►
Description
Speakers: Matt Pfeil, Vice President of Customer Solutions at DataStax; Rick Branson, Infrastructure Engineer at Instagram; Adrian Cockcroft, Cloud Architect at Netflix
In today's world, data is growing faster than ever. For online apps, two things matter more than anything else for the database: uptime and performance. The intersection between data growth and online requirements results in interesting technology choices. This panel will discuss the implications - and approaches - to maximize revenue via technology decisions.
A
A
So
to
really
kick
that
off,
you
know
at
a
high
level,
the
old
world
around
strong
consistency
versus
eventual
consistency
is
strong
consistency
and,
ladies
terms
of
guys
jump
in
whenever
you
want
me
to
stop
talking
what.
If
there
was
a
give
a
masculine
sleep
architecture
right
and
in
general-
and
you
always
knew
if
you
put
something
in
you,
got
that
thing
back,
but
there's
a
really
strong
limitations
to
that,
because
the
world
to
come
to
global
nothing's
like
now
I'm
actually
happen.
An
eventual
consistency
can
solve
for
rooms,
so
perfect,
no
Mike,
I.
B
Think
I
think
it
was
really
good.
People
tend
to
think
that
this
the
choice
there
is
something
that
is
sort
of
binary
in
a
way.
It's
like
you're,
weak
or
strong
is
really
more
of
a
spectrum,
and
actually
eventual
consistency,
isn't
technically
consistency
model.
It
just
means
that,
like
a
date,
it's
a
way
to
describe
a
system
like
this
android
are
like
dynamo
or
RBI
and
scored
on
bae
systems
and
I.
Think
God's
enough
point.
The
idea
is
that
it
opens
up,
is
new
area
of
actually
thinking
about.
B
You
know
how
your
latency
and
your
availability
at
impact,
your
business
decisions
and
actually
started
work
with
your
business
people,
instead
of
saying
than
just
saying
having
this
one
way
street
where
they
say
it
has
to
be
strong,
that
split
is
and
how
the
business
can
actually
be
informed
by.
It
is
work
with
the
technical
side.
While
these
consistency
issues
I
think
it's
actually
people
that
are
going
to
sort
of
making
these
relationships
work
are
going
to
be
able
to
use
these
systems
to
their
to
their
fullest.
You
actually
get
a
business
advantage.
C
I
guess
my
dad
Judas
you've
never
really
had
strong
consistency
in
the
first
place.
If
you
have
one
perfect
database-
and
maybe
you
do-
everyone
has
more
than
one
database
database
is
out
of
it.
So
you
end
up
having
to
do
some
kind
of
consistency
check
at
some
point
anyway,
if
you're
building,
you
know
anytime,
you
have
q's
between
things
or
go
for
your
story,
David
more
than
one
place.
You
know
it's
not
consistent
as.
B
A
lot
of
the
joke
that
you
actually
think
you
know
instagram.
I
seem
to
work
with
one,
and
I
was
all
thursday.
Those
Iowa's
developers
probably
know
more
about
web
developer
because
they're
separated
by
this
net
network
and
not
only
a
network
but
a
mobile
carrier
network,
which
is
usually
terrible.
So
they
better
learn
to
build
up
those
those
gaps
and
treat
that
as
it
really
isn't
distributes.
If
you
ask
track
to
even
like
emotional
use
days
very.
A
A
Actually
really
started.
Was
you
one
thing
that
you
can't
overdraft
successfully
as
a
gift
heart
because
they
don't
know
who
has
it
in
which
case?
Hence
the
title:
it's
an
actuarial
decision
where
you
just
go
work
backwards
to
the
price
of
your
products
and
crucify
the
percentage
needed
to
cover
the
loss,
whatever
that's
not
just
gift
card
I
can
also
feed
you
to
them,
stolen
product,
etc.
So
there
are
actually
very
very
few
things
in
life
that
actually
package
you
strongly
consistent.
I
should
don't
know.
B
Goes
back
to
the
pole,
I
bring
your
business
people
of
getting
involved
in
these
discussions
because
I,
for
instance,
the
whole
I
give
carpet
like
a
better.
I
think
a
better
example
that
generalizes
a
little
bit
better
is
what
happens
if
your
inventory
counts
are
wrong.
Like
you,
so
you
know
you
can
have
like
real
fears
real
world
about
it.
You
drop
your
hidden
hand
off.
B
It
is
somebody
actually
living,
because
if
we
drop
the
device
that
they
were
using
to
scan
things
the
warehouse
and
they
lost
that
data,
so
then
now
you've
actually
got
inventories.
You've
lost
track
ups
of
your
inventory
and
you
then
try
to
ship
that
inventory.
It
doesn't
exist
if
we're
going
told
a
customer
you're
going
to
ship
it.
So
what
do
you
do
is
that
a
lot
of
people
will
do
things
like
I
will
just
go
to
listen
to
our
gift
card.
So
then,
obviously
the
things
like
that
happen
across
the
board.
B
C
So
the
thing
is
really
about
what
what's
the
cost
of
being
wrong
and
the
thing
about
it
from
a
software
developers.
Point
of
view
is
the
things
they
stop
not
one
end
of
the
deal,
so
they
don't
have
to
deal
with
the
fact
that
it
might
be
wrong.
So
they
want
to
pretend
that
it's
consistently
because
they
can
have
a
model
in
their
software
which
pretends
it's
consistent
than
they
don't
have
to
worry
about
it
and
everybody
gettin
fail
because
something
the
reality
doesn't
conform
to
them
over
there.
They
can
write
software,
that's
simpler.
C
D
C
B
Trying
to
understand
like
exactly
what
you
know
if
you
can
download
all
these
things
and
kind
of
contain
all
these
variables
predict
them
now,
obviously
helps
actuarial
decision
actually
in
a
league.
Ok,
it's
maybe
ten.
It
costs
us
as
much
and
compare
that
to
you
know
what
what
this
could
be
able
to
compare
the
trashier
risks
for
things
that
are
inconsistent
or
slower
or
faster,
very
certain
impact,
because
consistency
does
is
a
little
bit
tangled
with
the
whole
agency
get
to
read
from
you
know,
to
rapid
ezza
or
in
nature.
A
Shored
up
there,
though,
you
send
out
a
medicine.
Consistent
systems
are
actually
good
for
the
business.
Let's
work
that
actor
the
technology
side
and
when
you
accept
an
eventually
consistent
system,
because
it
is
anyways,
there's
some
very
blatant
obvious
technology
things
whether
it's
round
availability
in
the
system,
it's
very
hard
to
do
business.
The
system
is
importing
all
the
way
through
to
performance.
Due
to
things
like
speed
of
light,
the
water
yeah.
B
A
Two
very
published
studies
office
actually,
by
reference
way
too
much
Google
found
I'm,
sorry,
amazon,
town
for
every
100
milliseconds
of
latency.
They
introduced
into
the
search
experience.
They
lost
one
percent
of
revenue,
100
milliseconds,
one
percent
of
revenue
and
Google
battle-
that
for
every
500
milliseconds
of
latency
they
introduced
to
the
search
results
they
lost.
Twenty
percent
of
traffic
so
being
slow
is
exponentially
bad.
A
D
B
C
Got
me
aircraft
position.
They
build
lanes
without
any
boxes
with
lots
of
spaces,
so
they
make
sure
that
it
doesn't
need
to
be,
even
if
everything
fails
and
the
entire
flight
system
for
every
plane
keeps
flying
on
his
current
pastor.
They
did
anyways
thing
that
the
actual
system
is
not
offended
right.
Safety
should
be
dependent
for
planes,
not
depression
food,
so
basically
they
do
get.
A
So
on
one
side
you
know
performance
matters
and
one
of
the
real
benefits
around
any
eventual
consistency,
so,
whether
a
sander
or
any
time
you
set
up,
for
example,
like
my
sequel,
replication
back
in
the
day,
which
was
a
synchronous
you
spend
asset
or
sometimes
because
that
s
just
feel
like
thing,
it's
really
hard
to
be
passionate
right.
So
if
100
milliseconds
is
a
magical
number
to.
C
A
C
So
we
repeated
some
testing
where
we
try
to
cut
the
stout
up.
Latency
for
our
applications
on
netflix
is
a
little
different.
Is
your
very
favorite?
So
you
know
it's
not
like
advertising
site.
Where
you
get
slightly
unhappy,
you
go
project
appetite
away,
you've
really
paid
for
it
be
trying
to
get
moving
stuff
and
we
optimize
the
startup
type
for
brazil
by
locating
some
some.
Some
machines
basically
fired
up
some
machines
in
and
because
kind
of
second
or
two
off
of
the
startup
time
to
play
movie,
but
it
only
affects
it
turns
out.
C
Most
of
Latin
America
ran
through
Miami.
To
get
to
anywhere
is
not
what
we
do
it
be
if
the
network
is
a
big
star
from
Miami
and
the
fact
that
the
ultimate
little
bit
of
Brazil
normally
roll
the
eyes,
peas
and
Brazil
go
to
Amazon
Brazil
directly.
Most
of
them
do
that
some
/
Miami
to
get
around
was
in
prison.
So.
C
C
C
So
we
basically
set
up
enough
interest
really
test
this,
and
then
we
decided,
after
we
could
say,
testing
workout,
shutup.
Okay,
not
that
so
that's
part
of
the
infrastructure
agility
that
dances,
try,
stuff,
learn
something,
and
this
is
a
google
loop
of
learning
or
orienting
yourself,
making
decisions
and
making
being
able
to
decide
and
all
these
things
faster
than
our
competition.
Who
is
throwing
out
struggling
the
school
trainer
only
with
80
about
whether
they
could
even
have
a
data
center
somewhere
anyway,
how
much
it
costs
to
ship
things
there.
C
We
just
divide
it
up,
but
before
you
can
have
a
conversation
for
it
completed
it,
so
that
that's
the
thing
for
us
that
saves
money,
the
ability
to
learn
very
quickly
and
in
very
small
incremental
ways,
and
yet
out
of
thee.
It's
all
about
agility,
rather
than
customs,
I
hope,
because
the
longest
pole
in
getting
something
done
now
is
developed
a
time.
Anything
do
to
reduce
developer.
Time
means
the
coho
competitive.
That
means
that
you
generally
make
more
money,
which
you've
got
more
money
to
spend
on
developers.
C
E
D
C
A
A
E
A
Cloud
ramazan
to
your
destiny,
the
assumption
that
was
here
in
terms
of
tale
of
two
different
companies
I
used
to
have
all
time
as
you
guys,
who
are
very
fit
into
the
cloud
bursts
of
them.
I
thinka
who
started
it
was
on
and
good
positive
Maggie
serves
you
guys.
Both
went
through
peer
to
business
risk.
A
little
shaky.
You
guys
just
spun
out
service
contracts,
David.
A
C
C
Year
or
more
formally
right,
but
German
and
they're,
incredibly
data,
centers
and
they're
doing
it
incredibly
fast
in
six
months
right,
most
people
so
you're
committing
early
and
that
early
commitment
is
longer
than
our
windows
of
the
future.
For
a
making
business
decisions
about
which
country
will
to
be
a
war.
Nevada
you're
going
to
choose
which
country
we're
going
to
go
into
less
than
six
months.
A
C
A
C
It
costs
optimizing
our
old
DVD
business,
which
is
slowly
shrinking
it
that
it
was
trying
to
read
that
as
efficiently
as
possible.
Right
there's,
no
new
features
there.
There's
no
new
customers
coming
in
equally
gradually,
switching
the
streaming
that
business
is
just
a
slowly.
Shrinking
mature
business
still
runs
on
oracle
still
in
the
datacenter.
We
don't
even
know
how
to
modify
the
schema
of
anyone
less
eventually
repentance.
So.
C
Try
not
to
break
it.
Well,
folks,
are
using
that's
that's
a
total.
That's
the
interest
model
or
a
mature
business
within
that
book
says
one
mature
business
where
we
do
something
profits
out
of
it:
fun,
expansion
everywhere
else,
and
one
fast
growing.
This
that's
what
was
sinking
money
into
it.
Ingenue
countries
yeah.
C
B
That's
terrible,
please
ma'am,
don't
you
got
nowhere
to
go
because
that
means
this
is
pretty
yeah.
So
the
idea
is
that
you
know
consider,
is
you
know,
fuels
with
some
of
the
rough
spots
you
gotta
think
you
have.
You
talked
about
the
fact
that
you're
sort
of
building
this
system
on
top
of
this
somewhat
unreliable,
somewhat
ephemeral
system,
and
so
the
country,
all
that
around
that's
what
we're
all
you're
talking
about
is
it
is
enabling
previous
technology
so
move
the.
C
Web
scale
was
before
we
came
in
right
now.
The
scale
at
which
the
clouds
phelps
be
usable
is
about
machines
sort
of
somewhere
in
that
range.
If
you've
got
more
than
hundred
thousand
machines,
you
really
need
to
be
running
around
cloud.
Most
of
the
facebooks
people
in
there
who
was
bigger
than
a
lot
of
the
banks
are
better
than
that
really
big
banks
or
top
or
back
so
they're.
E
C
Their
own
stuff-
and
they
should
be
running
there-
are
so
many
enough
people
to
do
that.
But
if
you're
the
sort
of
10,000
100,000
range
it's
kind
of
optional,
you
can
choose
to
do
your
own
or
you
can
choose.
No,
this
isn't
a
few
tens
of
thousands
gale.
The
other
thing
is
that
that
range
is
going
up.
You
could
argue
it's
going
up
doubling
every
year
because
I'm
from
the
doubling
in
size
every
year,
so
the
size
of
the
public
cloud
who
just
keeps
doubling,
then
we're
kind
of
good
to
go.
C
Cuz
we're
not
growing
fast
and
we're
going
to
stay
in
that
sweet
spot
I
mean
we.
There
are
times
where
we
have
outgrown
our
ability
to
deploy
and
come
back
to
the
center
funded
architecture
which
I
can
see.
Trivially
will
work
as
over
three
times
of
her
capacity
for
company
at
36
million
customers
through
100
million
customers
is
kind
of
idea.
So
before
this
gala
that
might
be
a
couple
of
tweets
around
the
edges
according
to
do
10
times.
My
current
number
of
customers
would
probably
done
something
architecture,
because
everyone
remaining
to
you
do
that.
C
But
ultimately
there
are,
we
have
the
center
clusters,
144
notes
right
now,
and
it's
just
linear
if
we
can
see
that
going
but
get
more
out
of
you
know
as
well.
So
it's
really
fundamental
to
the
scalable
architecture
that
have
a
data
store,
that
you
can
just
throw
things
at
that
benchmark
around
two
years
ago.
Canadian
knows
it
was
linear.
Okay,
we're
gonna
be
there
for
about.
Three
years
later
we
have
quite
hit.
They
were
so
a
half-scale
running
more
performance
than
that,
so
that
is
really
was
funded.
C
A
E
C
The
replication
now
so
the
question
is:
are
at
maintenance
and
how
do
we
do
that?
You
go
to
a
distributed
system
with
replicas
in
it,
and
we
just
read
is
Jenkins
drones
that
rotate
through
our
cluster
and
at
any
point
in
time
we're
replacing
a
note
it
in
plastic,
I
mean
continuously.
We
go
through
every
machine
couple
thousand
machines
and
once
we
go
through
replacing
those
fancy
machines,
we
pick
a
new
version.
C
Cassandra
will
go
to
a
test
environment
and
we
systematically
replace
every
Cassandra
instance
in
our
testing
Cassandra
and
once
we
get
the
end
of
oxygen
bar,
we
got
a
production
and
you
do
that
again.
We've
been
running
that
patent
for
a
year
two
years
now
has
the
way
to
do
so.
There's
always
a
machine
that
have
been
to
be
upgraded.
But
if
one
instance,
you
know
6
200
notes
per
cluster.
So
it's
basically
the
record
kept
away
the
fact
that
you're
using
critical
replication
apply
it
and
still
have
a
core
succeed.
That's
50!
C
E
C
Using
the
sender
to
replicate
all
the
data
that
we
need
in
a
region
to
another
region,
so
what
we're
trying
to
do
is
get
to
a
point
where
deputies
USA
runs
out
of
the
u.s.
East
and
US
west
regions,
and
we
can
send
traffic
to
either
one
and
all
your
base
will
be
there
and
within
a
hundred
milliseconds
of
latency.
It's
consistent
across
both
of
those
you
are
asynchronously
replicates
to
be
able
to.
C
E
Amazon
wish
my
head
was
God,
they
assume
some
point.
You
choose
the
size,
welcome
on
extra
large
instead
of
lawyers.
How
do
you
make
the
economic
determination?
One
is
more
valuable
to
start
putting
better
hard
work
place
goes
through
all
functioning
versus,
adding
more
those
they're.
Stealing
out.
You
see
enough
personnel
I.
B
B
D
D
B
C
Is
happening
which
is
now
at
a
historic
photo
and
you're
not
going
to
go
horizontal,
so
the
note
was
the
request,
the
time
it
takes
to
replace
that
note
to
help
pass
industry
later
in.
But
what
happens
on
the
high
14
/
is
not
attended
at
network,
so
they
actually
screened
at
arena.
30S
works.
You
can
put
more
data
on
the
engine,
run
fewer
options,
so
this
was
nice
at
a
benchmark.
From
last
year
before
dat
nose
into
our
classrooms
for
the
15
node
SSD
quested
talk
to
matt
three
times
they
were
a.