►
From YouTube: Eventual Consistency != Hopeful Consistency
Description
Eventual Consistency != Hopeful Consistency
Speaker: Christos Kalantzis, Engineering Manager, Netflix
This session will address Cassandra's tunable consistency model and cover how developers and companies should adopt a more Optimistic Software Design model.
A
Okay,
so
let's
get
started
so
I
am
on
the
intro
slide.
So
today
we
are
we're
going
to
be
talking
about
eventual
consistency
and
how
you
should
embrace
an
optimistic
design,
persistence
layer.
So
today
it's
going
to
be
more
of
a
high-level
scuffle
and
not
really
welcomed
into
Cassandra
internals,
although
the
eventual
consistency
discussion
and
will
center
around
Cassandra
a
lot
of
these
concepts
also
carry
over
to
other
distributed
systems
and
not
just
databases
next
slide,
so
Who
am
I.
A
B
A
So
let's
get
started
next
line,
so
Cassandra
replication
and
let's
do
a
Cassandra,
replicates
and
inconsistency,
read
that
cassandra
is
an
eventually
consistent
database.
That
means
your
data
will
get
there
eventually.
So
eventual
consistency
is
a
new
term
that
that's
been
corn
the
last
few
years,
basically,
and
it's
replicated,
even
even
in
your
old
master
slave
reputation
of
my
sequel.
That
was
an
eventually
consistent
model.
A
Not
just
you're
not
just
replicating
from
a
master
to
assume
you're
replicating
to
multiple
peers
at
the
same
time
and
they
are
replicating
to
you.
So
what
is
it
eventually?
Consistent?
I
mean
it
doesn't
mean
it.
Data
will
get
their
day
from
now.
It's
not
going
to
get
there
a
minute
for
more.
It's
not
even
it's
not
even
going
to
get.
There
second
come
in
most
cases
in
over
over
ninety-nine
percent
of
the
cases
your
data
is
being
replicated
in
milliseconds,
so
I'll
go
to
term
the
term.
It
sounds
kind
of
scary,
eventual
consistency.
A
A
B
A
But
Cassandra's
a
consistency
level
is
tunable,
so
you
can
tell
a
statement
to
only
return.
After
all,
the
nodes
have
gotten
have
gotten
the
data.
So
in
the
case
of
a
replication
factor
of
three,
so
every
right
you
want
it
to
be
the
three
different
locations.
You
can
say
a
only
return
when
it
all
comes
along.
Well,
that's
that's
a
no-no
in
distributed
computing.
You
can
do
it.
There's
it's
fine,
I'm
sure
there's
use
cases
and
we'll
all
make
sense,
but
you're
giving
up
a
lot
of
availability.
B
A
You're,
giving
up
a
lot
of
performance,
a
lot
of
people
have
looked
at
coram,
which
is
corn,
is
replication
factor
divided
by
2,
plus
1,
as
kind
of
a
sweet
spot
or
or
a
a
best
of
both
worlds.
If
you
wish
fine,
okay
I
mean
that
still
allows
you
to
use
of
the
nodes
in
the
Cassandra
cluster
and
still
in
still
achieve
core
them,
but
again
you're
waiting
for
more
than
one
mode
to
return.
A
The
statement
before
your
application
can
move
forward.
Again.
It's
a
man,
it's
not
a
no,
but
it's
not
the
ideal
situation
and
then
there's
replication,
factorable
warm
replication
factor
of
1
I.
Think
people
should
embrace
it
more,
especially
with
the
massive
amounts
of
signals
who
are
collecting
these
days.
You
know
forcing
an
application
to
wait
while
you
well
while
you're,
but
while
you're
waiting
for
your
database
to
write
to
more
than
one
node
seems
seems
a
little
excess
to
and
then
we'll
talk
more
about
them
a
little
so
back
in
the
early
2000s
crees.
A
B
A
A
So
so
we
came
up
with
these
architectures
and
so
that
applications
will
write
to
a
master
and
read
for
multiple
slaves,
the
next
slide,
but
the
slaves
could
lose
transactions.
It
was
a
feeling
that
that
statement
based
replication
and
I'll
use
my
sequel
as
an
example
that
statement
based
replication
was
flawless
and
you
can
never
lose
12.
All
that's
untrue.
I've
managed
many
among
my
sequel
quote-unquote
clusters
and
there
were
instances
where
the
slaves
lost,
who
didn't
didn't,
get
every
transactions
and
for
the
master.
A
Wanting
to
know
and
collect
every
piece
of
data
and
that
can
possibly
flow
through
their
system,
this
model
just
doesn't
cut
it.
You've
got
a
bottleneck.
The
bottleneck
is
that
one
master
collecting
all
the
rights.
You
want
a
distributed
database
that
that
can
share
the
responsibility
of
collecting
rights.
That's
where
a
distributed
database
like
Apache
Cassandra
excels,
because
any
node
in
the
cross
will
can
accept
right
and
and
as
a
matter
of
fact,
not
only
will
it
accept
right
it
will.
A
B
A
A
B
A
A
Facebook
is
a
huge,
my
sequel
user,
with
with
with
that
master
slave
model,
and
it's
working
for
them
and
they're,
starting
to
use
more
and
more
distributed
databases,
but
the
core
functionality
is
still
running
on
my
sequel:
I
stop
photos
zappos,
you
send
it
now
called
I
tale
Samantha.
These
are
all
companies
that,
although
are
currently
researching,
distributed,
databases
and
that
Apache
Cassandra
they
are
my
sequel
shops
who
are
using
that
master
slave
replication
and
and
again
you
can
do
it.
A
One
quote
said:
I
want
high
consistency.
In
my
reads:
rights
just
like
I
had
in
my
are
DBMS.
Oh
well,
I
I
just
showed
you
that
that
not
only
can
you
lose,
can
you
loop
it,
and
I
just
showed
that
not
only
is
the
master
in
the
slave,
not
necessarily
one
hundred
percent
consistent,
because
there
is
a
lag
from
between
from
when
the
master
the
master
transaction
gets
death
sent
to
the
slave.
A
Sometimes
also
statements
don't
get
there
and
I
remember
working
at
a
previous
company
where
then
we
would
write
to
a
master
and
we
on
the
sleigh.
But
then
we
had
to
introduce
asleep
between
the
right
and
read
because
again
it
wasn't.
It
wasn't
fully
consistent
between
that
master
in
that
slave
prez,
it
eventually
had
to
replicate
and
and
higher
higher
the
amount
of
right
front
foot
was
bigger.
That
sleep
had
to
be
so
I
mean
if
you're
going
to
scale
for
the
web.
A
If
any
of
this
audience
has
ever
worked
with
rails
rails
or
other
MVC
frameworks,
where
you
model
a
the
code
in
your
model,
is
then
translated
into
SQL
statements
to
two
then
run
on
your
are
DBMS
you'll
notice.
They
never
create
foreign
teams.
We've
been
turning
off
foreign
keys
for
years.
Why?
Because,
every
time
you
write
to
a
child
table
before
me
and
there's
a
or
include
bit
string
linking
it
to
a
parent
table,
it's
always.
It's
always
doing
three
needs
to
make
sure
that
and
that
the
parent
ID
exists
that
increases
latency.
A
So
people
dbas
have
been
turning
boring
things
up
for
years
now
to
try
and
eat
out
more
performance
out
of
their
database,
and
the
next
question
is:
can
I
trust
that
Cassandra
will
replicate
my
data
when
write
it
when
writing
at
CL
1?
Well,
okay.
So
let's
address
that
last
one
next
slide,
so
that
looks
didn't
experiment.
A
A
B
A
Center,
all
in
this
particular
toss,
all
the
records
were
read
successful,
so
I
mean
that
means
you
can
trust
next
slide.
So
next
slide
shows
a
graphical
representation
of
of
the
of
the
experiment,
so
you
have
48
nodes
in
each
data.
Center
I
was
sending
50,000
operations
per
second
on
each
side,
just
to
show
that
there's
load
on
on
the
cluster
and
in
on
the
right.
B
A
A
So
all
this
brings
us
to,
should
you
be
designing
optimistically
or
pessimistically
if
you
design,
pessimistically
and
assume
everything
should
get
there
and
I
should
cover
every
single
edge
case.
Well,
you're,
going
to
punish
ninety-nine
point
nine
percent
of
your
users
and
what
do
I
mean
by
ball?
Yes,
things
will
go
wrong,
let's
say
point
one
percent
of
statements,
the
boat
or
things
will
go
wrong
point
one
percent
of
the
time.
A
Do
you
really
want
to
punish
the
rest
of
your
customers
for
more
than
90
day
of
the
rest
of
your
ninety-nine
point,
nine
percent
of
your
customers.
If
you
increase
the
consistency
layer,
you're
increasing
your
latency
and
you're
diminishing
that
user
experience,
which
today
is
the
most
important
and
I'll,
give
you
an
example.
I
use
this
example.
Quite
often,
let's
say
you're
on
the
netflix
side
and
you
just
added
a
movie
to
your
list
or
the
interesting
tube
it
used
to
be
called
in
three
cubes
called
my
list
more
ninety-nine
percent
of
the
time.
B
B
A
Percent
and
I'm
being
quite
conservative,
yeah
quite
liberal
with
this,
with
this
with
one
percent
of
the
time
you're
not
going
to
see
that
movie,
but
when
you
refresh
it
will
be,
it
is
a
customer
going
to
leave.
That
looks
because
of
that
I
would
argue
know,
and
and
and
and
so
we've
embraced,
eventual
consistency.
B
B
B
So
we
will,
we
will
share
these
slight.
All
I
think
we
just
got
a
question:
how
can
we
get
a
slide
to
your
sharing,
so
we
always
post
to
planet
Cassandra,
the
slides
and
the
recording
from
each
webinar,
and
we
do
that
within
about
24
hours
of
the
webinar
ending
so
check
back
tomorrow
and
the
slides
and
the
recording
will
be
there.
A
B
A
B
B
A
A
But
really
you're
just
you're,
just
penalizing
the
majority
of
your
customers
for
those
edge
cases
where
something
might
go
wrong
and
and
and
you're
just
diminishing
the
user.
Experience
next
slide
in
an
optimistic
design.
You're
trusting
your
day
of
school,
and
you
know
your
business
and
your
application
that
you're
always
asking
yourself.
Is
it
really
wrong
important
and
you
can
handle
edge
facing
sleeping
teens
and
sequins?
A
I
understand
a
lot
of
times,
engineers
work
as
consultants
and
they're
just
brought
in
to
solve
a
problem,
and
they
really
can't
be
as
invested
as
possible
and
find
that
happens,
but
but
really
the
architects
of
these
applications
really
need
to
be
invested,
not
just
in
the
technological
aspect
but
William
in
business,
and
so
I've
already
shown
that
you
can
trust
Apache
Cassandra
with
delivering
the
data.
At
the
end
again,
your
mileage
may
vary
depending
on
the
quality
of
the
network
between
data
centers
or
within
a
data
center.
A
A
My
inventory
system,
I
mean
I,
can't
sell
widgets
I
won't
have
I'll,
be
a
laughingstock,
well
I
believe
there's
two
examples
where
Amazon
in
outlook
I
usually
use
Amazon
as
an
example
for
this,
but
just
this
week
my
wife
bought
something
on
Oh
to
look
that
that
she
believed
was
available
for
sale,
ultrabooks,
oltre,
parole,
yet
eventually
contacted
and
said
pun
intended,
eventually
contacted
her
and
said:
oh
we're.
So
sorry
we
unfortunately
do
not
have
that
installed.
B
B
B
A
Through
credit
towards
my
next
purchase,
you
end
up
with
a
very
satisfied.
Well,
you
end
up
with
a
customer
that
is
not
upset.
The
experience
was
very
smooth
because,
because
of
because
of
a
very
fast
persistence
layer,
however,
things
did
go
wrong.
You
know,
and
my
wife
might
have
been
v,
1
million
customer
and
at
1
million
you
know
this
issue
happened.
A
They
handle
with
a
angled
it
right.
Let's
look
at
another
example,
my
favorite
one
I'm
in
banking.
Thanks,
you
know,
banks
need
to
be
consistent.
You
can't
not
be
consistent
in
a
bank.
Well,
banks
have
been
eventually
consistent
since
the
beginning
of
time.
The
banking
system
is
the
ultimate
is
the
ultimate
eventually
consistent
system
and,
as
you
see
here,
I've
got
a
quote
from
that
file.
Co-Founder
of
data
stacks,
who
says,
banks
make
a
lot
of
money
off
of
eventual
consistency
if
I
overdraft
they
charge
me
now.
B
A
A
Next
line
hurdles
face
well
it
other
than
the
technological
hurdles.
For
today
for
this
presentation,
the
hurdles
faced
by
rule
face
you
audience
who
are
convinced
of
the
merits
of
eventual
consistent,
see
you
will
face
hurdles,
and
some
of
that
is
no.
Your
engineering,
colleagues
or
maybe
stubborn
I,
mean
you
know
one
plus
one
equals
two
to
them,
not
eventually
too.
So
you
know
you're
going
to
have
to
convince
some
people
that
this
is
the
right
thing
to
do:
middle
management
Wow.
A
So
some
of
the
son
of
some
of
the
most
some
of
the
hugest
hurdles
I've
ever
faced
work.
These
middle
managers,
whose
mantra
is
nobody
ever
got
fired
for
wine?
You
know
fill
in
the
blank
IBM
Oracle,
whatever
look
I
mean
they
there.
If
that
is
the
way
they
are
evaluated,
is
uptime
uptime,
based
on
a
set
of
throughput
and
and
a
number
of
errors,
and
and
because
the
game
is
stacked
that
way
against
them.
A
Well,
they
see
the
adoption
of
eventual
consistency
and
a
new
technology
like
Apache
Cassandra
as
something
that
could
potentially
lower
their
evaluation.
Even
you
need
the
middle
management
needs
to
be
convinced.
Well,
it
is
in
your
best
interest
to
to
go
this
role
and
an
embrace
eventual
consistency,
and,
quite
honestly,
some
some
companies
are
just
going
to
have
to
change
the
way
they
they
evaluate
their
their
success
and
what
success
means
to
them.
A
You
need
to
engage
the
product
teams
to
implement
some
time
of
content.
Some
type
of
contingency
pause
like
like
I
mentioned
earlier.
You
you
have
to
care
about
your
product
and
I'm.
Talking
to
you
engineers,
you
have
to
care,
you
have
to
understand
the
business
and
if
you
don't
understand
the
business-
and
you
can't
explain
why
you
need
to
do
this
in
business
terms
to
the
product
teams,
then
that's
gonna
be
a
huge
fertile
to
convince
them
that
it's
okay
to
sometimes
sell
a
widget
that
you
don't
have
any
swamp
next
slide.
A
So
how
to
overcome
those
hurdles
for
the
other
engineers
prove
it
in
a
POC
show
them
that
you
know
you
can
throw
a
whole
bunch
of
data
to
Cassandra
and
it'll,
make
it
there
and
and
and
and
and
if
it
doesn't
well
what
percentage
of
it
won't
make.
It
show
them
the
benefits
of
you
to
expect
the
children
that,
if
you
do
it
a
quorum
or
at
all
that
that
your
throughput
will
go
up
significantly
as
you
lower
that
consistency
level.
A
I
believe
google
released
a
white
paper
back
in
january
about
a
super
consistent
distributed
database
and
they
were
able
to
to
achieve
a
throughput
of
the
hundreds
hundreds
of
transactions
per
second
and
in
the
conclusion
of
the
white
paper
world's.
We
came
to
super
high
consistency
in
a
distributed
system
and
expect
a
group-
oh
it's
just
the
fact,
and
and
lastly,
an
after
after
trying
to
convince
engineers
trying
to
convince
middle
management
and
and
and
the
product
team,
and
we
still
don't
get
it
well.
B
A
B
B
A
So
even.
A
And
it
talks
about
polyglot,
polyglot
persistence.
There
is
no
one
solution.
One
size
fits
all
in
the
persistence
in
the
persistence
world
and
again
I
get
I
go
back
to
the
intro
of
my
presentation,
and
you
know
there
are
going
to
be
use
cases
where
you
need
a
super
high
consistency
and,
in
that
case,
by
all
means
use
something
that
is
that
meets
your
needs.
Your
compliance
needs
your
technological
needs.
B
B
A
A
B
A
Then
you
can
read
it
and
weep
when
you
write
to
Cassandra
it
is
it's
not
a
sequential
writing
to
the
slaves
it
is
going
there.
It
is
going
to
all
nodes
of
the
replica
set.
At
the
same
time,
the
only
difference
is
at
a
lower
consistency
level.
The
coordinator
is
waiting
for
fewer
nodes
to
confirm
the
receipt
of
the
of
the
right
before
it
returns
the
statement
to
the
application
with
so
in
essence,
all
the
nodes
got
them.
B
A
Absolutely
so,
when
I
talk
data
centers
in
the
AWS
world
and
for
those
who
don't
middle
Netflix
and
runs
all
their
Cassandra
buses
in
AWS
across
zones
and
across
regions,
we
do
we
do
replicate
our
clusters
across
multiple
regions
and
and
it's
also
the
nature
of
our
of
our
business.
Where
you
know
yes,
so
the
data
will
take.
Let's
say:
50
milliseconds
to
replicate
from
us
is
to
to
EU
West,
however,
and
remember
that
a
user
writing
in
EU
wish
to
who's
at
the
edge
geographically
is
hitting
API
servers.
A
We
do
is
never
going
to
go
across
the
pond
to
read
his
or
her
record,
so
I'm
still
within
a
the
active
records
for
that
user
are
still
within
the
same
region.
Now,
if
that
user
was
to
do
something
on
a
plane,
the
travel
across
the
Atlantic
and
and
then
and
then
try
and
read,
let's
say
their
list.
The
movie
list
I
have
extremely.
A
Okay,
so
this
is
more
a
Cassandra,
informal,
scroll
stream,
so
data
data
first
gross
for
memory
in
some
mem
tables
and
then
periodically
and
I
won't
get
into
to
all
the
different
things
that
can
flush
data
from
from
a
mem
table
2
to
the
SS
tables.
However,
periodically
data
will
be
flushed
from
memory
to
this.
B
Dimitrios
a
couple
of
questions
here:
how
many
knows
you
run?
What's
the
data
size
of
a
single
node?
B
A
Yes,
we
we
currently
are
in
the
final
stages
of
our
grade
from
1172
datastax
enterprise
we
can
fly,
which
is
overly
viscous.
End
belong
dot-to-dot,
something
underneath.
So,
yes,
we
are
running
the
effects
enterprise
are.
We
have
2500
plus
nodes
in
production
across
about
a
hundred
Cassandra
pastas,
the
average.
A
Me
more
important.
Our
average
size
cluster
is
12
to
24
nodes
and
our
biggest
cluster
is
200
in
really
late
nodes
and
our
smallest
cluster
is
three
node
cluster
and
our
and
from
from
data
center
replication.
We
do
anything
from
one
single
data
center
or
EWS
region
to
as
many
as
for
AWS
regions
or
gives.
B
A
B
A
Let
me
add,
let
me
answer
the
first
one
at
any
any
one
time.
Cassandra
is
not
more
that
the
data
is
consistent.
Everything
is
best
effort
and
and
and
so
that
answers
that
part.
So
there's
no
like
master
state
somewhere.
A
keeping
track
of
of
records
are
consistent
and
available
in
every
replica
set.
The
Christian
can,
you
repeat,
part
1,.
A
A
What
there
is,
then
is
is
a
tool
called
repair
which
it
will
then
go
and
analyze
data
in
a
mode
and
who
it's
replicas
should
meet
and
see.
Hey
I've
got
this
rep.
You
have
that
I'm
oversimplifying
this.
Of
course,
a
given
record
eh,
hey
all
you
people
who
should
have
a
popular
model.
Yes,
it's
the
one.
I
have
the
latest
one.
No,
it's
not!
Oh
you
have
the
lake
swamp.
Then
it
will
copy
the
latest
one
to
all
the
three
members
of
the
replica.
A
B
Great
here
we
go,
we
have
a
specific
argument
in
our
business.
We
need
information
available
to
customers
as
close
to
real
time
as
possible.
Other
vendors
have
stated
this
can't
be
done
with
a
no
sequel
solution
due
to
eventual
consistency,
so
I'm
not
sure
that
what
maybe
we're
mixing
and
matching
real
time
versus
consistency,
but
maybe
you
can
talk
a
little
bit
about
that
and
then
the
other
thing
I
was
thinking
christos.
Is
it
worth
talking
about
lightweight
transactions
in
in
two
dot?
Zero?
For
example?
Oh.
A
B
A
A
A
Did
you
know
what
I'm
not
going
to
venture
to
I'm
probably
going
to
stop
at
state?
It's
just
a
database
like
anyone
else.
You
write
something
in
there.
Then
you
get
it
out
where
eventually,
eventual
consistency
comes
out
is
not
the
speed
at
which
inform
all
accessories
day.
Well,
it
is,
it
is
more
the
be
as
the
data
is
everywhere
and
replicates
across
your
cluster.
How
consistent
will
every
read
do
as
you
read
it
out.
A
B
A
A
Was
doing
it
back
then
so
are
we
so,
but
as
we
decided
to
go
to
the
cloud
we
needed
something
that
can
run
on
commodity
hardware
that
could
be
distributed
that
could
replicate
across
multiple
regions
and
and
being
a
Java
shop.
We
wanted
more
control
on
the
on
the
persistence
layer.
We
wanted
to
look
at
the
crowed.
We
didn't.
B
A
Something
with
the
million
bells
and
whistles.
We
wanted
something
simple
that
acted
as
a
persistence
layer
and
then
we
can
move
budget
higher
up
in
the
applications
fall,
but
the
real,
so
being
open
source
was
important
for
being
open.
Source
in
Java
was
great
because
more
pictures
of
java
actual
but
the
real
kicker
was
its
multi-directional
multi
data
center
reputation
that
that
was
and
tools
in
it
to
correct
itself
if
need
be,
run,
repairs
and
make
and
make
the
cluster
consistent.
A
B
B
Over
the
last
I'd
say
six
to
nine
months,
we
actually
see
companies
who
have
very
very
small
data,
sets
in
no
way
shape
or
form.
Could
you
say
they
are
big
data
at
all,
but
this
requirement
for
always
on
availability
across
multiple
data.
Centers
is
a
is
more
and
more
driving
force
for
why
we
see
people
transitioning
from
relational
databases
or.
B
A
A
Cassandra
performs
really
really
really
really
well
you're,
not
running
Cassandra
because
of
tes
tes
rated
transactions
per
second
you're
running
it
for
high
availability,
you're
running
it
for
scalability
you're
running
it
you're
running
it
because
you
can
embrace
the
eventual
consistency
and
have
and
have
that
huge
amount
of
liquid
coming
room.
Cassandra
will
scale
linearly.
A
We
have
shown
whether
there
are
articles
on
on
netflix
is
a
tech
blog
showing
how
Cassandra
scales
linearly.
So
it's
comparing
one
node
of
Cassandra
with
one
note
of
couchbase
or
three
node
updates
and
remote
for
Sandra,
really
you're.
So
comparing
apples
to
apples
at
this
one,
you're
you're
you're,
doing
yourself
a
disservice
with
your
pitiful
model,
because.
B
A
B
B
B
A
A
A
A
It
wasn't
like
SQL,
so
that's
why
other
other
other,
no
sequel,
databases
like
who
had
a
much
much
cleaner
API,
you
know
won
the
hearts
and
minds
of
developers
early
on.
So
that's
where
the
Apache
Cassandra
community
said
any
booze,
ralston
and
c
ql
cql
came
along
now.
You
can
create
tables,
khuri
tables
and
use
a
lot
of
the
SQL
syntax
apology,
you're
familiar
with
today,
with
Cassandra,
using
the
soup
to
a
local
place.
Now,
of
course,
they're
still
no
joy!
B
B
B
If
you
want
to
carry
on
your
Cassandra
training,
some
of
the
questions
we
got
today
are
from
you
know:
people
new
to
Cassandra.
We
do
have
a
free
online
training
course
available
and
the
link
is
on
your
screen
right
now
and
then
please
join
us
for
our
next
webinar,
which
is
building
apps
on
Python,
with
the
Cassandra
Python
driver,
Eddie
sassily,
also
a
great
proponent
of
the
Cassandra
community.
So
Christoph's
will
let
you
get
back
to
your
vacation.
Have
a
fan
tire.
A
No,
not
a
recruit
already
did
I
already
blood
that
may
14th.
If
you
are
in
the
Bay
Area
the
next
datastax
Cassandra
meetup
was
the
South
Bay
Cassandra
meetup
will
be
at
Netflix
and
Ruffo
will
be
talking
when
you
had
question
before
about
replication
across
multiple
AWS
regions.
Her
topic
is
going
to
be
how
we
achieved
active
active
across
two
three
four
data
centers
using
Apache
Cassandra.
So
for
those
in
the
area.
Please,
we
will
be
recording
it,
so
we
will
be
making
it
available
to
be
to
the
whole
community.
A
Also,
if
you
are
in
Montreal,
I
will
be
speaking
at
Tuesday,
the
29
at
the
RPM
center
on
guy
street
at
630pm.
It
won't
be
this
presentation.
It
will
be
a
very
specific
proudest.
Netflix
used
Cassandra,
a
kind
of
presentation
and
for
people
who
want
to
talk
more
tech
that
one
that
one
you
might
be
interesting
for
you
as
well,
and
that's
it.
That's
all
my
announcements.