybe
one
of
the
vertex
guys
to
pop
and
slap
me
in
the
face
for
saying
that,
but
that's
the
way
I
look
at
it.
What
does
it
do?
That's
neat!
It's
got
a
synchronous.
Oh,
it's
got
hot
reloadable
modules,
which
will
kind
of
get
into
in
a
bit
and
it's
got
on.
The
modules
can
be
written
in
groovy
Ruby,
Java
JavaScript,
whatever,
because
vertex
is
almost
like
an
actor
model.
If
you
would
the
pieces
are
very
disc,
then
they
move
data
between
these
cues.
So
what
did
we
do
so?
A
And
then,
then,
that
kind
of
fellow
off,
so
why
HTTP?
Why
do
we
use
HTTP?
Well,
because
I've
worked
at
a
company
where
you
go
to
do
a
live
demo,
when
your
software
doesn't
use
HTTP
and
the
demo
doesn't
work
because
that
person's
firewall
provider
doesn't
let
anything
on
HP
through
so
HTT
HTT
P
is
very
universal.
A
It's
easy
to
secure!
You
know.
Some
of
these
things
were
like.
Oh
thrift
and
you
see.
There's
a
ticket
open
for
thrift,
security
right,
there's
no
ticket
open
for
age
to
be
security.
It's
called
SSL
compression.
Let's
compress
seek
you
out
again
HTTP
someone
already
figured
that
out
as
I
mentioned
it's
the
de
facto
way
to
do
everything
you
see
you
see.
Apps
that
should
be
like
you
know,
should
be
a
visual
basic
app,
but
someone's
trying
to
do
them
with
jquery
in
real
time
I
reduce
HTTP,
so
the
goal
of
really
introverted.
A
What
we
try
to
do
is
remember.
I
mentioned
those
round
trips
where
you're
sending
a
request
waiting
response
was:
can
we
do
this
with
less
round
trips
and
not
be
so
worried
about
trying
to
optimize?
How
big
is
this
slice
request
going
to
be
it's
10
bytes?
Can
we
make
it
eight
bytes?
It's
not
really
that
important.
If
you're
going
to
have
to
send
10
requests
to
get
the
data
you
want,
and
then
why
do
we
choose
Jason?
A
We
wanted
to
do
simple
things
like
in
certain
in
certain
cases
like
the
the
signature
for
batch
mutate
has
one
argument
and
get
return.
Something
and
slice
return
something
else.
The
concept
was
can't.
We
just
use
lists
and
maps
everywhere.
We
can,
and
this
way
there's
not
a
huge
deal
with
someone
looks
at
the
JSON
payload.
What
is
that?
How
do
I,
how
do
I
write
one
of
these
requests
we'll
get
into
that?
A
minute
request
is
composed
of
n
operations.
Each
operation
has
a
timeout.
A
A
On
top
of
that
again
we're
trying
to
limit
round
trips,
so
yeah
Jason's
got
a
little
more
packet.
You
know,
then
thrift,
oh
well,
it's
not
trying
to
be
mega
terms,
and
you
know
why
not
use
lightning,
fast,
serialization
library,
action
think
I
went
over
this
thrift
and
cogeneration
works
printing
job,
it
doesn't
work
in
PHP,
doesn't
work
with
nodejs
same
problems
that
row
TCP
dump
a
lot
of
times.
A
We
are
troubleshooting
Cassandra,
we're
like
what's
going
on
here,
you
know,
and
if
you,
if
you
can't
TCP
dump
the
protocol,
it's
not
always
easy
to
figure
that
out.
You
know
we
have
never
convince
and
when
they
want
to
check
the
sondra
like.
How
do
I
check
a
Sondra?
You
say
well:
go
to
maven
and
pull
down
like
seven
thousand
different
jars
to
do
a
set
request
and
you're
like
well.
That's
pretty
dumb
I!
Don't
want
to
do
that
so
again,
we're
limiting
round
trips!
A
A
This
feature
look
at
that
someone
should
be
able
to
look
at
this
JSON
and
figure
out
what
it
is
so
step.
One
is
pspace
set
to
is
the
column
family
step.
Three
is
slice
and
kind
of
take
all
the
byte
buffers
and
packing
composites
out
of
the
equation.
It
just
kind
of
make
Cassandra
ease
and
use
like
ma.
Go
there
I
said
it
right,
yeah
so
and
then,
when
you
get
back
as
a
result
again
something
simple,
so
you
know
easy
to
parse.
I
guess
I
mentioned.
A
Let's
use
list
in
maps
so
down
here,
so
you've
got.
You
know
the
slice
returns
and
map
of
results,
and
that's
really
what
the
rift
does.
Anyway,
when
you
say
get
slice,
it
returns
you
a
list
of
columns
and
in
c
ql
when
you
query
it
returns
you
a
list
of
rows
which
are
really
just
maps.
So
so
what
so?
What's
neat
about
it,
like
obviously
I,
can
just
turn
anything
into
J
Sensei.
It's
really
cool
but
like
what
did
you
add
so
I
kind
of
clicked
up
this
example.
A
So
say
we
have
this
list
of
beers
right
and
again
Sandra.
You
have
a
row
key
column
and
a
value
right.
That's
that
those
are
0
column,
the
specific
name.
How
so
we
have
these
three
beers
and
say
someone
wants
to
just
get
the
results
at
our
breakfast
downs
right.
It
doesn't
seem
like
an
exorbitant
request
right.
So
what
are
the
solutions
out?
There
write
a
copy
of
all
the
data
organized
by
serial.
Well,
that's
like
great.
A
That's
like
I
need
to
have
n
copies
of
everything
for
every
way
that
someone
wants
it,
which
in
some
cases,
makes
a
lot
of
sense.
But
in
some
cases
you
could
say,
storage
is
cheap
and
you
can
compress
it,
but
hey
you're,
storing
everything,
two
or
three
or
four
10,000
times
right.
It's
not
always
appropriate
and.
A
The
client
parse
it
out
right
and
that's
kind
of
got
its
own
little
problems
like
because,
if
you
are
supporting
more
than
one
client
like
nodejs
and
pearl
now,
you're
pushing
with
that
burn
on
the
client.
Cassandra
only
does
the
slice,
but
to
take
it
that
final
bit
of
the
way
you
got
to
make
a
library
for
it.
So
how
do?
How
do
we
kind
of
set
this
up
right
wouldn't
be
great?
A
So,
hey
guess
what
you
could
do
a
job
right
now,
I
wouldn't
suggest
doing
it
this
way,
but
it's
an
open
source
project,
so
everyone
throws
in
stuff
and
you
can't
be
like.
No.
That
would
be
terrible,
but
there's
a
good
reason
like
javascript
is
nice
and
easy.
So
actually,
this
is
actually
functioning
code
where
you
could
create
a
filter.
I'll
use
my
laser
pointer.
So
essentially
you
could
throw
inline
JavaScript
in
here.
A
Well,
it's
probably
efficient
and
we're
using
the
Rhino
library
that
does
the
server
side,
so
you're
actually
doing
server-side
JavaScript
on
the
results
of
the
Cassandra
filter,
which
is
kind
of
neat.
It's
very
ad
hoc
or
polygons
right.
But
of
course
you're
more
just
a
performance
should
be
doing
the
same
thing
in
groovy,
so
they
said
it's
the
same
result
here,
but
we're
just
doing
that
filtering
with
the
groovy
closure
and
you
could
do
with
a
Java
class
as
well
and
how
this
works.
Basically,
the
user
sends
a
request.
A
A
You
could
extract
complex.
This
is
the
main
reason
I
did
it,
because
a
lot
of
time
people
would
say
just
put
something
in
Cassandra
in
JSON,
right
store
the
whole
JSON
blob
it
and
then
you're
always
pulling
the
whole
JSON
blob
out.
But
what,
if
you
know,
in
particular
the
format
of
a
column-
and
you
want
to
do
a
specific
Jason
path-
query,
that's
something
you
can
do
with
this
type
of
a
setup.
A
So
one
of
my
favorite
books.
If
anyone
ever
read
just
Java
now,
it's
a
classic
there's
this
light
relief,
section,
so
I
kind
of
came
up
with
this
idea.
You
know
Nathan's
here
he'll
be
talking
later.
You
should
go
to
Nathan's
talk,
but
he
always
goes
like
over
the
top.
So
I'm,
like
I,
got
closer
so
I'm
gonna
have
the
first
tech
comedy
jam
and
it
might
be
the
first
and
last
because
my
deliveries
probably
suck
so
so
who's
heard
of
dynamo.
A
A
Really
hating
this
tough
it
out.
So
these
three
young
no
see
colors
go
into
a
bar
and
the
bartender
says
of
them.
What
kind
of
no
sequel
do
you
want
and
they
say
hbase
and
then
two
minutes
later
they
all
walk
out
the
bar
really
sad
and
they're,
like.
I
can't
believe
you
check
for
my
three
zookeepers
and
to
name
notes,
but
I
do
like
hbase
like
I'm
an
age-based
mailing
list.
A
lot
and
I
see
these
questions
they're
very
similar
to
Cassandra,
like
how
should
I
to
my
flush
settings
and
I,
like
usually.
C
A
You
should
freaking
flush
each
based
on
the
toilet
and
just
which
to
Cassandra,
and
you
know
where
cassandra
is
already
pitted
against.
For
some
reason
you
know
model
has
ever
everyone
ever
heard
of
University
yeah,
so
my
friend
decided
to
take
a
class
of
my
own
University
called
like
scaling.
You
know,
skelling
web
applications,
I
decided
to
take
it
like
go
to
Community
College
and
take
a
class
on
basket
weaving
at
the
end
of
the
thing.
I
could
weave
a
basket
and
he
still
couldn't
scale
an
application
brutal,
but
is
not
all
bad.
A
A
So
there's
this
other
piece
in
introvert
called
multiprocessor.
We
kind
of
went
over
filters
which
take
a
row
and
they
yes
or
no
certain
columns
or
the
whole
row
entirely.
Well.
What
else
can
we
do
so?
I
have
some
red
SMD
right,
brightest?
You
probably
can't
see
it
from
here
but
read
it.
Has
all
these
options
like
that?
I
mentioned
before
union
and
intersection.
They
don't
all
seem
very
hard
just
that
we
can't
ask
Cassandra
to
do
them,
because
cassandra
has
a
thrift
language
and
has
a
priori
language.
A
A
So
this
one
kind
of
looks
like
this:
this
one
flows
slightly
differently
than
the
filter,
because
the
filter
could
just
have
essentially
work
row
at
a
time,
but
this
is
a
little
more
complex.
So
we
do
a
slice
right
so
request.
Oh
I'm,
missing
a
lie,
no
request:
does
a
slice
comes
back
to
the
handler
to
another
slice
cool?
Then
it
says:
okay,
Union,
these
things
awesome.
Thank
you,
and
so
it's
not
really
very.
It's
very
easy
to
an
introvert,
because
introverted
has
this
really
vertex?
A
Has
this
actors
model
where
each
each
handler
is
no
end
point
and
you
could
pass
things
between
endpoints?
So
that's
the
way
it
works.
Slices
just
moved
from
one
end
point
to
another
until
they're
finally
done
so
what
could
you
use
this,
for
you
could
do
a
unions
I
feel
like
I,
said
Union
like
100
times
its
presentation.
A
You
can
see
intersections
and
you
can
do
some
type
of
join
situations
again
now
we're
talking
about
slices
that
have
like
a
fixed
size,
we're
not
trying
to
union
or
intersect
columns
with
10
million
rows
or
20
million
rows,
because
that's
something
pretty
hard
to
do
a
real
time,
but
four
slices
that
for
results
that
could
fit
in
memory.
We
can
do
this,
so
this
is
another
thing.
A
This
is
most
like
the
triggers
feature,
I
guess
if
you
could
put
it
this
way,
but
so
cassandra
has
this
fat
client
would
let
you
do
all
this
stuff?
It's
really
the
thing
under
this.
That
does
the
guts
and
the
puts
in
the
batch
mutates,
and
some
people
have
written
apps
living
in
Cassandra
in
choosing
the
fact
line,
because
they
could
take
away
all
that
overhead.
A
They
don't
need
to
go
through
thrift
and
have
thrift
sterilized
and
deserialize
something
they're
just
working
right
there
on
the
Cassandra
served
with
the
fat
client
and
they
do
what
they
need.
So
I
actually
took
this
from
I.
Don't
think
he's
here
today,
but
Josephine
is
a
really
cool
guy
new
york
tech
guy.
So
if
you
look
at
this
link
at
the
bottom,
you
got
his
presentation.
It's
a
really
good
one.
It
was
called
using
cassandra
as
a
central
nervous
system
for
your
I
believe
analytics
and
in
his
situation,
he's
getting
these
events
right.
A
Thing
about
it
is,
it
has
two
properties
called
events
has
a
play
percentage
and
some
actually,
it
has
n
attributes
it's
a
map
of
attributes.
So
this
is
a
thing
like
so
imagine
if
you
would
oops
sorry,
I
missed
the
slide
that
you
need
to
insert
this
data
like
15
or
20
different
ways
which
is
kind
of
what
his
his
case
is
doing.
He
needs
to
take
this
data
and
he
needs
to
build
the
reverse
index
of
of
users
by
platform.
He
needs
to
build
another.
C
A
Events
by
taste
he
needs
to
build
another
one,
so
the
only
kind
of
so
what
he
kind
of
came
up
with
was
the
system
that
will
this
is.
This
is
called
Skeletor
it's
on
github,
full
Skeletor,
it's
a
it's
a
Java
library,
it's
a
Java
client
for
skele
and
he's
showing
how
his
software
basically
takes
these
events
and
inserts
it
into
like
10
or
12
different
column
families.
A
A
What
I'm
showing
here
is
that
this
is
what
the
interface
looks
like
and
how
you
would
accomplish
the
things
so
we're
studying
the
parameters
once
and
for
for
the
parameters
and,
however
many
inserts
we
need
to
do.
We
do
the
most
server
side.
So
it's
like
a
badge
but
a
little
different,
and
how
does
this
work
well,
this
one?
Basically,
you
have
a
request
and
you
have
a
service
processor,
which
is
just
a
huge
blob
of
code.
It
could
have
anything
of
storage
proxy,
it
could
have
Cassandra
internals
metadata.
A
It
could
have
your
code
and
essentially
just
returns
an
ok.
So
I
need
you
to
insert
this
data
to
17
different
tables.
I.
Did
it
okay?
So
how
are
we
doing
on
time?
Oh
man
get
it
done
early.
So
so,
where
is
this
at
obviously
me
and
if
you
were
working
on
it,
we'd
have
liked
to
get
a
1
point
0
before
I
came
here,
but
it's
an
open
source
project.
So
it's
going
lots
of
different
ideas,
different
directions,
we're
still
one
point
0,
but
we
wrote
a
lot
of
dock.
A
So
it's
not
just
like
Oh
dive
into
the
code
and
hope
you
figure
it
out.
We
tried
to
explain
all
the
features
and
document
how
to
use
all
the
features.
The
documentation
isn't
always
super
hundred
percent
up
today,
because
we
are
modifying
stuff
changing
stuff,
but
at
least
it's
there
and
we
tried
and
the
cool
part.
C
A
Of
the
things
we
wanted
to
do
was
we
didn't
want
to
just
like
say:
introverted:
does
both
cql
and
drift.
So
if
you
write
a
filter,
you
can
apply
it
to
both
types
of
results.
You're,
not
really
locked
into
saying
this
thing
only
works
with
thrift,
and
that
thing
only
works
with
cql.
So
there's
a
nice
basically
trying
to
keep
those
pieces
that
a
lot
of
people
like
to
thrift
alive
and
one
of
the
things
we're
going
to
add
as
Virgil,
which
I
mentioned
before
we're
just
going
to
build
it
in.
A
So
it's
basically
there'll
be
a
Cassandra
instance
that
has
a
running
virgil
and
it
also
has
all
this
cool
stuff.
I
just
showed
before
one
of
the
neat
things
that
is
like
all
of
the
source
projects,
as
we
still
offer
each
other
right
and
I
mentioned
like
doing
really
large
rows
of
millions
of
columns.
So
HBase
handles
us
with
a
scanner
where
the
client
actually
has
an
iterator
and
they
move
through
the
data
kind
of
like
you
would
do
now.
A
C
A
The
question
was:
is
the
data?
What
happened
it
was
the
data
computed
distributed
Lee
and
then
know
we're
doing
the
Cassandra
slice
or
the
cql
query,
we're
getting
that
whole
results
and
then
we're
passing
it
to
this
piece
so
because
that's
essentially,
where
we're
hooking
in
we're
not
doing
like
low-level
internals
we're
using
stuff.
That's
already
there
so
seek
you
out,
Cory
we're
just
calling
CQ
berry
and
doing
some
aftermarket
hookup
yeah.
A
Well,
this
is
so
you
asked:
is
there
anything
that
for
a
faster
indexing
engine
and
no
because
we're
not
trying
to
give
a
data
format
or
a
feature
where,
just
if
you
have
essentially
we're
trying
to
we're
trying
to
like
add
on
more
effects
to
what's
already
there
Cassandra
like,
we
could
take
a
slice,
we
could
slice.
This
license
license
license
license
now,
when
you
do
these
things,
but
we're
not
trying
to
make
you
store
data
in
a
certain
way
or
use
a
certain
feature.
So
no.
A
What
you
should
do,
your
first
so
you're
saying
if
you
should
do
your
first
query
against
the
index
and
then
filter
it.
Yes,
there's
a
little
more
work
that
we
have
to
do
on
on
the
index,
because
to
read
from
an
index
and
draft
is
called
get
index
slices
and
we
haven't
worked
through
that.
Yet
so
I.
B
E
E
A
You
suggest
the
Jason
for
my
deal
with
different
data
types
and
yeah.
What
we're
using
is
we're
using
Jackson's
mapper,
so
we're
just
taking
these
these
operations,
which
start
with
java
objects
and
we're
letting
Jackson's
mapper
decide
what
to
do
with
that.
Given
object,
we
definitely
to
do
more
in
terms
of
I.
Haven't
tried
with
time
of
you
IDs
to
be
honest,
but
we
do.
A
F
A
G
A
It's
embedded
with
Cassandra,
so
if
you're
consigners
down
this
is
down
and
which
is
an
interesting
thing
because
you're
you
are
adding
more,
you
know
you're,
adding
this
extra
libraries
into
Cassandra.
So
you
do
have
to
account
for
that.
Like
now,
you're
going
to
have
more
objects
in
perm
gen
I
mentioned
the
reload
ability
there
that's
one
of
the
cool
pieces.
The
vertex
is
designed
to
be
able
to
insert
different
code
and
load
and
reload
at
runtime.
So
we're
trying
to
keep
that
of
not
you
shouldn't
have
to
shut
down
Cassandra
because
it.
G
G
A
So
that's
interesting
question:
what
Nathan
ass
is?
Will
it
perform
better
performance?
Questions
are
always
like
super
super
loaded
right.
So
really,
if
you'll
notice,
if
you,
if
you
saw
the
first
test
first,
the
first
CQ
ultimate
test.
Thrift
was
faster
than
seek
you
out
and
they
said
we'll
use
prepared
statements
right,
because
a
lot
of
the
cql
overhead
was
parsing.
A
The
statement
and
the
preparing,
so
that's
kind
of
the
concept
we're
going
with
here-
is
that
there
is
more
overhead
in
JSON
for
sure,
but
in
cases
where
you
could
turn
one
network
round
trip
or
something
it
would
have
been
ten
round
trips
into
one
performance
is
better
from
the
client
standpoint,
because
they
can
do
one
request,
block
and
one
request
get
work.
One
response,
rather
than
have
to
do,
10
things.
A
There
is
overhead
performance,
wise
and
then
each
of
these
on
each
of
these
event,
handlers
and
actor
models
have
to
deserialize
the
result
to
pass
it's
the
next
one,
it's
very
fast,
but
that
is
a
cost
right.
So
before
Cassandra
was
moving
the
request
through
a
thread
pool
which
was
just
moving
references
right
now.
A
So
I
don't
know
how
the
JavaScript
heart
works,
but
for
the
groovy
bits
it's
compiled
once
so,
when
you
create
the
when
you
create
the
filter
or
the
processor,
it
comes
in,
we
compile
it
and
then
we
use
that
reference
from
there
on
and
we
don't.
Actually
we
can
use
Java
too.
So
we
can
use
a
URL
class
loader,
for
example,
so
you
could
say:
I
have
this
class
and
that's
already
compiled
so
you're,
not
paying
any
groovy
overhead
to
use
that
class.
But
you
know
it's
not.
A
F
F
A
That's
a
valid
way
to
deploy
it.
I
think
I
mean
the
concept:
is
that
Cassandra
already?
Does
that
work
in
the
background
like
if
you
issue
a
get
request
to
a
given
server
and
that
server
doesn't
have
that
particular
key
that
Cassandra
forwards,
I
request
on
unless
you're
using
s
s
in
acts,
which
has
this
feature
which
tries
to
guess
the
right
note
for
the
key
that's
happening
anyway,
so
so
there
there
could
be
more
benefits
and
trying
to
more
tightly
co-located.
A
But
the
way
this
is
layered
on
we're
not
down
into
the
code
that
deep,
the
benefit
we're
trying
to
save
is
like
I
mentioned
the
virtual
case
right
if
you're
a
living
virgil
on
a
separate
web
app
server
somewhere,
that's
a
request
to
the
virtual
server
right
and
then
a
request
to
cassandra
and
then
back
and
back
at
least.
If
we
do
it
in
one
JVM,
we
took
out
all
those
extra
requests.
We
can't
make
it
any
faster
than
it
is,
but
at
least
we
can
take
away
all
the
overhead.
B
A
D
A
A
You
essentially
everything.
Let
me
see
if
I
can
flip
through
we're
going
way
back.
You
know
what
let
me
just
say
this
way:
everything's
in
that
right,
so
maps
could
have
optional
properties.
So
it's
just
like
kassandra's
column
as
a
TTL
right.
You
can
specify
it
or
not.
So
if
there's
no
TTL
and
even
the
example,
I
gave
they're
holding
a
15
minute,
sighs
I'm
going
to
be
done
waiting
for
that.
So.
B
C
A
C
A
B
B
D
D
D
A
Deploy
it,
what
you
have
to
do
is
you
drop
the
jar
of
this
project
in
your
Cassandra
live
and
you
you're
just
calling
a
different
job
of
Maine.
So
for
reference,
what
Cassandra
does
this
is
a
good
question
because
so
cosigner
used
to
just
have
thrift
and
then,
if
you
look,
if
you
dig
through
Cassandra
startup,
it
basically
says
start
the
rift
server
starts
eql
server.
So
all
we
have
is
a
similar
class.
That's
a
start.
Thrift
server
starts
eql
server,
start
introvert
server
the
end
so.
G
C
A
B
A
E
A
A
It
would
be
great
if
we
knew
each
filter
and
exactly
how
much
time
they're
taking
it
adding
additional
e
to
the
slice
time,
but
now
I
mean
there's
a
lot
of
the
way
I
do
that
personally
is
we
have
a
lot
of
kind
of
I
forget
the
name
of
them,
but
CCM
tools,
tools
like
debuggers
that
by
default
sample
the
time
use
of
everything,
so
I
don't
always
well.
It
is
important.