►
Description
SlideShare: http://www.slideshare.net/JonHaddad/crash-course-intro-to-cassandra
This is a crash course in the Cassandra storage model. We will start with how data is laid out on disk to understand why it's fast for real time workloads, and finish with an introduction to CQL, which we'll use to query Cassandra. If you have not used Cassandra previously this will bring you up to speed for the talks that follow throughout the day.
Jon Haddad is an Apahce Cassandra Evangelist at DataStax; prior to DataStax, he worked as a Senior Architect at SHIFT and is one of the maintainers of CQLengine. Jon has spent the last decade at various startups in the LA area. His previous projects include Answerbag.com and Livestrong.com.
A
B
B
B
Anything
like
that.
All
the
ring
is
distributed
throughout
learning,
hopefully
evenly
as
a
result
of
this
there's
no
single
point,
failure
which
is
pretty
cool.
You
don't
have
to
worry
about
and
we
failover
master
slave
promotions
either
election
and
things
like
that.
That's
just
not
concern
here,
primarily,
for
example,
built
on
top
of
or
high
availability
and
linear
scalability.
So
everything
every
similar
design
is
in
court.
B
Assigned
position
was
filled
with
this
in
mind,
so
things
that
may
seem
a
little
bit
weird
or
different
from
previous
databases
that
you
needed
to
use
before
their
build
with
the
single
role
of
this
once
well
with
pre-service,
and
this
also
works
well
with
300
circles
from
the
ground
up
built
with
multi
PC
in
mind.
So
this
is
part
of
our
mobility
thing.
We've
got
potentiated
Center
on
west
coast
and
on
the
East
Coast
Green
Europe.
When.
B
Seen
that
did
this
go
beyond
and
as
a
result,
works
really
really
well,
you
have
entirely
one
thing
to
keep
in
mind.
Is
that
because
the
way
it
works
is
every
girl
is
going
to
find
data
by
the
key?
So
whatever
your
primary
key
is
that's
how
you're
going
to
look
up
where
I'm
a
little
bit
more
into
detail
a
little
bit
how
it
works
with.
B
B
We
can
see
here
this
is
an
example
of
patterning.
It
will
be
distributed
evenly
throughout
these
its
nodes
in
this
cluster.
This
is
how
the
posters
were
traditionally
set
up.
Ring
would
get
part
of
a
slice
of
data
right
here.
People,
newer
versions
of
the
Stanford,
it's
broken
into
smaller
chunks.
It
is
distributed
instead
of
having
a
big
chunk
of
data.
Here,
it's
distributed
all
around
make
certain
operations
a
little
bit
easier
that
weren't
religion
to
see
didn't
do
right
now,.
B
A
B
To
additional
confidence,
so
this
is
working
out
your
reputation
file.
So
if
you
say
all
right
equals
three
tentative
3d
servers
and
none
of
them
have
a
master
there's
just
there's
one
server
and
other
replicas
are
things
you
can
tell
Cassandra
the
rapper,
where
you
can
say
well
in
that
case
into
it's
best
to
not
put
two
copies
of
your
data.
It
would
be
someone
on
the
same
graph.
So
in
case
that
you
have
dr.
Taylor
or
there's
a
failure.
Amazon,
it's
totally
fine
impossible.
B
So
when
she
think
I'm
ability,
there's
always
gonna
be
some
trade-offs,
because
we
want
this
thing
to
be
able
to
be
highly
available
with
a
single
face
on
and
the
multi-tiered
Center.
The
concession
that
we
make
is
about
consistency
when
there
are
problems
with
network
partitions
and
their
problems.
B
You
have
queries,
you
have
a
tunable
consistency
so
that
you
can
say
if
I
have
artists,
verification
factors.
Three
and
I'm.
Looking
for
information,
I
can
say
it's
okay.
This
informations
will
automate
just
give
me
whatever.
Whatever
comes
back
from
the
customer,
that's
totally
fine,
the
first
server
to
return
some
of
that
data.
B
You
can
also
flip
up
your
consistency
level.
You
can
say
for
I
want
two
out
of
three
to
agree
upon
what
the
right
answer
is.
It's
not
weird
what
you
can
say:
I
want
all
the
machines
to
agree
on
this
particular
strongly
consistent
and
give
up,
or
you
can't
kind
of
consistent,
pretty
consistent
and
way
higher
example.
B
C
B
You
can
do
multiple
reason
it
so
you
can
to
Sanders
built
so
that,
instead
of
thinking
about
one
giant
cluster
with
each
let's
say-
hey,
I'm,
sorry
he
was
how
the
cluster
and
then
cluster
in
West
Indies
and
they
smile.
So
when
you
do
it
right,
you
could
say
a
little
warm
and
in
this
case,
would
be
they
go
to
900
plus
room.
You've
got
these
back
to
three
to
nine
clusters.
Two
out
of
the
three
agree
over
here
and
then
asynchronously.
The
data
is
replicated.
B
B
B
B
B
B
B
E
B
So
when
you
have
multiple
Grossmont
partition,
this
is
actually
where
this
Tantra
is
extremely
useful.
This
is
a
usable
amount
of
high
security
data,
you're
collecting
around
their
specific
point
with
a
sensor
or
something
like
that.
What
you
say,
I
always
want
to
get
would
say
those
the
last
thousand
sensor
readings.
B
B
We
know
that
we
always
have
to
put
whatever
and
when
you
create
this
table,
the
effectively
photo
ID
a
photo.
Id
is
an
appeaser
partition.
P
comment:
ideas,
Computer
Club.
What
this
does
is.
It
means
that
all
the
comments
for
particular
photo
store
governmentís.
So
if
you,
how
many
million
photos
hitting
the
first
1000
is
not
going
to
be
nothing?
Compare
this
with
a
relational
database
or
something
like
or
really
anything
else.
Your
data
is
going
to
be
spread
across
a
lot
of
different
sections
of
disk.
B
B
B
F
F
B
F
B
So
in
that
case
it
is
you
itself,
a
second
second
table
with
the
name
as
the
partition
P
and
the
comment,
or
even
the
full
comment
or
just
a
comment:
I
team
as
they're
clustering.
So
that's
right
now
have
the
whole
comments
that
you
just
haven't
and
then
you
can
either
have
additional
all
the
information
is
that
if
you
want
to
have,
if
you
want
to
do
one
query
and
say,
give
me
all
those
comments.
B
Our
limited
to
things
that
steal:
well,
you
wind
up
quite
beautiful
one
of
the
things
that
pop
out
will
visit
and
why
rights
with
this
Amoroso
house,
literally
the
fastest
mayonnaise
rates,
are
so
fast
that
it's
actually
cheaper
to
do
a
whole
bunch
of
Rights
and
then
make
your
beats
faster.
Then
try
and
do
like
one
right.
F
F
B
So
they're
there,
the
last
question
that
was
asked
earlier
over
here.
This
is
one
of
the
tubes,
but
when
you're
designing
your
scheme,
is
you
want
to
keep
in
mind
that
it's
possible
that
you
might
have
data
center
at
less
than
data
center
nice
and
it's
very
very
possible
that
the
link
between
the
two
will
get
severed.
D
B
E
E
B
B
B
B
A
B
C
B
B
This
is
going
to
be
only
the
first
thing
right
here
is
your
personal
bank.
This
is
beautiful,
though,
if
you
had
a
set
of
parentheses,
and
it
was
around
this
exercise
practice.
This
whole
thing
would
be
the
partition
to
get
this.
You
have
to
look
it
up
five
boat
boat
away
in
time,
because
the
whole
part
that
both
keys
now
come
in
place
of
where.
C
B
B
B
Effectively,
wood
efficiency
look
query
against
the
secondary
index.
It
you
should
slowly
it's
going
to
clear
everything
possible,
so
under
notes,
your
notes
in
volume.
That's
the
reason.
This
secondary
indexes
are
managed
by
each
note
and
in
in
surrounding
so
when
each
where
is
going
digital
pipe
because
I
got
off
the
Atlantians.
So
it's
not
like
a
globally
next
one.
You
know
which
server
you're
going
to
get
a
query.
It's
terrible!
So
do
you
know.
E
B
B
B
B
B
E
B
B
B
B
B
B
So
the
right
path,
the
reason
why
I
can
say
they're
so
fast,
because
we
do
right
considering
like
super
long
and
memory.
So
your
basic
about
why
the
speed
under
this,
as
far
as
how
the
masculine
right
to
me
without
any
distance
II
rights,
are
actually
attached
up
and
my
fault
every
10
seconds.
You
can
actually
change
that
one.
Your
data
is
a
lot
more
durable,
but
most
people
don't
because
you
have
three
copies
your
data,
because
all
three
machines
dying
very
low
weight.
That's
their
contents.
B
The
mint
table
made
a
structure,
that's
a
memory.
Kuryakin
B,
plus
two
disk,
an
SSD
SS
tables,
are
a
beautiful
day
of
miles
because
of
the
nature
of
exam.
You
can
actually
have
one
partition,
spread
across
multiple
SS
cables
for
performance
reasons.
Desa
stables
are
compacted
so
you'll
have
let
say
here
as
the
stables
they'll
be
purged
together
at
one
table.
Partitions
will
be
together
so
effectively.
What
the
goal
is
number
of
reads
and.
B
B
B
B
B
B
A
B
B
B
One
thing:
that's
really
cool
about
this
I.
Don't
you
guys
learn
how
to
really
large
data?
It's
like
in
my
seat
or
something
like
that.
Backups
Jews
paint
the
try.
Some
of
this
is
actually
in
essence,
tables
those
are
written
to
just
and
that's
a
legitimate
practice,
because
there
are
compacted
because
they've
never
rethinking
the
files.
B
Updated
even
though
you
might
have
terabyte
cluster,
so
let's
start
what
would
be
happening.
This
is
the
minute
tell
us
live,
but
basically
the
same
pillows
right,
applicants
orbit
we're
in
support.
Here
again
you
have
chosen
where
it
will
go
to
the
right
now,
as
I
mentioned,
the
same
run
over
partition,
the
across
multiple
SS
tables,
all
that's
written
in
birch
together,
because
the
nature
of
SS
tables
you're
gonna
have
to
do
and
it
will
murder
them
seats
still.
So,
even
if
you've
gone,
let's
say
you
messed
for
a
thousand
things.
B
And,
what's
important
to
note
is
that
thing
called
reverb
there
if
you're
clearing
insistence
a
little
less
than
all,
which
is
gonna,
be
all
the
time
to
do
this?
It's
crazy
to
do
it.
Otherwise
there
is
a
chance
10%
chance
by
the
fall
that
note
we'll
go
ahead
and
do
all
the
other
notes
have
that
particular
piece
of
data
and
determine.
If
is
the
most
data
point
it's
not.
It
will
go
ahead
and
do
it
if
and
sundry
personal
so
effectively.
B
G
B
B
The
server
so
the
server
we
little
exam
server,
a
B
and
C
you
use
read
to
make
consistent
one.
It
gives
you
a
response
to
my
hand,
train,
but
I
actually
have
given
my
Hamilton
the
other
servers
over.
Yet
it
will
implement.
Do
repair
and
other
servers
agree
that
my
team
is
actually
Frank
enough
talking.
Then
we
will
fix
it,
so
it
will
temporarily.
B
B
B
B
Sure
this
time
anything
consistent
service
to
write
a
TV
into
NASA.
If
you
do
not,
if
you
do
not
want
a
TV,
your
server
will
be
told
in
jail.
Do
not
ever
run
the
cluster
or
any
server
production
without
a
TV.
It
is
it's.
A
total
nightmare
you'll
end
up
with
weird
issues
where
you
can
beat
it
up
and
that's
gone
before
you
can
meet
something
you
can
serve
something
else.
B
So
the
yes,
the
server
is
down.
They
do
require
all
then
you'll
get
an
error.
If
you
give
away
from
all
three
course
long,
it's
whatever
discrepancies,
there
are
beans,
still
give
you
one
day
to
find
that
in
the
the
query,
consistency
that
actually
aren't
like
it's
a
little
bit
succeed
of
this
consistent.
So
if
you
say
at
all-
and
one
server
was
down,
the
other
two
hoods
will
actually
still
get
that.
B
H
B
B
B
So
we've
actually
already
talked
about
some,
my
probe
data
modeling
stuff.
How
do
I
clear
my
data?
If
I
can
query
the
Hebrew
denormalized,
restore
things,
ten
different
weights
and,
let's
say,
look
we're
really
about
rights,
so
go
ahead
and
write
the
ten
legs.
The
pairs
and
use
the
bass
rights
just
pick
up
burgers
multiple
hit.
It
really
really
not
read,
and
you
have
a
system
scale
to
thousands
and
my
understanding
is
our
blogger,
author,
Steve
Altemus.
So
this
works.
G
B
So
the
question
is:
are
these?
Are
the
different
settings?
I
talked
about
controlled
that
client
summer
server
things
like
this
I
often
talk
to
my
assistant,
I
feel
a
lot
of
guessing.
That's
the
server
setting.
Other
things
like
often
is
repair.
That's
actually
said
it's
people
know.
So
you
would
say
this
table
was
25%
chance
to
read
most
the
only
thing
very
controlled
and
query
level
or
a
connection
level.
D
B
Yeah
so
the
question
the
person
with
regard
to
picking
your
partition,
if
you
eat
clear
by
two
of
those
fields
to
create
one
tables
without
you
know
five
different
ways.
One
query
that
data
pretty
great
five
days
when
you
do
an
insert
you're,
actually
using
insert
of
multi
tables
and
then
very
efficiently
apartment.