►
From YouTube: Creando cargas de trabajo flexibles gracias al autoescalado - Jorge Turrado, Docplanner Tech
Description
Kubernetes es una pieza de tecnología que nos aporta una gran facilidad para tener cargas de trabajo flexibles y adaptables según las necesidades. Esto se puede comprobar en como sin que hagamos nada nos ofrece posibilidad de autoescalar usando la CPU y la memoria que están consumiendo nuestros pods.
Normalmente esto es suficiente y funciona bien, pero en entornos de microservicios, donde cada vez más la comunicación es asíncrona, o en escenarios con estado, como pueden ser las bases de datos, kubernetes se queda un poco corto...
¿Y si me quedo sin nodos?
Durante la charla veremos como podemos extender las capacidades de Kubernetes usando herramientas como KEDA (Kubernetes-based Event Driven Autoscaling), Vertical Pod Autoscaler o el node autoescaler de los proveedores cloud.
¿Te lo vas a perder?
Contacta a Jorge en:
- https://www.linkedin.com/in/jorge-turrado-ferrero-64b70371
- https://twitter.com/JorgeTurrado
A
A
A
A
B
B
Ok.
Well,
let's
go
to
the
mess,
as
all
of
you
who
are
here
will
know
we
are
going
to
talk
about
how
we
can
use
what
vernet
is
so
that
our
workloads,
well,
it
turns
on
in
a
certain
way,
flexible
and
adapt
to
the
needs.
We
can
save
costs.
We
will
delve
into
it
a
bit.
First
of
all,
who
am
I,
I
am
Jorge,
Turrado
I
am
and
will
be
in
two
Länder.
Take
I
am
Microsoft.
Developer,
Technologies
and
I
am
keep
in
stay.
B
B
It
is
a
small
agenda
of
what
we
are
going
to
see.
Today
we
are
going
to
talk
about
that
scaled
car.
We
are
going
to
see
different
scaled
car
models.
I
am
aware
that
not
all
of
them
are
there
I
think
that
if
I
remember
correctly,
there
has
already
been
a
talk
by
the
Hannity
project,
probably
being
able
see
it,
but
you
have
probably
considered
that
it
even
allows
auto-
scaling.
B
B
Scaling
in
google
net
is
auto.
Scaling
simply
means
that
our
workloads
will
be
able
to
adapt
to
the
needs
they
will
be
able
to
adapt
both
in
number
of
instances.
That
would
be
horizontal
scaling
or
hp
to
horizontal
by
auto
skyler,
which,
basically,
depending
on
our
amount
of
resources,
used,.
What
is
suffering
the
distances
that
already
exist,,
we
are
going
to
add
instances
to
the
sides
or
add
more
instances,
and
one
and
two
miami
instances
are
not
capable
of
processing
all
the
traffic
that
they
are
receiving.
What
we
are
going.
B
B
B
Who
made
this
slide
because
I
think
it's
wonderful
to
explain
what
auto
scaling
consists:
of,,
that
being
said,
the
second
model
of
auto
scaling
that
we
have
is
virtual
I,
stop
virtual
vp
a
see,
filter
I
can
auto
stayler
instead
of
scaling
by
adding
different
instances.
We
are
going
to
scale
by
increasing
the
size
of
the
instances
and
a
third
auto
installed
model
that
is
not
necessary.
A
B
No
expansion,
what
the
autoscale
cluster
will
do
is
expand
and
reduce
the
number
of
nodes
under
demand
in
a
way,
and
we
are
going
to
be
able
to
add
extra
nodes
to
collect
those
resources
and
be
able
to
deploy
our
workloads.
We
are
going
to
see
it
a
little
more
in
detail,
one
by
one,
simply
to
see
what
it
offers
us
and
what
is
not
worth
ht
to
horizontal
by
car
skyler.
This
is
probably
what
everyone
sounds
like:
horizontal
scaling,
adding
and
removing
distances
based
on
different
types
of
metrics.
B
This
has
native
support
and
guber
net.
Is
there
is
a
horizontal
controller
for
control
in
control?
Sorry,
it
was
whoever
It
is
in
charge
of
managing
all
the
objects
of
type
hp
to
that.
Yesterday,
the
cluster
of
executing
the
queries
against
the
different
resources
of
the
metrics
that
we
find
that
we
are
defining
and
scaling
the
instances
based
on.
B
The
results
of
that
value,
for
example,
I-
can
say
that
I
want
that
my
workload
and
in
the
first
element
a
have
an
average
use
of
cpu
of
50%,
hp
itself
to
control
through
defining
it
in
this
object
of
horizontal
type
by
auto
skyler
by
auto
center.
Is
that
I
remember
What?
Is
this
called?
I'm
going
to
look
at
it
personally.,
Auto
Styler,
to
say
it
right,
I'm
going
to
be
wrong
from
the
beginning.
What
it
's
going
to
do?
Is
it's
going
to
make
an
agreement
with
the
metrics.
B
B
Tell
it
when
you
go
over
70
percent.
Add
servers
when
you
go
below
20%,
for
example,
remove
the
non
gubernatis
through
the
7th,.
What
he
is
going
to
try
is
that
the
target
value
is
always
close,
that
is,,
if
I,
tell
him
that
an
average
cpu
consumption
of
50
percent
and
add
a
link
and
give
instances
to
try
to
make
the
value
that
the
value
that
it
always
has
is
as
close
as
possible
to
that
objective
value
later
on.
B
When
we
talk
about
want,
we
will
talk
about
how
the
hp
as
work
and
what
they
support,
but
it
is
important
to
clarify
that
we
have
three
types
of
metrics
that
manage
hp,
a
metrix
gastón
metrix
and
external
metrix,,
which
are
usually
by
default
in
most
cloud
providers
through
the
metrics
server
that
comes
by
default
in
most
key
servers
and
in
the
only
one
found
in
the
clusters
of
the
ruler
of
different
klaus
The.
Only
one
that
finds
something
does
not
have
that
double
I
know,
but
they
say
it
very
clearly
how
to
add.
B
It
are
cpu
and
memory
with
which
the
most
common
model
of
the
hp.
As
for
the
vast
majority
of
workloads
is
useless
to
be
used,
but
my
memory
is
worth
another
model
vp
to
bill
body
of
auto
skyler.
What
it
will
allow
us
is
to
expand
and
reduce
the
size
of
the
resources
assigned
to
the
por
Let's.
Imagine
that
I
have
a
stateful
workload
which
I
need
to
give
it
more
power.
B
I,
don't
need
to
download
it
to
try
optimize
the
use
of
resources
that
I
have
in
the
cluster,
because
the
objective
of
autoscaling
is,
on
the
one
hand,
to
respond
to
the
demand
of
our
users
and,
on
the
other
hand,
it
is
to
reduce
the
consumption
of
resources
when
they
are
not
necessary,
because
in
the
end
the
resources
are
stopped,
paying
with
which
the
vp
to
what
it
is
going
to
do.
Let's
imagine
our
connected
workload
that
needs
to
auto
scale,
for
example
a
database
in
the
database.
B
Normally
we
take
this
carefully
because
it
depends
on
the
database
engine
and
the
configuration,
but
normally
adding
for
substances
does
not
improve,
for
example,
in
our
writing
speed
if
the
new
master
is
saturated,
adding
more
slave
nodes
will
not
allow
our
database
to
do
better
in
write
operations.
In
these
cases
it
could
make
a
lot
of
sense.
A
virtual
I
can
auto
stayler.
What
it
is
going
to
do
is
when
we
are
exceeding
the
allocated
resources
and
we
need
more
resources.
B
The
virtual
for
all
this
English,
what
it
is
going
to
do
is
modify
our
deployment
werder
weight,
our
our
manifest,
where
the
test
to
allocate
those
extra
resources
from
So
I
can
start
having
one
agent
and
end
up
having
four,,
always
within
the
limits
and
what
interests
us,.
But
this
progress
will
allow
us
to
modify
not
the
number
of
posts,
but
the
resources
assigned
to
the
posts.
B
and
finally,
in
the
auto
skyler
cluster.
Basically,
this
is
important
to
know
because
if
you
have
a
hp
and
if
you
are
using
any
scaling
model
in
general,
if
you
are
using,
it
is
highly
recommended
that
the
autoscale
cluster,
according
simply
what
it
will
do,
is
add
and
remove
all
on
demand
if
I
need
to
add
20
instances,
20
after
my
deployment
is
obviously
not,
and
if
what
I
want
is
to
save
resources
by
scaling
on
demand,
it
doesn't.
A
B
B
Be
able
to
add
new
posts
to
be
able
to
restart
our
application.
Our
new
instances
of
the
applications
and
The
question
is
hp,
a
or
vph,,
which
is
what
we
will
have
to
use,,
what
interests
us
and
what
can
give
us
the
most
performance,
okay,
well,.
There
really
isn't
something
set
in
stone
here:
either,
the
fact
that
we
use
one
will
depend
on
our
needs.
or
we
use
another.
The
main
advantages
in
hp
are
the
support
is
native
simply
by
installing
a
metrics
server
in
case
our
cluster
does
not
have
it
installed
by
default.
B
It
is
fully
extended
the
The
rulers,
documentation,
exemplifies
it
very
well..
It
is
the
most
common.
Another
advantage.
Is
that
if
our
real
application
is
this
English,,
we
will
not
have
a
problem
adding
new
instances.,
Front
horizontal,
open
to
the
sides
and
then
close,,
because
we
do
n't
care,,
we
are
stable.,
an
application
does
not
receive
traffic
and
that's
it.
There
is
no
problem.
The
spots
will
appear
and
be
destroyed
arbitrarily
no,
but
we
need
them.
C
A
B
12
3,
for
example:
now
it
will
put
5
and
then
it
will
return
to
the
3
or
the
ones
you
need
and
also,.
As
I
was
saying
a
moment
ago,
it
has
support
for
gastón
metrics,.
Apart
from
the
metrics
server
itself,,
it
is
available
in
green.,
You,
see
the
faces
with
metrix
and
the
extras,
bad
metrix..
There.
B
B
Going
to
be
able
to
cover
one
or
the
other,
for
example,,
a
very
typical
one
that
usually
exists
is
propitious
adapter
propitious
adapter
is
no
more,
that
an
adapter
that
exposes
metrics
through
custom,
mind
the
custom
metrix
api
and
that
allows
us
to
throw
qualys
to
a
province
server
to
know
data
related
to
auto
scaling,
for
example,
or
we
could
be
talking
about
the
truth
room
or
there
are
different
adapters
that
we
can
use
and
that
are
normally
going
to
be
exposed
through
the
iwate,
castor,
metrix
server,.
The
vp
is
worth
the
ub
uses.
B
The
advantages
it
has
is
that
it
allows
you
to
work
in
a
simulation
mode
on
a
hp,
so
I
tell
you
to
work,.
You
are
not
going
to
tell
me
the
ideal
number
of
instances
for
what
you
need
is
no,.
The
checkbook
in
wheat
will
directly
take
us
to
those
pieces,,
the
vp
a,
on
the
other
hand,.
It
will
allow
us
to
do
so,.
It
does
have
a
recommendation
mode
in
which
it
will
tell
us
the
value
What
I
would
put
to
your
workload.
B
Is,
well,,
a
minimum
of
all
of
us,
Jesus,
is
going
to
allow
us
to
play.
In
fact,.
The
VP
is
a
tool
that,
if
it
is
not
true,,
it
is
not
a
silver
bullet,,
it
does
not
solve
our
problems.
They.
Do
not
see
anyone
now
saying
It
is
that
I
have
seen
one,
although
it
says
that
with
the
vp
a
we
can
calculate
the
resources
of
the
posts,
but
if
it
can
give
us
an
indicative
idea
of
where
our
application
can
move,.
B
Yes,
it
does
coincide
that
we
have
the
virtual,
with
the
virtual
I
can
auto
stayler
deployed
in
our
cluster.
We
can
play
with
that
and
take
advantage
of
this
situation
to
check
what
are
the
metrics
that
could
interest
us.
Not
everything
is
not
gold.
All
that
glitters
we
are
going
to
have
to
install
the
virtual
by
auto
skyler.
Now,
if
I
remember
correctly,
The
following
slide
is
a
demo
with
which
I
would
take
advantage
of
the
fact
that
we
change
the
screen
and
they
show
you
where
you
could
find
Skyler's
act.
B
B
Crisis
problem,
who
does
not
work
with
potholes
when
opening
to
the
anp
and
from
kuwait,
to
put
what
he
is
going
to
do
is
subtract
artear
the
bots,
and
this
is
important
because
it
is
our
responsibility
when
We
deploy
a
V
to
make
sure
that
we
are
not
going
to
cause
too
much
downtime
in
our
application.
For
example,.
Let's
imagine
that
we
are
talking
about
a
master
slave
database
where
we
have
a
primary
replica
and
several
secondary
replicas
to
accommodate
that.
If
we
need
our
model
to
be
tolerant.
B
to
failures,
we
could
at
any
given
time
without
any
problem,
kill
a
new
master,
sorry
kill
without
noticing
the
new
master,
the
message
that
contains
all
master
to
increase
the
resources
it
has,
but
what
we
cannot
do
is
kill
them
all
at
once.
Yes,
at
once,
we
kill
the
master.
What
our
system
is
going
to
do
is
look
for
another
master
and
keep
responding,
keep
responding.
We
will
not
have
a
down.
It
is
in
our
application.
B
Our
service
is
unprepared
in
high
availability,
and
our
application
should
be
prepared
to
be
able
to
manage
small
micro
outages
that
can
occur
during
the
election
of
a
new
teacher,
but
if
we
kill
all
the
substances
at
once
to
subtract
them
from
the
new
resources,
that
would
not
be
a
solution.
What
is
the
solution?
How
can
we
avoid
this?
How
could
we
get?
Who?
Who
we
don't
have
these
small
damages,
because
we
see
that
all
the
instances
have
been
restarted
at
the
same
time?
It
is
worth
for
this.
That
vernet
is
also
the
tricks.
B
It
is
not
the
European
one,
but
it
is
the
ruler
himself.
Who
is
the
one
who
allows
us
to
manage
this
situation?
The
European
one?
What
is
going
to
always
respect
the
so
that
the
pot
disruption,
but
it's
a
bit
of
irruption
madrid-
is
an
object
of
the
property
asks
that
vernet
is
that
for
those
who
do
not
know,
basically,
what
it
allows
us
is
to
make
a
declaration
of
intent
of
what
is
our
minimum
our
minimum
resource
to
function?
B
We
can
say
it
is
valid
for
this
deployment
or
for
this
workload
it
does
not
matter
if
placement
or
any
other
workload
goes
up
for
this
workload.
When
you
go
to
do
an
operation
such
as
node
trainer
or
cloud
to
cloud
scaling
at
most,
you
can
have
x,
number
of
instances
or
x
percentage
of
instances
down
at
the
same.
B
A
C
B
A
note
lm
a
budget
instruction
message.
That
would
say
this
workload
cannot
lose
more
than
one
instance
per
cluster
maintenance
processes,.
But
once
again
it
is
our
responsibility,,
the
default
European
one
will
not
take
any
care,,
it
will
take
care
and
when
it
has
to
restart
a
bit,,
it
will
restart
it.
A
B
A
B
To
do
is
I
am
going
to
take
this
little
tab
here
to
show
you
the
volpe
to
the
vp
a
you
are
not
rather
your
cuber
better
project
is
inside
hoover,
net's
own
record.
It
is.
It
is
a
folder
inside
of
the
auto
skyler
challenge
of
the
establishment
that
vernet
is,
and
here
it
is
true
what
color
they
give
you.
If
you
need
it,
they
give
you
some
ideas
that
you
can
look
at
examples
of
how
they
work.
They
work
with
their
own
service
ashton.
It
is
also
a
division
so
that
you.
B
It
a
chaste
summary
gastón,
jesús
definition,
is
nothing
more
than
adding
support.
Adding
new
supported
manifests
to
our
green
cv.
Plaster
is
that
someone
is
going
to
control,
is
going
to
monitor
and
act
accordingly,
for
example
the
virtual
itself
by
three
pillars.
Who
is
going
to
do
that
work
here?
Okay,
said
this:
we
're
leaving
at
the
same
time
I
hope
it
won't
have
a
mount
effect,
because
these
things
don't
happen
and
a
lot.
The
first
thing
we're
going
to
do
is
test.
B
It
chat
a
hp
to
simple
here
this
this
code
to
stabilize
that
the
end
of
not
it
I
will
make
the
repository
public
to
the
public,
and
so
you
can
try
it,
but
basically
we
are
going
to
deploy
one
and
I
space
where
we
are
going
to
put
a
deployment,
nothing
very
crazy
with
some
resources
worth
a
service.
This
all
like
this
everything
so
far.
Well,
what
we
are
going
to
do
is
we
are
going
to
display
a
horizontal
with
cars
and
dear
that
it
is
an
object
within
life.
B
B
Tell
you
the
maximum
number
of
instances
that
I
want
and
the
minimum
and
the
hp
a
controller
is
going
to
move
based
on
the
metrics.
You
get
from
this
app.
It's
going
to
come
back
in
the
number
of
instances,
so
that
our
cpu
target
value
always
stays
at
50.
And
here
what
I
have
is
a
little
simulator
with
bombardier,
where
I
am
going
to
put
traffic
so
that
it
is
forced
to
scale
so
that
they
consume
resources
for
which.
B
B
What
I'm
going
to
do
is
okay,
right
now
we
have
a
replica
of
our
night,
but
if
I
look
at
the
hp
who
we
can't
look
at
through
the
we
can
consult
it
through
the
toy
l
or
through
the
leds,
to
consult
the
first
thing
that
the
ex
from
manchester
and
the
greenspace,
which
in
this
case
is
that
we
see
to
check
if
I
write
it
correctly,
it
will
be
wonderful
and
it
tells
us
look
here.
You
have
a
single
replica.
This
is
normal.
B
When
a
new
instance
starts,
the
metrics
server
still
do
not
have
detectable
associated
metrics.
It
is
normal
that
When,
the
entire
workload
appears
from
the
beginning,.
It
appears
from
scratch,.
It
does
not
have
associated
metrics,.
This
is
what
Linz
will
tell
us
something
similar,.
It
will
tell
us,
hey,.
It
does
not
have
a
metric
for
the
pop,
because
of
course,
first,,
which
has
to
exist.,
then
it
will
generate
metrics
and
then
they
will
be
available,
but
we
can
forget
about
it
and
what
we
are
going
to
do
is
our
our
load
generator,
well,.
B
We
are
going
to
put
this
in
three
instances,
if
they
were
an
application,
of
course,,
we
start
from
the
scenario
in
which
it
enters
a
lot
hit
load.
Therefore,
it
is
likely
that
we
would
have
if
this,
not.
If
we
configure
this
like
this
first,
it
is
likely
that
we
would
have
a
small
down
line
because
we
are
suffering
our
application.
Of
course,
at
h,
you
have
not
be
able
to
go
from
an
instance
to
receive
a
lot
of
traffic
educate
only
once
or
it
will
die
here.
We
see
that
it's
worth.
B
Not
instantaneous,
it
is
not
a
I'm
going
to
request
it
attached
to
the
node
and
I
see
what
there
is,
but
scraping
is
done
every
x
time
with
which
it
is
normal
for
a
few
seconds.
But
if
I
remember
correctly,
it
will
be
30
seconds.
What
is
done
is
to
create
ping
of
the
metrics.
I
am
not
one
hundred
percent
sure
add
more
load
and
make
it
faster
here.
Our
ex
engine
is
pulling.
B
I
was
complaining
about
that,,
it
can't
cope
and
we
do
n't
give
it
more
as
a
problem
but
well,
it
has
already
arrived.
If
we
see
it
again
to
check
simply
to
see
what
it
says
from
here,
it
says
better.
I
am
200%
short
of
the
value
that
you
have
put
and
you
are
using
200%
of
the
values
that
you
have
put
in
the
rich,
and
it
is
important
the
hp
as
work
with
the
value
that
we
have
put
in
the
request
section
in
this
one,
not.
B
A
B
Fp
1,
because,
let's
remember
it
is
about
the
answer,
there
are
scenarios
and
it
depends
on
what
interests
us
and
the
areas
only
They
will
work
in
case
we
obviously
have
the
result
section
defined,
in
case
we
don't,,
they
will
fail,,
it
will
give
us
an
error
and
we
will
have
to
go
see
the
disk
prime
to
see
what
it
opens,
but
well,.
Here
we
see
what
it
tells
us,
look
at
me.
B
I
have
four
instances
and
now,
for
whatever
they
say
with
four
instances,
I
still
have
a
lot
of
traffic
and
it
has
escalated
to
8.
Once
we
removed
the
traffic,
it
would
scale
back
down
without
any
problem
here.
An
important
thing
to
keep
in
mind
is
the
hp.
A
It
is
very
aggressive
to
add
instances
and
I
do
n't
know
how
very
protective
to
eliminate
them,
that
is,.
B
We
are
going
to
scale
by
adding
instances
very
quickly,,
but
we
are
going
to
reduce
the
number
of
instances
that
we
have
much
slower,,
very
protective
bases
in
that
sense,
it
is
protective,.
It
will
try
to
maintain
or
adapt
as
best
as
possible
to
give
support
to
the
users.
Okay,
well,,
we're
just
going
to
see
this
era
and
we're
going
to
remove
it.
I'm,
going
to
display
the
second
demo
with
the
auto
skater.
A
B
Same
it
is
very
similar.
We
have
a
space,
a
deployment
with
some
resources.
We
have
a
service
in
which
value
changes.
First,
we
have
the
virtual
for
the
virtual
by
auto
styler
sorry
vertical
I
can
auto
skater
I
always
always
get
confused
if
I
say
virtual
to
think
that
it
is
vertical
that
they
generated
has
once
again
confused
the
vertical
with
all
this
money
in
which,
in
this
we
are
simply
going
to
say,
look
auto.
B
Am
only
going
to
update
one
at
a
time.
has
registered
that
it
will
be
here,
auto
sky
link,
virtual
auto
skater,
really
registered
register
the
checkpoint
and
register
the
virtual
potato
styler,
and
here
it
is
already
telling
me
what
are
the
values
that
it
has
assigned
without
doing
anything.
It
is
telling
me
today
in
your
workload,
it
should
be
at
587
and
probably
if
I'm.
A
B
A
B
B
And
a
tax
or
another
updater,
and
we
come
up
through
the
blogs.
Well,
it
tells
us
It's
good
that
the
workload,,
let's
see
where
it
is
here
first
than
yesterday,
and
that
there
has
been
one
that
could
not
be
updated
because
it
does
not
comply
with
the
duty
that
violates
the
update
and
above
here,
well,.
It
would
tell
us
what
that
it
has
taken,
that
it
has
passed
percent
so
as
not
to
pass
us
by,
because
time
is
going
to
run
out.
B
B
A
B
Adapt
to
certain
scenarios
where
we
need
to
increase
iron,
the
scenario
where
adding
instances
does
not
solve
our
problems.
What
we
need
is
to
add.
Having
said
this,
we
are
going
to
Going
back
to
the
presentation
I'm
going
to
speed
up
a
bit
because,
in
the
end,
I
won't
have
time
to
talk
about
anything.
Else.,
These.
A
A
B
Amount
of
work
we
have
to
do
because
we
can
have
a
thousand
messages
in
a
queue
quickly
for
many
hours
or
in
any
type
of
korda,
and
not
be
worth
less,
because
we
have
a
df
consumption
but
very
low
and
still
the
size
of
the
col
increases
to
solve
this
type
of
problem.
There
is
a
lot
to
adapt
is
that
they
allow
us
specifically
for
different
things.
B
A
B
By
side,
a
company
in
which
the
company
in
which
one
of
the
members
of
the
online
trainers
board
worked
and
is
going
to
allow
us
to
install
based
on
effects
for
him
that
it
is
an
event,
an
event
is
a
message
in
a
queue
now,
when
we
see
the
scribes,
we
will
see
more
in
detail.
What
are
the
events,
but
basically
what
it
is
going
to
do
that
through
its
own,
its
own
serves
and
its
own
case
of
its
definition
is
we
are
going
to
define
a
series
of
scaler.
A
B
Connection
to
a
data
source,
we
are
going
to
tell
him
and
I
want
you
to
connect
to
this
rabbit
and
get
the
number
of
messages
from
these
highlighted.
The
I
want
you
to
go
to
a
service
bass
and
read
the
number
of
messages
in
this
topic.
I
want
you
to
go
to
an
average
and
execute
this
query
so
that
it
is
made
up
of
two.
B
A
A
B
Do
not
need
to
touch
the
hp
a,.
We
do
not
need
to
know
how
it
works,
it
is
below,.
We
simply
use
the
style
of
the
crm.
That
is
call
where
there
are
other
more
advanced
uses
with
is
kim
jong,
but
we
use
the
shear
crd,
ok
and
left
it's
going
to
create
a
hp,
a
and
the
mercy.
Metrics
server
is
going
to
expose
that
metric.
So
we've
created
the
metric,
we've
created
the
hp,
a
that
points
out
the
metric,
and
it
is
hp
itself
to
control
it
is
who
is
going
to
scale
the
loads.
B
B
B
If
I
remember
correctly,
with
version
2.5
that
came
out
on
Thursday,
there
are
about
40
scalers
that
we
can
use
for
us,
it's
transparent
in
the
sense
that
its
trailer
is
defined
once
documentation,
and
it
is
enough
that
we
see
how
to
use
it
and
integrate
ourselves,.
It
does
not
have
much
more
complexity
than
that
and
we
go
with
the
last
voice.
B
The
same
in
space
deployment,
and
here
what
we
change
is
the
part
of
the
axes
of
the
steiner
car
of
the
sheet
metal
instead
of
defining
a
pea
width,.
What
we
are
going
to
do
is
define
a
style
that,
for
this
to
progress,
I
am
going
to
apply
it
while
I
explain
it.
What
we
are
going
to
define
is,
in
the
same.
B
Our
minimum
of
replication,
what
is
our
maximum,
but
it
changes
what
that
we
are
going
to
use
trailers.
Trailers
are
each
one
of
those
scalers.
We
can
have
several
trivial
because
our
workload,
your
scale,
based
on
different
origins
in
this
case,
what
I
am
doing,
is
and
I
am
installing
in
the
deployment
auto
skater
deployment,
based
on
the
number
of
instances
that
are
in
this
link
space
with
this
selector.
This
is
a
styler.
B
B
And
I'm
sorry,
I
have
one
and
I
have
others..
You
are
climbing,.
It
will
lead
to
the
same
number
of
instances
that
have
this
and
it
will
do
so
because
it
remains,
it
will
be
in
charge
of
managing
I'm,
going
to
put
10
instances,
here,
I'm,
going
to
scale
it,
okay,
and
now
we're
going
to
see
how
first
it
already
has
one,
and
now,
when
it
starts,
we'll
see
how
it
arrives.
B
The
hp
itself
to
control
in
general,
the
value
is
connecting
in
this
case
to
the
y
of
q
vernet
of
the
day
is
saying
sorry
and
string
test.
How
many
votes
do
you
have
the
stronismo
space
with
this
selector
to
be
worth?
I
have
10.
If
we
come
here,
we
will
see
what
it
says
that
It
has
legs
and
I
have
10.
Now
it
has
gone
down
because
in
its
assets
with
which
I
have
an
average
of
two
and
five
for
each
I
have
four
points.
B
Therefore,
I
have
a
guide
and
who
is
going
to
give
us
this
for
this
next
simplification
of
components,
it
is
true
that
left
It
is
not
the
only
solution
that
it
gives
for
some
scenarios
it
may
be
the
best
and
for
others
it
does
not
want.
The
only
thing
it
does
is
auto
install.
There
are
other
more
complex
solutions
in
the
auto
scaling
part.
It
is
part
of
the
solution
and
they
give
you
a
more
complex
solution
than
whatever.
B
The
only
thing
you
are
going
to
be
able
to
get
is
an
auto
scaling
solution,
nothing
more.
It
is
not
limited
to
do
anything
else,
but
from
my
humble
point
of
view,
it
is
better
than
using
them
adapter,
because,
with
a
single
metrics
server,
we
can
configure
a
lot
of
different
scalers
Let's
see,
by
the
way,.
The
url
of
the
project
is
left
sh
in
case
you
want
to
take
a
look
and
just
to
see,
for
example,
we
could
scale.
In
fact,.
It
occurred
to
me
that
an
input
could
be
created
to
put
it,.
B
A
B
Am
using
audio
agents
the
foster,
as
we
would
see,
From
the
previous
talk.
I
could
say:
okay,
if
I
don't
have
any
pending
work,
it
doesn't
make
sense
to
have
10
instances
and
if
I
have
10
pending
jobs,
it
doesn't
make
sense
to
have
one
instance.
It
makes
sense
to
have
a
sufficient
number
of
distances
to
support
the
work.
We
have
talked
about
this
styler
As.
There
are
many
others,.
The
truth
is
that
it
can
be
done,.
B
A
C
B
Want
to
generate,
we
tell
it
the
qualy,
and
this
will
do
all
the
work
of
connecting
to
the
server
and
pulling
the
string
impromptu
l
to
recover
it.
It
will
expose
it
to
the
benfica
server
without
us
having
to
worry
about
anything
and
to
finish
it
off
very
quickly,
because
time
has
gone
by,.
The
truth
is
that
I
get
winded
up
a
lot.,
please
the
perfect
slide
from
the
left.
We
are
working
on
a
software,
a
real
http
scaling
with
first
level
support.
What
does
this
mean?
What
do
we
mean
by
this
right
now?
B
B
Be
all
I
don't
need
or
may
not
want
to
observe
averages
to
be
able
to
scale
based
on
cpe
to
sorry
for
more
than
the
amount
of
http
traffic
of
dryness.
We
are
working
in
a
year
that
of
first
level,
support,
that
is
to
say,
create
a
certain
infrastructure.
You
can
see
a
detailed
description
in
the
repository
of
the
organization
of
I
have
the
repository
of
http
add-
with
basically
that
it
is.
B
Your
load
balancer
and
your
application,
your
service,
that
monitors
the
amount
of
traffic
so
that
you
can
load
based
on
its
undamaged
external
dependencies
that
may
not
interest
you
and
here,
just
to
finish
small
disclaimer
use
that
bass
at
your
own
risk.
It
is
a
beta
project,
it
has
no
official
support.
This
is
working
from
me.
There
is
a
lot
of
momentum,
but
it
should
not
be
used
in
production,
not
yet
use
each
bass
at
your
own
risk.
B
That
is
all
the
truth.
Thank
you
very
much.
To
the
people
who
have
been
here,
I,
don't
know
if
you
have
any
questions,
something
you
want
to
know,,
otherwise,
I
practically
say
goodbye,
so
I
don't
forget,
because
I
always
tend
to
forget
in
the
plans.
We
also
always
look
for
talent.
You
can
more.
You
can
search
by
planner
point
tech,.
B
B
B
B
C
B
C
C
B
Two
and
a
half
three
years
ago,
2019
or
so
it
started
as
a
sandbox
and
has
already
graduated
and
you're
at
is
sorry
graduated.
No,
it
is
not
a
graduated
project,
but
it
is
in
the
incubation
phase
and
you
can
use
it
in
production.
It
is
safe
to
use
in
production.
In
fact,
we
have
several
clients
or
several
clients.
Not
several
listed
users
are
free
to
view.
We
do
not
force
anyone,
and
the
truth
is
that
there
baba
cloud,
for
example,
is
one
of
the
first
to
appear
there,.
B
Expansion,
product
manager,
and
all
the
rents
and
so
on,
a
bit
of
the
part
of
server
mes
that
said
it
gives,
is
a
service
of
10.
Now
it
is
a
service
that
is
being
used
by
microsoft.
Spontaneity
is
in
apps
an
application
mode
that
came
out
server.
Les
that
allows
us
to
use
mammals
allows
us
to
deploy
potts
that
uses
a
tupper
and
that
it
gives
in
a
managed
way
so
good,
good
news,
because
there
is
a
battle
test.
C
If
the
truth
is
that
this
topic
is
very
interesting,
I
am
very
curious
to
try
keep
mentioned
the
outdated
one.
I
also
think
it
is
a
very
interesting
project
if
we
put
it
in
the
context
of
a
university
project
of
power
scale
in
a
way
to
imitate
a
real
scenario
where
you
can
scale
using
different
methods.
So
I
find
it
very
interesting,
note
that
it
is
interesting
to
have
an
academic
part,
because
sometimes
we
do
crazy
experiments
that
can
be
applied
to
real
production
cases.
B
You
have
a
service
distributed
in
which
you
want
to
send
emails.
When
you
have
to
send
emails,
it
does
not
make
any
sense
to
scale
based
on
the
cpu,
because,
obviously,
if
what
your
application
does
is
read
a
message
from
a
queue
and
it
sends
a
message
from
an
app
and
the
provider
mail
to
send
an
email,
the
atp
will
never
go
up,
but
if
the
cpu
goes
up,
it
is
not
a
cause.
B
It
is
a
consequence
you
can
escalate
to
mitigate
the
consequence,
but
the
cause
will
continue
to
be
that
you
have
too
many
pending
messages
being
able
to
use.
What
remains
you
can
choose
and
say:
no
Look,
I
want
each
instance
to
have
more
or
less
that
is
capable
of
handling
50
messages
per
minute,
for
example,.
So
I
want
this
instance
to
have
50
messages
and
for
the
cluster
to
take
care
of
saying
you
need
10
or
you
need
0,
because
it
adds
the
functionality.
Very
It
is
interesting.
B
B
Almost
when
it
came
out
as
consultants
I
worked
in
a
micro
network,
consultancy
on
Microsoft
issues
and
since
then
I
have
been
following
it,
because
I
think
that
solving
is
very
big
in
the
auto
escalation
in
which
Vernet
is
only
one
piece
that
solves
a
real
problem
situation
that
exists.
And
it's
really
me
when
they
proposed
to
me
to
be
the
most
proud
thing.
I
am
because
it
seems
to
me
a
critical
piece
that
can
save
money,
because
it
can
save
a
lot
of
money
without.
C
B
C
C
C
B
What
you
do
is
to
support
all
the
open
projects.
They
are
really.
There
is
work
behind
the
next.
If
it
collaborates
resources
and
it
is
important
and
it
is
appreciated,
but
hey,
it
always
has,-
it
is
always
good
to
have
some
money
to
be
able
to
use,.
You
need
to
have
a
cluster
in
blue
when
ws
to
be
able
to
test
the
twente
to
be
able
to
launch,,
then
it
is.