►
From YouTube: Apache Cassandra and Python
Description
Description
Using Apache Cassandra from Python is easy to do. This talk will cover setting up and using a local development instance of Cassandra from Python. It will cover using the low level thrift interface, as well as using the higher level pycassa library.
Abstract
Very brief intro to Apache Cassandra
What is Apache Cassandra and where do I get it?
Using the Cassandra CLI to setup a keyspace (table) to hold our data
Installing the Cassandra thrift API module
Using Cassandra from the thrift API
Connecting
Writing
Reading
Batch operations
Installing the pycassa module
Using Cassandra from the pycassa module
Connecting
Reading
Writing
Batch operations
Indexing in Cassandra
Automatic vs Rolling your own
Using Composite Columns
Setting them up from the CLI
How to using them from pycassa
Lessons learned
Speaker Profile
Jeremiah Jordan is a Software Developer at Morningstar, Inc. His team created an Operational Data Store using Python along with Apache Cassandra, MySQL, Google Protocol Buffers, ActiveMQ, ZeroMQ, and Zookeeper (all being used from Python).
A
Probably
skills
will
eventually
consistent
period,
structured
key
value
store
the
distributed
system
technologies
from
batam,
oh
and
the
data
model
and
big
table
like
dynamo,
was
eventually
consistent.
Unlike
big
table
is
deposited
on
family
day
stay
them
all.
Richard
Vanessa
pool
cue
value
systems.
A
A
Your
key
is
along
the
side,
so
keys
and
Cassandra
are
stored,
based
on
what
Kendra
calls
the
partitioner
and
the
partitioner
determines
how
data
is
spread
across
across
your
multi
machine
cluster.
The
usual
use
of
the
partitioner
is
what
they
called
a
random
partitioner.
So
keys
are
sorted
in
sorting
a
random
order
and
the
order
is
basically
the
md5
hash
of
the
key,
but
it
turns
to
a
beyond
basically
random.
So
you
can't
do
any
kind
of
ordered,
querying
and
cross
keys
in
the
normal
use
of
cassandra.
A
That's
going
to
split
all
of
your
data
evenly
Buster
very
easily.
There
are
ordered
partition
errs
that
you
know,
will
sort
the
keys
in
the
order
you
specify,
but
those
are
harder
to
use
in
terms
of
keeping
your
data
distributed
across
your
cluster
evenly
because
I've
uses
that
key
order
to
put
stuff
on
the
servers.
If
all
of
your
data
is
ordered,
you're
going
to
have
hot
spots
when
you're
inserting
that
data,
so
the
recommendation
really
is
use
the
random
partitioner.
A
Unless
you
really
really
have
to
not
use
it,
and
you
really
want
to
know
what
you're
doing
you
know
there,
look
at
columns,
on
the
other
hand,
are
sorted
in
column,
name
order
and
the
column
means
there's
a
bunch
of
different
types.
You
could
have
four
column
and
they
can
be
strings.
They
can
be
numbers,
they
can
be
time
basis
you
IDs,
which
will
sort
in
time
order
and
that
lets
you
do
interesting
things
with
the
indices.
A
So
if
you
use
a
time
to
do
ID,
you
can
actually
query
give
me
all
the
columns
insert
is
between
five
o'clock
and
six
o'clock.
You
know
if
your,
if
your
column
name
is
the
time
based
thing
or
purse
strings,
you
can
query.
You
know
give
me
all
of
the
columns
between
a
and
B
and
it
will,
you
know,
do
a
string,
compare
and
all
the
column
names
and
give
those
back
to
you
for
the
game.
A
A
A
A
A
A
Use
their
packaging
for
free,
they
sell
services.
How
do
you
run
it
so
hard
about
development
instance
here?
So
basically,
you
install
it.
There
are
two
configuration
files
really
that
you'll
need
to
touch.
There's
one
file
that
says
where
to
put
the
data
directories
and
the
other.
You
know
where
it's
going
to
store
data
and
then
there's
the
logging
file
that
specifies
you
know.
What's
how
you
want
to
Center
the
log
and
those
are
two
things
causing
it
edit.
A
Once
you've
edited
I
really
drunk
just
run,
the
back
run
the
stark
script
you're
using
the
development
with
a
dash,
app
says
running
in
the
foreground,
so
it's
going
to
by
default.
It's
going
to
spawn
it
off
into
the
background
and
render
the
demon
and
Cassandra's
designed
such
that
to
stop
it.
You
run
the
till
command
on
it.
There's
no
stopping
managers
till
the
process
all
talk
to
that
so
some
stuff
when
you're
using
it
for
development.
A
What
I
usually
do
is
make
a
template
file
out
of
those
two
computer
files
and
then
modify
the
start
script
to
take
those
templates
figure
out
what
path
is
running
out
and
stick
the
data
under
there.
So
you
can
just
check
all
of
this
and
your
source
control
when
a
developer
pulls
it
out
of
source
control.
The
first
time
they
run
that
start
script
is
going
to
start
up
local
instance
of
Cassandra
on
their
machine
with
could
correct
pass
for
where
they
are
makes
it
really
easy
for
developers
to
start
up.
A
Once
you've
got
it
running,
a
person
you
need
to
do
is
set
up
your
tables
you're
keen
basis
in
your
column.
Families
so
hope
for
that.
Right
now
is
the
command
line
interface.
You
can
do
it
programmatically,
but
changing
ki
spaces
and
Colin
families
is
not
usually
a
good
idea
to
not
usually
something
you
want
to
do
on
the
fly,
because
you
want
the
controlled
environment.
A
Create
a
key
space
called
placement
strategy
which
basically
says,
tells
Cassandra
how
the
store
the
data
across
this
cluster,
the
simple
strategy
just
stores
data
in
key
order
across
the
moods
based
on
the
replication
factor
for
development
instance,
your
replication
factors
when
you've
won
there's
other
strategies
that
will
distribute
the
data
across
data
centers
network.
Aware
well
stuff.
You
can
do
with
that
use
your
key
space
in
creeks
all
families
inside
it
call
family
that
the
columns
are
at
key
in
another.
The
columns
are
hot
new
IDs
for
home.
A
To
talk
to
it
from
Python
there's
a
few
different
ways:
you
can
do
it
there's
the
straight
bris
client
cassandra
is
written
using
the
thrift.
Our
pristine
mechanism,
rich,
is
another
Apache
project
that
basically
lets
you
create
configuration
files
for
how
a
real
procedure
call
is
going
to
work
and
they
have
generators
for
50
different
languages.
For
being
able
to
talk.
For
that,
you
don't!
Actually
you
don't
really
want
to
use
that
directly.
That's
a
very
low
level
interface
and
it's
pretty
ugly
to
use
and
then
there's
the
picasa
client,
which
is
a
Python
client.
A
A
A
A
Republican
X
pasta
has
connection
pools,
do
all
the
stuff
that
connection
pools
do
tell
the
key
space
and
go
to
the
server
list.
It's
going
to
do.
Failover
between
the
server
list
read
create
a
thread
pool
and
if
you
have
multiple
threads
going
to
end
up
mentions,
you
know
one
each
thread
so
that
your
connections
are
balanced
across
the
different
servers
so
that
we
are
going
to
get
workin's
concurrency.
A
A
A
A
Nervous.
Most
people
probably
is
more,
is
batting,
so
one
at
a
time
is
always
slower
than
pop
to
the
time.
So
the
batching
interface
is
very
simple.
You
just
do
multi-level
dictionaries
instead
of
the
key
in
lowell
dictionary,
so
you
in
a
multi-level
dictionary
that
the
first
level
of
the
dictionary
has
to
keep
in
it
and
the
second
level
has
the
colonies
of
the
values.
A
And
then
there's
also
a
streaming
batching
interface
in
thigh-high,
positive
that
you
can
give
an
optional
to
size
2
if
you
do
it
the
queue
size
and
every
every.
So
in
this
case,
every
ten
operations
it's
going
to
send
that
off
to
the
server
would
even
make
one
of
these
passing
through
some
process.
That's
just
running
through
David
calling
concern
insert
insert
this
is
going
to
you
know,
fill
them
up
and
send
them,
or
if
you
don't
specify
a
queue
size,
you
can
have
an
infinite
queue
sizes
willing
to
do
it.
A
When
you
call
ten
and
you
can
always
call
send
early,
you
do
it,
but
you
said,
and
in
certain
routes
in
in
one
batch
that
books,
if
you
get
it
off
the
column
family,
this
is
all
better
than
one
column
family.
But
you
can
also
do
multi-column
penalty
batches.
Where
you
create
this
mutator
object,
and
then
you
specify
or
every
insert
or
remove
you
just
give
it
a
column.
Family
object.
A
If
it's,
which
column
filling
that's
going
to
operate
on
and
when
it
sends
it
off
to
the
server,
that's
going
to
be
set
off
with
a
server
and
one
request
across
those
column,
families
that
will
succeed
or
fail
across
this
call
family
good
for,
if
you're,
inserting
the
data
value
and
about
that
value
into
value,
you
can
put
those
together
and
go
back
to
send
them
off
together
until
your
data
and
your
index
is
update,
are
going
to
happen.
At
the
same
time,.
A
A
A
A
A
A
Specify
types
optional
so
and
you
have
actually
practice
group
icon,
so
you
have
two
ways
to
do
this:
you
can,
if
you've
got
that
you've
set
these
schema
information
up
in
the
database
itself,
then
you
don't
have
to
do
this
here.
If
you
have
it,
you
know
picasa,
when
it
connects,
will
read
that
information
and
automatic
setting
stuff
up.
But
if
you
have
it
set
it
in
the
database,
but
you
want
to
put
a
scheme
on
top
of
it
through
your
client.
A
You
can
talk
about
Casa,
eight
column
is
going
to
be
an
integer
hit,
columns
going
to
be
a
float,
and
then,
when
you
do
your
insert,
you
can
just
use
the
integer
32
or
the
6.1,
and
it's
going
to
convert
that
into
bytes
and
insert
it
up
lights
into
the
database.
And
then,
when
you
read
back
out
it'll
give
it
to
you
back
out.
It'll,
take
those
bikes
and
go
back
to
a
number.
A
With
that
typing
functionality
has
a
concept:
that's
similar
to
kind
of
like
object,
relational
model
build
relations
with
object
model
where
you
can
specify
objects,
and
you
can
tell
it
what
the
key
for
that
object
is
going
to
be
replaced
and
you
can
specify
attributes
and
what
what
age
they
are,
and
it's
going
to
take
these
attributes
and
column
column
names
and
this
key
if
we're
looking
it
up
and
then
you
create
a
column.
Family
map
object
from
the
users
that
this
was
the
other
to
create
with
the
type
of
information
in
it.
B
A
A
user
object
says
he
said
the
attributes
in
it
with
their
different
types,
the
time
object
and
click
on
and
then
insert
it
and
Picasa
will
take
all
that
stuff
and
convert
it
down
to
the
binary
format
of
those
things
and
answer
them
into
the
database
and
then
read
back
with
that
key
and
it's
going
to
return
you
one
of
those
user
objects.
That
already
has
all
that
should
be
populated
with
the
data
from
that
road,
and
you
can
also
do
multi
yet
with
multiple
keys
and
then
removing
you
passing
a
user
object.
A
A
The
disease
really
easy
use
like
that.
You
say:
hey,
Cassandra,
alia,
to
index
the
H
column.
Now
I
want
you
to
index
the
state
column,
and
then
you
can
actually
do
query
just
so.
You
know
give
me
everybody
from
Illinois
here
giving
everybody
was
22
and
filtering
queries
as
long
as
one
at
one
column
in
the
is
an
equality
in
that
query,
you
can
actually
do
greater
than
less
than
queries
and
a
couple
other
operations
use.
A
You
could
say,
tell
me
everyone
in
Illinois
sins
over
the
20,
and
you
can
actually
do
that,
but
they're
not
recommended
for
hike,
aren't
even
part
nality
values,
because
Cassandra
under
the
scenes
is
basically
building
a
hidden
column
family
for
that
index.
That's
you
know
turning
that
index
into
a
key
value
pair
queried
under
the
scenes.
A
So
if
you
don't
want
to
do
blow
out,
if
you
blow
out
the
number
of
columns
that
are
a
number
of
unique
values
that
are
being
indexed,
it's
going
to
make
your
query
slower,
because
now
it's
got
to
look
at.
You
know
that
index
with
a
huge
number
of
a
billion
things
in
the
row.
I
think
Cassandra
has
a
maximum
gross
eyes
that
make
two
billion
so
I
mean
you
can
get
a
lot
of
things.
A
lot
of
columns
in
there.
A
Also
fuse
that
built
indexes
they're
there
make
sure
right
slower,
because
for
every
right
the
server
has
to
do
a
read
before
right
to
check
if
it
has
to
remove
something
from
index
team
in
when
answer
into
the
index,
we're
as
a
normal
right
just
purely
goes
to
a
commit
law
and
returns.
So
it
doesn't
have
to
do
any
kind
of
reading.
A
A
A
Here's
how
you
using
the
native
indexes
from
within
picasa,
but
you
create
an
index
expression,
so
I
care
where
they
find
everyone
whose
state
illinois
and
the
filtering
so
who's
off
the
booze
age
is
greater
than
20,
and
then
you
just
they
constantly
index
slices
from
this
clause,
and
it's
going
to
return
you
key
data
pairs
for
everybody.
That's
in
that
through
that
and
do
whatever
you
want
and
actually
uses
a
python
generator
behind
the
scenes
and
does
internal
batching.
So
it
does
paging
to
the
server.
A
Now,
when
you
do
that,
you
have
to
deal
with
the
read
before
I
problem
yourself,
but
if
you're
only
inserting
new
data,
you
don't
have
to
take
that
penalty.
Know.
If
you
know
it's
new,
you
don't
have
to
do.
Read
you
just
right
place
and
it's
going
to
be
fast
and
those
are
you
enroll
yourself
is.
You
can
make
your
index
into
a
denormalized
query
instead
of
just
an
index,
so
you
know
you
pull
that
next
and
you
actually
get
your
value,
all
your
values.
A
You
want,
because
you
stick
the
data
in
size
of
the
result
there
and
then
you
don't
have
to
read
the
index
and
then
read
the
road,
and
then
there
are
other
techniques,
so
you
can
kind
of
roll
your
own.
You
can
do
something
kind
of
like
when
you
want
to
wear
you
index
on
this
field
and
on
the
other
field
it
has
the
concept
of
composite
columns,
or
basically
you
can
say
this
column.
Is
these
three
things?
A
I
mean
it's
an
integer
and
a
string
and
whatever
that
can
be
one
colony
and
you
can
actually
use
the
built-in
indexes
with
that
and
search
on
those
individual
fields
in
order.
So
if
you
use
the
dolton
index
with
the
composite
column,
you
know
this
guy's
a
number
in
a
string,
and
so
you
can
say
give
me
all
of
the
columns
where
the
number
is
20
and
it
will
do
that
work
in
Yaletown,
where
the
number
is
20
and
the
string
is
that.
A
Some
lessons
learn
but
I've
had
was
using
this
change
over
the
last
year
and
a
half
use
indexes,
don't
iterate
over
the
keys
in
your
database
because
that's
just
going
to
be
slow.
We
had
so
we
had
some
data
that
we
thought
that
isn't
going
to
be
so
big,
we'll
just
no
run
through
it
and
do
it
unless
you're
doing
a
MapReduce
operation
of
profit-taking
index.
A
B
C
A
A
The
connection
pool
there's
actually
offers
on
it
so
when
it
connects
there's
options
to
tell
it
to
tell
picasa
to
after
you
connect
query
the
database
to
get
a
server
list,
so
you
can't
have
a
high
casa
go
ask
the
database
what
all
the
servers
are
so
that
it
can
round
robin
across
them
or
you
can
have
it
just
only
use
the
servers
that
tell
you
I,
think
the
belting
aquarium
stuff.
If
you
do
go
into
like
a
multi
data
center
model,
I,
don't
think
picasa
yet
understands
the
concept
of.
A
I
want
to
ask
the
database
for
the
servers
in
my
current
data
center.
So
if
you
do
use
that
multi
data
center,
that's
going
to
sending
stuff,
but
it's
okay,
there
are
functionalities
for
clearing
Cassandra
and
asking
you
know.
What's
your
server
list,
another
thing
you
pushing
pretty
simple
when
it
connects
a
query
and
when
you
connect
to
a
column
family,
it
actually
queries
the
column
family
for
its
schema,
which
one
has
been
set
so
in
automatically
based
on
colonies.
A
Like
us,
isn't
it
I,
don't
think
Picasso?
Does
that
I'm
not
sure
if
they
do
any
of
that
under
the
hood
netflix
me
easily,
because
Cassandra
is
written
in
java
and
netflix
client
that
does
that
is
written
in
java,
and
so
it
just
uses
the
java
jar
file
from
Cassandra
to
ask
the
Cassandra
library
itself.
A
Where
is
it
this
key?
Go
you
know
so
that
you
know
they
don't
they're,
not
reimplemented,
those
algorithms.
You
know
they're
using
the
library
from
Cassandra
to
do
that,
so
they
don't
have
to
worry
about,
though
Cassandra
decides
to
change
that
algorithm.
Now
our
stuffs
broken
or
well
going
to
the
wrong
server
first,
so
there
you
thought
you
didn't,
have
an
extra
hop
and
asking
for
data.
A
A
True,
so
that's
true,
so
you
can
specify
it
picasa
I'll,
remember
what
it
defaults
to
either.
I
think
it
defaults
to
quorum
made
it
fall
to
one.
But
so
when
you
another
parameter
that
all
of
your
queries,
typical
Sanders,
you
can
specify
how
many
servers
do
you
want
to
respond
to
this
query
before
you
consider
that
the
answer
Valerie?
So
if
you
said
your
replication
factor
is
three
you
can
tell
it.
I
just
want
to
talk
to
one
node
and
the
first
note
that
tells
me
an
answer.
A
I'm
going
to
use
that
as
my
answer
or
you
can
tell
that
comment,
all
of
the
nodes
to
respond
and
I'm
going
to
use
the
conflict
resolved
answers
which
one's
the
newest.
So
if,
like
one
server
was
down
by
you
got
written,
you
know
when
you
asked
for
that.
Guy
may
have
stale
data.
So
if
you
use
one,
you
talk
to
him
you're
getting
a
Dell
data.
A
But
if
you
say
everybody,
you
know
you're
going
to
get
the
conflict
resolution,
that's
going
to
get
you
back
the
right
answer
and
also
the
concept
of
quorum
where
it's
you
know:
endo
n,
over
2,
plus
1
nodes
respond.
So
if
you,
if
you
read
and
write
with
quorum,
you're
going
to
have
consistent
data
for
successful
group,
successful
reason
successful
ways,
it
form
whether
it
be
consistent
for
their
fail
and
tell
you
there
aren't
enough
servers
up
to.
A
What
if
there
is
we're
using
it
for
an
operational
data,
store
and
we're
keeping
history
tracking
in
it
as
well?
So
we
store
current
value
of
beta
and
we
store
changes
to
data
for
the
past
90
days
and
we
use
the
the
time-to-live
features
in
Cassandra.
So
whenever
we
insert
something
into
our
historical
table,
we
give
it
a
90
day
time
to
live
and
Cassandra
will
automatically
throw
away
the
old
data.
D
A
Not
really
I
most
of
our
stuff,
where
we're
inserting
sign
for
the
most
part
we're
inserting
everything
by
heat
I
mean
for
the
most
part.
We
know
the
key
and
we
have
a
few
different
indices.
For
you
know,
we
keep
indices
of
what
data
points
are
set
for
what
keys
and
stuff
like
that,
you
can
that
we
actually
started
this
before
cassandra
has
a
built-in
indices,
so
we're
not
using
any
votes,
didn't
exist
when
we
started.
A
Like
the
biggest
thing
with
Cassandra
was
their
multi
data
center,
where
replication
and
querying
know,
so
you
can
insert
something
into
Cassandra
and
say
insert
it
consistently
into
my
data
center
and
asynchronously
into
the
other
car,
or
you
can
say,
insert
this
insert
this
consistently
every
data
center
connected
to,
or
you
know,
there's
all
kinds
of
different
queries
stuff
for,
depending
on
how
consistent
you
want
your
data.
How
fast
you
want
your
query
to
be
what
you
can
do.
That's
one
of
the
biggest
features
the
column
based
model
where
every
key.
B
A
So
so
that
so,
when
you
create
the,
when
you
create
your
key
space
and
your
column,
families
right
you
had
that
placement
strategy.
So
one
of
the
placement
strategies
is
basically
a
network
aware
of
placement
strategy
and
there's
actually
a
bunch
of
different
ones.
Well,
where
you
give
it
a
configuration
file
that
has
like
the
list
of
IP
addresses,
and
then
you
tell
it
so
what
actually
is
aware
of
racks
as
well?
A
So
if
you
want
to
say
this
rack
versus
that
rack
within
a
data
center,
what
racks
on
things
in
what
data
center
something
in
something
in
and
this
change
will
use
the
rack
awareness
to
do
closeness.
So,
if
I'm
asking
for
something
with,
you
know,
consistency
level
of
you
know
three
or
creating
and
I
say
Ellen
from
two.
You
know
local,
ok,
wish
note
is
in
my
same
rack
and
asked
him
for
the
data,
and
then
it
also
has
the
data
center
awareness
stuff,
where
you
can
set
your
replication
factor
actually
distinct
between
data
centers.
A
A
And
then
it
has
like
the
querying
ability
when
you
query
stuff,
you
can
say
you
know,
give
me
a
consistent
level,
one
from
current
data
center.
So
just
give
me
this
fast
as
you
can,
or
you
can
actually
say,
you
know,
give
me
a
consistency
level
of
all
from
every
data
center.
So
it's
going
to
make
sure
you
have
the
newest
value
from
every
day
to
send
her
out
there.
It's
just
going
to
take
a
long
time.
D
A
On
what
caching
you
have
on
you
know,
a
lookup
can
take.
You
may
only
be
one
disk
seek
to
get
the
data
off
disk
if
it's
on
disk
has
a
different,
Isaac
levels
available
for
storing
you
know
on
this
data.
Cassandra
has
kash's.
You
can
set
up
caches
with
it
Cassandra,
so
the
data
memory
stuff
like
that,
so
that
you
know
you
cap
on
top
of
this
andrew,
because.
D
A
C
A
C
A
Nothing
in
the
only
myself,
Cassandra
so
Cassandra
when
it
writes
data
to
disk
Cassandra,
only
ever
writes
files
to
disc
once
so
once
those
are
red,
like
your,
your
OS
is
going
to
start
catching
those
files
and
we've
actually
done
that's
a
very
big
performance
game.
If
your
data
files
on
disk
fit
in
the
OS
disk
cache
Cassandra
don't
go
changing
those,
so
they
really
did
or
have
issues
with
that,
so
they
get
into
the
OS
disk
cache.
And
then
your
queries
are
basically
out
of
memory.
D
A
D
A
B
A
A
Know
so
if
you
know,
if
it
fails,
you
know
every
succeeded,
everything
succeeded
and
the
changes
to
a
given
row
are
going
to
be
the
comment
now
across
rows.
Your
changes
are
an
atomic,
but
your
your
batch
maybe
will
succeed
or
fail
together.
But
if
you
have
something
reading
those
rows
at
the
same
time,
you
know
row
one
may
be
robots.
A
You
can
you
can
so
for
someone,
so
I'd
be
a
setup
retries
on
different
on.
So
with
the
server
list,
you
can
have
a
Jew
automatic
failover
if
something
kinds
out
or
or
not
any
Cassandra
rights
to
a
Cassandra
database
are
usually
our
unemployment.
So
you
know
you
have
the
time
Sprint's
get
a
timestamp
assigned
to
them
and
as
long
as
you're
writing
with
that
same
time
stamp,
you
know
a
new,
a
new
right
with
a
higher
timestamp.
A
Even
if
you
write
the
old
value
of
the
lower
timestamp
desire
timestamps
going
to
win,
so
you
can
keep
replaying
that
right
until
it's
a
deep
and
it's
not
going
to
hurt
anything
and
you
can
tell
and
pie
pops
up
by
default
will
do
that
if
it
gets
like
a
timeout
exception
or
an
unavailable
exception.
That
says
you
know
I
would
either
you
know
what
a
node
was
down
or
renewed
was
not
responding.
Whatever
it'll-it'll,
we
can
keep
trying.
C
A
Stamp
you
don't
have
to
by
default
it's
going
to
use
current
system
time
in
microseconds,
but
you
can
also
specify
it,
and
you
can
also
tell
if
you
want
it
to
return
the
time
stamp
and
then
the
values
when
it
gives
you
the
values
back.
It
will
give
you
back
a
tuple
that
has
the
value
of
the
time
stamp.
For
that.
A
B
A
They
talk
about
their
use
of
Cassandra
a
lot
they've,
actually
open
source
and
released
a
lot
of
their
management
tools
and
Cassandra
there.
Their
client
libraries
that
they've
developed
for
using
Cassandra
from
Java
actually
Cecily
interesting,
backup
and
recovery
tools
that
they
created
for
backing
up
to
Sandra
databases
that
are
running
in
Amazon,
spinning
back
up
incremental
backups
to
Amazon,
s3
and
being
able
to
restore
from
those
incremental
backups
all
kinds
of
interesting
box
tools.