►
From YouTube: Delivering TV Everywhere with Cloud Native Solutions - Arnaud Caron & Jerome Champetier, Mediakind
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
Delivering TV Everywhere with Cloud Native Solutions - Arnaud Caron & Jerome Champetier, Mediakind
Traditional TV players are facing huge challenges from the rapid growth of emerging video services such as Netflix, Amazon Prime and YouTube TV. TV service providers must modernize and accelerate their operations, moving from dedicated hardware in datacenter to embracing hybrid cloud technologies, DevOps and orchestration, while considering cloud frameworks constraints which aren’t natively designed for real-time media - and issues such as high processing & storage performance, network (latency, bandwidth, multicast), high availability and reliability. We will present our journey since 2015, showcasing how our architecture migration to microservices / containers / Kubernetes & Helm helped to lay the foundations for our global development team. We will share our architectural designs and explain how we overcame challenges such as K8s deployments, CNI, monitoring and packaging with Helm.
https://sched.co/MPYR
A
Hi,
everyone
I'm
sure
you
all
willing
this
session
to
be
ended
so
that
you
can
go
back
to
home
or
whatever.
So,
let's
start
so
sit
back
and
relax,
and
this
is
all
about
TV.
This
is
nice
topic
for
once
at
least
that's
what
I
think
so
make
yourself
comfortable
and
not
too
comfortable,
though
right
quickly,
we
decided
to
split
these
presentations
into
four
parts.
First,
we're
looking
at
what
we
do
as
media
kind
and
what
we
I
don't
why
we
notify
that
building
cloud
native
solution
is
relevant
for
our
industry.
A
Unfortunately,
it's
not
an
easy
journey
and
we'll
come
back
to
the
challenges
and
how
we
overcome
overcame
sorry
that
challenges.
If
we
have
time
and
at
the
end
of
the
day,
the
decision
just
to
keep
you
here
will
show
you
presentation
the
demo
of
the
360
spot
videos
moving
on.
My
name
is
Anakin
French,
as
you
can
tell
I'm
in
video
industry
for
more
than
15
years
now
through
multiple
leading
companies,
starting
from
development
architecture
and
product
management,
I'm,
now
driving
the
cloud
transformation
of
the
company
portfolio,
I.
B
A
Thanks
so
quickly,
under
the
view
of
media
kind,
we
are
global
leader
company,
the
media
space
and
we
have
a
large
focus
on
R&D.
As
you
can
see
across
the
globe,
multiple
locations
for
R&D
centers.
We
are
hiring
as
well.
By
the
way
our
customers
are
the
mostly
network,
TV
services
operators,
everyone
that
is
TV
in
the
world.
Today.
You
have
some
indicated
number
on
the
bottom
right
and
really
going
into
that
now.
A
I
would
say
that,
in
summary,
if
you
guys
watch
a
live,
TV,
even
sports,
on-demand
OTT-
something
there's
very
big
chance
that
somehow
the
video
feeds
the
data
feeds
has
been
through
a
media
guide,
a
software
at
some
point
or
hardware
at
some
point,
okay
get
getting
into
the
into
the
flow
well,
it
means
we
start
from
the
capture
of
devices
to
source,
to
cameras.
If
you're
familiar
with
this,
you
know
about
this.
They
are
deployed
in
the
in
the
venue
in
the
stadiums.
A
What
we
aim
at
doing
is
to
deliver
that
videos
down
to
the
virus
displays
TV
set
phone
tablets
whatever,
and
what
we
do
is
that
we
operate
in
the
middle
of
the
stream
ingestion,
the
decoding,
the
processing
storage
distributions
services
and
so
on
for
for
the
for
a
TV
that
enables
a
large
variety
of
use.
Cases
live,
TV,
replay,
TV
on
demand,
video
experience
and
so
on
and
so
forth.
A
Just
a
bit
of
context
on
the
industry
where
we
are
coming
for
me,
I
think
it's
important
to
understand
the
the
transitions
were
making
as
part
of
this
industry
I'm
not
going
to
talk
about
jean-bernard,
which
is
the
inventor
of
the
TV
back
in
1929.
We
don't
need
to
do
that,
but
I
just
want
to
illustrate
this.
That
we
have
industry
has
a
big
roots
into
the
electronics
science
and
not
really
into
the
computing
and
the
software
technologies.
B
A
Very
important
to
understand,
because,
ultimately,
this
is
one
of
the
reasons
why
we
take
some
time
to
do
this
kind
of
transitions
to
set
a
bit
the
scene
on
the
challenges.
For
years
and
years
we
must
live
focused
on
the
performance.
We
were
trying
to
increase
our
quality
of
services,
in
this
case
video
services
in
squeezing
more
and
more
TV
channels
into
some
network
constraint
networks
with
limited
power,
not
compromising
on
the
very
high
reliability
that
we
expect
in
the
industry.
A
Unfortunately,
we
didn't
care
much
about
the
flexibility
of
the
solution,
so
that
was
a
bit
of
an
issue,
but
that's
ok
and
we
didn't
care
about
running
on
standard
IT
equipment.
Mostly
quality
was
the
main,
the
main
focus
as
you
can
guess.
This
is
changing
now
the
main
changes
came
up
in
the
last
last
few
years.
So
probably
you
know
about
Netflix
about
Amazon
Prime.
Well,
YouTube
TV
usually
have
some
of
these
applications
on
your
smart
phone
today,
and
this
is
called
OTT.
A
Ogt
stands
for
over-the-top
delivery.
Basically,
it
means
distributing
media
and
streaming
services
directly
to
the
viewers
on
their
devices
on
top
of
different.
On
top
of
the
networks
that
are
deployed,
volume
network
operators,
whether
it's
a
physical
line
at
home,
DSL
line
or
wild
line
up
in
the
air,
4G
3G
whatever,
although,
unfortunately,
these
network
providers
were
using
tube
and
all
the
TV
services
with
the
network
services,
they
were
able
to
get
monies
and
bundle
subscription
package.
Based
on
this.
A
This
is
one
of
the
high
pressure
that
we
got
in
this
industry.
The
other
one
is
is
actually
you
it's
actually
as
the
consumer,
we
change
the
habits.
We
heavily
changing
our
video
consumption
habits
in
the
last
year,
as
you
can
imagine,
we
all
have
these
TV
applications
on
the
phone,
and
the
reason
is
that
we
want
to
be
able
to
watch
TV
on
the
go
in
the
train
in
the
streets.
Maybe
someone
is
watching
TV
right
now
in
this
session.
It's
bit
boring,
we
want
also.
A
We
don't
want
anymore
to
watch
TV
in
the
couch
at
home.
That's
not
the
only
usage
we're
going
really
beyond
that.
At
the
same
time,
we
want
to
be
able
to
watch
TV
anytime
any
times,
meaning
not
necessarily
exactly
when
the
show
is
on
air.
I
want
to
play
this
whenever
I
want
I
want
to
do.
Video
on-demand
I
want
to
do
a
subscription
video
on
demand,
as
well
as
a
cool
features
versus
watching
TV.
A
When
it's
on
air
right
now,
of
course,
you
still
have
to
live
sports
events
that
are
very
relevant
for
on-air
timing
deliveries,
but
for
everything
else,
the
series
alone,
no
one,
is
watching
that
anymore
live
stream.
The
last
point
is
that
we
also
requesting
a
personal
experience.
You
know
about
nice.
Netflix
features
like
because
you
watch
this
I
would
like
to
get
access
to
this.
That's
a
new
way
to
get
more
content,
more
relevant
content
individually.
A
So
how
this
is
industry?
How
do
we
adapt
this
sooner?
Why
do
we
go
to
clone
native?
Ultimately,
so,
on
our
side
four
years
ago,
we
start
looking
at
removal
removing
these
barriers
related
to
the
flexibility
if
I
summarize
it.
This
is
all
about
flexibility
and
Jerome
will
go
into
details
about
that
in
a
couple
of
minutes.
A
The
first
and
the
important
bearings
is
the
infrastructure
media
domain
traditionally
care
about
building
their
own
hardware
or
their
own
appliance,
as
we
call
it
their
own
broadcasts
in
the
video
physical
links,
connectivity
even
their
own
chipsets,
for
processing
not
using
the
standard
CPU
but
building
their
own
chipsets.
Coming
back
to
the
roots
and
the
electronic
signs
to
move
to
the
next
level,
we
need
to
be
able
to
use
off-the-shelf
hardware
and
we
need
to
be
able
to
use
a
private
public
cloud
hybrid
cloud
data
centers.
A
This
is
the
only
way
to
move
forward
and
get
a
unified
management
and
two
sets.
Besides,
we
also
need
to
deliver
new
services,
so
I
can
click
at
the
internet
speed.
So
this
is
all
about
operations,
as
you
can
see
up
there.
In
the
past
we
were
doing
the
configuration
changes
in
the
front
panel,
just
typing
some
buttons
and
having
this
small
LCD
screen
to
change
the
configurations
of
the
device.
This
is
not
possible,
it
is
not
flexible
list
and
that
helped
us
at
all.
A
Analyst
yep
we're
trying
to
be
efficient
in
terms
of
Finance
money.
This
industry
is
trying
to
get
some
money
out
of
that
stuff,
so
we
need
to
adapt
to
the
business
model,
commercial
model
and
the
flexibility,
and
we
need
to
do
some
savings
in
the
past.
This
was
not
really
a
constraints.
It's
getting
way
more
constraints
now.
A
I'm
going
into
it
a
bit
deeper,
we
have
multiple
types
of
workloads
and
that's
why
it's
a
bit
important
to
understand
these
differences
versus
the
traditional
e-commerce
websites
that
we
can
see
everywhere
here
and
the
foundation
of
this
cloud
native
solutions.
We
have
some
traditional
transactional,
sorry
use
cases
and
workloads.
Typically,
when
you
log
on
to
watch
your
TV
to
get
your
port
or
to
get
access
to
sign
up
for
and
as
VOD
services
and
so
on.
There's
a
portal
here.
A
This
is
running
in
the
cloud,
doesn't
matter
that
the
transactional
services,
typically
something
that
the
cloud
native
solutions
and
Kuban
it
is
doing
very,
very
well
today,
in
term
of
scalability
and
efficiency.
The
second
workload
is
more
at
the
Florentine
born.
Here
we
talk
about,
live,
TV,
typically
live
TV,
something
that
kubernetes
and
the
others
are
not
really
less,
are
not
really
good
at
and
we
need
to
adjust
how
it's
in
our
solutions,
we
have
pretty
high
constraints
of
the
network's,
as
you
can
see
in
time
of
pain,
we've
in
term
of
latency.
A
We
use
mostly
stateful
applications
in
this
case
because
when
you
do
live
in
coding,
you
cannot
stop
the
encoder.
Otherwise
you
don't
have
TV
anymore
and
we
have
a
high
uptime
and
high
reliability
requests
as
well.
Again,
we
don't
want
to
stop
TV.
It's
not.
This
is
not
an
option.
The
third
workload
is
more
about
the
storage.
A
Typically,
when
you
press
record
on
your
remote
control,
then
something
is
records,
whether
it's
locally
now
most
of
the
time
it's
up
in
the
air
in
the
cloud
or
in
data
center
somewhere,
so
that
we
can
playback
this
recording
at
another
time
in
your
train
or
here
at
Capcom
as
I
mentioned
earlier.
This
is
good
cloud.
Dvr
solutions,
cloud
area
request
requires
or
a
lot
of
petabytes
of
data
that
needs
to
be
accessible
instantly.
Wherever
you
are
located
again,
kubernetes
and
vehicle
system
is
not
very
great,
as
it
decided
to
build
a
TV
service.
A
B
You
so
I,
don't
know
more
saying,
is
really
important
to
understand
you.
All
these
workloads
have
got
to
work
together,
because
when
you
watch
television,
you
want
all
of
these
stuff
to
work
together.
You
you
want
to
press
record
to
watch
to
watch
your
particular
live
TV
broadcast
later
and
so
on.
So
it's
quite
an
important
point.
B
So
really,
that's
really
what
we
call
contribution,
networks
or
production
network,
so
really
move
to
the
cloud
and
the
bits
on
the
right
there
is
is
to
do
with
access
networks
is
to
do
with
what
you
and
me
used
to
watch
television
on
either
your
mobile
phone
or
your
home
broadband
or
your
cable
or
your
cable
television
subscription
type.
So
that
also
is
now
we
call
we
call
by
the
cloud
downlink
more
and
more,
as
opposed
to
being
a
an
access
network.
B
So
that's
really
quite
a
progressive
change
that
I'm
going
to
take
you
through
the
technicality
of
that
change.
So
yeah
put
a
couple
of
years
ago,
three,
three
years
ago,
ready
we
started
looking
at
taking
our
appliances
taking
our
monolithic
software
and,
like
most
of
you,
should
know
cutting
into
the
micro
services
and
trying
to
package
it
and
deploy
it
somewhat.
So
we've
actually
adopted
adopted
docker
as
a
way
to
be
low
containers
and
deployed
above
both
communities.
So
that's
the
kind
of
stack
we
support.
B
Then
we
moved
on
to
OpenStack
played
that
without
a
little
bit
and
then
we
now
support
all
pretty
much
all
public
cloud
providers
and
same
thing,
so
we
have
to
change
our
application
layer
quite
a
bit.
What
we
used
to
call
our
controller
layer,
we
we've
we've
had
to
adapt
that
to
to
this
new
world,
and
that
goes
along
with
analytics
of
my
dream,
which
I
will
talk
about
later.
B
So
what
about
infrastructure
so
I?
Don't
make
the
point
that
television
started
with
more
electronics
as
a
really
good
point,
because
we
started
way.
I
started
really
in
that
world
is
working
on
dedicated
appliances
where
we
had
specific,
embedded
processors
with
Linux
on
it,
which
had
one
sole
purpose,
was
video
processing
or
satellite
modulation
or
satellite
demodulation,
with
terrestrial
broadcasting
that
kind
of
stuff.
These
were
dedicated
appliances,
they
start
to
connected
via
ethernet
for
the
control
layer,
but
really
what
we
saw
over
time.
B
They
started
moving
to
more
and
more
integrated
into
servers.
So,
as
you
saw
server
processing
power
going
up,
then
they
were
able
to
actually
take
take
the
kind
of
workloads
we
were
doing
on
dedicated
appliances
where
we
had
dedicated
ships
to
do
that.
Now
we
see
that
one
more
international
servers,
so
it's
an
example
of
PCI
Express
card.
B
You
can
find
it
on
a
service
that
we
use
for
interconnecting
with
more
broadcast
networks,
kind
of
a
satellite
demodulator
and
so
on,
and
then
of
course,
like
any
businesses
or
customers
or
TD
operators
or
telcos
have
business
purgatives.
So
they
want
to
give
be
more
efficient
cost
wise,
so
they
started
pushing
us
to
look
hey
guys.
You
should
look
at
cloud
like
address
the
Google
cloud
as
a
way
to
deploy
your
workload.
So
that's
been
a
core
bit
of
an
infrastructure
change
for
us
coming
from
purpose-built,
electronics
to
general-purpose
infrastructure.
B
So
part
of
the
infrastructure
which
is
important
is
networking
so
networking
also
we
did
things
our
way.
We
can
see
here
a
picture
of
a
coax
cable
with
a
lovely
BNC
connector.
The
back.
That's
was
our
network
for
many
many
many
years.
It's
an
example
of
ASIO
SDI,
the
special
sale
interface
we
used
for
televisions
they
still
use
today.
Unfortunately,
we
have
to
cope
with
that.
B
Just
to
give
you
an
idea
about
numbers,
I
think
this
was
a
3GS
D
Isis
v,
gigabits
per
second
for
one
TV
channel
one
camera
feed,
4k
television
moves
to
12
gigabits
per
second,
the
kind
of
interfaces
and
the
24
gig
version
is
still
in
standardization.
So
it's
not
dying
off
till
cope
with
it
and,
of
course
that's
what
I
betrayed
and
for
the
compressed
kind
of
domain
with
all
our
stuff
has
moved
to
IP,
where
50
people
reused
the
typical
switch
routers
in
infrastructure.
B
But
what
we
found
is
more
and
more
that
the
standards
we
used
to
find
in
the
old
classical
world
are
moving
toward
the
world
of
IP,
which
is
which
is
quite
nice.
But
what
we
also
see
is
that,
while
that
transition
happens,
we
don't
just
forklift
what
we
did
in
the
appliance
world
in
the
broadcast
specific
world
and
transition
to
IP.
We
are
trying
to
make
the
most
of
IP.
For
example,
these
cables
use
to
embed
with
your
video
and
metadata
in
one
cable
in
one
particular
feed
with
IP.
B
You
can
imagine
that
everything
splits
in
beautiful
essences
you
can
send
the
video
to
a
particular
production
house,
the
audio
somewhere
and
the
subtitle
somewhere
else
again.
So
it's
it's
kind
of
position
is
actually
beneficial
for
the
whole
industry.
But
while
we've
done
that,
we
doing
that
progressively
by
the
way
and
cloud
bring
its
own
set
of
challenges
because,
of
course,
you're
dealing
with
public
networks,
you
can
have
direct
connections,
I
know,
but
it's
still
their
season
progression.
They
still
efforts
to
tantalize
how
we
do
cloud
up
inks,
for
example,
okay.
B
There
are
supposed
operation
of
doing
that
by
the
way.
So,
when
we
transition
from
broadcast
specific
to
IP,
as
I
said,
we
were
try
not
to
forklift,
so
we
try
to
use
IP
as
best
as
possible
and
we
did
it
is
we
try
to
break
away
from
certain
the
constraints
we
have
in
the
old
networks?
So
there's
something
called
multicast.
B
If
you
don't
know
what
Rita
caste
is
it's
the
ability
for
a
sender
to
send
one
packet
over
the
network
and
that
particular
packet
is
replicated
by
the
networks
to
all
the
receivers
that
subscribe
to
that
feed?
It's
great
loved
it.
We
built
entire
broadcast
plants
that
way,
unfortunately,
when
you
go
to
containers,
it
doesn't
work.
So
we
because
most
container
networks,
if
you've
been
some
of
the
CNI
talks,
really
the
way.
B
The
whole
thing
works
is
really
unique
assets
and
the
reason
why
is
because
that
is
because
containers
initially
were
really
defined
and
designed
for
dealing
with
web
transactions.
Okay,
that's
the
traditional
architecture.
Really
kubernetes
is
really
really
good
at
so
yeah
gives
a
headache.
So
that
means
we
have
to
effectively
when
we
fire
data
to
many
many
receivers,
we
have
to
find
multiple
versions
of
the
packets.
You
know
to
go
across
these
networks.
B
Actually,
that's
not
the
only
thing
we
can
also
use
what
we
something
we
call
CDN,
so
I
think
you've
always
referred
to
Akamai,
and
so
the
kind
of
type
of
CD
ends
that
we
use
to
get
the
content
to
consumers
so
yeah
bit
of
a
headache.
So
we
have
to
convert
flows
bit
of
a
shame,
but
we
have
to
do
it.
You
know
to
get
to
this
to
these
kind
of
networks,
so
once
you've
converted
flows,
I
mean
we
also
use
various
of
the
mechanism
that
we
can
use.
Tcp,
segmented
type,
video
as
well.
B
That's
my
good
time
after
I
explain
a
little
bit
more
about
that
when
we
moved
to
contain
a
networking,
and
this
is
actually
very
specific
to
communities,
we
started
looking
at
host
networking.
So
the
reason
why,
because
it
was
really
easy,
just
like
one
line
host
networking,
true
in
your
pod
definition
and
you're
done
really
good,
but
it
has
a
couple
of
drawbacks.
The
problems
we
had
with
it
is,
first
of
all,
if
you
know
a
little
bit
about
node
port,
so
I
was
networking.
B
If
you
have
multiple
pods
on
the
same
hosts
and
they
open
the
same
ports,
you
start
clashing.
There's
no
ideal.
Port
clashes
is
really
not
something.
That's
really
really
well,
you
end
up
having
to
be
scheduled
bottom
different
nodes
and
so
on.
It's
just
and
also
because
of
the
environment.
You
saw
the
diagram
earlier
about
the
cloud
of
pink
in
cloud
downlink.
So
your
sources
effectively,
if
you,
if
your
pods
new,
hosts
the
IP
address
of
the
port
activity,
there
are
world
changes,
so
you
have
to
reconfigure
your
sources.
B
So
all
the
dynamic
stuff
just
goes
away.
So
it's
just
not
ideal.
So
then
we
tried
to
use
something
called,
not
all
services
which
is
not
a
feature
of
of
communities
which
is
better
because
it
means
that
your
ports,
if
you
don't
know
it's
the
for
that
particular
port-
is
export
across
all
the
nodes
in
your
cluster.
But
the
problem
is
that
the
traffic
is
referred
to
the
two
different
note:
if
the
pods
trying
to
address
is
actually
not
that,
so
that's
not
great.
B
When
we
talk
about
the
type
of
B
trees,
we
were
introducing
you
yeah,
you
know
that's
not
ideal,
and
finally,
things
comes
a
lot
with
many
many
discussions
with
our
customers.
Now
we
found
that
in
that
case,
forklifting
the
appliance
way
of
working
in
two
committees
might
actually
potentially
work
with
malicious
CNI
with
the
fugitive
human
triple
Nix.
So
we
actually
investigated
that
quite
a
bit
and
it
does
work
quite
nicely.
So
that
means
you
have.
B
Your
pod
has
got
multiple
interfaces,
one
for
control,
which
is
your
poor
Network,
is
your
final
calico
song
and
also
you've
got
multiple
Nicks
which
are
related
to
data
network.
So
what
we
do
is
very
often
in
the
appliance
world
we
separate
the
control
traffic
so,
for
example,
the
configuration
whether
you
want
to
nudge
the
UN
SD
channel
away
from
the
actual
video
carrying
Network.
B
So
we
normally
have
video
networks
that
are
dedicated
for
video,
and
the
reason
for
now
is
because,
of
course,
if
you
start
having
a
lot
of
configuration
all
of
upgrades,
you
start
disturbing
the
video
signal
and
that's
not
the
speed,
not
good.
When
you
watch
television,
especially
not
like
television,
so
that's
really
the
kind
of
idea.
It's
still
not
there
frankly,
because
either
journey
really,
because
when
you
have
multiple
NICs,
that
means
you
have
to
deal
with
flow
management.
I
know
you've.
A
lot
of
talks
about
is
Joe
and
Lincoln
D
and
so
on.
B
These
are
to
me
kind
of
flow
management
pieces
they
capable
of
working
out
in
a
large
cluster
how
to
route
traffic
across
services
services.
We
need
a
kind
of
service
mesh
for
video,
really
so
stds
so
I
mentioned
there.
We
call
OTT
over-the-top
television,
which
is
you
carry
television
over
someone
else's
network,
an
example
of
Life
to
Live.
This
is
an
example
of
live
television
as
systems
as
we
deploy
today
on
kubernetes,
so
we
take
in
sources.
We
send
out
to
an
encoder
which
nobody
does
compression
change.
B
B
Now,
it's
great,
so
we
converted
that
entire
spot.
So
what
community
is
allowed
us
to
do
here
is
by
using
the
concept
of
ports.
We've
fundamentally
reduced
the
kind
of
failure
domain
that
we
have
when
an
appliance
used
to
be
powered
down
for
some
reason
or
network
used
to
fail,
then
the
whole
appliance
without
report
you
can
be
used.
You
can
reduce
the
failure
domain
from
fairly
large
to
really
really
small
the
TV
service,
actually
a
single
step
of
transformation
over
TD
service.
So
this
is
really
really
nice
really
nice.
Also.
B
What
community
gives
us
is
the
ability,
Reese
pimp
odds
when
they
fail,
but
now
there's
a
little
bit
with
doing
this,
because
if
one
of
these
pods
in
that
red
arrow,
which
is
the
kind
of
video
traffic,
if
you
want
it
one
of
these
spot
dyes,
it's
actually
an
outage.
It's
actually
an
artifact
that
can
visually
see
it's
all.
Your
TV
channel
disappears,
which
is
really
not
great.
Some
of
our
customers
have
spent
lots
of
lots
of
money
to
actually
buy
content
rights
for
football
matches
and
and
so
on
and
movies.
B
So
it's
not
really
not
ideal,
so
we
can't
just
cope
with
the
I
guess
the
pod
life
cycle
that
humanity
gives
us.
So
we
actually
taken
the
same
mechanism
we
used
to
using
appliances,
we
create
what
I
could
1,
plus
1
or
n,
plus
1
n
plus
M.
We
actually
over
provision
really,
because
this
is
real
time.
We
have
to
actually
synchronize
lockstep
the
multiple
encoders
together
to
make
sure
that,
if
something
happens,
if
you
won't
put
dyes,
you
have
to
switch,
then
the
next
frame
in
the
video
is
actually
served
in
real
time.
B
So
we
call
that
one
thing,
so
we
actually
build
on
top
of
communities.
Okay,
so
it
works
quite
nicely,
though.
So
what
given
us
the
ability
also
to
ingest
content
in
multiple
abilities
zone,
so
especially
the
case
for
settler
downlinks,
where
some
more
customers
are
taking
content
from
the
from
say
the
same
source
/
ingesting
that
into
two
different
locations
or
three
different
locations.
That's
really
useful,
because
that
means
he
can
now
spread
ability,
easy
clusters
and
really
nice
to
manage
with
kubernetes,
okay,
okay
and
that's
pretty
much
it
so
application
deployment.
B
So
we
started
in
that
journey
as
a
say.
Yes
as
a
collection
of
containers,
and
so
that's
actually,
we
thought
we
started.
We
started
talking
for
that.
Then
we
moved
on
to
because
we
start
to
be
starting
a
lot
of
containers
like
pretty
much
I'm
sure
all
of
you
in
this
room.
We
started
using
helm,
so
we
started
putting
things
like
the
functions,
the
parameters,
whether
you
wanted
SD
HD
or
4k,
into
charts,
which
is
quite
nice.
B
We
also
use
chart
museum
as
a
spot
about
deployments
both
on
primary
cloud
and
then,
of
course,
we
started
having
lots
and
lots
and
lots
of
different
combinations.
Every
customers
got
a
slightly
different
way
of
working
a
slightly
different
way
of
operating
the
system,
so
we
had
to
come
up
with
common
solution,
blueprints
that
we
could
refer
to
and
effectively
give
to.
Our
services
team
gave
to
our
solution.
B
Architects
be
able
to
deal
with
all
that
stuff
and
finally,
we've
noticed
a
lot
of
customers
start
to
also
asking
the
same
thing,
and
we
do
the
same
thing
in
development.
By
the
way
we
start
having
multiple
environments,
virtually
of
staging
testing
environments
and
production
environments.
Actually
they
also
have
development
development
environments
too.
B
That's
all
great,
but
I,
think
is
still
in
progress,
so
we
still
have
to
do
things
like
what
to
do
about
upgrade
what
to
do
about
downgrade
system
expansion
and
I
can
also
so
we
investigating
fundamentally
the
operator
framework.
You've
done
probably
have
a
lot
of
talks
about
about
this
through
the
course
of
the
week.
So
that's
the
kind
of
things
were
looking
at
also
a
good
old
NFS
who
just
use
that
for
developing
on
appliances.
For
me,
15
years
ago,
it's
still
out
there.
B
We
use
that
today
and
we're
trying
to
get
rid
of
it's
really,
not
really
replicated
not
ideal.
So
we
investigating
multiple
options
that
okay,
so
now
what
about
resource
requests,
so
we
quotes
with
quite
hungry
for
resources.
These
appliances
that
we
talking
about
were
designed
using
custom,
chipsets,
FPGA
zigs,
a
kind
of
stuff.
So
that
meant
that,
because
we
was
purpose-built,
we
knew
exactly
what
they
were
capable
of
doing.
If
you've
bought
a
box
from
us,
it
was
like
12
channels
of
SD
or
four
channels
of
HD.
That's
it.
B
You
know
exactly
what
to
have
now
in
the
feeds
you
have
really
with
abilities.
We
had
to
do
that
dynamically,
so
we
start
developing
an
algorithm
where
we
were
taking
known
configurations,
no
one
TV
channel
costs,
and
we
fed
that
to
a
machine
learning
algorithm
that
gave
us
the
ability
to
do
it,
of
course,
small.
So
how
many
achieve
a
TV
channel,
depending
on
his
creation,
were
dealing
with,
was
actually
used
to
process
that
channel
and
then
we
we
try
now
to
look
at
whether
how
to
ship
that
model
to
to
our
customers.
B
B
So
she
you
guys,
have
some
ideas
about
how
to
do
this
in
a
bit
generic
way
and
find
us
later
interested
to
hear
what
you've
got
to
suggest
so
as
gbit/s
is,
as
can
swallow
the
whole
of
the
television
world.
What
is
production
and
so
on,
and
especially
because
we've
broken
these
appliances
into
multiple
micro
services.
B
We
have
to
deal
with
a
lot
of
the
monitoring
of
it
and
looking
at
what
comes
out
of
these
micro
services
into
the
video
is
actually
quite
a
challenge,
because
you
can't
put
a
person
with
eyes
to
look
at
what
comes
out
of
a
cloud
you
you've
got.
Actually
you
look
at
it.
What
is
there
so
we've
had
to
do
some
some
efforts,
so
we
started
looking
at
watches.
B
We
have
done
already
has
really
done
what
you
think
is
actually
breaking
focus
on
a--
we
built
to
display
video
thumbnails
because
that
didn't
really
exist
before
and
effectively
is
trying
to
make
IT
tools
more
television
aware.
So
it's
nice
variety
to
swallow
the
whole
of
the
television
ecosystem
entirely,
but
like
voice
over
IP.
It's
not
gonna
happen
without
any
change
in
either
side,
so
we
are
changing,
but
there
are
teas.
The
rest
of
us
is
alluding
to
change
I.
B
B
So
now
we've
done
this
move
to
to
the
cloud.
By
the
way
we
can
do
a
lot
more
stuff
with
it.
So
that
means
we
can
free
ourselves
and
the
constraints
we
had
effectively
dealing
with
just
2d
pictures
and
2d
frames.
We
actually
can
do
a
lot
more
than
what
we
used
to
do
before.
Actually
so
this
is
being
developed.
Recently,
we
don't
write
down
a
few
events.
We
actually
law
called
a
couple
basketball
matches
on
this.
B
Using
this
technology,
we've
created
a
effectively
a
brand
new
thing
which
is
about
this
live
for
60
video,
augmented
experience.
So
this
is
not
about
replacing
the
you
know,
the
experience
that
you
and
me
are
used
to
to
watch
television
on
the
big
screen.
This
is
kind
of
to
offer.
Second
screen
experience:
okay,
you're,
not
the
way
I
go
goes
with
physics,
video
for
about
an
entire
football
match.
I,
don't
think
it's
really
the
thing
about
it.
B
So
if
we
provide
a
different
viewpoint
on
an
event,
so
we
take
in
content
from
these
physics,
video
cameras,
which
are
film
income,
filming
out,
really
really
high
resolution.
We're
talking
about
6
k
8k.
So
it's
a
lot
of
pixels
to
encode
right.
We
are
processing
this
into
Google
Cloud
and
we're
sending
that
to
various
this
place,
so
that
could
be
just
basic
tablet
experience
or
what
you'd
find
on
a
on
the
VI.
It's
it
and
really.
We've
we've
done
this
in
partnership
with
many
companies,
because
this
is
brand
new.
B
There's
a
lot
of
lot
of
interest
into
this,
so
we've
partnered
with
the
Steilacoom,
with
an
imaginative,
er,
app
Feld
media
for
the
packager
and
the
tiling
of
the
content
and
in
VR
for
the
camera
and
film
for
the
production.
So
that's
really
really
cool
as
to
kind
of
really
do
cool
things.
We
can
do
with
these
with
these
environments.
B
So
couple
of
example,
things:
we've
done.
We
looked
at
live
sports,
so
that's
the
Dutch
Telecom
basketball
match
if
you're
interested
I've
got
a
big
demo
after
the
Tolkien
come
and
have
a
chat.
I
can
show
you
what
we've
done
and
live
music
concerts
as
well.
It's
trying
to
give
you
the
best
seat
in
the
house
right
in
front
of
the
seat
right
in
front
of
the
the
scene
or
even
next
to
the
singer
as
well,
really
cool
stuff.
Any
sports
coming
coming
along
quite
quite
impressively.
B
B
So
I
know
there
was
a
talk
earlier
in
a
keynote
this
morning.
That
said,
you
know,
you
shouldn't,
prohibits
all
stateful
tasks
and
work
clothes
from
kubernetes,
but
really
it's
not
really
there
not
really
making
use
of
what
we've
got
available
so
again
important,
so
we're
trying
to
move
from
stateful
encoders
to
completely
stateless,
including
jobs.
So
that
means
that,
rather
than
actually
having
to
encode
everything
as
one
big
video,
we
actually
chop
the
video
into
tiles,
both
especially
in
the
same
frame
and
temporary
over
time.
B
So
that
means
that
we
can
now
make
sure
that
all
of
these
tiles
are
actually
small,
encode
jobs,
stateless
single
jobs,
I
carry
that
configuration
their
purpose
and
where
they
belong,
when
it'll
to
be
able
to
reconstruct
them
on
the
output,
and
that
really
I
think
is
something
eyes.
Reader
because
got
great
potentials
in
to
actually
change
the
world
television
work.
B
We
started
using
that
a
little
bit,
but
the
problem
is
what
scalars
are
normally
based
on
metrics,
like
CPU
consumption
and
so
on,
which
is
really
isn't
I
mean
we
need
really
to
have
used
custom
metrics,
just
more
specific
to
television.
So
that's
also
an
area
of
or
development,
where
we
need
to
look
at
a
bit
further.
Okay
and.
B
C
Thanks
for
a
good
presentation
regarding
this
networking
journey,
you
have
shown
have
you
looked
here
in
the
conference
and
network
service
mesh
because
you
asked
for
a
video
service
mention
because,
but
may
be
more
general,
they
are
two
layers
really
networking
could
help
you.
If
you
had
a
sandbox
project.
B
Yes,
the
answer
is
yeah
and
we
still
need
to
look
at
a
bit
further.
The
way
we
we
track
with
track
transformation
in
our
systems
today
is
very,
very
similar
to
what
it
would
to
water
service
mesh.
We
do.
The
only
problem
is
currently.
We
know
when
we've
seen
with
service
meshes
that
you're
dealing
with
mainly
something
less
transactions,
so
in
order
for
us
to
make
to
make
use
of
all
the
service
mesh
tools,
even
a
networking
service
match
is
that
we
have
to
somewhat
change
the
way
the
video
itself
is
carried
over
the
network.
B
So
currently
it
is
really
fine.
Forget.
Ok,
you
fire
packets,
you
fire
the
content
to
the
particular
place
in
most
cases,
not
all
the
time,
but
in
most
cases,
release
fine
and
forget,
so
we
have
protocols,
some
of
them
open
source
dealing
with
be
able
to
do.
We
choice,
for
example,
if
if
the
packets
lost
in
transit
or
if
either
the
receiver
or
the
Senate
on
something
wrong
with
the
transmission,
but
that
is
to
move
all
forward
to
into
the
service
Metro
babe,
that's
task
on
potentially.
C
D
B
Showed,
for
example,
some
graph
wanna
plugins
are
those
open
source?
Not
yet
is
something
we've
done.
A
bit
of
an
innovation
projects
is
addressing
to
node
tools.
We've
done
it'll
do
to
get
to
get
us
going
really,
but
yeah.
That's
that's
one.
Potentially
one
area
we
could
give.
Actually
we
could
explore
yes,
because
it's
it's
quite
important.
A
And
it's
quite
a
bit
quite
simple
in
the
end,
maybe
I
complain
on
this
I
think
traditionally,
the
broadcast
industry
and
media
industry
is
not
open
source.
This
is
a
man
chats
mindset
to
change
here.
Definitely
we
want
to
get
there.
We
want
to
get
part
of
the
community
and
so
on,
but
there
is
some
barriers
here.
So
yes,
we'd
love
to
put
that
in
open
source,
because
just
video
plugins,
but
as
challenges
you
I've.
B
D
B
It
did
I
think
latency
is
got
an
important
parameter
in
what
we
do,
so
it
is
important
to
end-to-end
really,
but
it
affects
different
different
areas
of
the
of
the
of
the
of
the
life
cycle
of
the
content
differently.
So
if
you're
talking
about
say
in
an
sport,
environment
and
interview,
for
example,
that
the
latencies
ask
God
to
be
as
low
as
possible,
so
yes,
that's!
That's!
That's
incredibly
important.
B
We've
measured
different
levels
of
latency
with
different
CNI
plugins,
the
ones
that
we
stabilize
for
for
video
carrying
content,
has
got
the
minimal
number
of
minimum
number
of
buffering.
That's
that!
Yes,
it
did
influence
a
lot
yes,
so
we
don't
want
to
have
a
necessary
number
of
hops
into
into
the
into
the
system.
B
Also,
an
important
thing
to
know
is
that
we've
noticed
a
lot
of
the
CNI
buggies
do
have
certain
CPU
consumption,
especially
we
took
about
a
lot
of
data,
so
that
has
an
impact
not
just
on
latency,
but
also
how
much
resource
we
have
left
to
do.
Video
processing
at
the
back
of
it
all
your
processing
on
the
back
of
it.