►
Description
Speakers: Michael Figuiere and Patrick McFadin, Principal Solutions Architect at DataStax
Slides: http://www.slideshare.net/planetcassandra/cassandra-summit-data-stax-java-driver
Cassandra 1.2 finalizes CQL3 and introduces a new binary protocol for client/server communication. These two components are the foundation of the new line of drivers developed by DataStax. Based on years of experience with Cassandra, these new drivers for Java, .Net and Python come with an asynchronous and lightweight architecture, a clean and simple API, a standardized way to discover nodes and to manage load balancing and fail over. This presentation will give an in depth look at these new drivers which will make your Cassandra-based applications even more robust, efficient and simple to write.
A
A
So
we
we're
going
to
give
you
a
little
bit
of
insight
about
this
Java
driver
and
in
general,
actually
about
the
general
kind
of
architecture
we
try
to
develop
for
our
brand
new
drivers.
It's
a
fairly
a
new
vision
for
for
forecasting
wha,
it's
a
long
term
investment.
So
we
don't
expect
any
of
you
guys
to
switch
from
your
current
Swift
drive
your
Swift
in
switched
based
investment
like
white.
Now
it's
it's
just
really
a
long
term
investment
in
the
long
term
vision
for
for
Cassandra
to
make
it
simple
to
use
so.
A
Yeah,
so
we
can
start
by
actually
this
world
vision
of
making
casimir
simple,
simply
just
start
from
secure
before
cql.
You
are
the
fairly
unconsciousness
interface
forecasting.
Well,
you
know
with
some
different
way
of
expressing
queries,
whether
you
were
in
Java
in
c-sharp
or
in
in
the
CLI
interface.
It
was
not
really
exciting.
It
was
very
confusing,
very
complex,
but
at
least
it
was
a
cheap
way
for
the
customer
project
to
boots
drew
up
some
some
driver
for
every
language.
So
now
we've
got
this.
A
This
new
vision
that
that
come
with
cql-
and
I
guess
most
of
you
guys
were
at
Sylvan
Stoke
white
before
so.
We
won't
spend
too
much
time
introducing
cql.
Basically,
cql
is
not
just
actually
a
qui
language,
as
the
name
implies.
It's
also
a
fairly
updated
a
data
model
on
top
of
cass
anwar
with
scene
abstraction
layer
on
top
of
what
you,
on
top
of
the
classic
data
models
that
you
used
to
use
in
kaznia.
So
you
know
casting
was
a
column
oriented
database
in
in
c
ql.
A
What
you
manipulate
is
actually
actually
just
a
bunch
of
genome
life's
table
just
like
those
that
you're
used
to
using
SQL
in
your
my
sequel,
relational
database.
The
only
difference
here
is
that
you
don't
have
any
joints
and
you're
not
able
to
do
any
kind
of
aggregation
in
your
queries.
So,
as
you
will
see
in
in
the
example
layer
and
I
guess,
most
of
you
are
familiar
anyway,
where
it's
definitely
a
sql-like
query
language.
The
goal
is
really
to
leverage
all
knowledge
and
all
the
knowledge
of
your
chi
minh,
of
developers
out
there
in
SQL.
A
So
it's
not
schema-less
anymore.
We
believe
that,
even
if
schema-less
sound
is
pretty
20
ins
and
no
sequel
world's
nowadays,
we
believe
that
actually
anywhere,
you
always
end
up
having
some
schema
definition,
even
if
it
actually
end
up
being
some
schema
written
in
Python
or
in
would
be
or
whatever,
because
in
your
in
your
language,
in
your
application
logic,
you
actually
need
some
kind
of
awareness
of
the
actual
scheme
as
a
to
manipulate.
A
A
So,
as
you
know,
Swift
is
an
RPC
framework
and
as
a
mountain,
it's
a
fairly
cheap,
where
it
has
been
a
fairly
cheap
way
for
us
to
to
give
you
some
some
driver
for
pretty
much
every
language
for
free,
because
the
Apache
customer
project
was
just
delivering
some
igl
and
then
you
were
able
to
compile
some
client
for
every
pore,
pretty
much
every
programming
language
out
there.
So
it
was
great.
But
now
that
customer
is
getting
mature,
it
start
to
be
fairly
limited.
It's
actually
Swift
is
just
a
week
with
weapons
protocol.
A
So
it's
not
really
great,
because
we
want
to
move
to
move
forward
with
some
streaming.
We
want
to
have
some
notification.
We
want
to
do
a
lot
of
things
in
kasson.
Wha
am,
and
this
swift
protocol
is
actually
studying
to
be
a
bottleneck
for
for
for
progress
in
casimir,
so
and
just
to
stress
it,
because
it's
very
important.
It's
not
because
there's
this
new
investment
that
we
will
start
to
drop
any
kind
of
Swift
support.
We
totally
respect
all
the
investments
that
have
been
done
in
the
Swift
landscape
for
four
years.
A
We
know
that
there's
a
bunch
of
business
and
a
bunch
of
of
applications
that
run
on
top
of
this
interface.
So
it's
simply
a
brand
new
interface
for
on
which
we
start
to
walk
and
invest
for
for
now
in
the
future.
But
sweet
will
will
keep
being
women
will
remain
being
supported
for
for
some
years
for
sure,
you're
totally
safe
with
your
investment,
so
that
week
will
actually
give
you
a
little
bit
of
insight,
but
waste
like
to
to
die
mode
or
some
stuff
in
this
brand-new
cql
land
scale.
Ok,.
B
So
do
we
warranty
remodeling
about
that,
but
everybody
was
in
my
data.
Modeling
talk
earlier
well
and
I'm.
Just
gonna
cover
this
really
quickly,
but
kind
of
in
the
perspective
of
our
cql
driver.
So
if
you
talk,
we've
talked
quite
a
bit
about
cql
from
a
data
modeling
perspective,
and
really
it's
because
we're
talking
about
your
applications,
hey,
you
know,
cql
is
a
great
way
to
describe
that
data
model.
So,
of
course,
with
your
data
model
you
can
stand.
Are
living
closer
to
your
application.
B
We're
trying
to
describe
things
you
know
I
want
to
do
something
on
my
webpage,
like
I
need
some
object
or
some
sort
of
information,
so
we're
trying
to
get.
This
is
closer
to
your
application,
where
it's
less
of
a
generalizable
pattern.
So
Cassandra
doesn't
you
know,
of
course
doesn't
do,
joins
it's
not
as
good
for
seeking
things
like
that,
it's
better
whenever
you
can
predetermine
your
data
model
beforehand.
So
that's
that's
what
it's
about
what
I
say
it
expresses
your
applications.
Intent
is
really
from
the
top
down
from
the
application
level.
B
Now,
when
I
do
my
data
modeling
talks,
I
talk
a
lot
in
now.
In
c
ql,
which
is
so
much
easier,
I
think
in
a
lot
of
ways
when
I
did
him
in
thrift,
it
was
harder
because
I
would
I
would
use
like
CLI,
but
there's
no
description.
So
there's
nothing.
That
would
say-
and
this
is
what
your
data
model
looks
like.
B
It
was
more
of
descriptive
where
and
we'll
store
some
time
series
data
in
a
column
family
later
there's
nothing
that
said
that
in
actual
written
out
data
model,
so
in
the
data
models
of
what
we
know
that
I've
gone
through
now,
we
can-
I
can
describe
them
to
you
and
now
we
can
use
this
driver
because
that's
what
it's
using.
So
we
have
kind
of
the
one-two
punch
here.
If
you
were
at
my
talk
earlier
today.
B
All
of
that
I
have
implemented
in
an
example,
in
example,
code
and
so
that's
sitting
out
there
on
github
somewhere
I
haven't
made
it
all
public,
but
so
that
it
just
gives
you
that
kind
of
okay
I'm,
taking
from
the
conceptual
stage
now
down
to
the
model,
state
or
actual
usage
stage,
so
I
think
that's
enough
about
data
modeling.
If
you
want
to
know
more
I
got
plenty
more
resources
out
there,
but
I
think
that's
enough
for
this
talk
so.
A
For
so
so,
as
I
mention
in
c
ql,
everything
is
just
a
table
just
like
as
your
you.
You
were
used
to
manipulate
in
SQL.
So
it's
fully
there's
no
white
walls,
anymore
white
walls
are
actually
broken
up
in
civil
wars
in
in
this
static
table,
just
like
you
were
used
to
seeing
some
genome
life's
table
in
SQL
or
maybe
in
some
ways
all
set
on
some
of
some
joint
query
in
in
my
sequel
or
week
or
whatever.
A
So
what
we've
got
is
actually
this
table
that
once
again,
is
just
like
rational
table,
but
the
only
difference
is
that,
instead
of
defining
a
single
primary
key,
we
define
you
can
actually
define
a
single
primary
key,
but
you
can
also
define
when
you
need
to
store
some
more
complex
data,
you're
able
to
store
some
to
define
a
composite
primary
key
that
is
basically
made
of
partition
key
and
of
one
of
civil
clustering
keys.
So
it's
all
about
that
locality
here.
A
So
basically,
the
primary
key
is
just
in
order
to
define
on
which,
on
which
node
you
want
all
your
data
to
be
colocalized
to
be
co-located
and
the
clothes
Twinkie
is
them
to
define
how
you
want
your
daughter
to
be
co-located
on
the
disk,
so
its
first
to
co-locate
the
data
on
on
a
single
node
and
them
out
co
located
on
the
disk.
So
actually
what
this
example
can
simply
be
queried
with
this
very
simple
SQL
ish,
secure
statement.
So,
as
you
can
see,
the
learning
curve
for
cql
is
definitely
sugar.
Small.
A
So
I
don't
want
to
dig
deeper
in
well,
because
I
guess
you've
heard
about
it
a
lot
of
time
doing
this
conference.
So
let's
have
a
look
at
the
actual
new
driver
architecture,
the
architectures
that
we
want
for
our
new
drivers.
So
we
we
truly
want
to
leverage
this.
These
brand
new
protocols
that
come
as
a
replacement
for
swift,
and
we
also
want
to
leverage
cql
and
to
to
make
sure
that
you
don't
have
some
noise
with
any
legacy
interface
from
Swift.
So
that's
why
we
come
with
a
brand
new
architecture.
We
don't
try
to.
A
You
know
to
merge
or
mix
any
kind
of
old
and
new
style,
because
we
want
something
that
is
really
clean
and
simple
for
for
newcomers
as
well
as
for
experienced
users.
So
everything
is
based
on
this
new
protocol
as
there's
no
more
Swift
dependencies
and
we
will
will
will
work
to
provide
for
every
major
language
this
week,
kind
of
API.
A
So
what
I
call
a
sequel
API
in
this
architecture
is
a
basically
a
fully
dedicated
an
API
fully
day
page
for
situation
is
designed
for
every
very
specific
features
of
CTL
that
that
is
able
to
support
all
it's
very
cool
features.
Gb
API
is
just
unwrapped,
a
choice
that
we
will
try
to
provide
for
all
this
for
all
this
language,
to
integrate
with
the
standard
DB
API
such
as
jdbc,
DB,
API
for
item
and
everything.
It's
not
meant
for
any
prediction
usage.
A
It's
just
meant
for
some
simple
integration
pass
with
some
existing
tools
or
ESB
each
year,
whatever
tools
or
middleware
or
library.
You
want
to
integrate
with
so
Ananse
again,
it's
pretty
cool
for
us
to
to
to
why
this
kind
of
adapter,
because
what
we've
got
right
now
is
some
table
or
wizard
set
along
with
some
twink
worries,
so
that
actually
totally
fit
in
jdbc
and
also
standard,
DBA,
p
I.
So
that's
pretty
cool
to
a
level
which
is
this
similarity
and
last
but
not
least,
will
provide
an
object,
mapping
interface
for
for
every
language.
A
We
are
not
talking
about
any
kind
of
RM,
any
kind
of
object,
relational
mapping
or
whatever
we
will
talking
about
being
able
to
map
from
an
object
to
aquariums
and
from
a
wizard
set
to
an
object
just
for
productivity
from
a
Productivity
standpoint.
It's
a
we
don't
want
to
get
into
the
middleware
and
former
business.
You
know
so,
let's
up
deeper,
look
at
this
Java
driver
by
itself,
so
it
come
with
a
bunch
of
very
interesting
features
that
might
by
themself
make
you
want
to
switch
to
this
new
architecture.
A
Actually,
so
we
will
consider
it
as
our
weapons
implementation,
as
this
driver
has
been
implemented.
The
first
Java
as
a
java
community
is
by
far
the
biggest
customer
users
community.
We,
the
driver,
is
entirely
based
for
its
network
io
on
Nettie
I
guess
most
of
you
guys,
I've
heard
about
it.
It's
a
it's
a
network,
I
OU
libraries
that
has
been
written
by
j
bus,
which
is
fully
a
synchronous
based
on
some
non
working
areas.
We
wish
you
poor
people
statement
from
the
very
first
version,
so
same
issues
as
the
way
Shinra
world
same
solutions.
A
We've
got
some
text
query
and
we
don't
want
to
carry
some
text
of
the
wire,
especially
for
large
scale
applications.
We've
got
a
fully
automatic
and
standardized
a
failover
strategy
across
all
our
drivers,
so
we
want
to
avoid
the
issues
that
you
used
to
have.
That
is
whenever
you
were
using
ecto
astyanax
by
casa
and
some
any
other
drivers
in
your
architecture.
It
was
pretty
tough
to
know
what
was
the
failure
worst
wedgie
of
these
driver
and
what
was
the
load
balancing
strategy
of
this
also
driver.
A
So
when
you
had
to
do
some
troubleshooting,
it
was
pretty
tough
to
to
make
it
so
here
we
will
really
try
to
have
and
not
just
try.
Actually,
we
will
really
work
to
have
a
standardized
architecture
for,
and
some
dyes
way
to
deal
with.
All
these
issues
same
goes
for
not
discovery.
We,
the
driver,
will
automatically
discover
all
the
nodes
you
just
provide
some
scene
nodes
and,
and
the
driver
is
able
to
to
receive
the
complete
list
of
of
nodes
in
your
case
on
cluster.
A
We
just
send
a
pipeline
of
requests
to
the
server,
and
then
we
receive
the
West
bones,
possibly
out
of
order,
and
we
totally
manage
that
because,
along
with
each
way
spawns
in
each
week
west,
we
receive
some
swim.
I
we've
got
some
IDs
that
let
us
dispatch
the
way
spawns
to
the
white
sweat
that
is
waiting
for
for
it.
A
Notifications
is
a
pretty
cool
features
as
well,
because
up
to
now,
most
of
the
most
advanced
heed
client
libraries
for
caswell
add
actually
to
to
pull
some
castle
nodes
in
order
to
which
weaves
the
wall,
topology
or
all
the
schemas
or
everything,
and
to
keep
it
up
to
date
with
the
time
because,
obviously
is
topology
is
meant
to
do
too
to
change
with
time.
Maybe
if
you
lose
a
node
or
if
one
of
your
herbs
is
adding
some
more
load.
A
So
here
the
client
can
just
register
with
one
of
the
node
for
some
even
for
some
notification,
and
it
will
receive
sorry
some
even
as
soon
as
a
node
has
been
added
or
no
that's
been.
We
move
from
the
cluster
so
that
we
can
update
our
our
load
balancing
policy
very
quickly,
so
the
driver
is
fully
a
synchronous.
That's
something
very
important
for
us.
It's
not
to
start.
The
games
is
long
tour
about
synchronous
versus
a
synchronous.
A
It
just
all
executed
as
synchronously
through
some
non-blocking
I/o,
so
the
driver
maintain
a
bunch
of
connection
per
node.
The
amount
of
connection
depend
on
on
the
current
pressure
on
this
node
and
each
connection
is
able
to
make
a
tool
to
enroll
bunch
of
of
pending
requests.
So
by
default
we
can
order
as
much
as
1
128
pending
requests
per
connection,
which
clearly,
which
is
the
pressure
on
the
amount
of
connection
we
need
to
keep
their
node,
which
is
actually
pretty
great.
B
This
fixes
that
problem
I've
actually
put
this
into
production
where
it
really
made
a
difference
and
I
have
to
thank
you
for
that,
because
it
made
my
life
a
lot
easier
because
I,
what
was
the
answer
before
hsha
and
that
was
the
half
half
sink?
Half
a
cig
connection
on
Cassandra,
which
is
a
compromise
which
I
don't
like
to
make
so
I
from
a
real-world
perspective.
You
think
about
what
we're
doing
with
our
applications,
where
Cassandra
can
just
start
scaling
like
crazy.
You
need
to
have
something
keep
up
with
it.
A
So
to
have
a
look
at
what
actually
happened,
let's
look
at
the
world
weed
and
white
pass
for
a
single,
lock,
phooey,
that
a
client
sweat,
for
instance,
will
try
to
push
on
on
the
node.
So
a
client's
way.
Is
you
a
week
west
and
push
it
against
the
driver?
The
driver
will
switch,
slowed
bouncing
while
G
and
sue
its
its
internal
logic
pick
up
a
node
and
then
pick
up
a
connection
on
which
to
push
this.
This
request
the
node
will
receive
it
will
even
to
the
answer
to
this
request.
A
A
You
can
just
add
in
your
clowns
with
a
bunch
of
concrete
you
object
and,
and
the
driver
will
just
dispatch
with
the
incoming
response
to
every
of
this
future
object,
so
you
will
quickly
get
used,
which
is
due
driver
to
use
and
abuse
of
all
this
concurrent
programming
of
all
this
concurrent
requests
execution.
You
don't
have
any
more
to
try
to
pack
as
much
as
possible
or
your
request
or
whatever
you
can,
because
we've
got
some
requests
by
planning.
A
You
can
totally
push
a
bunch
if
you've
got
to
to
fetch
for
on
the
words
of
ID.
Instead
of
having
a
meal
ticket
request,
you
can
simply
push
a
bunch
of
weak
words
that
will
be
dispatched
all
over
your
cluster
without
creating
any
any
odd
spot.
You
know
so
maybe
that
we
can
have
a
look
at
the
API
okay.
B
About
some
real
code
yeah
here,
I'm
gonna
move
around
because
it's
kind
of
late
the
day
right.
Let
me
just
move
our
heads
a
little
bit
so
yeah.
This
is
the
trouble
being
the
last
talk
of
the
day
right
because
everyone's
like
man,
two
days
of
this,
so
all
right.
This
is
going
to
be
simple
because
it
is
simple.
So
what
I'm
doing
here
is
I'm
just
creating.
So
all
this
codes
up
on
github
right
now
by
the
way
so
I'll
tweet
out
the
link
later.
B
So,
if
you
want
to
follow
me,
then
you
can
get
the
link
for
this,
but
so
what
I've
done
is
created
a
working
example
and
we'll
go
through
several
of
these
just
will
go
through
some
points
about
how
it
works
and
then
I'm
going
to
show
you
some
actual
examples.
So
if
you've
never
done
a
connection
to
Cassandra
before
this
is
going
to
look
pretty
easy
if
you've
done
it
in
using
other
drivers.
This
is
gonna.
Look
really
easy.
B
So
really
what
I'm
doing
here
is
I
just
have
a
list
of
connection
points
and
Mikey
spaces.
Video
DB,
which
is
my
example,
database
that
I
used
everywhere,
but
it's
just
creating
a
cluster.
That's
it
I'm.
Creating
a
cluster
I.
Add
the
contact
points.
Those
are
the
different
IPS
and
I
say:
build
done,
walk
away
so
want
to
make
to
connect.
That's
when
it
actually
fires
up
everything
in
to
that
particular
key
space.
B
If
you've
ever
used
other
drivers
that
that
was
a
few
more
steps
in
there-
and
there
was
a
few
more
things
to
worry
about,
but
not
as
much
once
I
have
this
session
I'm
done.
I
now,
I
can
start
using
that
session
for
actual
interaction
with
the
database,
and
that's
it
next.
So
the
this
is
going
to
be
a
basic
right
now.
This
looks
really
ugly,
because
I
had
to
use
a
string
string
buffer
to
do
this.
B
But
what
I'm
going
to
just
tell
you
right
now
is
this:
is
the
not
preferred
way
to
do
this?
So
when
I
wrote
Oracle
code,
one
of
the
things
that's
security
people
would
do
is
tell
me
you're,
never
going
to
insert
a
string
you're,
never
going
to
do
a
answer
with
a
string
because
you
just
open
up
the
door
for
a
sequel
injection.
B
So
what
I'm
doing
here
is
I'm
creating
a
sequel
injection,
so
I'm
giving
you
the
easy
example
of
like
we
could
do
this
if
we
want
to,
but
I
have
other
examples
that
will
keep
you
from
doing
this.
This
is
a
non
preferred
method,
but
basically,
what
I'm
doing
is
I'm.
Taking
that
cql
statement,
the
insert
into
I'm
taking
this
API
of
a
user
object,
I'm
taking
the
first
name
last
name
blah
blah
blah
and
then
I
just
say
session,
not
execute.
B
C
A
B
Now
we're
doing
the
reading
now
we're
doing
a
raid
okay.
Thank
you
late
in
the
day.
For
me,
too,
okay,
it's
been
a
crazy
couple
of
days.
So
the
read
same
thing,
but
if
you
were
called
for
my
data
modeling
discussion,
so
I'm
going
to
do
is
select
star
on
the
user
table
and
I'm.
The
username
is
the
row
key.
So
now
I'm
going
to
just
get
a
bunch
of
cql
rose.
B
B
I
create
a
get
a
result
set
of
rows:
I
just
iterate
over
those
rows,
even
if
the
actual
storage
implementation
is
it
multiple
columns
in
the
same
row,
c
ql,
giving
us
multiple
rows
as
its
abstraction
will
make
this
a
lot
easier
than
trying
to
iterate
over
multiple
columns
and
a
standard
column
family.
So
this
is
demonstrating
just
that.
It's
that
simple
you
get.
That
particular,
you
know
you
select,
you
get
rosette
and
just
iterate
over
the
rose
and
pull
the
things
out.
That's
done
and
announce.
B
A
B
So
this
is
the
way
we
prefer
you
do
it
and
there's
there's
a
more
expanded
way
to
doing
this,
but
this
prepared
statement
so
earlier
what
I
had
done
is
created
a
a
statement.
It
was
just
an
insert
with
question
marks
and
so
that
that's
what
that
a
zit
user
is
all
about.
I
didn't
show
all
that
other
code.
I
know
that's
what
your
question
was.
The
complete
example
is
on
github
every
bit
of
this,
so
it's
abbreviated
because
I
have
this
much
space
and
you
know
you
don't
want
to
have
to
you
know
it.
B
We
actually
had
a
hard
time
trying
to
cram
that
much
into
it,
because
you
really
want
to
make
it
so
you
can
read
it.
If
I
put
my
entire
code
base
up
here,
I
could
read
it
from
here,
so
we
don't
want
to
do
that
so
anyway,
what
happened
before
is.
I
had
I
created
an
insert
statement
with
question
marks
in
it,
and
those
question
marks
are
now
being
replaced
by
the
user
name
first
name
last
name
and
when
I
do
us
when
I
bind
and
then
I
do
a
set
string.
B
This
is
just
like
what
we
did
with
Oracle
my
sequel,
jdbc
drivers
where
we
do
a
bind,
and
then
we
have
a
prepared
statement
and
setting
values
inside
of
it,
but
once
we're
done
execute
done,
I,
don't
think
I've
ever
okay,
the
asynchronous
read
now
we're
getting
into
the
cool
stuff,
and
this
is
I.
Think
I've
talked
to
several
you
about
this.
B
Where,
when
you
have
a
cluster
of
Cassandra,
you
don't
want
to
just
read
it
one
at
a
time,
because
what's
the
point,
if
your
data
spread
out
randomly
around
your
cluster,
doesn't
it
make
more
sense
to
read
things
in
parallel
and
an
asynchronous
read
is
where
this
driver
really
starts
to
make
sense?
And
so
what
I'm
going
to
do
here
is
I'm
doing
this
get
videos
by
username.
So
there's
a
lot
of
stuff
going
on
here.
B
Try
to
gonna
try
to
explain
what's
going
on,
but
essentially
what
I'm
doing
is
I
have
this
I'm
getting
a
bunch
of
video
titles
by
a
single
user?
So
me,
a
user
p
mcfadden
I
got
like
20
videos
out
there.
I
have
a
prepared
statement
that
will
go
get
the
ids
for
all
of
those,
but
what
I'm
doing
is
I'm
creating
this
futures
list
and
so
I
bind
it
to
my
username
I.
B
Do
a
session
execute
so
now
I
have
it
and
as
I'm
iterating
over
each
one
of
those
video
IDs
I'm
creating
a
future
for
that,
so
I
do
a
future
sad!
So
as
soon
as
I
do
that
it's
running
that
command
it's
going
in
parallel
and
running
those
out
boom
boom
boom
boom
boom,
as
I'm
working
through
my
code.
So
these
are.
These
are
what's
happening
in
the
background,
so
when
I
get
down
here.
Finally,
I'm
going
to
iterate
over
all
those
futures
that
I
created
in
the
future
is
just
something
that's
running
right.
B
Now,
it's
on
a
thread
when
I
do
a
future
dot
get
and
I
did
an
uninterruptedly.
If
you
do
it
without
that,
then
you
have
to
catch
the
potential
threat
interruption
because
there's
threads
but
I'm
getting
it
uninterruptedly.
That's
when
we
block
at
that
point
and
I
just
iterate
over
that
now.
The
intent
is
that,
with
the
correct,
what
the
correct
roki
going
out
to
the
different
servers,
all
those
things
are
going
to
be
happening
in
parallel.
B
All
the
like,
if
I
had
25
videos,
it's
going
to
make
25
parallel
calls
around
my
ring
and
that's
what
I'm
hoping
for
is
a
get
that
parallel
task
running
for
me,
if,
let's
say
I,
add
100
nodes,
if
I
had
200
notes,
I
would
want
to
take
advantage
of
that
fact
and
run
everything
in
parallel.
So
what's
going
to
happen,
is
the
LA?
The
slowest
query
is
going
to
be
the
slowest
it's
going
to
be
the
slowest
part
of
the
entire
thing.
B
B
It
would
just
be
the
one
that's
the
slowest,
so
this
is
much
more
efficient
whenever
you're
grabbing
a
lot
of
data,
and
that
happens
if
we,
if
you
use
one
of
my
index
patterns
for
data,
modeling
you're
going
to
get
back
a
bunch
of
ids,
and
I
want
you
to
go,
get
all
those
ids
from
that
index.
This
is
how
you
can
do
it
and
then
do
it
very
efficiently.
B
Next,
oh
yeah,
now
this
one's
tough
I,
really
rare,
I'm
gonna
go
through
this
and
Mikael
is
probably
gonna
correct
me
because
he's
much
better
concurrent
programmer
than
I
am,
but
this
is
one
that
I
really
do
want.
You
I'm
gonna
go
through
a
little
bit,
but
I
really
do
want
you
to
go.
Look
at
the
example
if
you're
really
interested
how
this
works,
because
this
is
just
a
snippet
of
that
code.
B
So
what
we're
doing
here
is
something
very
similar
where
we
were
running
we're
creating
futures,
but
in
this
case
I'm
actually
creating
background
threads
that
are
running
independently
all
at
the
same
time
and
there's
some
craziness
going
on
here
and
I
have
a
lot
more
comments.
In
my
code
too,
but
very
similar
I'm
just
grabbing
all
my
video
IDs
and
then
I
create
this
runnable,
and
this
is
just
another
thread
and
I'm
executing
this
background.
B
But
what
this
background
is
doing
is
iterating
through
all
of
these
potential
video
IDs
and
grabbing
them
all
in
separate
background
threads
independent
from
each
other.
This
is
that
concurrency
thing
that
really
is
going
to
be
fast
if
you're
doing
it
right,
but
as
it's
grabbing
each
one,
it's
putting
it
into
a.
What
is
that
it's
I?
Don't
even
have
it
up
there,
it's
some
sort
of
a
hash
map
or
something
yeah.
It's
a
hashmap
and
it's
creating
this
long
list
of
all
the
IDS.
B
So
this
is
again
like
where
you
have
an
index
where
you
grab
a
bunch
of
ids
and
you're,
going
to
suck
them
all
in
and
hopefully
do
these
things
in
parallel.
This
is
what's
executing
in
the
background
one
at
a
time
and
with
the
listener
you're
doing
that
without
any
blocks,
and
it's
all
happening
in
the
background
so
again,
you're
trying
to
take
advantage
of
the
the
speed
and
the
efficiency
of
doing
things
in
parallel.
Anything
to
add
on
that
it's
hard
to
add:
ok,
we'll
down
shift
a
little
bit
here.
B
That
was
a
little
crazy.
The
query
builder,
the
query
builder,
is
again
it's
along
the
line
of
where
we
want
you
to
do
things
instead
of
using
just
creating
a
string
and
then
dumping
that
directly
into
it
is
you're
actually
building
the
query
through
API.
So
the
query
builder
is
pretty
interesting,
so
I
create
a
query
and
you'll
notice.
It
looks
very
similar.
You
know
I'd
select
at
all
from
I
grabbed
it.
B
You
know
from
the
column,
family
and
the
the
key
space
and
then
we're
then
I
have
an
equal
claws
and
I
put
my
limit
at
the
end.
The
query
builder
I
think
from
my
I
understand
a
lot
of
the
security
people
do
auditing
on
security.
They
require
this
and
we've
talked
to
some
customers.
That
say
no
I
have
to
have
a
query
builder
I
can't
even
do
this
any
other
way.
I
can't
even
use
a
prepared
statement,
so
I
need
a
query
builder.
This
has
gone
through
quite
a
few
checks.
B
Now,
one
of
the
things
that
I'm
doing
other,
along
with
this
is
I,
also
have
the
ability
to
set
the
consistency.
Level.
I've
told
you
probably
if
I've
ever
done
in
consulting
on
site
is
a
consistency.
Level
is
something
you
set
per
request
that
doesn't
go
away:
thrift
and
c
ql
both
how
the
way
you
can
set
the
consistency
level
I
want
this
request
to
be
at
coram
or
at
one
or
at
all,
and
so
in
this
case
I.
For
this
request,
I
set
my
consistency
level
21
same
thing.
That
happens
after
that.
A
It
you
so
so,
let's
get
back
to
to
some
architecture
with
some
poser
very
interesting
features
in
in
this
driver.
I've
mentioned
that
we
we
were
all
laying
failover
load
bouncing.
So
let's
have
a
look
at
actually
how
we
organize
everything.
So
in
this
Jaguar,
ms,
actually
showing
you
what's
kind
of
stepped,
we
have
a
node,
cooee
execution,
a
process
you
know
in
our
driver.
So
whenever
the
client
is
pushing
some
requests,
it
will
be
routed
to
school.
A
Some
internal
load,
balancing
policy,
which
you
will
see
in
few
slides
that
you
you're
able
to
customize
and
fully
we
implement,
or
you
can
just
pick
up
the
load,
balancing
policies
that
you
want.
Actually,
this
load
balancing
policy
will
then
pick
up
one
of
the
node
to
which
we
should
woods
a
week
west
if
this
nose
turns
out
to
be
unavail
that,
like
it's
the
case
in
the
in
this
number
two
awards
here,
we
are
trying
to
which
a
node
which
is
not
variable
them.
The
load.
A
Balancing
policy
will
actually
pick
up
the
next
node
in
the
list
of
node.
We
wish
to
to
push
queries
against
that
the
case
in
the
third
rows
in
this
diagram.
Actually
it
actually
works
perfectly,
and
so
this
note
is
able
to
push
a
week
west
back
to
our
drivers,
and
this
this
response
will
goes
who's,
a
witch.
Why
polity?
This
we
try?
Polity
will
actually
figure
out
whether
it
needs
to
to
give
it
a
try
or
not.
A
Of
course,
it's
simple
for
you
to
customize
this
behavior
another
and
there's
mistake
in
this
title.
So
here
we
are
actually
end
up
doing
a
media
center,
a
load
bouncing
pretty
better
a
token
of
where
one,
just
like
the
one
you
you
could
have
in
aesthetics,
that
is,
we
will
would
our
our
request
to
some
nodes
that
are
actually
webcast
at
the
directory
somewhere,
quick
as
false
a
query.
A
B
This
is
the
same
connection
that
I
created
before,
but
in
this
case
what
I'm
doing
is
I'm,
adding
one
more
thing
with
my
own
and
load
balance
or
policy.
Now
this
one
is
a
default
load.
Balancer
policy
I
can
create
my
own
load
balancing
policies,
maybe
if
I
wanted
to
go
into
certain
racks
of
equipment
or
I
want
to
bias
against
certain
servers
or,
however,
I
want
to
load
balance,
that's
fine
whatever
it
is,
you
can
do
it
yourself
and
the
default
ones
are
probably
going
to
be
fine.
A
B
This
is
a
retry
yeah,
that's
the
next
one.
I
have
yeah,
so
the
retry
policy
I
have
written
my
own
retry
policy
and
we
were
trying
to
figure
out
why
we
would
do
it,
but
I
actually
wrote
a
retry
policy
and
how
it
downgrades
the
connections
is
different.
So
a
lot
of
people
I
tell
hey
whenever
you're
writing
at
quorum.
B
If
you
want
to
maintain
your
maximum
uptime
really,
you
should
fall
back
and
try
retry
at
local
like
a
local
one,
and
so
you
do
local
quorum
and
then
you
do
one,
that's
how
you
can
fall
back
and
at
least
to
that
one
level
where
you
have
a
server
that's
available,
so
you
could
just
automatically
build
this.
In
with
your
own
retry
policy,
I
created
my
own
retrying
policy
that
it
was
like
during
a
time
of
day
it
wouldn't
retry
or
something
I
can't
figure
out
a
use
for
it.
B
B
They're,
real
okay
object,
mapping!
I'll!
Let
you
cover
you
want
me
to
cover
this
real,
quick,
so
the
object
mapping
is
a
future
thing.
This
is
going
to
be
in
the
next
version
right,
one
to
one
yeah,
one
that
way
and
it's
not
jpa,
but
it's
ish.
So
in
my
example,
where
I
was
setting
and
getting
user
and
video
and
this
I
would
create
my
object
for
user
I
can
put
in
annotations
for
each
one
of
those
each
one
of
those
properties
and
using
this
now,
I
don't
have
to
go
and
set
string,
get
string.
B
A
So
you've
probably
seen
this
Quinn
shot
during
the
keynote
of
gelatin.
Yesterday,
that's
something
very
important
for
us
in
our
strategy
to
move
forward
with
a
great
experience
for
Cosmo
developers,
we've
got
brand-new
a
query
language.
It
says
this
much
more
efficient
transport
layer
that,
let
you
simply
drop
Swift.
If
you,
if
you
want,
we've,
got
some
new
drivers
that
make
it
very
simple
for
you
to
work
with
Casanova
and
that's
the
last
piece
at
least
for
now
of
our
win.
We
nude
experience
for
you
with
Casanova.
A
We
will
provide
you
this
to
all
that
this
iges
that
will
actually
let
you
execute
some
cql
statement
blows.
Your
boo
is
your
data,
explore
everything
you've
got
in
castro
muah
and
much
more
into
future.
You
know
we.
We've
got
a
bunch
of
videos
in
stock
in
mind
for
for
this
tool.
We
want
to
make
it
really
really
simple
for
you
two
to
double
your
and
troubleshoot
your
cash,
no
application
in
development
mode
and
anything
else.
So
yeah
keep
up
from
stay
tuned
for
this
very
interesting
tools.
A
A
B
So
if
the
choice
of
fail
is
not
good,
take
it
down
to
one
well
that
should
up
to
this
point,
I've
been
having
people,
do
it
programmatically,
like
in
their
own
code,
catch
the
exception,
which
is
bad
and
then
deal
with
it.
Well
now
what
we
can
do
is
just
plug
it
in
as
a
policy,
and
then
it'll
just
automatically
do
that,
and
you
can
insert
that
into
any
anytime
and
say,
oh
and
for
these
requests,
I
want
to
use
this
retry
policy
and
then.
A
A
B
C
B
A
B
Alright,
no
that's
a
valid
question,
and
so
this
is.
This
is
resiliency.
So
in
normal
operating
procedure
you
have
let's
leave
a
four
node
ring
and
you
have
a
replication
factor
of
two
or
three
all
right
quorum
is
what
I
would
recommend
always
because
you're
always
going
to
get
better
consistency.
That
way,
and
that's
just
the
default
to
one
is
where
I'm
making
the
decision
that
let's
say
that
it's
something
like
a
user
data
on
my
screen
on
my
web
page.
That's
I!
B
Don't
want
my
web
page
to
go
down
if,
for
instance,
I
lose
two
nodes.
I
can't
satisfy
quorum
at
that
point,
but
I
can
satisfy
a
seal
of
one
so
I'm,
making
a
decision
in
my
application
code
to
be
able
to
do
that.
So,
if
you're,
not
ok
with
that,
then
don't
use
a
retry
policy,
but
what
I'm
doing
is
I'm
trying,
there's
nothing
wrong
with
being
trying
to
be
consistent
and
quorum
all
the
time.
It's
just.
B
B
A
So,
actually,
you
can
configure
your
which
white
to
add
some
logging
in
the
end
it
all.
Actually,
I
keep
in
mind
also
that
the
load
balancing
puts
interface
and
the
wheat
wipers
interface
out
pretty
trivial
and
that
actually,
if
you
want
to
implement
your
own
one
or
simply
excellent
one
of
ours,
it's
pretty
simple
for
you.
So
if
you've
got
some
very
specific
exotic
business
logic
to
integrate
in
your
decision,
it's
it
simple
for
you
to
make
it
yeah.
Oh
here.
D
D
B
Alright,
so
we
don't
have
any
official
benchmarks,
so
what
it
ends
going
to
do
is
gonna.
Do
a
multi
gun,
which
is
a
API
call
on
the
server.
So
the
coordinators
job
is
going
to
be
that
it's
going
to
go
out
and
gather
that
data
and
then
stream
it
back
to
you
or
bring
it
back
to
when
you're
doing
it.
This
way,
with
your
threads
you're
doing
it
from
a
client
perspective
so
which
one's
faster,
which
one
do
you
think
it's
faster.
A
D
D
B
D
B
A
Was
a
prepared
statement,
maybe
well,
you
cannot
prepare
a
statement
with
in
for
the
name
so
right
now
its
limitation
and
actually
we're
comfortable
with
that,
because
we
don't
want
to
push
something
that
we
don't
continue
as
best
possible
practice
right
now.
We
don't
want
to
to
encourage
you
to
to
to
white
some
massive
milky
gate,
because
we
believe
that
it's
a
it's
a
better
approach
to
try
to
dispatch
to
spread
your
your
get
as
much
as
possible
all
for
your
closer
species.
A
D
A
Normally,
you
should
have
some
fairly
interesting
result
with
just
putting
a
bunch
of
a
simple
people
statement
over
your
cluster.
Once
again,
you
can
also
consider
if
you
got
really
like
10,000
of
ID's
to
fetch.
You
can
also
consider
sending
some
some
select
statement
with
encloses
with
maybe
two
three
four,
maybe
10
ID
is
not
more.
You
know
just
for
you
two
to
make
some
small
batch
know
of
phage,
but
yeah.
D
B
A
So
dotnet
we
will
be
the
final
version
of
the
dot
and
Driver.
These
bones
live
these
months.
Probably
we
plan
to
open
source
the
Python
driver,
so
you
can
expect
a
final
version
much
later,
because
we
we
keep
up.
We
take
our
time
to
stabilize
the
API,
because
we
really,
as
I
said
before
it's
a
long
term
investment.
A
Obviously
we
cannot
support
all
the
non
wedges
because
we
there's
a
bunch
of
smaller
community
or
smaller
exotic
language
that
we
want
Bob
between.
We
want
to
have
some
plans,
either
by
ourselves
or
together
with
as
a
community,
to
provide
some
great
option.
You
know,
following
this
kind
of
architectures,
that
we've
that
we
try
to
promote
for
all
the
major
languages.