►
From YouTube: GoLab 2019 - Francesco Romani - KubeVirt: bringing virtual machines in a Kubernetes world
Description
🔔 FOLLOW GOLAB CHANNEL 🔔
KubeVirt extends Kubernetes by allowing users to run virtual machines side to side to containers in their Kubernetes clusters. We will have a whirlwind tour of the architecture, the challenges, the technologies, the components that make this possible.
In this talk we are going to look how KubeVirt allows general purpose management of virtual machines with Kubernetes, alongside application containers. We will see the ways KubeVirt bridges the gap between standard, well known cloud VM management and the brave new world of container orchestration. We will dive in the implementation of KubeVirt, to examine the architecture, and some of the challenges integrating virtual machines with the Kubernetes; we will see the components that make KubeVirt, and some of the tradeoffs involved. We will also peek in two of the most active development topics: the deployment of the whole platform and the monitoring integration with the standard Prometheus infrastructure.
This channel is dedicated to the videos of the GoLab conference.
💛 Follow us on TWITTER: https://twitter.com/golab_conf
💜 Follow us on FACEBOOK: https://www.facebook.com/golabconf
GoLab is the first Italian international conference on the Go programming language, organized by Develer.
Develer is not just an Italian company projecting and releasing hardware and software solutions for the industrial environment, but is also an ensemble of people sharing their great passion for new technologies and how they can be applied to your everyday life.
🔸 Follow DEVELER on INSTAGRAM: https://www.instagram.com/wearedeveler/
🔹 Follow DEVELER on FACEBOOK: https://www.facebook.com/we.are.develer/
🔸 Follow DEVELER on TWITTER: https://twitter.com/develer
🔹 Follow DEVELER on LINKEDIN: https://www.linkedin.com/company/114426/
🔸 Follow DEVELER on TELEGRAM: https://t.me/wearedeveler
🔹 Follow DEVELER on TIK TOK: https://www.tiktok.com/@wearedeveler
A
Hey,
hello,
thanks
for
attending
we're
going
to
talk
about,
indeed
about
this
cube
root,
project,
I
work
with
and
specifically
how
our
our
experience
in
which
using
golang
in
our
project.
So
it's
going
to
be
a
we're
in
tour,
pretty
high
level
so
few
words
about
myself
I'm,
a
software
engineer,
I
work
for
reddit
from
few
years,
I'm
involved
in
this
project,
which
brings
together
cuban
natives
and
the
veterans
asian
stack.
So
my
daily
job
involves
golang,
which
I
very
interested
in
learning
more
by
the
way.
A
So
what
is
this
cube
root
about?
We
want
to
bring
together
cute
machines.
Containers
I
told
you
already
so
what's
a
little
machine
very,
very,
very
quick
recap:
bit
machine
is
a
program
which
allows
us
to
virtualize
complete
computer
system,
so
it
presents
it
allows
to
run
another
operating
system
on
top
of
our
operating
system
and
we
present
virtualized
hardware
operating
system
to
the
guest
operating
system.
So
we
can
configure
how
this
boot
order
could
look
like
and
we
will
have.
A
We
will
also
can
leverage
some
celery
some
features
of
the
order,
for
example,
recent
Intel,
AMD,
CPUs
or
specialized
that
hardware
like
network
cards.
An
important
topic
is
the
hypervisor,
which
is
the
manager
which
allows
better
machine
to
run
an
improvisor
could
be
who
could
replace
an
operating
system
or
could
be
a
component
of
operating
system.
For
example,
in
Linux
we
have
the
KVM
subsystem,
which
is
an
app
advisor
and
not
to
allows
us
to
access
the
acceleration
features
of
the
CPUs.
A
On
the
other
end,
the
component,
which
will
tries
the
hardware
properties
called
qmu.
Well,
there
are
many,
but
one
of
the
biggest
and
most
important
is
qmu,
which
indeed
presents
which
reside
order
to
the
guest
operating
system.
Here's
a
picture
which
lay
presents
how
the
software
components
are
stacked
out,
so
we
have
our.
We
have
the
art
where
we
have
the
operating
system
or
the
hypervisor.
In
our
case,
we
have
the
operating
system
with
the
vaporizer
module
KBM.
A
Hopefully
it's
visible,
and
on
top
of
that
we
have
our
regular
programs
and
our
virtual
machine,
which
is
the
bigger
box
which
is
on
one
way.
This
is
another
program.
On
the
other
end,
it's
pretends
to
be
another
art
another
box
altogether,
so
we
have
the
butyl
Hardware
the
guest
operating
system
and
eventually
the
guest
applications
running
inside
the
other
end
containers.
A
This
is
the
other
half
of
the.
What
qubit
is
about.
Containers
are
a
big
topic
in
the
last
few
years,
but
containers
are
actually
for.
Linux
are
actually
an
aggregation
of
feature
of
the
Linux
kernel,
so
we
have
a
lot
of
layers
and
a
lot
of
components.
We
need
to
interact
with
each
other
to
create
this
isolation
and
make
the
application
running
in
the
container
pretend
to
be
the
only
one
of
the
system
using
those
kind
of
confinement
technologies.
You
can
limit,
for
example,
how
much
memory
or
much
CPU.
A
A
So
you
pack
your
application
with
all
the
dependency
dependencies,
maybe
resource
files
or
whatever
else
is
needed
to
run
and
you
ship
it,
and
hopefully
it
will
work
the
first
try,
but
to
run
those
contrast,
application
on
a
fleet
or
for
compute
and
Aust.
So
you
may
may
very
likely
need
management
platform.
Kubernetes
is
probably
the
most
famous,
so
when
it
is
allows
to
run
this,
you
contain
its
own
fleet.
Let
them
scale
you
can
monitor
them.
You
can
hold
everything
you
need
and
probably
more
and
here
comes
cube
vert,
which
is
an
add-on.
A
You
will
install
in
your
kubernetes
or
openshift
cluster
and
you
get
the
capability
to
run
virtual
machines.
Community
sees
cube
here,
it
is
built
on
top
of
kubernetes
feature
and
brings
with
it
the
utilization
stuck
in
a
transparent
way.
You
just
gain
a
new
tile,
new
kubernetes
type,
which
is
a
better
machine
and
few
others,
and
you
run
with
a
machine
contrast
on
your
on
your
cluster
side
by
side.
Otherwise,
you
will
need
probably
to
to
control
planes.
The
nice
thing
is
that
you
now
have
just
one
control
plane
here.
A
Is
that
how
it
could
like,
if
you
take
a
look
on
the
on
a
single
host
which
run
them
both?
You
see
that
you
have
now
a
virtual
machine
layered
on
top
of
the
container
and
time,
because
you
have
a
butyl
machine
running
inside
the
container,
so
this
is
basically
it.
This
is
what
cube
root
is
about
now,
Qbert
is
built,
it's
meant
to
be.
A
A
part
of
the
equipment
is
ecosystem
in
entities,
so
it
is
the
cuban
and
the
container
in
kubernetes
ecosystem
is
written,
most
lingo
and
indeed
Qbert
is
written
in
go
so
on
once
on.
One
end
right
this
project
and
go
is
pretty
much
force
adoption,
because
when
you
have
such
a
rich
ecosystem,
you
want
to
make
use
of
it.
On
the
other
end,
this
is
not
really
an
issue
by
an
amine,
because
go
is
awesome.
We
know
at
Golub.
So
now,
let's
talk
about
which
features
we
the
particular
and
which
features
of
the
language.
A
We
particularly
found
interesting
and
helped
us
or
helped
the
last
a
little
less
than
we
expected.
So
the
usual
suspect
when
we
learn
about
go,
is
one
among
the
of
the
very
first
thing
we
know
about
go
is
oh
how
some
the
compiler
is,
how
fast
it
is,
and
how
would
that
impact
us
or
impact
most
of
developers
it
well
completion
times
are
faster,
so
you
have
quite
a
fast
turnaround
time
totally
true
totally
awesome,
but
we
will
see
that
in
a
moment.
Sometimes
it's
not
really
that
shiny.
For
other
reasons.
A
So
when
you
have
your
big
artifact
I
mean
like
hundreds
of
Meg's,
then
your
layers
become
big.
So
again
you
gain
something,
but
you
lose
something
because
you
have
bigger
than
a
lot
of
times
and
somehow
the
developer
flow
shocks
a
bit
again.
So
this
is
not.
This
is
not
something
that
it
is
not
solvable.
There
are
indeed
mitigation
and
way
to
circumvent
this
issue,
but
still
it
is
something
that
gets
in
the
way
and
you
have
to
deal
with
it.
A
So
this
is
pretty
much
what
I
just
said
and
yes,
they
are
another
really
nice
feature
of
the
go.
Is
the
C
interoperability?
We
have
not
sure
other
session
mentioned
or
Dell
with
that,
but
still
we
have
this
Siegel
module,
which
is
totally
great
because
you
can
just
consume
C
libraries
and
how
that
helped
us,
because
the
component
of
the
brutalization
stack
are
very
big,
mature
projects,
multi-million
slides
code,
though
big,
and
we
wanted
to
use
them
from
the
zero
and
we
did
because
using
C
go.
A
You
can
just
consume
the
the
those
libraries
and
and
easy
access
to
client
libraries,
and
so
we
can.
We
are
able
to
consume
like,
for
example,
we
built,
which
is
the
memory
the
management
layer
built
on
top
of
qmo
from
the
zero.
Another
option
is
always
to
rain,
implement
the
client-side
impure
go,
and
it
is
often
possible,
for
example,
when
the
server
API
is
accessing
across
the
network,
so
using
natural
protocols
but
having
the
possibility
to
consume
the
client
libraries
using
very
tested
code
was
nice
and
we
enjoyed
from
day
zero.
A
A
So
first
story
is
about
validating
the
confident
configuration.
So
let's
have
a
step
back.
Why
we
need
to
talk
and
to
deal
with
that
in
the
first
place.
Butts
are
relatively
simple,
simple
objects
when
you
need
to
configure
them.
Pods
are
the
basic
building
blocks
of
kubernetes.
If
someone
doesn't
know
so,
when
you
configure
your
objects
running
on
containers,
you
have
relatively
few
knobs
to
talk
to
go
to
tune
on
the
other
end.
A
Theater
machines
do
have
a
lot
of
configuration
options,
and
this
is
actually
a
good
thing
and
you
want
to
keep
them
problem.
Is
some
of
some
of
those
fields
depend
on
of
each
other
and
by
depend
on
each
other,
I
mean
they
have
values
that
are
legal
or
illegal
depend
on
other
fields
in
the
VM
configuration
area
or
they
become
optimal
or
suboptimal,
depending
on
the
odd
on
those
other
fields.
A
Here
is
an
example
taken
from
a
screenshot
of
viewer
of
viewing
a
graphical
interface
to
manage
bit
machine.
This
is
actually
withdraw
the
machine
manager
it's
hard
to
convey
how
many
options
you
want.
You
have
a
new
one
to
have,
because
you
want
to
have
a
very
fine
control
on
how
the
virtual
are
cooler
via
trailered
will
look
like
and
how
the
Victrola,
where
is
mapped
on
the
real?
A
So
you
have
those
plenty
of
options
you
want
to
keep
them,
but
you
want
to
preserve
the
queue.
Burnett
is
look
and
feel
you
want
to
present.
You
want
somehow
help
you
user,
you
want,
do
you
don't
want
to
hover
or
welcome
them,
and
you
want
to
bridge
the
gap
between
these
rich
set
of
features
and
the
same
relative
simplicity
of
the
containers.
A
So
we
had,
we
use
templates
templates
are
pretty
much
like
the
name
suggests.
Are
the
blueprints
of
objects
with
relatively
few
or
few
fields
you
can
tune,
and
we
present
them
to
the
user,
for
example
using
graphic
web
using
interface.
So
you
have
fewer
fields
to
talk
to
to
tune.
Of
course
you
can
have.
You
can
have
access
to
the
full
range,
but
you
can
have
a
simplified
view
problem.
However,
we
still
have
the
problem
of
interacting
fields.
A
For
example,
again
you
can
expose
a
multi
cue
network
interface
card
and
you
want,
for
example,
to
tune
how
many
quem
any
cues
you
have
so
many
real
concurrency.
You
can
have
your
real
peril,
ISM
you
can
head,
but
these
fields
for
this
field.
This
value,
for
example,
depend
on
how
many
virtual
CPU
you
have.
So
we
have
interdependent
fields
and
those
fields
depend
our
express
how
they
are
express.
The
professor
using
JSON
pattern.
A
Notation,
for
example,
you
have,
let's
say,
domain
dot,
dot,
dot,
CPU
count,
it's
not
a
real,
it's
not
a
real
deal,
but
you
get
close
enough
to
have
an
idea.
So
you
want
to
express
those
relationship,
but
there
is
another,
but
because
those
the
the
data
structure,
which
represent
a
VM
configuration,
is
deeply
nested
again,
because
how
the
configuration
is
made
is
deeply
complex.
A
So
when
you
translate
them
these
two
data
structure,
disease
gets
very,
very
complex
and
nested
and
you
can
have
default
objects,
so
some
objects
could
maybe
not
exist
because
they
are,
for
example,
expressed
as
pointers.
Pointer
is
nil.
That
means
this
is
the
four.
So
when
you
take
a
JSON
path,
the
fat
from
the
root
object
to
the
end
object.
This
path
can
be
broken
because
a
nil
object
could
be
in
between.
A
So
how
do
out
sorry?
How
do
you
distinguish
between
broken
path
and
an
illegal
part?
What's
an
illegal
part
apart
is
a
patoot
object?
Called
net
could
not
possibly
exist
if
the
fubar
device
does
not
exist,
it
could
not
possibly
exist
in
the
tree
in
the
tree
of
your
object
in
your
object
on
the
other
end,
you
maybe
have
a
default
device,
so
it
is
not
yet
there,
but
it
is
legal.
The
package
we
use
to
deal
with
that
cannot
make
this
distinguish,
which
is
also
pretty
hard
in
the
general
case.
A
So
we
use
the
reflect
package
to
fully
explore
dynamically,
explore
the
data
structures
representing
our
VM
configuration
and
dynamically,
create
a
lot
of
the
missing
object
along
the
way
and
insulate
them.
So
we
get
the
nice
benefit
that
we
can
now
traverse
this
tree
of
objects
and
if
an
object
exists,
then
the
pot
is
legal.
So
the
rule
we
have
is
ok
is
great,
is
legal
and
we
get
fulfilled
0
object.
Take
away.
This
is
actually
pretty
easy.
To
do
is
less
than
to
200
lines
of
code,
and
it's
surprisingly
easy.
A
Sometimes
the
reflect
package
gets
a
bad
reputation.
It's
hard
to
use
it.
It
turns
out
you,
you
write,
a
read
a
few
lines
about
the
the
documentation
is
not
a
dart.
One
may
rightfully
ask
why
just
not
patch
the
upstream
package
with
something
we
are
up
to
again.
The
handling
the
general
case,
which
they
are
rightfully
focusing
on,
is
not
really
that
easy.
A
Another
story
about
monitoring,
so
we
need
to
monitor
in
VMs.
How
do
you
behave?
How
much?
How
much
resources
they
consume,
how
they
are
doing
so?
How
many
out
the
period
CPU
are
used?
How
the
virtual
nikkor
use
it
again
stop
and
pause.
You
you!
You
read,
know
that,
because
instead
I
mean
we
said
in
beginning
running
inside
the
containers
and
the
container
and
time
like
hibernators
provides
us
all
kind
of
metrics,
but
we
want
another
set
of
metrics
relied
about
what's
happening
inside
the
container,
not
just
outside
the
container,
the
container
boundary
won't.
A
A
Okay,
just
to
provide
some
context,
we
these
pens
at
all
because
viateur
machines
do
you,
do
use
the
shared
storage,
so
the
the
storage
in
the
virtual
that
individual
disk
sorry
they
are
not
backed
by
a
hostile,
Austin,
local
storage.
We
need
to
talk
and
to
access
the
storage
across
the
network,
because
we
want
to
mute
a
machine
to
be
migrated
ball
and
to
be
able
to
me
to
make
them
maker
migrate
ball.
A
We
need
to
have
to
access
the
storage
from
both
sides
of
the
immigration,
so
this
is
the
what
makes
it
possible
and
again
why
to
bother
in
the
first
place
because
live
migration
is
a
key
feature
of
VMs.
We
want
the
VMS
to
be
alive
and
well,
even
even
if
we
need
to
move,
for
example,
the
Austin
which
they
were
running
in
the
first
place
in
the
in
maintenance
in
comparison
to
containers
VMs
the
vm
uptime
is
a
various
something
we
deeply
care
about,
so
back
to
the
virial
business
virtual
machine
to
monitor
them.
A
You
we
have
synchronous,
possibly
blocking
API
and
for
blocking
I
mean
blocking
for
minutes,
or
even
hours
is
doing
something
that
should
not
should
not
happen
but
could
happen,
and
we
need
to
deal
with
that.
We
cannot
just
file
an
upstream
bug
and
we
be
done
with
it
and
just
wait
for
the
fix
to
happen,
because
this
unfortunately
have
deeper
design
issue.
This
is
actually
a
long-standing
bug
open.
So
how
to
deal
with
that
again.
A
A
So
this
is
problem
we
solved
before
we
saw
before
because
they
work
it
for
a
work
in
theater
machine,
monitor
management
solution
and
indeed
this
problem.
This
is
a
known
issue
and
we
have
issue
solution
to
this
problem.
So,
let's
quickly
go
through
the
possible
solution.
Of
course
event
loop
comes
to
mind,
but
it
is
not
really
possible
there
because
of
the
or
how
the
API
looks
like.
So
we
need
threads
long,
so
short.
Well,
one
thread
per
VM
is
not
really
possible,
doesn't
really
scale
so
again.
A
The
most
common
solution
is
a
thread
pool,
but
managing
those
kind
of
issues
with
a
thread
pool
gets
hard
really
fast,
because
you
want
you
can
you
cannot
afford
to
have
one
misbehaving
VM
to
block
the
entire
thread
pool,
so
you
need,
for
example,
to
deal
with
blockage
threads
in
your
thread,
pool
to
replace
to
replace
block
threads
to
have
deadlines.
Timeouts
is
not
like
rocket
science,
but
it
is
hard
to
do
to
do
properly
in
hard
to
to
tune
how
go
long
helps
with
that.
A
We
just
throw
the
routines
after
the
problem
and
they
work,
and
it
is
simple
because
the
routines
are
cheap.
The
routines
are
fast
to
spawn,
so,
instead
of
trades,
they
scale
up
really
well
and
instead
of
threads,
you
can
just
spawn
a
goroutine
to
run
your
monitoring
if
it
blocks
it
blocks.
You
just
need
to
do
some
minimal,
bookkeeping
and
so
far,
so
good.
A
So
really
that
good,
no,
because
you
still
have
to
do
some
bookkeeping,
for
example,
truck
and
responsive
VMs,
meaning
if
monitoring
goo
routine,
doesn't
respond
quickly
enough.
You
have
to
track
it
and
be
aware
of
that,
because
you
don't
want
to
treat
to
throw
go
routines
that
and
impossible
for
the
moment
task
to
waste
them,
and
you
need,
for
example,
to
to
do
some
traveling.
But
why
trotting?
Because
we
need
to
conform
to
the
parameters
model
parameters
is
the
standard
solution
to
monitoring
kubernetes
clustered
so,
and
parameters
is
a
pool,
complete
pool
you
need.
A
You
are
supposed
to
to
extract
fresh
data
from
your
parameters
endpoint
every
time
you
are
requested,
so
long
story
short,
you
are
supposed
to
spawn
goroutines
every
time
you
get
requests
from
tomatoes,
so
again
minimal,
bookkeeping,
but
much
much
simpler.
Something
like
one
order
remark
did
or
magnitude
simpler
and
safer,
and
this
is
pretty
much
it.
So
what's
the
takeaways
I
can
give
you
well.
A
The
takeaways
is
that
I
will.
Just
I
will
like
to
highlight
two
things.
First
thing
is
again:
do
not
fight
the
deco
system.
I
mentioned
before
we
went.
I
was
involved
in
a
bit
machine
management
solution
which
is
Oviatt,
which
is
great
but
and
somehow
I
had
the
temptation,
I
must
admit
to
Rea,
implement
the
old
patterns
in
the
new
in
the
new
language
in
the
new
ecosystem
in
kubernetes
ingo.
We
always
really.
We
should
not
do
that,
but
sometimes
it
creeps
in.
Do
not
do
that.
A
Try
out
to
reevaluate
your
choice,
your
design,
choice,
choices
and
the
tool
you
are
using
and
conform
to
the
ecosystem.
It
is
much
easier,
do
not
fight
the
stream,
not
swim
against
the
stream,
and
the
other
thing
I
would
like
to
highlight.
Is
that
is
related
to
this?
One
is
that
you
have
to
reevaluate
your
design,
choose
and
choices
depending
on
the
tools
you
have
available,
but
this
road
is
goes
in.
A
Both
directions
become
because
you
want
to
reevaluate
your
different
design
choices
to
see
work
best
with
the
tool
you
have,
but
you
can
also
pick
the
tool
you
have,
depending
on
that
InDesign
choose
design
choices
and
visa
versa.
So
it's
a
circle.
Actually
the
direction
is
bought.
Anyways
again
goes
both
ways.
A
A
We
totally
do
we
totally
use
senior
CNI
and
other
than
that
time,
really
data
expert
in
their
terror.
So
we
cannot,
unfortunately,
answer
to
properly
to
the
second
part
in
the
few
question.
But
yes,
we
do
and
we
are
totally
trying
to
conform
as
much
as
we
can
to
the
tool
ecosystem.
So
if
you
want
to
have
a
look,
somehow
I
forgot
to
mention
that
this
the
website
is
keeper
dota,
yo,
very
simple,
so
you
can
start
from
there
and
get
all
of
the
links
you
want.
A
A
A
Another
important
things
to
know
is
that
we
consume
the
Siegel
way
useless
going
directly
so
with
the
I'm
not
aware
of,
for
example,
intricacies
or
issues
in
the
indigo
side.
So
because
it's
third
party
is
the
deliberate
client
libraries.
So
if
they
had
an
issue,
I,
don't
I,
don't
know
exactly,
but
from
our
perspective
brilliantly.
A
A
Question
is
yeah.
Sorry,
sorry,
if
interrupted
question
is,
if
we
had,
we
want
to
send
the
patch
upstream
to
work
here
to
fix
the
validation
issue,
mention
it
problem.
We
are
trying
to
contribute
our
back
as
much
as
we
can
to
abstain
up.
Sim
first
problem
is
that
our
use
case
it's
not
well
aligned
to
the
use
case.
We
use
the
this
package
for
is
not
really
a
really
well
aligned
to
dare
use
case.
So
it's
not
a
bad
per
se.
A
Not
at
all
we
we
run,
we
want
to
run
the
we
want
to
run
the
real
vectorization
stat
you
everyone
knows
about
like
on
k,
vm
and
the
is
the
same
vectorization
stack.
It
is
used
by
other
solutions
own
unchanged,
so
everything
runs
in
the
virtualization
like
in
qmo.
For
example,
we
kb
m
will
run
on
our
on
our
own
t.
On
top
of
cuba.
Nettie's
know
some.
Some
operating
system
are
more
supported
in
others,
but
nothing
is
explicitly
unsupported
or
discouraged
to
use.
A
The
question
was:
if
we
support
new
Matt
apology
or
if
we
allow
something
that
it
is
not
yet
supported
or
exposing
Becky
Burnett
is,
and
yes,
no
mine
is
a
good
example,
because
we
don't
know
we
don't
support
yet
because
we
were
still,
we
still
want
to
figure
out
how
to
do
properly
so
as
well.
This
is
what
I'm
I'm
aware.
I'm,
not
aware
of
recent
I
mean
last
week
developments,
but
as
far
as
you
know,
no
my
will
be
there,
but
not
there.
Yet.