►
Description
Speaker: Matthieu Nantern, Software Engineer at Xebia
Slides: http://www.slideshare.net/planetcassandra/c-summit-eu-2013-delivering-christmas-gifts-in-france-since-2012
Every year more and more people buy their Christmas gifts online and that gifts are delivered by the postal service of France "La Poste". At the end of the infrastructure a (not so) little MySQL was struggling for survival against that overwhelming load. Then, in 2011, MySQL hit its limit... Come hear the true story of La Poste switching its parcel management, PHP-based application, from MySQL to Cassandra in 3 weeks. You'll be taught about the details of the project constraints, how to use Cassandra from PHP, the migration plan, how to manage resilience testing, deploy your Cassandra with Puppet, and all the wonderful knowledge we accumulate through this project.
A
My
name
is
matching
of
them
and
I'm
a
software
engineer
at
zip.
Your
friends
I
like
to
build
software,
deploy
it
with
pipette,
principally
and
during
the
last
couple
of
years.
I've
worked
for
the
Postal
Service
of
friends,
it's
called
lab
bust
and
it's
there
that
I
discovered
Cassandra.
When
we
moved
from
MySQL
to
Cassandra.
A
Little
overview
of
what
we
will
see
today,
first,
the
context
of
our
application,
then
why
we
had
to
switch
our
alternatives
in
a
third
part.
I
will
detail
our
Cassandra
implementation
and,
finally,
our
we
deployed
our
solution
in
production
that
will
take
30
minutes
and
I
will
keep
10
minutes
at
the
end
for
Q&A
all
right.
Let's
go
for
the
context.
A
A
Parser
distribution
is
one
of
the
increasing
business
of
La
Poste,
thanks
to
the
growth
of
online
purchasing
at
the
end
of
the
year,
for
example,
because
everyone
is
buying
their
gift
for
Christmas,
the
Lord
of
our
application
is
doubled,
sure
that
almost
everyone
here
had
already
buy
things
online
influence.
It's
lapis
that
handles
a
substantial
part
of
the
delivery.
More
than
4
million
of
parcels
can
be
distributed
a
day
by
70,000
postman.
A
Yes,
Westminster
and
after
the
purchase
every
parcel
is
then
routed
to
the
to
the
final
customer,
bypassed
man
or
kept
15
days.
If
nobody
claims
it
after
15
days,
the
parcel
is
sent
back
to
D
to
the
original
sender.
Do
the
daily
workflow
of
a
Bassman
is
first
it
scans.
Every
parcel
that
he
needs
to
distribute
then,
before
going
on
to
a
prince,
is
list
of
parcels.
A
A
A
You
remember
the
step
chew
on
my
postman
store
when
the
bassman
prints
a
list
of
Foley's
parcels
that
operation
on
a
busy
day
of
November
2011
could
take
up
to
15
minutes
every
post
man
in
France.
We
are
waiting
15
minutes
before
starting
detrol
and
and,
as
you
can
see
on
the
graph
every
year,
the
busiest
day
is
thirty
percent
bigger
higher
than
the
previous
year.
A
That
was
because,
when
a
postman
twins
is
list,
we
need
to
modify
every
parcel
on
that
list
in
our
database.
That's
four
million
of
Reed
&
4
million
of
right
in
a
very
short
period
of
time,
and
that's
just
for
the
step
two
of
our
postman
straw.
At
this
time
we
are
the
mysql
server
with
24
thurs
disk
array,
but
that
was
not
enough
to
to
absorb
30
percent
of
growth
every
year.
A
A
A
A
That
was
the
easy
solution,
because
we
we
know
MySQL,
we
use
it
everywhere,
but
the
main
design
that
here
was
that
we
had
to
do
everything
by
hand
like
the
shouting,
for
example,
that
represent
a
lot
of
code
rewriting,
and
maybe
that
will
prove
insufficient
in
two
or
three
years
because
of
the
thirty
percent
of
there
is
a
solution,
and
that's
why
I'm
here
today
is
was
to
switch
to
another
system
Cassandra,
we
like
the
the
scalability.
If
we
need
to
serve
more
requests
in
the
following
year,
then
we
just
add
more
servers.
A
A
A
In
MySQL,
we
used
a
very
simple
structure
to
store
our
data,
one
column
for
40
ID,
another
one
for
the
data
in
XML
format
and
four
columns
for
data
on
which
we
searched,
and
we
did
the
same.
In
Cassandra.
We
just
made
a
single
change
in
our
data
structure
in
MySQL.
If
a
request
needs
to
data
as
secondary
indexes,
for
example,
or
creation,
debt
and
group
ID,
then
we
add
two
columns
in
Cassandra.
A
We
grouped
we
merge
those
two
columns
into
one
that
will
contain
group
ID
plus
creation
that
we've
done
that
for
performance
reason,
because
we
we
care
about
our
data.
We
use
a
replication
factor
of
three.
That
means
that
every
data
inserted
in
our
cluster
is
replicated
three
times
and
we
wait
the
response
of
two
servers
among
three
on
each
request:
that's
quorum.
That
provides
us
consistency
for
our
data
right,
because
the
software
processing
parcel
distribution
is
is
written
in
PHP.
We
had
to
curl
Cassandra
in
PHP
to
do
so.
A
Using
Cassandra
with
PHP
casa
is,
is
really
straightforward.
It's
really
easy
and
first
you
create
a
pool
that
will
connect
to
Cassandra
instances.
That's
the
second
parameter,
an
array
have
all
your
Cassandra
service,
then,
with
that
pool
you,
you
get
your
current
family
by
name
and
now
you
can
insert
a
row
with
multiple
columns
by
using
the
insert
function.
A
A
That
was
the
easy
part.
As
we
said
earlier,
we
needed
to
be
able
to
query
data
with
secondary
indexes
to
do
so
with
PHP
kasam.
You
need
first
to
create
an
index
expression,
that's
the
first
line
and
you
pass
to
that
function
and
index
name
the
name
of
your
column.
You
will
be
using
as
secondary
index.
A
Then
an
index
value,
it's
the
value
you
will
search
into
your
colon
and
finally,
an
operator
operator
can
be
equal,
greater
than
or
lower
than
or
equal
anything
you
want.
After
creating
one
or
more
index
expression,
we
use
them
to
build
something
that
is
hold
an
index
closed.
That's
the
second
line
at
least
one
index
expression
in
every
index
closed
as
two
to
use
an
equality
match,
and
then
you
can
query
cassandra
with
secondary
indexes.
A
A
The
TTL
TTL
is
for
jolliffe
and
it's
the
TTL
in
seconds
with
that
and
colon
1
and
cullen
two
will
get
the
same
teacher,
but
we
needed
more.
We
needed
a
way
to
to
set
different
TTL
for
each
column,
as
Cassandra
was
able
to
so
because
PHP
casa
is
an
open
source
software
I
added
support
for
TTL
by
Colin.
A
A
A
To
do
so,
we
used
a
pipette
to
provision
our
12
servers,
running
ubuntu
10
for
those
who
don't
know
Babette,
it's
an
automation,
software
that
helps
manage
your
infrastructure
from
provisioning
to
configuration.
We
put
that
you
can
easily
automate
repetitive
task
like
installing
cassandra.
For
example,
the
masterless
architecture
of
cassandra
and
the
community
debian
package
helped
us
to
deploy
all
servers
of
our
cluster
with
the
bet,
just
by
doing
that,
with
those
three
lines
per
pet
will
install
cassandra
on
all
server
of
our
cluster.
A
The
only
parameter
that
changes
in
a
naga
cluster
in
Cassandra
configuration
is
the
initial
token
to
set
the
initial
taken
with
Babette.
You
can
use
Ruby
inside
inside
template
to
calculate
the
initial
token
for
every
server
of
your
cluster,
and
so
with
a
formula
like
that,
you
can
install
easily
Cassandra.
A
A
Each
new
version
of
our
data
model
generates
a
file
called
Cassandra
migration
and
a
number,
and
we
put
all
those
files
into
a
dead
end
package,
and
we
tell
pipette
to
install
that
package,
always
with
the
three
same
lines
and
when
per
pet
installs
our
package.
The
package
will
first
queries
Cassandra
to
get
the
current
migration
number.
If
the
package
number
is
higher,
then
we
apply
our
remaining
migration
file
in
order,
for
example,
here
we
have
a
current
immigration
number
of
two
and
a
migration
number
of
four
into
our
package.
A
A
After
installing
our
cluster,
we
we
made
performance
testing
on
it.
We
all
like
to
play
with
our
new
toy.
So
during
that
test,
we
simulated
the
true
of
around
4000
postman
that
that,
while
scanning
continuously
passes
with
MySQL
our
reference,
that
was
messy
with
approximately
200
requests
per
second
and
the
test
that
lasted
two
and
half
hours.
A
A
A
That's
helped
for
the
CPU
consumption
and,
as
I
said
earlier,
instead
of
making
requests
with
three
or
four
secondary
indexes,
we
chose
to
merge
those
indexes
in
order
to
have
at
the
most
one
secondary
index,
for
every
call.
We've
done
that,
because
request
performance
was
not
good
enough
when
using
more
than
two
or
three
secondary
indexes
in
Cassandra.
A
Right
we,
which
owes
a
Cassandra
for
scalability
native
data
replication,
but
also
for
for
tolerance
capabilities
to
to
try
that
part.
We
we
run
the
same
test
as
in
performance
testing,
but
after
some
time
we
we
kill
one
or
more
cassandra
sorrow
the
results
realm
line
with
the
one
expected.
The
first
colin
is
our
reference.
A
All
servers
are
up
and
running,
losing
one
server.
That's
on
Cullen
too
we
kill
one
server
does
not
impact
the
overall
performance.
Norio
shows
up
losing
two
or
more
servers
among
three
consecutive
servers.
That's
current
three
and
four
implies
no
more
access
to
two
data
on
those
nodes,
because
we
we
use
quorum
on
each
request,
so
we
need
at
least
two
answers
for
our
cluster.
A
The
main
constraint
was
that
no
downtime
was
allowed
and
that
immigration
has
to
be
done
before
mid-september
2012.
When
the
load
start
to
write.
The
migration
plan
was
what's
important,
first,
rewrite
our
new
data
into
Cassandra.
If
we
need
to
read
data,
we
read
it
first
from
mysql
and
if
nothing
is
found,
then
we
thought
into
cassandra
after
15
days.
All
data
are
now
into
Cassandra.
A
As
I
said,
Cassandra
was
released
in
production
in
non
September
2012.
Just
before
the
start
of
of
end
of
year
rush
every
day
we
are
looking
at
up
Center
response
times.
Users,
feedback
and
everything
was
working
as
expected
response
times
were
good
users
were
happy
to
just
like
us
after
two
really
odd
years,
where
first
men
were
waiting
for
the
system
to
go
on
to
now,
the
system
was
wait
for
them
as
it's
expected
in
2013
and
follow
first
year.
No
work
is
expected
on
the
performance
side
of
parcel
distribution
in
France
thanks
to
Cassandra.
A
A
A
A
A
Yes,
we
we
have
looked
at
each
base,
but
we
didn't,
we
didn't
choose
it
because
it
was
more
art
to
set
up
and
to
to
operate
the
the
the
operation
chilled.
The
operational
team
of
lapis
is
a
five
or
six
people,
persons,
so
that
that
that
was
not
an
option
each
base
we
also
looked
at
MongoDB,
but
because
I
were
dated,
don't
fit
in
memory.
A
A
A
A
A
We
we
deployed
our
cluster
in
September
2012
and
we
didn't
have
any
issue
until
februari
2013,
where
the
some
server
we're
not
responding
anymore.
So
we
just
add
two
to
reboot
them.
It
was
probably
a
memory
leak,
but
that
happened
just
once
on
februari
2013
and
now
every
month
we
have
a
short
period
of
time,
15
minutes
and
we
restart
we
reboot
our
servers.