►
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
Large Synoptic Survey Telescope: From Cloud Native to Dark Energy - Fabrice Jammes, LSST Corporation & French National Institute of Nuclear and Particle Physics (IN2P3)
Relying on the largest camera sensor ever designed, the Large Synoptic Survey Telescope (LSST) will revolutionize astronomy. The LSST data processing pipeline requires computational power of thousands of processors and several petabytes of data storage capacity per year. Celestial objects physical properties are stored in a database which will include trillions of entries. With a volume of many petabytes, this catalog will play a major role in research from identifying asteroids to understanding dark energy and matter. To meet those needs, a highly distributed software called Qserv, is being developed by an international team of engineers from USA (Stanford University) and France (IN2P3) Fabrice Jammes (IN2P3) will present the Qserv architecture, the challenges it must address, and how this disrupting database efficiently leverages Kubernetes.
https://sched.co/MPft
A
All
right,
okay,
hello,
good
evening,
everyone
our
good
afternoon,
so
my
name
is
su.
Goo
I
am
the
track
host
for
this
room
today,
I
am
excited
to
introduce
Fabrice
Zhang
and
who
is
a
scientist
and
he's
going
to
talk
about
how
science
and
containers
and
kubernetes
come
together?
How
to
containerize
science
projects
and
without
further
ado,
is
Fabrice.
B
Thank
you
so
much
hello,
so
nice
to
meet
you.
So
yes,
I'm
a
booth
and
I
work
as
a
seasoned
mean
and
so
far
developer
or
so
on.
The
Russian
optics
of
a
telescope
I
will
explain
you.
What
is
this
telescope,
but
the
goal
you
want
to
achieve
and
I
will
explain
you
also,
or
we
use
student
achieve
to
put
all
the
software
stack
to
manage
all
the
stuff
on
stack.
So
first
we
describe
the
telescope.
Okay,
then
I
will
talk
about
chooser,
which
is
the
database.
B
So
look
the
telescope.
It's
a
first
project
in
the
US
with
a
budget,
it's
1
billion
dollar
budget
and
then
it's
a
very
large
telescope.
You
can
see
the
CCD
so
compare
it
to
the
CCD
of
your
phone.
You
see
that
it's
much
bigger
and
for
your
information
when
pixel
is
a
it
will
get
1
football
soccer
on
the
moon.
B
B
So
you
see
the
telescope
gets
the
raw
data
with
the
CCD,
which
is
advanced
silly
silicon
component,
which
is
difficult
to
build
and
then
the
data
we
we
send
it
to
data
center.
So
we
have
a
data
center
in
France
in
yo
and
the
data
center
in
Urbana
Champaign
University
in
Chicago,
and
so
the
data
goes
food
to
from
Chile
to
to
the
data
center
telescope.
In
Chile
we
build
a
special
gigabit
network
connection
or
so
with
Chile,
and
then
we
process
to
image
in
the
data
center.
B
Okay,
then
we
build
the
star
and
galaxy
catalog,
which
is
the
main
component
for
science
catalog,
for
example,
the
first
one
was
built
by
Ptolemy
a
long
time
ago
in
Egypt
and
then
a
French.
Also
at
the
18th
century
Monsieur
we
build
a
famous
catalog
of
110
galaxies
and
stars.
Okay,
so
catalog
the
list
of
galaxies
and
star
with
their
physical
properties.
B
That's
what
we
extract
from
the
image
during
the
data
processing
of
the
image,
so
we
will
get
a
40
billion
galaxies
and
star
catalog,
which
is
the
largest
thing,
the
largest
size
we
have
for
all
telescope
vicia
for,
but
if
the
next
telescope
will
be
even
bigger,
it's
telescope,
you
get
a
new
new
scale.
So
it's
like
you
know
when
you
do
chemistry,
you
need
a
lot
of
of
people
to
know.
Is
the
chemistry
is
working,
the
medicine
is
working,
then
you
need
yeah.
B
So
is
the
data
so
the
image?
Okay,
it
could
be
a
t-38
petabyte
of
persisted
image.
We
will
back
up
okay
and
we
will
generate
as
an
exabyte
of
image,
because
you
know
we
need
to
do
Co
edition,
okay,
so
in
order
to
discover
stars
or
galaxies
which
are
not
on
any
image
week,
own
edition
so
image,
so
we
coalition
them
if
there
are
the
same
version
of
the
sky.
So
this
can
use
the
noise
okay
and
you
can
also
discover
new
object,
but
no
image.
B
B
For
example,
if
a
galaxy
you
say
antoinette
galaxy,
it
would
be
one
line
in
the
object
table
and
if
you
have
thousand
thousand
photo
of
the
galaxy
you
get
it
on
the
phazon
image,
you
will
have
nine
trillion.
Detection.
Excuse
me
one
thousand
line
the
detection
table.
Okay,
so
it's
large
data,
but
the
relational
model
is
relational.
Database
is
quite
simple
because
all
the
telescope
I've
been
basically
the
this
relational
model.
Doesn't
change
form
on
telescope
to
another
and
telescope
exist
since
a
long
time?
B
Okay,
we
need
also
to
keep
the
parlance
of
the
data,
so
the
provenance
which
are
going
then
was
used
to
to
discover
the
star
in
the
image
of
the
galaxy
and
also
which
computer
were
used,
which
compiler
we
use.
So
we
need
to
track
the
generation
of
the
data
from
the
photon
which
go
to
the
CCD
which
to
which
it's
the
CCD
to
the
bit,
which
is
taught
in
the
computer.
So
we
had
all
this
metadata
to
keep
so
now.
B
So
this
database
is
built
at
stanford
university,
slack,
the
linear
accelerator
Santa,
so
by
10:00
engine,
and
also
by
myself,
which
is
in
France,
because
France
collaborate
I
five
percent
of
the
budget
to
this
to
this
project
and
then
also
we
have
a
season
means
in
the
large
data
center.
So
in
CSA
it
twelve
thousand
odd
data
center
I
mean
to
Petri.
So
one
thousand
note
data
center
in
France
and
so
yeah.
We
have
of
development
platform
because
to
develop
and
test
a
small
scale.
B
We
need
at
least
50
nodes,
because
you
see
we
will
have
to
be
500
or
1000
odd
on
production.
So
we
need
a
cluster
to
test
okay.
So,
yes,
we
are
native
to
bonitas.
We
are
working
on
pepper
in
production
and
I
mean
public
collection.
So
also
we
do
monitoring.
We
are
preparing
the
prediction,
because
the
telescope
will
all
run
in
2022,
okay,.
B
So
in
the
team
of
that
access
it
has
a
lot
of
tasks.
Okay,
so
managing
the
image
manages
the
metadata
getting
the
alert
or
so
because
interest
in
astronomy.
When
you
find
something
like
a
supernova
or
transient
object
in
the
sky,
we
don't
have
a
definition,
so
we
will
when
they
detect
an
object
and
alert
and
also
telescope.
She
will
like
telescope
shine
space.
They
will
focus
on
this
object
to
get
its
property.
B
So
it's
a
collaboration
between
the
telescope,
so
yeah.
We
would
work
about
all
of
this.
Okay,
the
main
project
is
to
services
a
distributed.
The
database
for
supernova
tests
a
bit
the
same
concept.
So
look
we
speed
the
sky
in
twenty
thousand
part.
We
call
each
chunk
or
shard,
and
then
we
distribute
the
child
on
alone,
odd,
same
Isaac,
Maha,
DB
database,
or
so,
if
you
have
twenty
thousand
shot
and
one
thousand
Lord,
you
will
have
twenty
thousand
twenty
shard
on
each
node.
B
If
you
have
thought,
petabyte
of
data
will
have
five
terabyte
on
each
node,
so
five
terabyte
and
twenty
database,
one
Messick,
will
not
come
under
lift
that
can
manage
index
and
stuff
like
that.
So
we
distribute
this
and
then
the
user
has
a
magical
client.
We
launch
a
query
which
is
mass
equal
client.
It
go
to
a
C++
code,
okay,
which
is
a
custom-made
C++
code,
and
then
we
have
a
passer
we
analyze.
B
The
query,
for
example,
is
the
cuy
is
a
select
cone
star
from
objective
one
to
count
of
the
object,
a
very
simple
query.
Then:
we
understand
the
semantics
of
the
query.
We
rewrite
20,000
Chun
Kuen
for
all
the
chunk
so
which
comes
the
object
in
each
chunk.
We
distribute
it
with
the
middle
world,
which
is
called
extra
DS.
The
developer
is
embedded
in
the
team,
actually
its
manager,
so
petabyte
of
data
at
the
production
services,
while
working
Rob
is
middle
one,
and
then
we
distribute
this.
We
have
20
fathoms
head.
B
We
wait
for
this
20,000
count.
The
result
goes
back
to
what
we
call
a
temporary
result
table
on
the
master
we
have
220
in
this
table.
Who
has
a
20,000
count?
We
perform
an
aggregation
query
which
some
are
the
count
and
we
result
returns
the
result,
the
user
okay.
So
this
is
the
for
a
simple
example,
but
it
work
also
for
John
a
more
complex
query.
We
do
not
support
all
the
sequel,
but
we
see
we
try
to
support
all
the
sequel
that
are
required
by
astronomer.
B
So
a
visa
of
test,
wonderful
test
or
development
platform
see
it's
in
leo,
so
this
is
too
much
it's
15
odd.
It
was
offered
by
dell
five
five
years
ago,
the
list
of
Ponzo
and
then
yeah.
Now
it's
a
bit
older
or
odd
way,
but
we
perform
test
and
we
have
loaded
250
terabyte
of
data
inside
this
platform
and
we
one
performance
test.
It's
a
synthetized
data
because
real
data
set,
we
reached
a
big
galaxy.
It
doesn't
exist
at
the
scale,
so
we
we
need
to
generate
synthetized
data.
B
We
also
have
the
same
the
same
equivalent
infrastructure,
which
is
in
chicago
in
banner
Champaign,
and
you
know
we
load
a
real
data
from
the
wife's
telescope,
which
is
a
recent
telescope,
and
you
know
we
have
25
to
by
data
set.
We
will
data
and
astronomer
and
cosmologists.
They
are
launching
query
ago
against
this
data,
and
so
they
can
do
science,
the
first
cluster,
it's
a
leap
of
performance.
Nobody,
no
scientist,
use
that
because
you
can
do
science,
ok,
so
this
is
the
test
platform.
This
man
is
the
manager
of
the
world
project.
B
B
So
it
was
very
difficult
to
install
the
code
which
the
project
started
ten
years
ago
and
then
at
first
we
do
not
have
container
with
not
have
cabana
test.
We
have
to
reinstall
zoo
because
all
the
or
the
codon
sometime
we
have
200
nodes
from
Testim
200
node.
So
we
have
to
again
start
the
code
by
and
we
are
custom
procedure
and
honestly
it
was
one
more
first
one
more
work.
Each
time
it
was
very
long
to
rebuild
the
code
developer.
B
It
was
difficult
for
them
to
upgrade
the
code
because
to
push
everybody,
there
was
not
a
real
Porsche.
Do
everybody
was
time
to
act?
Okay,
so
it
was
difficult
to
go
to
one
or
last
res
kept
still,
and
it
only
was.
We
only
run
once
it
once
a
year,
okay,
but
then
we
go
to
cognitive
and
container,
because
it
brings
lots
of
flexibility
and
applying
the
software
and
well
in
development.
So
we
need
to
deploy
the
software
a
lot,
maybe
several
time
a
day.
B
So
what
we
do
the
developer
now
on
this
vocation,
if
you
push
the
code
to
the
github
repository,
the
code
is
open
source.
You
can
get
it
on
de
turbine,
the
LSS
teach
yourself
repository
and
then
push
the
code.
The
code
so
trigger
a
travesty
I
build
Travis,
it
build
the
containers,
it
installed,
mini
cube.
It's
an
occlusive
trust.
There
is
a
small
amount
of
data,
it's
checked
that
the
crew
are
working
and
if
the
query
are
working,
you
push
a
container
to
the
docker
hub.
B
Okay,
so
you
are
developer,
you
push,
you
get
a
coffee,
you
come
back.
If
you
say
Travis
is
green.
You
can
use
the
container
which
is
built
by
Travis.
There
is
the
version.
Is
the
ash
of
the
community?
It's
easy
to
use,
then
the
developer?
What
you
can
do
you
can
use
previously
OpenStack.
We
have
OpenStack,
but
the
engineer
and
left
for
private
company.
So
we
don't
have
no
more
OpenStack
maintain
in
my
lab,
but
now
we
use
g-cloud.
B
We
are
working
closely
with
you
with
Google
and
then
you
can
spawn
a
cluster
okay,
OpenStack
cluster
in
the
cloud
and
then
you
can
you
go
to
the
you
used
to
control
lawn
check
yourself,
dr.
script
and
you
you
will
spawn.
You
will
spawn
you
you
container
on
the
cluster
okay,
so
you
see
it.
You
have
two
action.
First,
you
push
your
card
so
gone,
you
start
the
cluster
and
you
launch
a
container
or
you
can
use
an
existing
cluster.
B
Okay,
so
yeah.
We
have
done
this
test,
so
we
are
running
it
on
OpenStack.
For
two
years
now
we
are
using
Google
because
it's
more
user
friendly,
no
maintenance
and
it
ship,
because
you
have
80%
discount
on
the
node,
because
you
use
Google
can't
even
show
it
to
you
all
day
for
between
oh
the
custard
Google.
So
it's
okay!
When
you
use
it,
if
you
don't
use
it,
you
don't
pay
and
then
at
Google
also
we
have
loaded
a
35
terabyte
data
set.
B
So
the
wide
data
set
I
was
talking
to
you
before
to
compare
performance
between
album
metal
platform
and
the
GK
okay,
and
we
were
running
also
experiment
on
local,
safe
with
OpenStack.
But
the
engineer
has
left
so
OpenStack.
We
have
no
mental,
no,
no!
So
we
appear
at
the
same
time,
so
we
switch
to
Google.
B
So
now,
if
you
are
in
your
data
center,
so
on
the
obey
metal
platform,
so
in
your
and
chick
and
Chicago,
then
the
user,
you
can
developer,
you
can
SSH
to
the
platform
and
the
push
is
even
the
same,
install
script
and
on
the
the
cloud
platform.
Okay
and
then
this
will
install
to
server
so
on
the
BAM
metal
with
a
very
same
install
script.
Okay,
so
it's
easy
to
easy
to
install
anywhere
the
record
with
the
same
Porsha.
Do
that's
what
we
Benitez
to
us,
whether
of
the
main
feature,
because
you
see
it.
B
B
If
you
go
to
the
CI
or
so
so
when
the
developer
push
the
card,
what
does
it
do?
12
is
so
it
build
the
container,
so
it
checks
build.
Is
it
working
it
tugs,
a
container
with
the
ash
of
the
commit?
Then
we
start
mini
chewed.
Okay
inside
12
is
pretty
easy
to
do
that
lots
of
documentation
and
you
internet.
Then
you
start
the
QCF
pod,
which
is
a
cluster
inside
mini
tube.
B
B
So
that's
what
good
native
bring
to
us?
Okay!
So
this
is
the
detailer
for
pod
and
stage
full
set
and
replica
set.
We
have
you
know
you
know
database,
we
have
multiple
component.
I
won't
go
into
the
detail,
but
what
is
big
to
us
here?
We
have
automatic
scaling,
so
we
were
at
Google
on
14
odd
and
we
wanted
to
test
on
15
odd,
to
have
the
very
same
clusters
and
on
the
data
center
and
yeah
to
second,
we
have
10
more
node
and
we
have
the
QSF
running
on
it.
B
So
it
was
pretty
impressive
compared
to
what
took
using
the
work
you
need
to
do
when
you
want
to
scale
inside
you
own
that
a
center,
so
first
thanks
to
Google
and
also
to
kubernetes.
We
have
a
sub
continuous
scheduling
we
are.
Are
you
drivin
IOPS
our
limit?
So
we
need
to
schedule
a
pod
aside:
each
volume,
each
chunk
of
the
sky
and
then
kubernetes.
It
can
do
this
very
well.
B
Okay,
you
just
have
to
to
put
it
in
the
configuration
of
your
pod
and
you
pod
get
aside
you
your
data,
see,
and
then
we
use
also
to
link,
because
when
you
have
15
odd,
you
know
some.
There
is
always
one
node,
which
is
broken,
which
is
not
feeling
well
and
when
you
skip
the
module
scale.
So
more,
you
have
problem,
and
this
this
is
useful
to
be
start
with
your
container.
Okay.
This
is
pretty
useful,
but
I
say
my
my
colleague.
B
You
don't
want
also
it
to
restart
because
it
could
I
bugged
so
in
production,
if
it
restart
with
you
the
time
one
time
on
six
months.
It's
not
a
big
deal,
but
if
you
saves
your
night,
you
do
not
have
to
remain
tense
to
in
the
night.
It's
okay,
but
yeah,
you
don't
want
it
to
restart.
We
want
the
software
to
work,
of
course,
of
course,
but
yeah
it
can
be
an
interesting
picture.
B
If
you
don't
use
it
a
lot,
then
we
have
also
continuous
deployment
as
I
show
to
you
so
manual,
but
we
are
going
towards
something
which
is
more
automated.
We
have
volume
management,
which
is
interesting
because
yeah
we
have
stateful
set
with
persistent
volume,
claim
template.
I
won't
go
inside
the
detail,
but
kubernetes
manager
just
wedge
and
when
you
move
from,
we
have
local
disk
on
a
cluster
and
we
go
to
Google
to
use
Google
persistent
disk,
which
is
San,
and
when
we
move
you
do
not
have
to
change
the
script.
B
B
Also,
the
monitoring
is
quite
easy.
We
will
try
to
deploy
on
now.
We
are
dependent
on
the
monitoring
of
the
data
center.
Is
data
center
ISIL,
as
is
unconstrained,
and
so
we
have
to
create,
will
fit
between
this
constraint
and
it's
difficult,
but
maybe
with
parameters
we
will
try
to.
We
will
provide
the
standard
stack
of
monitoring,
and
this
is
great
because,
if
you
install
to
a
new
customer
new,
you
can
also
deploy
the
monitoring
incident.
So
this
is
interesting
or
at
least
deploy
an
app
I
for
money.
B
I'll
check
also
is
interesting
because
you
have,
we
have
lots
of
services,
see
lots
of
puzzles
which
are
burning
and
sorry
so
always
but
process,
which
is
not
working
very
well.
It's
normal,
a
large
scale,
and
so
you
can
embed
the
pod
inside
the
pod
before
you
were
using,
for
example,
nodules,
and
you
have
to
check
the
measures,
instantaneous
maintain
Dodgers,
and
then
you
can
now
you
can
embed
the
probe
inside
the
pod,
and
then
you
know
what
is
working
or
not
working.
B
Only
by
asking
your
dns
you
see,
because
when
the
pod
is
ok
between
register
your
pod
inside
the
DNS
or
all
you
have
to
do.
Is
you
run
a
ping
on
the
name
of
the
pod,
for
example?
And
then
you
know
it's
ready,
so
you
know
to
check
if
everything
is
working,
you
have
the
clip
control,
so
it
it's
like.
If
you're
on
the
top
on
your
laptop,
you
see
all
the
Photoshop
when
you're
close
to
cluster,
but
in
one
place
and
also
you
can
use
the
DNS,
and
you
know
it's
funny.
B
So
with
kubernetes,
you
have
something
they
call
admission
control
and
if
you
do
nothing,
you
were
the
one
with
no
security
but
with
the
same
instant
script
you
can
against
well
configured
kubernetes
cluster
inject
security
and
the
pod
will
inject
GML
inside
the
pod
for
technical
details
and
the
pod
will
run
with
low
the
low
privilege,
and
so
the
developer
does.
The
same
stuff
deployed
on
the
development
platform
is
would
of
the
prediction
platform.
The
pod
is
very
secure
and
cubanÃa
tests
can
do
that.
B
But,
of
course
the
Cuban
is
a
Misrata
active
manager,
but
only
one
type.
The
word
recall,
put
security
policy
same
thing
for
the
network.
You
can
define
you
firewall,
the
developer
can
define
firewall
and
words.
We
want
visit
avid
access
to
the
machine
which
kubernetes
is
that
what
the
called
Network
policy,
so
you
can
bring
all
of
that
everywhere,
so,
but
the
killer
of
material.
We
found
that
possibility,
so
we
were
able
to
move
the
application
everywhere
with
no
work
and
before
we
have
moved
the
application,
we
were
left
with
a
lot
of
work.
B
That
was
my.
My
job
did
in
fact,
and
so
it's
time
there
was
to
install
the
application
and
a
new
cluster.
A
new
data
center
I
have
two
more
I
said.
My
management
I
want
to
only
focus
on
the
installation
and
it
was
red
alarm
and
we
were
installing
and
debugging,
and
now
last
time
we
install
it
was
an
GAE
because
Google's
of
orders
of
combustion
for
the
$100,000
to
test
we
have
collaborating
with
with
Google
and
gke.
B
B
So
now
to
finish,
I
will
make
a
comparison
between
compromise.
What
are
the
advantage
of
optimized
versus
public
cloud?
What
are
the
advantage
of
public
cloud
because
I've
tested
both
and
know
I
have
some
some
ID
of
that.
So
what
we
are
doing,
we
are
different
model,
so
the
first
model
was
running
or
the
stack
at
all
in
blue,
the
software,
the
Cuban
a
test
and
the
infrastructure,
and
then
we
Google
we
test
when
you
choose
set
up
on
top
Google
key
Benitez
in
Zion.
B
So
I
do
not
have
the
wizard
for
the
bigquery.
But
if
you
check
on
the
internet
or
host
Amazon
and
LSST-
and
you
will
check
here
as
there
is
a
blog
with
your
definition,
so
I
will
explain
what
is
GK
briefly,
but
all
I
see
GK
as
the
user,
but
I
think
the
same
thing
with
Amazon
or
Asia.
What
I
said
true
for
any
cloud
provider.
So
you
see
the
Gaffa
the
deployed
container
a
trash
can,
and
so
the
patent
is
a
knowledge
inside
kubernetes
and
no
and
your
first.
B
We
can
use
Q
Benitez
and
the
benefit
of
that
knowledge.
But
yeah.
We
have
to
install
Cubana
test
at
home,
okay
and
that
this
is
a
bit
difficult,
because
that's
all
the
thing
to
do
if
you
want
to
manage
given
a
test
at
home,
so
we
we
do
that
also
because
you
have
both
configuration
I
think
it's
very
interesting
job
to
manage
accumulated
but
yeah.
If
you
do
not
work,
it's
good
to
have
also
some
support
from
expect.
When
you
are
alone
at
home,
it
might
be
difficult.
So,
okay!
B
B
B
It's
easy
to
setup
development
cluster
and
if
you
data
is
not
private,
you
can
spawn
the
cluster
and
you
don't
have
to
manage
about
security,
because
you've
found
the
cluster
to
any
developer
or
any
tester
and
Instagram
will
work
and
Google.
Also
they
have
some
nice
operator,
a
feature
which
are
very
interesting
and
that
you
don't
see
it
will
be
difficult
to
implement
it
in
your
own
cluster,
like
dynamic,
node
provisioning.
So
if
the
cluster
is
too
low,
you
can
create
new
node
on
the
fly
so
also
dynamic,
storage
provisioning.
So
they
come
came
disk.
B
You
start
your
application.
Your
applications
require
storage
and
the
storage
schematically.
So
it's
an
interesting
course
that
we
found
it
very
expensive,
well
100k
for
own
one
more
test
for
1%
with
what
percent
of
the
data,
so
it's
very
expensive
and
I
think
we
can't
afford
using
Google.
It's
easy
also
to
get
stuck
with
proprietary
picture,
because
you
see,
for
example,
Google
and
Gino.
The
don't
sees
the
very
difference
between
what
provides
GK
and
the
pure
kubernetes,
and
so
they
can
tell
you
use
wish
feature
is
very
interesting,
but
it's
only
a
Google
feature.
B
So
when
you
go
back
at
home,
you
don't
have
the
feature,
so
you
have
to
change
your
script
to
go
back
at
all.
So
that's
why
it's
good
to
learn
Cubana
test
at
all
and
know
how
it
work.
So
when
you
go
to
the
vendor,
you
can't
be
used
only
the
standard
feature
and
not
the
other
one.
What
to
avoid
looking
and
so
also
it
I
can
it
is
internal,
so
you
don't
know
at
all.
B
Oh
it's
working,
and
so
sometimes,
if
you
want
to
do
performance
for
example
or
fine-tuning,
you
don't
know
how
to
do.
Ok,
it's
a
black
box
and
also
we
found
that
it
was
running
slowest
and
bad
metal
on
25%,
because
the
tip
normal
I
think
because
they
have
son,
team
of
storage
and
earth.
We
have
local
disk,
but
you
see
Google
is
very
flexible.
You
can
start
a
lot
of
not
very
easily
and
fine-tune
it
from
the
infrastructure
side.
B
But
if
you
want
like
local
disk,
this
is
not
possible,
so
it's
flexible
in
one
way,
but
not
in
another
way.
So
if
you
have
you
do
cluster
at
all,
you
will
have
flexibility
because
it's
a
do-it-yourself
cluster,
so
you
will
build
it
on
your
own,
so
you
can
fine
tune
here.
For
example,
we
can
use
local
disk,
ok,
but
it
will
require
still
engine.
It's
the
interesting
job
because,
like
you
Cuba
the
Installer
is
pretty
nice,
to
least
nice
or
so
so
you
don't
think,
and
it's
interesting.
B
But
for
example,
we
are
the
security
issue
in
January,
so
we
have
to
upgrade
Benitez
we're
on
some
2s
and
the
minor
upgrade
wasn't
working
because
some
to
SV
provides
a
package
with
it
wasn't
working.
So
we
have
to
upgrade
to
measure
my
job
great
1.14.
We
switch
to
call
their
nests
then
coordinates
it
was.
The
cache
was
not
working
well,
so
we
fight
to
intended
to
to
fix
the
cache
ever
so
you
see
it's
too
weak
a
bit,
stressing
at
Google
no
problem.
It's
always
work.
B
Okay,
so
yeah
both
have
four
and
constant
and
then
also
because
we
have
local
disk,
you
can
find
two
new
infrastructure
and
in
the
one
confessed
and
the
private
cloud.
This
is
because
we
need
look
at
this.
Okay.
We
need,
we
don't
know,
but
you
it
was
stopped.
What
design
to
an
undefended
so
course
is
that
is
difficult
to
achieve
the
global
cost,
because
we
are
to
maintain
the
building.
You
have
to
do
a
conditioning.
You
have
the
maintenance,
you
have
to
pay
see
that
mean
you
have
to
change.
B
Otherwise
you
have
to
pay
maintenance
for
the
add
one,
so
it
difficult
to
retrieve
the
global
cost.
So
at
Google,
it's
very
interesting
to
know.
Oh,
it
goes
pursue
and
it
required
on
manpower
also
for
scepter
maintenance
and
also
other
upgrade
and
difficult
when
you
want
to
change
now.
So
our
our
is
a
bit
deprecated
after
five
year
and
we
need
to
change
a
50
nodes
and,
honestly,
you
see,
we
all
need
some
three
thousand
cabbage
add
and
find
it
and
get
get
the
computer.
B
You
have
a
change
of
computer,
so
it's
a
Google,
you
click
and
you
have
to
craft
an
instrument.
So
you
see
it's
different
and
also
it
difficult
to
release
the
cluster
from
scratch
in
cognitive
sometime,
it's
very
useful
to
rebuild
everything
from
scratch.
So
if
you
have
programmed,
you
don't
find
the
solution.
If
you
start
from
scratch-
and
you
know
your
describe
your
infrastructure
and
you
have,
what
is
running
is
what
you
are
discarded
yeah.
This
is
poor
at
the
kubernetes
level,
but
Bureau
of
kubernetes.
The
infrastructure
is
difficult
to
then
start
from
scratch.
B
B
C
B
B
So
we
have
people
from
web
pay
from
country
were
phasing
infrastructure
of
the
query
who
live
and
fast
and
other
people
if
you
drove
on
through,
but
anybody
in
the
world
can
get
the
fast
petascale
data
set
and
under
lead
for
free,
of
course,
but
the
cost
is
inside
the
infrastructure
yeah
most
of
the
data
set
for
astronomers.
They
are
always
free
and
there
is
a
tick
elite,
virtual
Observatory
aliens,
so
sorry
standard
since
a
long
time.
B
So
you
can
cross
match
data
from
also
to
observer
to
it,
because
it's
a
requirement
in
cosmology,
so
you
can
check
for
an
object
and
you
want
to
get
all
the
assembled
of
the
other
object
for
all
the
telescope.
So
they
have
expand
out
and
we
were
complaining
to
say.
But
the
problem
is
that
yeah,
you
need
one
thousand
machine
to
under
the
data.
That's
what
it's
complex
and
anybody
can
to
provide
access
to.
B
We
hope
that
girl
are
made
when
we
will
open
the
telescope.
People
say:
do
it
a
lot,
so
the
database,
when
we
I
just
first,
that
I
will
need
you
in
a
few
days
and
we
will
like
crude
provider
to
sponsor
us.
Maybe-
and
you
see,
provide
free
access
to
everybody
to
get
the
data
between
two
or
three
days,
maybe
not
with
not
with
the
relational
database
but
something
more
simpler,
but
before
they
can
get
access
and
then,
when
the
peak
is
finished,
then
or
we
will
go
with
regular
to
could
be,
but
yeah.